m MSU LIBRARIES “ RETURNING MATERLAL§: Place in book drop to remove this checkout from your record. FINES wiII be charged if book is returned after the date stamped be1ow. .m—figfi g12 A 100 f 1 ”(5‘ .' W" a. .' V; . N‘- .41) ‘ N" - Mn LIV: III-Ill! .,¢F.b‘. ltrrfiy THE MEASUREMENT AND MODIFICATION OF ABSENCE BEHAVIOR AND OTHER EMPLOYEE TIME USES DURING IMPLEMENTATION OF AN ABSENTEEISM CONTROL INFORMATION SYSTEM: A TIME-SERIES STUDY BY Ian Alexander Miners A DISSERTATION Submitted to: Michigan State University in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY School of Labor and Industrial Relations C) Copyright by Ian Alexander Miners - 1984 - ABSTRACT THE MEASUREMENT AND MODIFICATION OF ABSENCE BEHAVIOR AND OTHER EMPLOYEE TIME USES DURING IMPLEMENTATION OF AN ABSENTEEISM CONTROL INFORMATION SYSTEM: A TIME-SERIES STUDY BY Ian Alexander Miners A substantive and empirical reconceptualization of the absence research paradigm, this project addresses limita- tions of the conventional paradigm reported in prior research by developing a detailed system for time-series analysis. To integrate concerns of conceptual consistency and measurement raised in the literature (Chadwick-Jones, Nicholson and Brown, 1983; Hammer and Landau, 1981), absence, attendance and time scheduled off have been recon- ceptualized in the present work as alternate forms of employee time-use. These have been operationally defined as a set CM? 13 interdependent moving values calculated daily across the duration of an absence control program. A time- series split-half reliability test was used to determine the reliability of these measures. Both cross-sectional T-tests and. a time-series correlation. technique ‘were applied. to investigate program impact and to test substantive hypoth- eses about absence behavior. For both the T-tests and cor- relations, hypotheses of two kinds were examined. In each Ian Alexander Miners case, the first kind focused on testing the relative merit of the procedures used to examine the second or substantive kinds of hypotheses. Since the absence literature postu- lates significant moderations of absence behavior by annu- ally recurring seasonal changes or' holidays, substantial attention was also devoted to distinguishing the extent of such intra-annual shifts so that they could be considered in the assessment of program impact. Major substantive hypoth- eses dealt with whether employee time-use changed from short-ternl to long-tern! absence and.:from illegitimate ‘bo legitimate forms of absence. The results generally showed the measures to be an order of magnitude more reliable than measures used in prior research (Breaugh, 1981; Hammer and Landau, 1981), although under small sample conditions the measures became less reli- able. The absence control program providing the multi-year database of this research did clearly reduce short-term non- legitimate absence significantly and a co-occurrent increase in time worked on the job for each employee also took place. There also occurred a powerful legitimacy effect, a shift of employee time-use from non-legitimate absence into long-term absence which was still legitimate after program installa- tion. Ian Alexander Miners On a methodological level, the evidence brought to bear on procedural hypotheses clearly supported the use of group level time-series analysis, and in particular supported the time series correlation technique employed here. The evi- dence also supported the contention that daily cross-sec- tions of individual absence data are distributed highly non- normally. Despite this finding, T-tests comparing' daily cross-sections of individual data over time still showed results that are consistent with the results of group—level time series analyses. The comparison of daily cross sec- tions has obvious practical merit since it allows estimation of costs and benefits. But for inferential tests to analyze the significance of trends in time-use behavior, conven- tional cross-sectional procedures are an: best tenuous and where possible these should probably be set aside in favor of the time-series procedures. On this theoretical level, a strong substitution of long-term legitimate absence for short-term non-legitimate absence (r = 0.90) occurred across all four different time intervals examined during the span of the program. The evi- dence strongly supported the legitimacy hypothesis. When Ian Alexander Miners the effect of legitimacy was controlled for however, no evi- dence of a group-level progression toward longer-term absence was apparent. With regard to program, impact, the absence control project itself was, by design, a managerial and technical system installed for reporting and guiding the control of employee absenteeism. It was a contractual, OD and research success for the consultants, and an administrative success for the DOT. However, the impact of the program was some- what diluted by substitution of long-term legitimate absence for short-term non-legitimate absence, and by an externally imposed increase in legitimate absence due to 80 hours of paid leave granted by the City Government in concession bar- gaining. In short, the consultants and the IXIP succeeded beyond the goal that had been set. But overall absence changed very little since the city authorities gave back part of the time saved, and the DOT long-term absence con- trol policies were not designed to prevent much of it from shifting into long-term legitimate absence. DEDICATION To Dad, who taught me to think in models and to build the models into things that work To Mom, who taught me to love books and to wonder about human behavior To Mike, who taught me his trade and gave me a clear path to learn Those gifts were well given. This one is for all of us. ii ACKNOWLEDGEMENTS A great many people have poured their efforts into this work. Lionel Little of The IDepartment of' Transport in Detroit and Tony Juliano of Moore and Juliano Associates were most helpful in providing access to the research data. Dr. Michael Moore of the School of Labor and Industrial Relations at Michigan State University guided the design and installation of the absence control system. It was Mike's insight into the need for precise detail in day-to-day data collection that created the detailed raw data analyzed here. And it was his support and guidance that nurtured the dis- sertation research process. Dr. Tom Patten has helped guide my progress from the begin- ning of the Ph.D. program. Both he and Dr. Dan Ilgen have contributed their time and critical insights to the struc- ture of this document and to refining its content. They have been very helpful about providing quick and sincere feedback. and the dissertation. is a :much. better product because of their help. Dr. Larry Foster and Dr. Mary Zalesny also contributed some helpful suggestions and special thanks are due to them for their encouragement and support. There are several special people in the Computer Services Department at The University of Toledo, Toledo, Ohio, whose support was invaluable. Wendy Fraker especially, and Donna, Gregg and Lance -- thank you for the personal attention you gave to the awkward problems this research posed. My friends and family have helped when they could and have tried to understand and share this work. There were many times when their faith in me made a great difference in the small, cold hours of the morning. Jan Tabernik, my editor, secretary, and friend, has typed every draft at every stage of this dissertation. Often she had to find a way to do the work at the last minute or at night and without any compromise of quality. Thank you, Jan. ' iii TABLE OF CONTENTS Chapter 1: IntrOduction O O O O O O O O O O O O O O O O O 0 O O O O O O O O O 0 O 0 An Introduction to the DOT Absence Control System .. Objectives 0......OOOOOOOOOOOOOOOOO...0.0.0.0.000... Procedural Objectives ........................... substantive Objectives O0.00000000000000000000000 Distinctive Features ................ ........ ....... Dissertation Structure .... ....... .................. Chapter 11: Literature Review and Explanation of the Hypotheses 0.0.0.0000...0.0...00.0.0.0...OOOOOOOOOOOOOO General Introduction to the Literature Review ...... Practices in Employment Relations Which Relate to Absence Management and Control .................. Conventional Measures of Absence ................... Methodological Issues .............................. Variation Due to Demographic Subgroup Differences .. Pre-Program Shifts in Employee Time-Use ............ Patterns of Cyclical Change in Absence Behavior .... Theories of Unintentionally Induced Over—Time Change in Absence Behavior: Internal Progression and Legitimacy EffeCts 00.000.00.00...0.0.0000...0.. ProgreSSion 0......OOOOOOOOOOOOOOOOOOOOOOOOOOOOOO Legitimacy 00......00......OOOOOOOOOOOOOOIOOIOOOO Program Evaluation: An Overall Long—Term Assessment Of Program ImpaCt .0.00......OOOOOOOOOOOOOOOOOOOOOOO iv 010101 \I ll 13 13 15 16 21 31 32 33 36 37 38 40 A Summary of Theoretical and Measurement Issues .... Chapter III: Data Collection, Sample Demographics, and the Derived Measures .................................. Access to the DOT and Data Collection Procedures ... The Consulting Intervention ........................ Preparing the Time-Use Database .................... The Sampling Procedure ............................. DOT Workforce Demographics ......................... Procedures to Generate the Derived Time-Use Indices Rules for Recategorization of the 45 Raw Absence Reasons over Time ......OOOOOOO......OOOOOOOOOOOOOOO Three Different Measures of Time-Use: Frequency, Severity, and Mean Severity per Occurrence ......... Chapter IV: Methodology and Operational Hypotheses ... IntrOducti-On ...0..........OOOOOIOOOOOOOO0.00.00...O Rearranging the Case Data by Relative Date ......... Procedures in A: Normality Tests of the Data Distributions 0......0...........OOOOOOOOOOOOOOOO... Procedures in B: T-Tests of the Cross-Sectional comparisons ........0.00.0.........OOOOOOOOOOO0....O — Pre-Program Incidental Shifts Using Measures Based on a 30—Day Period ............ - T-Tests of Pre-Program Changes on the IOO—Day Measures ............................. Seasonal Effects ....................... Holiday Effects ........................ Combined Seasonal and Holiday Effects .. T-Tests of Post-Installation Changes ............... - Inter-Annual Change .......................... -H011day Effects .....OOOOOOOOOOOOOOOOOOO...0.. - Seasonal Effects and Annual Renewal .......... Inter-Annual Shifts Due to Program Changes ......... 42 45 46 52 54 56 58 60 67 69 69 73 74 75 76 78 78 79 80 80 80 82 83 Explanatory Comment on Long-Term Substitution Effects ......O0.......0.......OOOOOOOIOOOOOOO0.0... A Note on Calculation of the Means and Selection of the Split-Halves ......OOOOOOOOOOOOOO0.000.000.0000. Time Series Correlation Estimates for Split-Half Reliability in the Measures .....OIOOOOOOOOOOOOOOOO. Conditions Conducive to Anomalously Low Reliability in the Time-Use Measures ............ Time-Series Correlations to Test the Substitution Hypotheses and to Investigate the Impact of the Absence Control Program Over Time .................. Correlational Tests on the lOO-Day Dataset ...... - Group-Level Progression Effects .............. ‘— Legitimacy Effects 0.0.0.000.........OOOOOOOOO - Combined Progression and Legitimacy Effects .. - Shifts Toward Work and Away from Non-Legitimate Short-Term Absence o o o o o o o o o o o o Time-Series Correlations Between Measures in the 365—Day Dataset 0.0.0.0000.........OOOOOOOO. Chapter V: Results and Discussion .................... IntrOduction O0.0.0.0.........OOOOOOOOOOOOOOOO0.0... Normality and Empirical Data Function Statistics for the (Fl-Test Data ......OOOOOOOOOOOOOOO Discussion of Results for the Normality and EDF Statistics 00............OOCOOOOOOOOOOOOOIO...... Results for the T—Tests of Cross-Sectional comparisons I.......OOCCOOOOOOOOOOOOCO......OOOOCOOO Discussion of T-Tests on Overall Pre-Program Shifts A Note on Internal Consistency and Relations Among Different Time 0585 ......OOOOOOOOOOI......OOOOOOOOO T-Tests for Pre—Program Intra-Annual Change ..... - Discussion of T-Test Results on Pre-Program Seasonal Shifts .............................. - Discussion of Results on the T-Tests on Pre- Program Holiday Shifts Observed in the 100- Day Measures ................................. vi 85 86 87 89 90 91 91 92 93 93 94 95 95 96 99 100 104 108 109 109 112 - Discussion of the T-Tests on Combined Seasonal and Holiday Effects in the Pre-Program 100- DaYMeasures ......OOOOOOOOOIOOCOOO0.0.0.0.... 115 T-Tests of Long-Term Change During the Time that the Program was Fully in Force .................. 117 - Discussion of the T-Tests for Year-to-Year Change in the loo-Day Frequency and Severity Measures During the Program Period ........... 117 - Discussion of T-Test Results for Hypotheses About Holiday Effects in the lOO—Day Measures During the Program Period .................... 121 - Discussion of Results for the T-Tests of Combined Seasonal, Holiday and Annual Renewal Effects During the Program Period on the 100- Day Measures ................................. 124 Comment on the Measures Used to Examine Intra-Annual T'TeSt ReSUlts 00.000000000000000......OOOOOOOOOOOOO 124 Discussion of Results of the T-Tests on Overall Program-Induced Change in the lOO-Day Measures .. 127 Discussion of the T-Tests on Overall Long—Term Program Induced Changes Indicted by the 365- Day Measures, A Direct Evaluation of the Program .0O.......00............OOOOOOOOOIOOO.... 128 The Split-Half Time-Series Reliability Correlations 135 The Reliability of Time-Series Measures for the Substantive Hypotheses .......................... 135 Anomalous Values for the Split-Half Reliabilities .. 138 Discussion of Time-Series Correlations of Sub- stitution Effects Using the Severity Indices ....... 140 A Comment on Frequency-Severity Relationships ...... 144 Chapter VI: Conclusions and Implications ............. 145 General COflClUSiOnS O.......OOIOOOOOOOOOOOOO00...... 145 A Detailed Evaluation of Long-Term Program Impact .. 149 Implications of the Present Research for Future work ............OIOOOOOOOOOOO......OOOOOOOOOOCO0.0. 152 Implications for Personnel Practice ................ 156 vii Appendix A: The Detailed Description of the Workforce Demographics 0.0.0.000...O......OOOOOOOOOO......OOOOOO UnpaCKing ......0............OOOOOOIOOOOOOICOOO.... Detailed Statistics for the DOT Workforce Demographics 0....0.0............OOOOOOOOOOOOOOO Brief Demographics for the Subsample Group ........ Demographics for the Sample from All DOT Employees ......OOOOOOOOOOO......OOOOOOOOOOOOOOO Appendix B: Derivation of the Raw Time-Use Data and Creation of the Derived Data Structure ............... Derivation of the Raw Time-Use Data and Creation Of the Derived Data Structure ......OOOOOOOOOOOOOOO Derivation of the Time-Use Data Values ......... Appendix C: Methodological Notes .................... Procedures for Assessing Normality and Distribu- tional Form ......IOOIOOOOOOOO......OOOOOOOOOOOOOOO A Methodological Note on the Time-Series Correlation Procedures ............................ The Time—Series Correlation Procedure Described Substantive Correlational Tests ................ Appendix D: A Review and Discussion of the Detailed Results O......OOOOOOIOOOOOOOOOOO......OOOOOOOOOOOOOOO A Review and Discussion of the Detailed Results ... Normality and EDF Statistics for the T-Test Data .0.........COOOOOOOOOCOOOOOOO......OOOOOOOO Summary of Results for the Normality and EDF Statistics 0.000.000...0.0..........OOOOOOOOOOOOOOO Results for the T-Tests of Cross-Sectional comparisons ......OOOOOOOOOOOO00.000.00.000.0.....O T-Tests on Overall Pre—Program Shifts .......... — T-Test Results on Pre-Program Seasonal Shifts - Results of the T-Tests on Pre-Program Holiday Shifts .............................. viii 159 159 162 172 173 176 177 183 191 192 195 196 199 208 208 210 216 218 221 225 227 — Results of the T—Tests on Combined Seasonal and Holiday Effects in the Pre-Program 100- Day Measures ................................. 229 - Results of the T-Tests for Year-to—Year Change in the lOO-Day Measures ............... 232 - T-Test Results for Hypotheses about Holiday Effects in the loo—Day Measures .............. 235 - Summary of Results for the T-Tests of Combined Seasonal and Annual Renewal Effects on the 100-Day Measures ............................. 237 Comment on the Measures Used to Examine Intra— Annual T‘TeSt ReSUltS O00.0.00.........OOOOOOOOOOOO. 237 Results of the T-Tests on Overall Program-Induced Change in the 100-Day Measures ..................... 238 Results of the T-Tests on Overall Program Induced Changes Indicated by the 365-Day Measures 0.00.00...OO0.000.........OOOOOOOOOOOOOO 244 The Split-Half Time—Series Reliability Correlations 245 Summary of Results of the Time—Series Split- Half Reliability Correlations ................... 247 Results of the Time-Series Correlations on Substitution Effects on the Severity Indices ....... 249 Summary of Time—Series Correlations Testing the Substantive Hypothesis, Using Severity Indices 0......................OOOOOOOOOOOOOOOOO. 251 A Comment on Frequency-Severity Relationships ...... 251 Bibliography ............ ..... .......... ......... ...... 253 ix LIST OF TABLES Table l: Normality and EDF Statistics for Selected 30-Day, loo-Day and 365-Day T-Test Data Distri- butions of All-Time-Absent ......................... Table 2: T-Test Results for the Hypothesis that Pre-Program Shifts in Time-Use Behaviors Occurred: 30-Day Data ........................... Table 3: T-Test Results on Hypotheses of General Seasonal Shifts in the Pre-Program IOO-Day Measures: October 20, 1978 to April 14, 1979 Table 4: T-Test Results of Holiday Shifts in the lOO-Day Pre-Program Time-Use Measures: January 5, 1979 to April 14, 1979 ................ Table 5: T-Test Results for the Combined Seasonal and Holiday Effects on the lOO-Day Pre-Program Measures: January 5, 1979 to June 15, 1979 ...... Table 6: T-Test Results for Hypotheses about Long- Term Year-to-Year Change in the lOO-Day Measures, Frequency Indices Only ........................... Table 7: T-Test Results for Hypotheses about Long- Term Year-to-Year Change in the lOO-Day Measures, Severity Indices Only ............................ Table 8: T-Test Results for Holiday Effects on the loo-Day Data, Severity Indices Only .............. Table 9: T-Test Results for Hypotheses of Combined Holiday, Seasonal and Renewal Effects on the 100- Day Measures: Severity Indices Only ............. Table 10: T-Test Results of Program Induced Change in the 365-Day Measures .......................... Table 11: Time Series Correlations for Split-Half Reliability ...................................... Page 105 110 118 119 122 125 136 Table 12: Time Series Correlations Testing Substantive Hypotheses for Severity Indices Only ............................................. 141 Table 13: Time Series Correlations to Inspect Relations Between Frequency Indices and Severity Indices ................................. 144 Table 14: Year and Month and Day Hired ............... 164 Table 15: Birth Date ................................. 165 Table 16: Job Class .................................. 166 Table 17: Race ....................................... 170 Table 18: Year and Month of Quit ..................... 172 Table 19: Normality and EDF Statistics for Selected 30-Day T-Test Data Distributions ................. 213 Table 20: Normality and EDF Statistics for Selected loo-Day T-Test Data Distributions ................ 214 Table 21: Normality and EDF Statistics for Selected 365-Day T-Test Data Distributions ................ 215 Table 22: T-Test Results of the Hypothesis that a Pre-Program Shift in Time-Use Behaviors Occurred in Time-Use Behaviors Occurred - 30-Day Data ...................................... 220 Table 23: T-Test Results of Seasonal Shifts in the Pre-Program loo-Day Measures: October 20, 1978 to April 14, 1979 ................................ 224 Table 24: T-Test Results of Holiday Shifts in the loo-Day Pre-Program Time-Use Measures: January 5, 1979 to April 14, 1979 ................ 226 Table 25: T-Test Results for the Combined Seasonal and Holiday Effects on the lOO-Day Pre-Program Measures: January 5, 1979 to June 15, 1979 ...... 228 Table 26: T-Test Results for Hypotheses about Long- Term Year-to-Year Change in the loo-Day Measures, Frequency Indices Only ................. 230 Table 27: T-Test Results for Hypotheses about Long- Term Year-to-Year Change in the loo-Day Measures, Severity Indices Only .................. 231 xi Table 28: T-Test Results for Holiday Effects on the loo-Day Data, Severity Indices Only ......... Table 29: T-Test Results for Hypotheses of Combined Holiday, Seasonal and Renewal Effects on the lOO-Day Measures, Severity Indices Only ......... Table 30: T-Test Results of Program Induced Change in the 365-Day Measures ......................... Table 31: Time Series Correlations for Split-Half Reliability ..................................... Table 32: Time Series Correlations Testing Substantive Hypotheses for Severity Indices Only ............................................ Table 33: Time Series Correlations to Inspect Relations Between Frequency Indices and Severity Indices ................................ xii 234 236 LIST OF FIGURES Figure 1: The City of Detroit Department Transportation Attendance Control Operations Guide .......................................... Figure 2: DOT Employees' Demographic Character- istics, for All Employees and for the Research Sample ................................ Figure 3: Summary of the Raw Absence Data As Recorded in the Department of Transport Absence Management Database or as Modified and then Used to Generate the Derived Data ..... Figure 4: Logical Relations Among Different Uses of Employee Time ............................... Figure 5: Derived Time-Use Measures ............... Figure 6: A Categorical Summary of Tests and Analyses ....................................... Figure 7: Approximate Graphic Profiles of Long- Term Program Impact on the Overall 365-Day Time-Use Measures: Relative Scales ............ Figure 8: The Demographic Data Variables .......... Figure 9: Distribution of All Employees in the DOT Workforce by Geographic & Functional Cost Centers ................................... Figure 10: The Raw Absence Data Tape Variables ..... Figure 11: Logical Relations Among Different Uses of Employee Time .......................... Figure 12: Summary of the Raw Absence Data as Recorded in the Department of Transport Absence Management Database or as Modified and then Used to Generate the Derived Data ................................... Page Figure 13: Derived Data Types Figure 14: The Individual "Derived" lOO-Day Data Types Figure 15: The Individual "Derived" Annual Data Types Figure 16: Approximate Graphic Profiles of Long- Term Program Impact on the Overall 365-Day Time-Use Measures: Relative Scales xiv CHAPTER I Introduction Current work in personnel research has a greater need than. ever before to find and show strong relationships between personnel practices and concrete dollar conse- quences. At first glance, absenteeism is one tangible con- sequence where costs and the control of them should be eas- ily measured and managed. Apparently this has not been the case. The field still does not have an effective response to the problem. In 1982 a rough estimate (an adaptation of Mirvis and Lawler, 1977) was updated using the 1981 CPI and national absence statistics, placing overall absenteeism costs at $53.6 bullion, while payroll related costs alone were about $23.6 billion. This burden alone is a compelling reason for reducing absence if possible. And if an effec- tive system existed with which to monitor and analyze changes brought about through absence control efforts, then a systematic knowledge could be developed about what works best to control absence. But absence research reported to date has not demonstrated an effective means for program evaluation and absenteeism management. Meanwhile, there exists a growing emphasis in the lit- erature (Breaugh, 1981; Hammer & Landau, 1981; Steers and 2 Rhodes, 1978) on reconceptualization of absenteeism (Chad- wick-Jones, Nicholson & Brown, 1983) and on refinements in measurement. Several neW' opportunities Ihave emerged for using innovations in data preparation (Stevens, 1951; Hamb- lin, 1974; Tukey, 1977) and analysis to enhance this work (Glass, Wilson & Gottman, 1975; Montgomery & Johnson, 1976; Bloomfield, 1976; Thrall & Engleman, 1981; Liu, 1981). The research reported below is an attempt to establish and apply a stronger paradigm for absence research and con- stitutes a significant departure from the conventional approach to absence measurement. Both the conceptual basis and the operational definition of the measured variables are new . An Introduction 39 the DOT Absence Control System During 1978 and 1979 a comprehensive system for identi- fying and reporting absenteeism and for implementing control practices to reduce non-legitimate short-term absence was installed at the Department of Transport in Detroit, Michi- gan. The project was implemented in the stages described below. Further detail is provided in Chapter III. (a) Top management support was secured for the project and extensive supervisory input into the determination of vari- ous forms of employee time-use was encouraged. Supervisory input was facilitated using a hierarchically sequenced series of structured problem-solving groups and management 3 training workshops. Data for July 1, 1978 to June 30, 1979 were collected with the payroll data for use as a pre-pro- gram baseline period in later analyses. During the prepro- gram installation period, nominal groups were used to focus employee, supervisor, and manager input regarding causes of absenteeism, solutions to the "absenteeism. problem," and incentives for attendance. These meetings were part of both a diagnostic and programmatic effort to raise awareness of absenteeimn at the DOT. It was begun roughly nine months before the computer aspects of the program were implemented. (b) A Real-time information system was created to computer- ize the flow of information about employee time uses so that timely and adequately detailed reports could be provided to supervisors. These reports were to become the crucial information base required for consistent supervisory imple- mentation of absence control policy. (c) The absence control program itself was implemented on a real and high-impact basis, coming on-line in stages during May 1980. In those last 30 days of the installation period (June 1980), absence control steps 'were actually’ imple- mented. Employees and supervisors were shown how the system worked and were given a chance to learn what impact it would have on them. (d) On July lst of 1980, an additional component of the absence control program went into effect. Each employee was given a "bank" of casual leave, to be used at his or her 4 discretion, as paid, legitimate, short-term absence. This bank of casual leave, plus holidays and annual vacation, were to be renewed each year in accord with the collective bargaining agreements with the several unions involved. In effect, the absence control system thereby provided an annual renewal for each employee of legitimate discretionary absence days on July 1 of each year. (e) As employees worked and the system ran, many employees prematurely used up their annual bank of casual leave and then began to accumulate incidents of non-legitimate absence. At that point, the corrective discipline system began to operate. Each successive incident resulted in a report to the supervisor advising (1) what step in the dis- cipline process had been reached, and (2) what specific action must now be taken. In addition to disciplinary action, an incentive bonus of about $100.00 was awarded to each employee whose absence level was still low enough to qualify at the end of the first year of the program. After two years of operation, the program was still running strongly and enjoyed continued supervisory support right up to the top management level. At that time the data for this present research were c0pied and released for anal- ysis. Objectives The central project in this dissertation has been designed to evaluate the impact of an absence control pro- gram on employee absence behavior. Four objectives had to be addressed in the process in order to build a clear chain of research procedure linking the raw data into a focused body of evidence that bears on each substantive hypothesis. Procedural Objectives: 1. Reliable and sensitive measures for the study of absence had to be created and tested empirically, both in terms of their reliability and for overall sensitivity in use to both long-term changes and intra-annual changes. 2. A controversy in the literature regarding the study of absence frequency data in preference to absence severity (or duration) was examined using the derived measures. It was hypothesized that when an appropriate test of reliability was used, severity measures would not be less reliable than frequency measures. Substantive Objectives: 3. Several substantive concerns about hypothesized shifts in absence behavior from one form to another in the presence of absence control programs were investigated. This was done so that any such shifts could either be distinguished from expected overall long-term forms of program impact or identified as unintended but signifi- cant program related changes. 6 If absence behavior shifted from short-term to long-term absences, showing progressively more severe forms of withdrawal over time (regardless of whether the short-term absences involved were legit- imate (allowed) absences, this would imply that the program induced. a. progression, effect. at a group level. If absence behavior shifted from patterns of time- use that were made non-legitimate by the absence control program to patterns that were still legiti- mate, this would indicate a legitimacy effect. Depending on what new forms of absence appeared and to what extent, a legitimacy effect might indicate success or partial failure. Due to the annual renewal of legitimate short-term absence days for all employees, these effects were expected to be restimulated in an annual cycle. If they accurred, there would then have been evidence for annual cycles of "seasonal" shifts as well as for long-term shifts that span the entire post-in- stallation period. Both have importance for theory testing and for program evaluation and absence con- trol policy. Also, the presence of such short-term shifts would have meant that the overall impact of the program could only be determined when the proce- dure controlled for renewal effects. 7 The overall impact of the absence-control program which provided the data for this research was evaluated. The consultants had specifically contracted to reduce short- term non-legitimate absence. Concerns noted above implied that this might have had unintended conse- quences. The progrmn was therefore evaluated both in terms of the extent to which reduced short-term non-1e- gitimate absence occurred, and in terms of what major concomitant changes also took place. Distinctive Features Six features distinguish this dissertation project from prior work. 1. Absenteeism was conceptually and operationally defined here as one form of employee time use behavior. Since the same rationale applies to any other form of employee time use such as time worked on the job, or time scheduled off ... this permitted an "internally compara- ble" set of measures for all forms of employee time use to be created. Due to the- internal comparability of the measures and the unusually high quality and detail of the unprocessed raw data, it was also possible to design procedures that recategorize all the daily values into their correct time-use types as these changed with duration of a given time-use behavior over time. For example, an absence 8 might at first be recorded as a short-term absence with- out leave. If it then persisted for more than five working days, it became a long-term absence. Measure- ment integrity requires that short-term absences which accumulate day-by-day to become long-term absences must then be recategorized from the first day of the occur- rence onward. The logic and sequencing of sorting pro- cedures employed in all such recategorizations done here is described in Chapter III. Once the consistent and corrected classification of employee time-uses over time was provided for, a new kind of time-series measure for both frequencies and durations of the various different employee time uses could then be derived from the cleaned raw daily data. Both 100-day moving values (short-term) and annual mov- ing values were calculated for each type of time use on each and every day of the program, for each employee, and for each selected subpopulation of employees. These derived daily estimates, based on data across either the past 100 days (sensitive to changes within the year) or across the preceding 365 day period (only sensitive to changes from year to year), are, by construction, pro- portionately less radically affected by short-term fluc- tuations than single-day or weekly based measures, and provide more stable estimates of long-term variation in the group time use patterns than the raw daily frequency 9 or severity values would, without being insensitive to daily fluctuation. The derived statistics that were the focus of all subs- tantive inquiry in this research project are primarily group statistics. In no case did any substantive analy- sis rely directly upon estimates of individual time use behavior except for the cross-sectional normality tests and T-tests. The research was unusual but not unique in this since others have used group measures of absence behavior (Chadwick-Jones, Nicholson and Brown, 1982). This group focus was central to two further innovations that are pivotal in subsequent applications of inferen- tial statistics. a. By calculating a daily estimate of the group mean value for each time-use measure, a problem characteristic of individual absence data dis- tributions could be substantially reduced by construction. Absence research. which depends directly upon the absence behavior of individual employees has been characteristically undermined by asymmetric data distributions that are both upwardly skewed and abruptly truncated at the lower end (Hammer and Landau, 1981). Under other circumstances, the asymmetry could be removed by simple transformation procedures, but the severe truncation makes this procedure inap- 10 propriate for individual absence data. Fortu- nately, the group means rarely equal zero and so they do not suffer from a truncation problem. Also, the group and not the individual is the appropriate research focus for both theory development and support of absenteeism control policy decisions. The daily group means also provide a theoreti- cally far more appropriate set of data values than individual values for estimating the reli- ability of the measures over time. Traditional absenteeism research has attempted to test the over-time stability of absenteeism data by repeatedly comparing two sets of absenteeism data values, each taken from the same set of individuals, but at two points in time. That procedure cannot show the over-time stability of absence measurement procedures 'because it is designed so as to discover only the stability of individual absence behavior between times although, if individual absence behavior itself were not variable, there 'would. never be any absenteeism to study. The dissertation used time series correlation of the daily group-means for one randomly selected half of the values for each measure, with the 11 daily group means for the other half over spe- cifically selected periods within the program, as the basis for its major tests of substantive hypotheses. 5. The distributional normality of the individual values used in the construction of T-tests was examined using the Kolmogorov Normality test statistic "D". As a con- sequence, the relative appropriateness of using T-test statistics can be shown. 6. The use of both a 100-day base period and a 365-day base period for calculating derived time-use values allows both a good sensitivity to short-term changes (using the lOO-day base), and an exclusive sensitivity to long-term inter-annual changes using the 365-day base. (Seasonal effects, holidays and intra-annual short-term effects that repeat yearly are simply excluded by construction.) Dissertation Structure The main portion of this dissertation is comprised of six chapters. Chapter I introduces the project and describes the structure and content of the dissertation. Chapter II reviews the literature and relates this to spe- cific hypotheses that were tested on absenteeism measurement and theories of absence behavior. A summary of the central issues, both in measurement and theory of absence behavior is presented on page 27. In Chapter III, the procedures 12 used for data collection in the present research are briefly described, sample demographics are summarized, and the logic and procedures for creating derived measures of absence behavior and other time-use patterns are explained. (See Figure 5 on page 48.) On page 47 these measures are summa- rized in a comprehensive list. Chapter IV briefly describes the research methodology and pmesents the hypotheses that were tested in eaCh step of the dissertation project. On page 57, Figure 6 outlines the sequence of steps included in data analysis and hypothesis testing. Chapter V reports the results of data analysis in both tabular and graphical form, with comments on the purpose and implications of each report. Chapter VI presents a brief summary and a discus- sion of the findings and the conclusions that follow and comments on directions for further research on employee time uses. The appendices to this report present greater detail on both the results and many of the methodological and proce- dural problems resolved in the present work. For reader convenience, a brief summary of these steps has been included in the dissertation itself, with references pro- vided. to the :more detailed. description included. in the appendices. CHAPTER II Literature Review and Explanation of the Hypotheses General Introduction to the Literature Review A discussion of the literature relevant to research on absence measurement and theory is introduced below and has been presented in greater detail throughout this Chapter along with accompanying hypotheses. In essence, two under- lying themes pervade this literature. 1. How can absenteeism be reliably and precisely measured? In recent years, the measurement problem has gained more and more attention. It is the subject of the first step in each phase of hypothesis testing for the present research. Prior to the use of T-tests, the underlying distributions were examined for normality. And prior to the use of time-se- ries correlations, the respective time-series measures were examined for reliability. 2. How does absenteeism behavior actually change over time from one kind of time-use to another? (This theoretical question was first raised by Hill and Trist (1952, 1953, 1962) in work that represents forms of absenteeism as different solutions to an employee's aliena- tion from work, an hypothetical need to with- draw from the job.) 13 14 A list of these themes, including several questions raised by the absenteeism literature and addressed by this dissertation, has been included toward the end of this Chap— ter on page 41. In the work to establish a viable paradigm for absence research, the problems of theoretical conceptu- alization and operational definition have been extensively intertwined, perhaps slowing the evolution of both. These issues have been the focus for specific treatment in Chapter IV through Chapter VI, and in the appendices. waever, a theory—based phrasing of all hypotheses tested in the pres- ent work has been prepared in the terms of their respective theoretical foundation. Throughout the review of literature presented. below, those theory-based articulations of research hypotheses have been presented. These appear in the order that their respective value to this research was established within the present literature review. The cross-references which accompany these theory-based hypoth- eses refer to the related but operationally defined hypoth— eses articulated in Chapter IV. A conceptual consistency in integrating these theory- based hypotheses with the procedural hypotheses had to be established in terms of the actual research setting. To support this link within the present research methodology, hypotheses concerning the concept of "annual renewal" of the employees' casual leave days have also been presented here. To establish a common ground for later discussion of the 15 literature, the disciplinary practices in employment rela- tions and the various measures of workforce absenteeism that are already established as conventions within the field have been summarized below. Practices in Employment Relations Which Relate to Absence Management and Control The discipline and labor arbitration practices in labor relations require that rules be enforced consistently, across all unionized employees, and without exceptions (Elkouri and. Elkouri, 1974). A rule must apply and Ibe applied to all employees to the same extent and in the same way. Furthermore, under Title VII of the Civil Rights Act of 1964 and related court decisions, the employer may not discriminate 111 employment against protected minority employees on an idiosyncratic and arbitrary basis. Just cause and fair and equitable policy and practice must be demonstrated. The application of discipline to control absence is an action which directly affects an employee's career opportunities, and it falls under the jurisdiction of laws protecting the rights of workers in employment. Accordingly, it is necessary that an absence control program implement control practices in a way that affects all employees in the same way and to the same extent. That is, the rules must be consistently applied in all cases. 16 Yet, consistent application of the rules to all absence occurrences, allowing corrective discipline procedures to be applied, requires that the company have a full and detailed knowledge at all times of each employee's absenteeism record. Prior to discipline, a supervisor must know exactly how many legitimate absence days the employee may still claim before corrective discipline begins, and which stage of corrective discipline the employee has reached, and what action must be taken the next time the employee violates the rules for absence control. In any company of even moderate size, or where employees do not always have the same super- visor, this detailed knowledge will not be available unless a systematic effort is made to provide it. Furthermore, since an absence control and discipline policy is vulnerable and can be undermined by allowing indi- vidual exceptions, i1: must be constructed.tx> exclude such exceptions and to provide a workable and reasonable policy for all employees as a group. As a result, the proper focus of program design and evaluation is at the level of the group, while those policies must be continuously implemented by uniform enforcement based upon detailed and current information at the level of the individual. Conventional Measures 9f Absence Absence has been operationally defined in various ways in the literature. Throughout these manifestations, three 17 forms adopted by the U.S. Bureau of Labor Statistics (B.L.S.) have achieved general acceptance. 1. The absence rate (Hedges, 1977; Miner, 1977; Taylor, 1981) is a percentage based on the duration of absence over time. The typical form is (absence rate) = (number of hours lost to absence per month) divided by (average number of employees multiplied by number of hours usu- ally worked), all multiplied by 100. The incidence rate (Hedges, 1977; Taylor, 1981) also is a percentage, but this one indicates overall frequency of employee absence in any given unit of time. The typ- ical form is (incidence rate) = (number' of workers absent) divided by (total employed), multiplied by (100). The time period used as a base is usually one week, although lOO—days, lOO-weeks, or one year have also been used. The severity rate is also a duration-based percentage rate, but this one is a percentage of time absent in a given period for the absent employee group alone. The term "severity" itself has been established in use by the absence literature to refer to severity or duration of absence as a form of withdrawal. Increased severity means longer duration of absence. The typical form is (severity rate) = (average number of hours lost by absent workers) divided by (average number of hours scheduled for work by absent workers) multiplied by (100). 18 These three standand B.L.S. measures of absence have been adopted and extended by Chadwick-Jones, Nicholson & Brown (1982) in a recent reconceptualization of research on absenteeism. Their extensions are of two kinds. The less pervasive innovation was the presentation of several meas- ures custom-tailored to address specific substantive issues. These are as follows. (a) The Blue-Monday Index (BMI) Difference Between Total Absence Levels on Monday and Friday BMI = --------------------------------------- (per week) Number of Employees in Sample (b) The Worst-Day Index (WDI) Difference Between Total Absence Levels on Best and Worst Days WDI: ——————————————————————————————————————— (per week) Number of Employees in Sample (c) Short-Term Index (STI) Total Number of One- and Two-Day Absences Starting in the Week STI = ----------------------------------------- (per week) Number of Employees in Sample (d) Percentage of "Cleans" of Total (CLEANS) Number of Employees With No One- or Two- Day Absences Per Year CLEANS = ---------------------------------------- (per week) Number of Employees in Sample (e) Mean Length of Absence in Days (MLA) Days Lost Due to Absence MLA = ------------------------------ Frequency of Absence Incidents 100 100 100 100 19 (f) Labor Turnover Rate (LTR) Number of Employees Leaving Per Year LTR = ------------------------------------ X 100 Number of Employees on the Payroll (g) Accident Absence Rate (AAR) Number of Accidents at Work That AAR = Caused a 3 Day or More Absence, Per 100 Employees Per 100 Weeks Although these seven measures are intriguing, nonethe- less, the second innovation in measurement presented by Chadwick-Jones, Nicholson & Brown has far more sweeping implications. In this case, the researchers varied not the substantive focus of measurement, but the time interval across which the observed behavior was to be aggregated. Three time intervals were used: a. rate per week b. rate per year c. rate per 100 weeks The resulting derived data allowed analysis of absenteeism patterns by comparison of over-time plots based on weekly rate levels. This focus on Time-Series analysis is a funda- mental requirement for research that asks questions about over-time shifts in the patterns of absence behavior. The use of several different spans of time as the base period for measurement also allows an "adjustment of focus" for the research to discover events that may be either broader in scope or greater in detail than a single "focus" can illumi- 20 nate. In the present research, the time-series measurement thrust has been expanded in two ways. First, the precision of' measurement was further refined from weekly to even shorter, daily and hourly intervals, and second, the time spans across which observations were to be aggregated were constructed. to use moving values based on. both 100-day totals and 365-day totals. The resulting database is suit- able for addressing the substantive and methodological issues posed. in the current research. It is also, by design, sensitive to daily changes which should enable it to reveal even the weekly patterns‘that the custom-tailored measures presented by Chadwick-Jones, Nicholson & Brown were derived to discover. Details are presented in Chapter III and further explained in Appendix B. Finally, the work of Glasser (1970) suggests that some measure of "time on the job in between absences" would be very helpful in distinguishing time-use by the non-absence- prone employee group from that of the absence-prone group. It would also help clarify where the employees' work or absence time is being shifted to due to program induced changes or substitution of one form of absence for another. A discussion of theories concerning how and why absence behaviors change over time follows the comments on subgroup differences below. 21 Methodological Issues The mainstream operational definition and conceptuali- zation of absence research has predominantly been an intra- individual one (Hill and Trist, 1953; Steers and Rhodes, 1978, 1980; Nicholson and Goodge, 1976; Breaugh, 1981; Ham- mer and Landau, 1981). Both the design of the measures and sampling procedures, and the use of cross-sectional tests of over-time reliability reflect this stance. From this intra- individual focus stem a number of problems for the tradi- tional research paradigm. Since those problems and the testing of means to overcome or aovid them constitute a cen- tral theme in the present work, literature and hypotheses concerning both theory and methodology have been reviewed. Various researchers (Lathmn & Pursell, 1975; Steers & Rhodes, 1978 and 1980; Breaugh, 1981; Hammer & Landau, 1981; Terborg, Less, Smith, Davis & Turbin, 1982; Chadwick-Jones, Nicholson & Brown, 1982; and numerous others) have noted the superior over-time test-retest reliability of absence fre- quency over absence severity data. Severity is a term used to describe the length of absence "spells" while frequency means how often these episodes occur. In brief, the corre- lations previously reported between sets of absence fre- quency data for a group of individuals at time 1 and again later at time 2 are almost without exception considerably higher than over-time correlations between datasets for the 22 same employees that are measures of absence duration. In one response to this instability, Latham & Pursell (1975, 1977) have proposed a change in focus from absence to atten- dance. Apparently, attendance statistics are more reliable (less variable) than duration of absence, and absence sta- tistics are often confounded in data collection by sick leave abuse, etc., so that the real reasons for absence may not be made evident from the customary self reports. This proposal has been challenged by Ilgen (1977), who notes that studying attendance will not tell us about the details of absence behavior and that, in any case, the customary test- retest reliability statistic is at best rarely (and then only under carefully controlled assumptions) an appropriate choice for a measure of absence behavior in which many researchers have reported considerable variation over time (Glasser, 1970; Garrison & Muchinsky, 1977; Nicholson, Jack- son & Holmes, 1978; Breaugh, 1981; Hammer & Landau, 1981; Markham, Dansereau & Alutto, 1982). Responding to the con- troversy in a slightly different approach, Smulders (1980) cites statistics on absence behavior in the Netherlands and notes that using values for frequency over severity would cause a gross underestimation of the seriousness of the absence problem and is, in any case, theoretically inappro- priate where what is of theoretical interest is the absence duration. 23 Another area of absence research (Cheloha & Farr, 1980; Nicholson, Brown & Chadwick-Jones, 1977; Nicholson & Goodge, 1976; Parasuraman, 1982; Mobley, 1982) argues for the iden- tification of specific subcultural groups of employees who tend to be more prone to absence or to attendance, respec- tively. Nicholson, Ehxnni & Chadwick-Jones (1977) specifi- cally promote the construct of "absence-proneness". This group basis for sampling is fundamental to their more recent research on absence (Chadwick-Jones, Nicholson & Brown, 1982). There is also a body of work accumulating which specif- ically examines the statistical properties of absence-prone groups (Glasser, 1970; Behrend & Pecock, 1976; Garrison & Muchinsky, 1977). Glasser applied stochastic techniques from epidemiology to assign employees to absence-prone and non-absence-prone groups for further analysis of their behavior. His work strongly supports the selection of such subcultural samples and also advances another absenteeism measure or index ... "scheduled time worked between absences". Behrend & Pocock (1976) have also noted clear trends in absence-proneness over time and problems of skew- ness in the data distribution. Another body of work which addresses absence-proneness (Garrison & Muchinsky, 1977) strongly supports the construct of absence-proneness but points out that membership in the 24 absence-prone group tends to change over time and may be a relatively short-lived condition at the level of the indi- vidual employee. (A substantial change in group composition over a 21 month period was reported.) This suggests that group composition should be re-established for each time period across which there may be a substantial shift in mem- bership. Finally, work by Hammer & Landau (1981) notes that in addition to serious unreliability in the severity measures, the absenteeism data distributions also suffer from charac- teristic problems of truncation at the lower end of the dis- tribution (a large cluster of very low absence employees), and from severe positive skewness due to the inclusion of a relatively few atypically large extreme values. This extreme value problem has been noted by others. Ilgen & Hollenbeck (1977) noted the _low reliability of absence severity data and censored (threw out) the extremely long sick leave absences to improve the symmetry of the distribu- tion and reliability of the measure. It can also be argued that both the truncation concern and the extreme value prob- lem are matters that stem from including in the sample of absence-prone employees, other employees: (a) who» don't belong because they are not absent enough to belong ... to be members of the absence-prone group, or (b) who don't belong because they are not present at work enough to truly influence or be influenced by membership in the absence- 25 prone group. In short, there are reasons of both theoreti- cal consistency and methodological rigor for separating these types of extreme values from the absence-prone group. Unfortunately, as the research of Garrison and Muchinsky (1977) indicates, such a selection process must be redone on a very frequent basis if the research is not to artificially exclude newly absent-prone workers or include others who have, in fact, "reformed". In the balance, it is conceptually reasonable to define absence as a condition of limited participation in employ- ment, characterized by neither consistent attendance nor extreme non-attendance, which are respectively the lower and upper limits on the range of employee time-use as it relates to absenteeism. Hammer & Landau (1981) have noted procedures for trans- forming data to reduce non-normality in data distributions due to the inevitable presence of extreme values in the dataset. But they recommended instead the use of Tobin's Probit analysis procedure, due to the severity of typical truncation problems. The data transformation procedures mentioned. by Hammer and Landau are now' customary tools (Tukey, 1977) for improving the symmetry of distribution in datasets used in the natural sciences. If the problem of truncation at the lower end due to directly relying on data from non-absence prone employees could be overcome, such 26 methods could and possibly should be applied to absence data. The group-focus which is basic to the time-series analyses in this present work conveniently circumvents the truncation problem by construction, except in the T-test data which remain cross sections of individuals' data and not group data. It also operates to drastically reduce the effect of extreme values on the group level derived measures used for time-series correlation. Consequently, testing the effect of such transformation has been left for later work. Yet, to analyze only data for absent employees would involve rejection of the information about related and alternative time-uses contained in the data for the excluded extreme sub-groups. Also, the analysis would represent only the respective subgroup, not the workforce which is the unit of theoretical and practical interest. An alternative pro- cedure was employed to evaluate change in absence behavior in the present research. The data for all employees were categorically sorted and aggregated across the sample group 1, and this never allows reconstruction of either the raw data or the subseq- uent summary statistics by transformation back toward the 29 original state. Use of frequency over severity is the least conservative power transformation that could be selected, removing along with some measurement error most of the information in the raw data. The stronger reliabilities which have traditionally been observed for absence frequen- cies as opposed to severities are partially indicative of whatever remains of the information in time-use severities after transformation to the zero power has occurred. Since most of tine observed variation tends tx> get removed along with that process of transformation, the correlations between two cross-sections of frequency data ought to be higher by construction. From a procedural standpoint, another problem also undermines the use of cross-sectional over-time comparisons to test reliability. Variation both from measurement or sampling error and from real changes in the absence behavior over time is grouped together in the respective cross sec- tions. Any differences noted between the two cross-sections must therefore be a combined effect of both measurement or sampling error and change in the observed behavior. And, these two contributing component effects are undistinguisha- bly combined with each other in the cross-sectional data so that any real changes observed over time will operate directly to reduce the apparent reliability -- the correla- tion between measures of respective individuals' absence behaviors over time. Once again, the apparently low reli- 30 ability statistic is to be expected -- an artifact of the analytical. procedure rather than. the measure of absence itself. To circumvent this problem, the present procedure has employed a time-series approach both to measurement itself and to analysis, instead of only using the traditional com- parisons of two daily cross-sections. As a result, the reliability of the measures over-time may be ascertained using split-half correlations between two time-series of group means. These means represent the daily values of time-use behavior for randomly selected halves of a sample of workers. The reliability statistics generated can be affected by sampling error and by measurement error, but will not be undermined by the inclusion of observed "real" changes along with measurement error in the reliability data. Details on the time-series correlation procedure are presented in Chapter IV and in Appendix C. Given the construction of the time-series measures used here, the concerns raised above about reliability could then be addressed through specific tests of the new time-series measures. This was done by testing the following two hypotheses. Hypotheses: The time-series correlations between the two sub- sample time-series of group means for both sever- 31 ity measures and frequency measures will be sig- nificant and positive and approach r = 1.0. (See Chapter IV, Hypothesis 10.) and, Neither frequency nor severity measures demon- strated a substantially greater time-series reli- ability than the other. Both were reliable meas- ures. (See Chapter IV, Hypothesis lO-a.) Variation Due to Demographic Subgroup Differences In a given workforce, a number of systematic variations in absence behavior attributable to population demographics of employees can be expected to occur. Mowday and Spencer (1981) noted differences in employee absence behavior asso- ciated with type of work and scope of job occupation. Steers and Rhodes (1978, 1980) offer a broad-based, theoret- ical model including a very wide selection of plausible demographic influences. Chadwick-Jones, Nicholson and Brown (1982) suggest that subsample groups selected on the basis of industry, sex, age, type of occupation, and pay practices may all be expected to differ in absence behavior. Further work by Hill and Trist (1955) suggests that length of ser- vice, individual accident rate, and time of shift worked, as well as a number of other issues may also determine employee subgroups that are characteristically different in absence behavior. 32 Demographic variations within absence behavior are of interest both because these would allow the effectiveness of measures to be examined under varying subcultural conditions and because analysis within these subgroups, exclusive of other employees, should result in less "error" of measure- ment. Nonetheless, comparative analysis by demographic sub- groups has been deferred for detailed attention in later work. The random selection of a sample to represent all employees in the IXH? workforce jpresumably precluded. any undue and excessive influence from over-representation of demographic subgroups. A comparison of sample demographics with the workforce demographics has been presented in Chap- ter III, Figure 2. Preprogram Shifts in Employee Time Use Since the absence control program included a heavy com- ponent; of' management. development and supervisor training prior to the start-up of the system, some anticipatory changes in absence levels were expected to occur prior to full enforcement of the control program. To check for and discover the general direction of such preprogram changes, the time-use data for the first 30 days of the baseline year were compared with the data for 30 days directly preceding full start-up of the program in its first complete year. What was really expected was a reduction of short-term dis- cretionary absence, but since backlash effects due to 33 employees taking advantage of a last chance to "get away with" absenteeism was another possible outcome, the follow- ing hypothesis was tested -- as a null hypothesis. Null Hypothesis: H0: There were no significant changes in any employee time-uses prior to program installa- tion. In particular, absenteeism levels will have remained steady. (See Chapter IV, Hypothesis 2.) or, H1: Significant preprogram changes in absenteeism occurred before full installation. Since the two time periods involved were both in the early summer and prior to full program imple- mentation, seasonal changes and other cycli- cal effects were not expected to be substan- tiated. More specific details have been presented in Chapter IV. Patterns of Cyclical Change in Absence Behavior Various studies on absenteeism have noted the need for longitudinal research (Markham, Dansereau & Adutto, 1982a, 1982b; Nicholson, Jackson & Howes, 1978; Breaugh, 1981; Ham- mer & Landau, 1981). To date, most such work has been con- ducted by using multiple cross-sectional comparisons and graphical plots of change over time. These procedures were 34 extended by Chadwick-Jones, Nicholson & Brown (1982) in detailed week-by-week graphical comparisons which reveal definite cyclical and .systematic time-series ‘patterns of change over time. The investigation of these phenomena, using precise and reliable day-by-day measures in a systematic time-series analysis is a necessary step in studying absence. There has been evidence presented by many p"ior researchers which did lead to an expectation of seasonal shifts and holiday effects in the present research, and which suggested a recurrence of such behavior in annual cycles. Where the research methodology depends on comparing daily cross-sec- tions (as in the T-tests here), these cyclical changes could easily be incorrectly incorporated into estimates of overall change. So, it was appropriate to examine the data for var- ious types of seasonal shifts prior to further analysis. Furthermore, the absence control. jprogram itself included an annual renewal of the number of casual leave and vacation days available to each employee, and this happened every year on July 1. As a result, the occurrence of cycli- cal effects was all the more likely to follow a pattern of annual renewal. Also, the seasonal and. holiday' effects noted above seemed very likely to be compounded by other short-term intra-annual fluctuations due to the unusually strong observed substitution effects discussed below. 35 Three general types of hypotheses about cyclical change were tested using T-test procedures between carefully selected points in time. But, because of the empirical con- ditions under which raw data were originally generated, the operational hypotheses presented in Chapter IV had to test for some of the above effects as combination effects, could only test for purely seasonal effects within the baseline year, and could test for annual renewal effects exclusively across periods within the post installation years. In short, the "theoretical hypotheses" below don't necessarily correspond directly to "operational hypotheses" as tested (see Chapter IV). In some cases, combined effects of two or more different cyclical shifts had to be made the focus of inquiry. Specific details of the circumstances surrounding the careful selection of daily cross-sections to be T-tested have been explained in Chapter IV. The following three theoretical hypotheses about cyclical changes were investi- gated, not to find all instances of occurrence, but to get some guidance for refining the interpretation of later over- all changes attributable to the program itself. Hypotheses: (See Chapter IV): Absenteeism levels will show seasonal differences, increasing in midwinter and midsummer, but decreasing in early spring and fall (Hypothesis 3, Hypotheses 5 and 8). And, 36 Absenteesim will tend to have increased at holiday periods in late December and early July due to a recurring holiday effect (Hypothesis 4, Hypotheses 5 and 7). And, Due to program installation and a legitimacy effect (discussed below), post installation levels of short-term non-legitimate absence will have been significantly higher prior to annual renewal rather than after it; and short-term legitimate absence will have increased after annual renewal (Hypothesis 8). Theories of Unintentionally Induced Over-Time Change in Absence BEhavior: Internal Substitution and Progression Effects Theories about internal shifts or substitutions between different forms of absence (an internal "loophole" through which employees could very possibly find "safe" ways to con- tinue their absenteeism and which would dilute the desired impact of an absence control program) have often been advanced in the literature (Hill & Trist, 1953, 1955, 1962; Chadwick-Jones, Nicholson 6: Brown, 1982; Nicholson, 1976; Braun, 1978). Two of these together, the legitimacy hypothesis and the progression hypothesis, posit shifts of absence behavior from illegitimate short-term to increas- ingly legitimate longer term. absence behavior under the 37 impact of an absence control policy. Substitution of one form of withdrawal for another has been discussed in terms of: accidents and sick leave substituting for regular work (Hill & Trist, 1953); absence plus overtime work for regular work (Allen, 1981); absence instead of other forms of with- drawal (Mobley, 1982); absence prior to turnover; and tardi- ness or sabotage or drug abuse or turnover as substitutes to absence (Parasuraman, 1982). Progression: Early theories (Hill and Trist, 1953; Herzberg, Maus- ner, Snyderman & Capwell, 1957) posit a fairly simple pro- gression of individual employee absence toward increasingly severe withdrawal from work. The focus in those studies was on the progression of individual employees into increasingly severe forms of withdrawal and not on a group-wide progres- sion effect. As a consequence, the time-series procedures in the present research could only have discovered support for the progression effect if the entire sample showed a general tendency toward increased withdrawal from work. Since that would have forced the DOT to increase their work- force, this was not likely. But, so that it would be possi- ble to check for such a group-level progression effect, the following hypothesis was tested. 38 Hypothesis: A pattern of progression from short-term legiti- mate absence into long-term legitimate absence will have occurred at the group level. A negative and significant correlation between these two forms of absence will be found when the two series of daily means are correlated (see Chapter IV, Hypothesis 11). Legitimacy: More recent work (Parasuraman, 1982; Hammer & Landau, 1981; and Nicholson, Jackson & Howes, 1978) posits a pro- gression in the pattern of substitution from uncertified and non-legitimate but less serious absence to certified and legitimate absence, and to longer term and more serious forms of withdrawal including turnover. The substitution effects would tend to occur in the same direction as the progression effects under the implementation of absence con- trol. A shift of employee time-use toward more severe but legitimate forms of absence from short-term non-legitimate absence could very likely occur. In the absence of specific practices to prevent these substitutions, such internal shifts are a highly plausible side effect of absence con- trol. Given daily precision of measurement and a high level of reliability in the measures, it was possible to distin- guish here between these two patterns in a few selected spe- 39 cific comparisons, either on the basis of timing of the behavioral shifts between the various measures across two cross-sections (the T-tests to compare means), or by using a time-series correlation technique. Procedures to do this have been briefly discussed in Chapter IV with a more detailed methodological note in Appendix C. The results are presented in Chapter V. The legitimacy effects were one very plausible overall result of time absence control program, zn1 alternative to working. The T-test procedure offered a simple but workable "estimate" of program impact which could be used to discover overall legitimacy effects under specific circumstances as detailed in Chapter IV. The following hypothesis was there- fore tested. Hypothesis: After program installation, significant year-to- year changes between corresponding dates will have occurred and, in. particular, non-legitimate absence will have decreased while legitimate forms of absence will have increased (Hypothesis 6 and Hypothesis 7). The time-series correlation procedure was specifically tai- lored to allow analysis of shifts within time periods between different forms of time-use. A discussion of the method is presented in Chapter IV and supplemented with fur- ther detail in Appendix C. By means of this procedure the following hypothesis was tested. 40 Hypothesis: After program installation the level of non-legit- imate short-term absence will have tended to decrease and the levels of long-term and short- term legitimate absence will have tended to increase over time. As a result, the time-series of daily mean levels for legitimate and non-legit- imate absence behavior will be negatively and sig- nificantly correlated (Hypotheses 12 and 13). Program Evaluation: An Overall Long-Term Assessment 9f Pro- gram Impact The absence control program was specifically intended to stimulate and control a reduction in short-term non-le- gitimate absence. It was not contractually specified but was generally presumed that this would produce a correspond- ing increase in time worked, Despite the possibility of substitution effects subverting this unspecified but intended. impact of the program, Iboth T-tests ‘using the cross-sectional data and time-series correlations using the time-series of daily mean time-use levels were implemented. The T-tests were designed to discover long-term, year- to-year changes between the baseline year and the post-in- stallation period by comparing mean time-use levels on cor- responding dates. The following hypothesis was T-tested. 41 Hypothesis: Short-term non-legitimate absence in the baseline year will have decreased significantly by the end of the first year after installation, and time worked will have increased in a corresponding amount. The same patterns of change will also have become evident in comparison of the baseline year and the second year after full installation (Hypotheses 6 and 9). To test for the same relationship as a continuing pair of interdependent trends over time, the following hypothesis was tested using the time-series correlation procedure. Hypothesis: After program installation, time-series correla- tions Ibetween short-term.:non-legitimate absence and time worked will be significant and negative (Hypothesis 14). It should be noted with regard to the various hypoth- eses above that the T-tests used to compare means of two daily cross-sections establish only the direction and size of effects brought about by a trend over time and provide an estimate of significance for those effects. The time-series correlations, on the other hand, only establish the strength, direction, and significance of the relationship that produced the effects, and say very little about their 42 magnitude. These two procedures complement each other. To the extent that both help establish the direction of change, there is some redundancy. A Summary 9f Theoretical and Measurement Issues From the above discussion, a number of key points may be drawn to summarize the issues raised from analysis of the absenteesim literature itself. 1. There are internal trade-offs between absenteeism and other forms of employee time-use, such as time worked between absences and scheduled time off. Spending time in any one of these activities excludes doing the oth- ers. Time-use behaviors such as absenteeism undergo patterns of continuous change over time so that careful study of them requires a time-series approach instead of a com- parison of cross-sections. There are short-term. weekly’ shifts and longer term cyclical and seasonal shifts in the patterns of time- use behavior that must be identified and separated from inter-annual, long-term changes in order to assess the impact of absenteeism control. Some forms of absence are more legitimate than others and employees tend to be absent for a variety of dif- ferent reasons. There may be a causal relationship between. the legitimacy’ of an absence type and its occurrence . 43 The occurrence of absence may also tend to vary somewhat depending on demographic characteristics of the workforce, notably: type of occupation, sex, age, industry, compensation practices, time of shift worked, length of service, and individual accident rate. Absence control policy decisions require data that describe time-use for the group. Absenteeism disci- pline policies must apply and be applied to all employ- ees equally if enforcement is to be equitable and legally defensible. For comparison with absence figures across samples from different work forces, absence rates have typically been designed as overall summary statistics, reflecting various patterns of group behavior independent of the numbers of employees involved. Duration of the absence, and frequency of occurrence are both popular objects of measurement for employee time-use behavior. There exists a lively controversy as to which is a better measure. Duration is more con- ceptually attractive. Frequency has been recommended as a preferred alternative because it has a higher test-retest reliability correlation. Data from absence behavior tends to distribute itself abnormally. There is a truncated lower end due to the inclusion of values for employees who are never or almost never absent. Theme is a strong positive skew 44 to the distribution which could be easily and conven- iently corrected if the truncation problem were not present. 10. Each employee has a characteristic short-term tendency to be absent-prone or non-absent-prone. This tendency changes a great. deal over time so that accurately determining membership in the absence-prone or non-ab- sence-prone group would require re-establishing the group composition daily. A set of time-use measures which is responsive to these insights suggested by prior research has been developed for the present work and described in detail in Chapter III and in Appendix B. The derived measures used in the disserta- tion are listed and annotated in Figure 14 and Figure 15. These data were used to test operational hypotheses based on the theory-based hypotheses presented above, to examine in turn the respective theories about absence behavior. The methodology and operational hypotheses have been presented at length in Chapter IV, with detailed methodological notes in Appendix C. Results and interpretations are summarized in Chapter V. CHAPTER III Data Collection, Sample Demographics, and the Derived Measures In order to address the conceptual and. measurement issues raised by absence research, a very particular kind of database was required. The details of access and data col- lection for such a sample have described below, followed by a description of the sample demographics. Perhaps the most complex refinement in the present work, and a central ele- ment of the research program, was the conceptualization and creation. of the 'various iderived. time-series measures of time-use by the workforce. In the latter part of Chapter III is a detailed presentation of the conceptual framework and procedures which generate the derived time-use measures. Access 39 the Research Site and Data Collection Procedures During 1977 and 1978, the Department of Transport (DOT) in the City of Detroit contracted with Moore & Juliano, Inc. of Lansing, Michigan (the consultants) to have them create and install a system for absenteeism measurement, reporting, and control. The data for the present research were gener- ated as an integral part of that system. With the coopera- tion of both the Moore & Juliano firm and the DOT, access to data tapes containing the entire database was made possible. The purpose of the present inquiry includes an evaluation of that project. Detailed access was therefore provided to 45 46 gather information about how the control system was installed, how the system for data collection was developed, how the data collection itself was administered, and how the program was implemented. A brief description follows. The Consulting Intervention: From. the beginning, top management in the iDOT was strongly committed to the development and implementation of the absenteeism program. Prior to full-scale program imple- mentation, most supervisory personnel in the DOT were given training in the use of managerial techniques for performance planning, planning-group communication, and group problem- solving. From this base, the managers chose absenteeism as the problem they wished to address. Ultimately, the con- tract to install an absence reduction program was estab- lished. In the process of working together to diagnose absenteeism problems through interviews, nominal groups, and historical research, a detailed, guided exploration and cat- egorization of their absenteeism problems occurred. From this base, through a hierarchical system of structured prob- lem-solving workshops, the DOT Personnel Department and the consultants determined a detailed set of types of employee time-use (or absence) for all DOT employees. This typology was to be used later in coding the daily reasons for time- use of all employees on the DOT payroll, every day of the program as part of the reporting and control system and as 47 part of the research effort to discover the extent of the problem as it evolved over time. In many cases the data collected were already in exis- tence for payroll purposes so that care and accuracy in data collection were mandated by law. For other types of time- use, particularly in. the detailed. absence categories, a technical classification decision might occasionally be required from the DOT about how to assign each specific day's worktime where some non-routine employee activity was involved. It was important to avoid forcing supervisors into a judgement-call position where they would be responsi- ble for an on-the-spot determination of how to interpret the rules regarding an employee's time-use. The earlier train- ing and problem-solving program had created a communication channel for propagating the necessary message. Supervisors were trained to recognize and make only the routine judg- ments already anticipated and defined in the context of the absence control program. They were to defer the non-routine decisions 1x: the personnel department. Detailed technical training of the personnel staff and payroll clerks was pro- vided. There, judgments could be made in a way consistent with the overall absence control policy and contractual com- mitments with the various affected unions. Indeed, the absence-control program was actually worked out with these unions before implementation, although man- 48 agement never solicited public support due to the political nature of unions. All unions accepted the program without grievances. The personnel department centralized the absence control policy decisions under the control of an absenteeism program coordinator who was directly responsible to the supervisor of the payroll and personnel departments, and who was trained in policy, labor relations, data entry, and related aspects of policy implementation. In addition to pre-program training and centralized administration of policy-making decisions, the program pro- vided weekly computer reports on departmental absenteeism directly to the department supervisors. Other reports of a monthly, quarterly, and yearly nature were provided to the personnel department, line :managers, and. top executives. Each supervisor was required to implement counseling and disciplinary procedures such as warnings and disciplinary days off as soon as these became necessary due to evidence of excessive, non-legitimate absenteeism for any given employee. Steps in the policy were based on the historical research study conducted by the consultants which determined average absence rates (days absent per year or month) and inactivity rates (percentages of the workforce absent per year or month). These are listed by organizational unit in Figure 1 below, along with a tabular summary of absence con- trol procedure. Steps in the jpolicy occurred. when the employee hit various percentage levels (i.e., Step Three at 49 50%, Step Four at 100%, Step Five at 150%, termination at 200%). A continuous track record was maintained for each employee and the record of violations was wiped clean every year on a rolling one-year basis. Each July 1, employees received a fresh bank of casual leave and vacation days as provided for by collective bargaining agreements. Based on each employee's track record, both disciplinary measures and an attendance bonus were awarded. The overall effect of all these procedures ‘was an unusually consistent and uniform implementation of both data collection and absence control. Accordingly, the data from July 1980 onward contain little or no contamination due to supervisors parlaying discipline decisions into political and personal favors, a problem present in the archived base- line data. These issues were structured out of the supervi- sor/subordinate relationship» and. out. of the supervisor's role in every way it could be done by utilization of time reporting systems and specially trained data entry person- nel. Careful pre-planning to prepare for data collection and training of the supervisors was vital. The vast bulk of time-use data were recorded directly from the payroll cards particularly for the baseline year (July 1, 1978 to June 30, 1979). 50 coma mu.~ «a. nozuo counuoouoocn .ocswasn can once: nm.n ~v.~ amus< mn.~ h~.~ acauuuuawcwaut 783m 2233:. so» Sou 035 85.0. cum moon» new ouuwueouoou ma aofiuoucououmou sodas. om.m m~.v descenuom vv.h ea.n oocdcoucuuz new» undue .um: lance an scamw>ao Anewmw>ao Houucou .puooo .cwcuooo . . HWCCOmuom. Hoccomuuev .ccouu4 .cmuu< meant we HA 5 v acacdnout Emacs: 335: a 933 £32 . Houucoo Hauucoo mouucou .331, a .25.; L3 . housewoucpozv .ucouuc ooaavcouu< .ucmuu< uuoaam uuuasm uoad>uoasm coxoe cognac pm.oa n . xooum gasses O O O owe woos sow .omuuo)x a a pd . aoo~ coma unaduOumM: 909 no ucoouom o.a ha.v sawuw>ao noduauuoomcnua .mwuuouwwm om.a mo.v sawnw>uo.ncuue usuaumsflpd ucapsaoxm wouozomwn oococcouuc undead: nodusoo a o a mcwccom aneuopcdx a couuwuz magnum: oouuoz o>soa “pouusomm . scamcoousm scancoousm vcooom :ouuuuz ~ moum dosuau acuuu< ovum ma.c~ n v .awuwa >00 huuwna a o a "haemaua «use: .uoo» nod a «ouch cod owe om ov oo>o~oem nod m>uov ovum use: unsound coumsonxm mama >uu>duoncn 3 3 m 33333-5: 233 3003.3. 32533-5: 00392 00302.. ow uuuwm Hosanna oucmuusooo snow ocaaomom onu meauso o n v m ~ A “doum nomouo>4 Macauouawz Houucoo oosopcouut The City of Detroit Department of Transportation Attendance Control Operations Guide 1: Figure 51 The reasons for unscheduled absence were limited to a fixed set of codes and each event was recorded by the employee's supervisor if the situation was routine and clearly required no judgement calls. Otherwise these decisions were referred to the program coordinator, primarily to preserve consis- tency for enforcement of the related disciplinary measures wherever use of discipline became necessary. Despite all the care given to proper coding of the time-use data, two kinds of irregularities did occur in the records of certain demographic data taken from the payroll cards. Some of the demographic ‘data that were collected were miscoded such that those categories of data were ren- dered unusable. Also, some of the demographics, including both salary and hourly wage data, were stored on the new data tapes in a machine-specific compressed format and are not easily accessible outside of the DOT system. Fortu- nately, the portion of demographic data which was coded in usable form is complete enough to provide an excellent sam- pling frame for the present research. More details on the demographics are presented below. The same type of irregularity in the data coding affected certain categories of the time-use data itself, especially in the baseline period (1978-1979). Again the effect was consistent. But, although its presence could be identified, this type of irregularity could not always be 52 pinned down to one unique reason for occurrance. So, in the program which generates the derived dataset, a procedure has been included to recode those ambiguous data as "missing values." Preparing the Time-Use Database A brief summary of procedures for recoding the raw data and deriving the time-use indices has been presented later in this Chapter, and a detailed explanation is included in Appendix B. In essence, there was an extreme abundance of raw data made available for the pmesent work ... some ten million characters of data stored in a datastructure which implicitly made use of the order of storage to keep track of the date to which each data value referred. These data had to be cleaned to correct any systematic miscoding, and then had to be reorganized as follows. The demographic data were separated out from the time- use data and edited to eliminate redundancy, and to find (and where possible, correct) any errors or omissions. The data on absence type and absence severity and date (time-use data) were also split out into a separate database. In the process, the date information was used to calculate a rela- tive date for each day's data, starting with day 1 on Janu- ary 1, 1978. At this time, the social security number for each employee was replaced with a four digit case label to preserve confidentiality, and this was included as a sorting 53 field in both the demographic data file and the file on rel- ative date, absence type and severity. Only those demo- graphics to be used in further research were included. Fur- thermore, the new case label, relative date, absence (or time-use) type, and absence severity were all written out to a separate tape file. This new file is unpacked, restructured so it can be used with more flexibility in the data analysis. It con- tains all the raw data and its respective relative date information, and.iJ: is bigger than the cmiginal, about 30 million characters in size instead of 10 million, even with- out the demographic data. Yet this absence data was still in very raw form. Much sorting and recategorizing, as well as the calculation of annual moving values for each measure on each time-use index, each day, had yet to be done. The rationale and procedures for creating derived data struc- tures have been explained briefly following the descriptions of sampling jprocedures and. workforce demographics Zbelow. For convenience of comparison, the sample demographics have been presented together with the workforce demographics in Figure 2. The sheer size of the detailed database for the entire workforcce and the amount of computer memory and pro- cessing time required to sort and prepare each case for the derived database made it necessary to analylze a representa- tive sample of the cases rather than the entire workforce dataset. As a result, the following sampling procedure was implemented. 54 The Sampling Procedure Selection of a sample had to satisfy a number of requirements. (1) The sample had to be randomly selected by a procedure which gave every one of the 2336 cases an equal chance of being selected into either half of the split sam- ple. (2) The two split halves had to be randomly selected at the same time such that both included about the same num- ber of cases and the total number of cases desired was about 160 (a rough estimate of the maximum number of cases the computer could effectiveLy process). (3) To maximize the amount of independent information brought from the overall workforce data and selected into the sample, no case was considered for inclusion more than once. Sampling was done without replacement. (4) For convenience in passing the resulting file of selected cases to the program which exe- cuted the next step in analysis, the data had to be kept as a fortran data file. The simplest tool to do this turned out to be a small, custom-designed Fortran program. The sampling program actually worked as follows. The computer read in three values given it by the fortran pro- gram: a randomly chosen large odd number, a value for the number of random choices the computer was to make (2336), and a sampling fraction (160/2336 in the present instance), which the computer was to use in deciding to include or reject a case from the sample. The program then used a for- 55 tran subroutine (called GGUBS) from the International Mathe- matical and Scientific Library (a bank of standard computer programs that the Fortran language users can include in their own software to solve special but commonly encountered problems). The subroutine which was used in this case took the three values given to it and generated 2336 numbers between 0 and 1.0, drawn from a uniform random distribution. Then the computer read in the first case of data from the demographic data file, compared the first random. number (between 0 and 1.0) with the sampling fraction (160/2336 to get about 160 cases out of the entire 2336 cases), and if the random number was less than or equal to the sampling fraction, the computer included that first case in the sam- ple. If the random number was greater than the value of the sampling fraction, the case was then omitted from the sample and the computer tested the second random number against the sampling fraction to decide whether to include or omit the second case. Then it tested the third random number and decided about the third case, etc. At the end of the pro- cess a total of 164 cases had been selected for inclusion. The four extra cases (164 rather than only 160 as planned) were presumably generated as a consequence of chance varia- tions in the values generated by the random number genera- tor. By this procedure the 164 case sample was selected. At the same time, the overall sample was split into two sub- 56 sets. In effect, the first case to be included was written onto a file for the first subset, then a second case was randomly selected from the cases that had not yet been con- sidered for inclusion and this second case was written into the second subset. Then a third case was selected from the remaining cases and written into the first subset, etc. In each case selection, no case was considered for inclusion which had already had a chance to be included in either half (or subset) of the sample. After all 2336 cases had been randomly selected and included in each subset, and since no case could be present in both subsets, the entire sample was composed of 164 unique randomly selected cases. The repre- sentativity of this 164 case sample has been examined by comparing demographics for those cases with demographics for the 2336 cases in the entire DOT workforce. Results are summarized in Figure 2. DOT Workforce Demographics The Department of Transport is a predominantly male, black workforce composed of about 60% bus drivers, 20% mechanics, and 20% distributed across various other occupa- tions. The total number of employees is variable, but over the course of the present research, 2336 employees partici- pated. A detailed summary of demographic characteristics was prepared for both the overall workforce and the sample drawn to represent the entire workforce. Those details are 57 presented in Appendix A. Both have been summarized below in tabular form to allow convenient comparison of the sample with the workforce characteristics. I I I Value for All Sample I I Characteristic Employees Value I l l I Number of Employees 2336.0 164.0 I I Median Seniority Year 1974 1974 I I Modal Seniority Year 1979 1979 I I Median Age in 1982 37 38 I I Modal Age in 1982 32 32 I I % Transportation Equipment | I Operators 57 55 I I % of Single Employees 83.0 83.5 I I % of Employees with no Dependents 68 73 I I % of Workforce by Race and Sex I | % Black 79.5 78.0 I | % White 18.9 22.0 | | % Male 84.8 87.2 I I I Figure 2: DOT Employees' Demographic Characteristics, For All Employees and for the Research Sample As Figure 2 indicates, demographic characteristics for the sample are generally quite similar to those for the workforce (Hi the characteristics listed. Furthermore, as the differences between the median and modal values for age and seniority year indicate, both the sample and the overall workforce are going through changes in composition, of the same direction and magnitude. 58 Procedures 1:2 Generate the Derived Time-Use I_n_dices (See Appendix B for more detail.) The raw absence data contains daily values for the case number, relative date, absence type, and "absence" severity in hours. The first two values are assigned to the second two and identify to which employee and on what day those last two data values relate. The meanings of these absence data values are dependent on several things. 1. Of primary importance is the reason for absence (or worked time) which was originally coded for the specific employee on a given day. 2. Second, the duration of the occurrence of the time- use over time may persist until it forces a recoding of the entire occurrence or at least of part of it. Indeed, the duration of an absence episode may initially be understood to be short and yet later may turn out to last for weeks, or even months. Reassignment of such time-use occurrences to corrected categories from their onset is the first of two main functions served by the Fortran program which generated the derived database. 3. Third in importance but critical to the integrity of the derived data are a number of systematic coding errors which had to be removed from the data before they could be interpreted and used. 59 The logical structure of relations between absence rea- sons has been used to create a pmogram that converts each primitive "absence reason" (Figure 3) and duration coded in the raw data into derived measures (Figure 5) for one or more of the 13 time-use indices. Figure 4 is a flowchart of the recategorization procedures. Data values listed in Fig- ure 3 were resorted according to a set of recategorization rules to determine the values of derived data items listed in Figure 5. The recategorization rules used have been explained following the flowchart. Read the flowchart (Figure 4) from right to left, fol- lowing the arrows. Note that where the condition associated with a given arrow in the flowchart was satisfied, the data values in the raw data categories (on the right) will have been processed as contributing elements and added into the associated higher-level "resultant" category of time use (on the left). "Resultant" or derived data categories are each marked with bold-faced alphabetic characters in the respec- tive box of the flowchart, and these alphabetic characters correspond to the alphabetic labels associated with each of the derived time-use categories in Figure 5. Similarly, there are bold-faced numeric labels associated with each raw absence or time-use category listed in the flowchart and these numeric labels correspond to the numeric labels for each Raw Absence Data Type listed in Figure 3. 60 In the original data, coding procedures describing absence, time worked or time scheduled off for DOT employees used. 45 reasons to categorize these various time uses. These reasons are listed on the following page in Figure 3. The actual language used to label these reasons is the offi— cial DOT terminology for categories of time use. Rules for Recategorization 3f the 4_5_ Raw Absence Reasons Over Time (Refer 39 Figure 4) The 45 categories of time-use listed as raw absence reasons in Figure 3 above were combined into the 13 derived categories listed in Figure 5 as follows. 1. Any employee time that was recorded as laid off, resigned, discharged for cause, retired, deceased, or trans- ferred, where the employee was not rehired in 365 days or less, remained categorized as termination time. (Box M in Figure 4.) 2. Termination of an employee may be reversed later. Depending on the reasons for reversal and the time between termination and rehire, this termination time had to be recategorized into a specific resultant category. A rou- tine, non-punitive layoff for five: days or less became scheduled time off (Box D). If it lasted more than five days but less than 365 days, it became a long-term legiti- mate absence (Box J). If the employee voluntarily quit and was rehired in five days or less, this was recategorized as 61 Type 9: Time Worked pp the Job Code Definition 1 Created for generating the de- rived database as a code for Day Present and Worked. 2 Off Day - Worked (employee was asked to work on a regularly scheduled day off instead of some other work day to correct a work scheduling problem for the DOT) Type 1: Scheduled pg Contractually Provided Absenteeism Code Definition 3 Regularly Scheduled Day Off 4 Off Day - Trade 5 Off Day - Adjustment 6 Casual Leave Day 7 Casual Leave Time 8 Vacation 9 Swing Holiday 10 Holiday 11 Death in Family 12 Occupational Injury 13 School/Training (DOT Related) 14 Leave of Absence 15 Military Leave 16 Jury Duty 17 Union Business 18 Conventions (DOT approved) 19 20 Laid Off Civil Service Business 21 Suspended 22 Not Scheduled 23 Court Time (Witness or DOT related) f—____—__———_——_'————_——_—_—_——'—fl—_———_——_——_———”—1 Figure 3: Summary of the Raw Absence Data Types as Recorded in the Department of Transport Absence Management Database or as Modified and then Used to Generate the Derived Data Continued on next page L__._..____.._____——_—_———_——_———____________.___ --__.______..._________________.____ _J 62 Figure 3 Continued r 1 I Type 2: Long-Term/Non-Recurring Absenteeism I I Code Definition I I 24 Long-Term Disability I I 25 Sick Employee (6 days or I I more) I I (These codes may be used to handle catastrophic * I I illnesses approved by the Personnel Division I I without the 6 day restriction.) I I 26 Resigned I I 27 Discharged I I 28 Retired I I 29 Deceased I I 30 Transferred I I 31 Long-Term with Injury I I 32 S & A with Injury : I Type 3: Excessive/Recurring Absenteeism I I Code Definition I I 33 Sick Employee I I 34 Sick Family Member I I 35 Death - Not Family I I 36 Wedding I I 37 Moving I I 38 Off with Permission I I 39 Absent without Permission I I 40 Single Miss (TEO's only) ** I I 41 Double Miss (TEO's only) ** | I 42 Strike I I 43 Tardy I I 44 Other I I 45 Unexcused Absence I L J Notes: 1. According to the DOT Attendance Control Policy, employee absenteeism was grouped into three meta-categories, with time worked on the job included, there are now four: Type 0 - Time Worked on the Job Type 1 - Scheduled/Contractually Provided Absenteeism Type 2 - Long-Term/Non-Recurring Absenteeism Type 3 - Excessive/Recurring Absenteeism 2. These types are reworked as shown in Figure 4 to create the various levels and categories of derived data. * Illnesses such as cardiac arrest, kidney failure, serious injury, etc. ** "TEO's" are Transportation Equipment Operators (bus driv- ers). 63 Figure 4: of Employee Not Rehired for 365 Days or More Termination Time ‘ OLaid or: 19 Termination eRcsiqned 26 Without Not Resigned, Nor Retired, eDischarged 27 I’ Rehire Nor Discharged. eRetired 28 M Pehired in 5 Days or Less. eDeceased 29 e Transferred 30 eScheduled Off Time 3 All 00ft Day - Trade 4 00!! Day - Adjustments ifxdgiid .___ e Swing Holiday 9 0 efloliday 10 Not F eStrike 42 Fired and Rehired Discharged 5 Rehired in S in 30 Days or Less Days to Punishment 365 Days Time E Suspended Workeer Day Hot Scheduled for This Worker22 G 3:13;“? or Missing Value "I I eSchool/Training (work Related) 13 I All eCourt Time (Witness or Work Related23 Time Other I OJury Duty 16 A Worked Work e Union Business 17 C Deonventions for Work 18 eCivil Service BusinessZO If ‘5 5 Days or Less Exits” ; Military LeavelS Than 366 Over 365 Days Days 5 eScheduled workday 1 w Time at es s A with Injury 32 work Time eLongterm Absent - InjurecBl 8 Worked OLongterm - Disabled 24 OSick - 6 Days or More 25 eLeave of Absence 14 eSingle Miss 40 eDouble Miss 41 eTardy Employee 43 Late Time Lost to Time _. Lateness L eUnexcused Absence 45 eAHOL 39 Short eSick Employee 33 eSick Family Member 34 Non- ODeath - Not Family 35 eWedding 36 ‘ Legit ‘— eMoving 37 00ft with Permission38 Absence eOther Unpaid/Non-Legitimate Short Absence 44 K All Time Absent :* J g " Longterm Legitimate Absence Short eCasual Leave Day 5 eDeath in Family 11 fitgrg: : :hgig Legitimate eCasual Leave Time 7 eOccupational InjurylZ .fi Y 1 mun“ eVacation 3 :¥ I Logical Relations Among Different Uses Time 64 a short-term legitimate absence (Box 1). Termination with- out rehire for 365 days or more remained a termination and the time use was assigned accordingly (Box M). Discharge for cause, followed by a rehire in 30 days or less was reas- signed as punishment time (Box E), and other disciplinary suspensions were also put into this category. Time for a worker recorded as deceased who later turned up on the record in any category other than a form of termination was reassigned as missing data (Box G). 3. Employee time that ‘was originally scheduled as scheduled off-time, an off-day traded from some other employee or for some other adjustment, or a swing holiday, public holiday, or strike, all were categorized as routine time off (Box F). 4. Punishment time and routine time off both contrib- uted to become "all-scheduled-time-off" (Box D). 5. Any day not scheduled for the worker and any miss- ing values were both reallocated as missing values (Box G). 6. Military leave for five days or less, school or training related to work, time spent in court as a witness or otherwise related to work, jury duty, union business, conventions for work, and civil service business were all recategorized as "other work" (Box C). 65 7. Military leave for 365 days or less was recategor- ized as long-term legitimate absence (Box J). Over 365 days, it became termination time (Box M). 8. Scheduled work days and off days that were worked, along with time worked on days when the employee arrived late, were all recategorized as routine time at work (Box B). 9. Both routine time at work and other work became components of all time worked (Box A). 10. Any time consumed by employees in arriving late was categorized as late time (Box L). Late days in most cases had to be split into late time and routine time worked (Box B). 11. In the absence of casual leave, instances of: unexplained absences, sick employees, a death outside of the immediate family, time off to move, absence without leave, time off to care for a sick family member or to attend a wedding, unpaid time off with permission, or other brief unpaid absences, all were categorized as non-legitimate short-term absence (Box K). 12. Casual leave days, casual leave time (part of a casual leave day split up), death in the family, time off due to occupational injury, and vacation time all became short-term legitimate absence (Box I). 66 13. Long-term employee time off due to sickness or accident or injury, long-term dis}iled employees, or a leave of absence for more than five days all were clustered under long-term legitimate absence (Box J). 14. Short—ternl non-legitimate absence, long-term legitimate absence, and short-term legitimate absence were components of all time absent (Box H). These 13 different types of time-use categories stand in respect 1x3 each other ixi'very specific interdependent relationships. There is some internal hierarchical cluster- ing among the different time uses, and because the use of time is an exclusive behavior, to choose any one fundamental time-use behavior precludes simultaneously emitting any other. As a result, the flow of time-use from one category into any other or combination of others can be examined quite conveniently. Following the paths and rules outlined in the flowchart (Figure 4), the raw absence reasons in Figure 3 have been recategorized to create the values for the 13 derived time- use measures below in Figure 5. For a more detailed description of the procedures used to recategorize the raw data and to generate the derived time-use values needed to test hypotheses in the present research, refer to Appendix B. 67 Three Different Measures pf Time Use: Frequency, Severity and Mean Severity per Occurrance After recategorization of the time uses into their cor- rected categories, three types of moving values were calcu- lated as measures for each time-use index, each day, for each employee's data: (1) "frequency" of occurrence for the employees on this time-use across the preceding specified Time Use Category/Index All time worked on or off the job Time worked on the job Time worked off the job All scheduled time off Time off due to punishment Routine normal time scheduled off or strike Missing value or unscheduled day Time absent for all reasons Short-term legitimate absence Long-term legitimate absence Short—term non—legitimate absence (except late) Late time Unended termination time (no rehire) Fit-H4530 "3121100013, Zt“ Figure 5: Derived Time—Use Measures length of time; (2) total duration or the sum of all "sever- ity" values for the employee on this type of time-use across the preceding specified length of time; and (3) "mean sever- ity'I per occurrence, which was calculated as a ratio of the severity over the frequency for the respective employee on 68 the specific day. Two different lengths of time were used as base periods for calculating the moving values. 365-day moving values were used to isolate the long-term inter-an- nual effects because these would exclude intra-annual cycles by construction. loo-day moving values were used to examine intra-annual patterns of time-use because these were: (a) short enough not to cancel out annual cycles such as renewal or seasonal effects, (b) long enough to provide a relatively stable base for the moving value, and (c) divisible by 100 so that an estimate of the corresponding daily time-use could be made at a glance. Moving values were also used instead of single-day values (as in much conventional research on absence) in order to make the means a convenient moving average suitable to smooth out sporadic highs and lows to be expected due to the variability of individual absence behavior. When. these values 'were then. averaged across their respective sample groups, they were to provide a larger aggregate or base of group behavior for study. And, from this base, a more stable and reliable estimate of the mean daily values on each time-use index was then calcu- lated for the respective sample group. These three types of moving values, for both the 100-day base and the 365-day base, and across all 13 catego- ries of time-use, became the substantive basis of the research reported in Chapter IV and Chapter V. A more detailed description of the procedure and the measures them- selves is presented in Appendix B. CHAPTER IV Methodology and Operational Hypotheses Introduction: This introduction is a brief survey of the framework of procedures and tests implemented here. Further detail on each aspect follows. The methodology and hypotheses described below are specifically designed to address ques- tions of program. impact and to investigate theories of absence behavior by analysis of group level measures. The steps in this analysis and related hypotheses follow exactly the sequence presented in Figure 6. Then the results gener- ated by implementing these procedures also follow the same sequence and these have been presented and discussed in that order in Chapter V. Because of the problems of non-normality typically encountered in data distributions like the ones used here in the T-tests, and because of the controversy over the use of severity versus frequency measures, it was first necessary to test procedural hypotheses concerning normality for the T-test data and reliability for the time-series measures, and to establish the relative merit of the present program evaluation paradigm. Then a selection of tests had to be executed to check for spurious short-term shifts such as seasonal differences. In addition to shedding light on some non-central substantive issues, that information was 69 70 Cross Sectional Analyses Time-Series Correlations A C Normality Tests Split~Half Time Series 1. The 30-day datasets Reliability Correlations used to assess pre- 1. On the lOO-day dataset program impact. a. during the baseline 2. The lOO-day short— year, after the 100~day term Tntest data. moving values were 3. The 365-day long-term developed from day 281 T-test data. to 546. b. during the post-instal- lation period from day 982 onward when the 100; I day moving values were T;Tests again fully developed. 1. On the 30-day datasets 2. On the 365-day dataset testing for pre-pro- from day 1247 onward, gram overall shifts. the span during which 2. On the 100-day dataset moving values of the testing for pre-pro- long-term annual mea- gram patterns: sures were fully dev- a. seasonal effects eloped. b. holiday effects i c. combined seasonal & holiday effects D 3. On the lOO-day dataset Time-Series Correlational testing for change Tests of Substantive during the program. Hypotheses a. year to year change 1. On the lOOsday dataset within the program ex- to see the strength of cluding seasonal and gorrelations indicating holiday effects and short-run: intra-annual program a. progression effects induced effects b. legitimacy effects b. holiday effects but c. combined progression 8 not seasonal or annual legitimacy renewal effects d. Shifts toward or away c. seasonal effects and from work or absence. annual renewal. but * These will include co- compensating for holi- occurrent intra~annual days shifts. 4. On the loo-day dataset 2. On the long-term 365- to examine overall day dataset to see the inter-annual shifts strength of long-term due to the program. inter-annual effects, 5. On the 365-day dataset exclusive of annually to determine overall repeating seasonality, program impact on the holidays and renewal annual time-use effects. indices exclusive of a. progression seasonal effects and b. substitution intra-annual substi- c. combined progression & tution effects and substitution holidays. d. shifts into or away from work or absence Figure 6: A Categorical Summary of Tests and Analyses 71 needed to modify the interpretation of subsequent substitu- tion hypotheses and the evaluation of the program. A description of the normality testing procedure and the related rationale is presented here, with further details in Appendix C. Also in Appendix C is a methodological note on the jprocedures for ‘time-series correlations. These 'were used to calculate split-half reliabilities for each of the time-use measures and 1x) test substantive hypotheses con- cerning progression and legitimacy effects and shifts from absence toward more time worked. The rest of this Chapter presents four clusters of methodology and related hypotheses. First, the data distri- butions for the T-test data had to be examined. Prior research has shown such distributions to be generally non- normal. In order that the relative merit of using the pres- ent data for T-tests could be assessed, the degree of non- normality le these data cross-sections required investigation. This investigation includes the steps out- lined in Box A of Figure 6. Second, the rationale and structure of the T-tests themselves were presented with related hypotheses (Box B of Figure 6). Three sets of T-tests were implemented. (A) Using the data for 30-day periods at the start of the base- line year and at the end of the pre-program installation interval, a set of general tests were done to see whether 72 anything changed between those two time periods, prior to the absence control program itself (see Figure 6, Box B, Item B(l). (B) Similarly, the lOO-day data from the base- line year were used to find out whether any seasonal or hol- iday effects were evident in the data since these data would have, by construction, revealed such changes, but would not be contaminated by later program induced effects. Like the pre—program overall changes, seasonal influences could have moderated any changes induced by the program and contami- nated the T-tests for program impact. If no such changes were revealed, then changes later induced by the program as annual renewal of casual leave days on each July 1 would be interpretable as effects more likely due to the program than such unintended effects. The procedures involved are Steps B-Za, B-Zb, and B-2c in Box B of Figure 6. A third, fourth and fifth set of T-tests were implemented to test substan- tive hypotheses designed to examine the data for impacts of the program. These are Steps B-3a, B-3b, B-3c, B-4 and B-5, outlined in Figure 6. Third, due to the controversial unreliability of absence measures reported in prior research, it was neces- sary to establish the reliability of measures designed for the present work. The procedures used for this have been outlined in Box C of Figure 6, and a detailed methodological note on the time-series split-half correlation procedure has been presented in Appendix C. 73 Fourth, a similar time-series correlation procedure was used to test substantive hypotheses of both short-term and long-term change. These tests have also been outlined in Figure 6, Box D. In all the following detailed presentation of methodol- ogy and hypotheses, the sequence of procedures and problems outlined in Figure 6 has been adhered to. In Chapter 5, the same sequence of steps has also been followed for presenta- tion of the results. Rearranging the Case Data by Relative Date Once the thirteen derived time-use measures were cre- ated, the new database had to be sorted ... rearranged for group-level analysis so that all values for the 39 time-use indices that pertained to a given day were grouped together regardless of which employee they previously referred to. This was done for both the lOO-day measures and the annual (365-day) measures, for the 82 cases in each random half of the sample representing all DOT employees, and the data files were then merged for these two randomly selected halves to make a third overall 164-case sample. Each of these three sorted files was written out onto a new tape file for use in later analyses for both the 100-day measures and the annual measures. In the discussion which fellows, the term intra-annual effects means "cyclical and annually recurring effects based on.ai IOO-day period," while inter- 74 annual effects means analysis" means "based on a 365-day base period and relatively unaffected by intra-annual cycles." The letter labels A, B, C and D refer to specific clusters of steps outlined in Figure 6. Procedures in A: Normality Tests of the Data Distributions As a precursor to interpreting the T-tests that were to check for seasonality, substitution and preprogram change effects, or to interpreting the T-tests for overall program impact, the normality of the daily cross-sections of data had to be assessed. using the SAS "univariate" procedure, the datasets for both the 100-day measures and the annual measures were analyzed to determine the normality of the data distributions on the dates chosen to be the focus of T-test calculations. The Kolmogorov D—statistic was used as well as more traditional statistics describing the empirical data functions. See Appendix C for details. This step was applied only to the 164 cases of data generated on selected days for the entire sample from all employees. The follow- ing procedural hypothesis was tested. HYPOTHESIS l: Asymmetry and non-normality in the data distribu- tions as shown by the Empirical Data Function (EDF) statistics will be severe enough that the applicability of normality—based inferential sta— tistics must be questioned. 75 Selected results of the normality tests are summarized in Chapter V, Table 1. Procedures in B: T-tests of the Cross-Sectional Comparisons Introductory explanation: For exploratory purposes and to assess consistency with established practice and past findings in absenteeism research, T-tests of the signifi- cance of differences in means over time were to be done, together with an interpretatbmi of the EDF statistics and D-test results. These cross-sectional comparisons were done for five substantive reasons: 1. To spot-check for possible pre-program shifts in time-use attributable to anticipation of the absence control program by comparing the 30-day data cross sections; 2. To examine the patterns of time-use over time and test the significance of predicted pre-in- stallation seasonal and holiday effects; 3. To test on the lOO-day measures the substitu- tion hypotheses that various specific shifts in employee time-use did occur from one form of absence to another during the period of the absence control program itself; 4. To test for significance of program-induced change using the lOO-day dataset; and 5. On the 365-day dataset, to assess overall pro- gram impact on the annual time-use indices, exclusive of annually repeating patterns. 76 B-l. Pre-program Incidental Shifts Using Measures Based on a 30-Day Period: Since there was only one year of baseline data, the derived data do not really constitute a fully annual set of moving values until the last day of the baseline year, nor do they constitute a legitimate and fully developed part of the lOO-day measures until day 281. Prior to that point, the cumulative values of the time-use indices indicate both (1) observed patterns of time-use and (2) an incidental increase in these indices due to the fact that accumulation of annual values for the behavior had not yet had a full year to occur. On the other hand, essentially the same con- dition held true for the first year of data collected under the program, and there was one month of data collected between June 1 and June 30, 1980, which was immediately prior to full enforcement but after the start-up of the com- puterized absence-control program itself. It was quite pos- ‘ sible that some very real shifts in time-use behavior might have occurred due to pre-program training and planning activities and due to employees anticipating the program. At the same time, seasonal and/or holiday-induced shifts were not expected between June and July prior to the start- up of the absence-control program. It appeared reasonable to compare time-use indices from 30 days into the baseline year (Day 211) with time-use indices after 30 days of data collection, just before the program started with a fresh bank of casual leave for each employee (Day 912). 77 Time-use indices for July 30, 1978 were accordingly compared with time-use indices for June 30, 1980, to spot- check for pre-progrmm changes. Because of the hmevity of the span (30 days) over which these data were being col- lected, and some possibility of seasonal contamination (due to a l-month difference in timing), the results would have been be somewhat suspect, even if EDF statistics indicated normality in the distributions. Nonetheless, the following hypothesis was tested. HYPOTHESIS 2: H-O: Significant changes will not be observed in the mean levels of employee time-use between the end of the first month at 30 days into the beginning of the baseline year (July 30) and the end of the pre-program period (June 30, 1980). The findings with respect to preprogram change were interpreted to determine the kind and extent of anticipatory socialization effects involving any of the 13 categories of time-use. In order to distinguish later findings on subseq— uent tests from anticipatory preprogram shifts, the extent of preprogram change had to be assessed. Results for Hypothesis 2 have been presented in Chapter V, Table 2. 78 B-2. T-tests of Pre-Program Changes on the lOO-Day Meas- ures: To distinguish program impact and other shifts from incidental cyclical changes it was needful to assess the extent of holiday related and seasonal differences. This was done as follows. B-2a. Seasonal effects: "Seasonal" shifts in employee time-use were likely to co-occur with hypothesized progression and substitution effects during the post-installation period. Accordingly, the simplest way to examine the data for seasonal changes was to focus on pre-program data, and for the same reasons described above (duration of the observation period being less than one year), this meant that only the lOO-day meas- ures could effectively be used. The most dramatic evidence of seasonal shifts seemed likely to surface by comparing time-usage at the end of the winter months with time-usage at the end of the summer months, exclusive of data for the 4th of July holiday period and the Christmas season. Accordingly, the following hypothesis was tested for each type of time-use measured. HYPOTHESIS 3: There will be significant differences in the mean of the lOO-day time-use indices between April 14, 1979, and October 20, 1979. Specifically, there occurred significant seasonal change in mean lev- 79 els of employee time-use across the 6-month sea- sonal interval. The lOO-days preceding these dates do not co—occur with major holidays which would tend to have an erratic short- term effect, except for Labor Day and Easter respectively, balancing each other's effect. Results for Hypothesis 3 have been presented in Chapter V, Table 3. B-Zb. Holiday effects: An hypothesis was also tested that there would be a "holiday effect" evident in comparison of the data including the Christmas season with data for the period immediately following. Time-use indices for April 14, 1979 were com- pared with those for January 5, 1979, on the theory that by early’ January any increase in. the .moving 'values. due to absences related to Christmas holidays would become evident in the indices, and these would have been smoothed into the lOO-day moving values by April 14. The hypothesis was as follows: HYPOTHESIS 4: Prior to program installation and absence control, there will be significant differences in employee time-use behavior between April 14, 1979, and Jan- uary 5, 1979, a significant holiday effect. Results for Hypothesis 4 have been presented in Chapter V, Table 4. 80 B-2c. Combined seasonal and holiday effects: To check for significant changes in short-term time-use patterns between dates when the combined seasonal and holi- day conditions were quite different, a comparison was made between values of the 100-day measures near the end of the baseline year and the values of the measures just after the Christmas season, on January 5, 1979. HYPOTHESIS 5: There was a significant pre-program change in the lOO-day time-use indices between late spring/early summer (June 15, 1979) and late fall/early winter (January 5, 1979). A combined holiday effect and seasonal shift will have occurred in the lOO-day pre-program time-use measures. Results for Hypothesis 5 have been presented in Chapter V, Table 6. B-3. T-tests of Post-Installation Changes: Within the span of the program (after installation) a number of changes were expected. These were examined in three sets of hypothesis tests as follows. B-3a. Inter-annual change: Differences between identical dates in 1981 and 1982 would, by construction, be primarily indicative of year-to- year change rather than cyclical variation. Some such long- 81 term impact was expected. Accordingly, differences between the measures were T-tested across the following intervals. -- March 10, 1981 March 10, 1982 -- September 8, 1980 September 8, 1981 -- October 20, 1980 October 20, 1981 -- April 23, 1981 April 23, 1982 HYPOTHESIS 6: Significant year-to-year changes will have occurred. in the IOO-day' measures of short-term absence and so will significant changes in the lOO-day measures of long-term absence between com- parable dates in 1980, 1981 and 1982. The IOO-day measures of both short-term absence and long-term absence will. show that. inter-annual shifts occurred in levels of time-use as absence across the post-installation duration of the program. In particular: legitimate and long-term absences will have increased, non-legitimate and short-term absence will have decreased, and there will have been a significant decrease in the overall absence index, plus an increase in the levels of time- worked. Results for Hypothesis 6 have been summarized in Chapter V, Table 6 and Table 7, for frequency and severity measures respectively. 82 B-3b. Holiday effects: Changes will occur in levels of absence and other time uses in association with major holidays across periods when seasonal and annual renewal effects are not expected to have much effect. This was T-tested by comparing values on the lOO-day measures for early winter (which includes Christmas) with those of late winter, and by comparing values for early summer (which includes July 4th) with those for late summer/ early fall. Accordingly, differences between the measures were T-tested across the following intervals. -- March 10, 1981 - April 23, 1981 -- March 10, 1981 - April 23, 1982 -- September 8, 1980 - October 20, 1980 -- September 8, 1981 - October 20, 1981 HYPOTHESIS 7: Holidays will have a significant impact on time- use behaviors. Levels of time-use at work were lower and levels of time-use in absence will have increased in association with holidays. Results for Hypothesis 7 have been summarized in Chapter V, Table 8. B-3c. Seasonal effects and annual renewal: Because of the timing of annual renewal, seasonal effects and annual renewal should have both operated in the 83 same direction, inducing high levels in short-term absence and lower levels of time worked. On the assumption that the Christmas and July' 4th holiday' periods would. both have effects of about the same magnitude and direction, differ- ences across carefully selected dates would show primarily this combined seasonal and renewal effect when similar holi- days were set in comparison between winter and summer. To test for combined seasonal and renewal effects, differences in the 100-day time-use indices across the following inter— vals were tested. -- March 10, 1981 - September 8, 1980 -- March 10, 1982 - September 8, 1981 -- April 23, I981 - October 20, 1980 -- April 23, 1982 - October 20, 1980 HYPOTHESIS 8: There will have been significant changes in the time-use indices due to combinations of annual renewal, holiday, and seasonal shifts. Results for Hypothesis 8 have been summarized in Chapter V, Table 9. B-4 & B-S. Inter-Annual Shifts due to Program Changes: T-tests were also implemented on both the lOO-day and the 365-day measures to determine overall program impact on time-use indices where they could not, by construction, be influenced by annually repeating renewal, holiday, or sea- 84 sonal effects. For the lOO-day measures, this meant select- ing identical dates for analysis from year to year. Because the 365-day base, by construction, automatically disincludes a day's worth of data 365 days before every time it includes one more day, any such annual cycles would have had no effect on the 365-day indices once those indices were fully developed and included a full 365 days of time-use behavior. Only inter-annual long-term change should remain evident in those 365-day indices ... the long-term program impact. The inter-annual shifts due to substitution and progression can be better' detected 'using' time-series correlations. But, overall change in the means under the impact of the program should be made evident by comparing levels of the 365-day indices from year-end to year-end. This was done by compar- ing the following dates. June 30, 1981 - June 30, 1979 June 30, 1982 - June 30, 1979 And, for program impact continuing after the first program year, June 30, 1981 - June 30, 1982. HYPOTHESIS 9: There were significant overall reductions in non- legitimate absence and, increases in legitimate absence and time-worked as a result of the pro- gram. 85 An unexpected result, due ix> the workers being granted 80 extra hours (2 weeks) of paid time off, made the lOO-day results somewhat misleading. As a result, only the 365-day results have been reported in detail here. The 365-day results for Hypothesis 9 have been summarized in Chapter V, Table 10. A short note and discussion of the lOO-day results also appears there. A graphical summary of the T-test results in presented in Chapter V, Figure 7. Explanatory Comment on Long-Term Substitution Effects Overall shifts from short-term to long-term. absence behavior will tend to occur over time across the intervals from the baseline period to the first year of the program, and from the first to second year of the program, and from the second to the third year of the program. Because there was a. practice in force of annually renewing for each employee the bank of casual leave days (which could be taken at will and which were to cover all short-term legitimate absence), there would presumably be an annual recycling of the substitution effect. The practice was to renew the bank of casual leave days on July I each year. Accordingly, dif- ferences in time-use patterns about that date should have provided evidence of any intra-annual substitution/progres- sion effects that occurred. These, however, would be excluded by construction from differences in the 365-day indices, and evidence for substitution evident in the 86 365-day data would, therefore, be indicative of a long-term and inter-annual overall trend, not a cyclical intra-annual effect. _A Note £1 Calculation 3f the Means and Selection 9f the Split-Halves In order to test the reliability of the time-use meas- ures using the split-half correlation technique, two compa- rable sets of the sample data had to be compared by corre- lating the time-series of their means. That, in its turn, required for each index that two equivalent samples repre- senting the workforce be selected and that their two compa- rable sets of means be calculated. This was done as fo1- lows. For both the 365-day data and the lOO-day data, each of the three subsets of data (164 randomly selected cases from the DOT database, and the two halves of that dataset split out in samples of 82 cases each by selecting every second case) were then subjected to the SAS "MEANS" proce- dure. More explicitly, that means a total of six datasets were analyzed: 164 cases of annual data and two subsets of 82 cases of annual data; 164 cases of lOO-day data and two subsets of 82 cases of lOO-day data. The means for all 39 time-use indices for each dataset eadh day of the program were written out day by day into a separate tape file for each set of data. These means provided the basis for all 87 further analysis beyond the level of the T-tests, analysis based upon time-series comparisons. C. Time Series Estimates for Split-Half Reliability in the Measures. C-l. To test reliability of the measures for both the lOO-day and the 365-day datasets, the series of daily mean values for each time-use index was compared with its count- erpart in the other split-half by time-series correlation of the mean daily level of the index for all employees in the one half with that for the other half ... over all days of the baseline period or program for which the means actually represented the fully developed state of the measure. Since the relationship this procedure is designed to test for is one of near-identity where any differences between the halves would be artifacts of measurement or sampling error and not of substantive differences, these correlations should ideally approach r = I. (See Appendix C for a note on the time-series correlation procedure.) This generated three sets of correlations for the 39 different time-use measures. The three periods were: (a) For the lOO-day data, during the baseline period, from day 281 to day 546; (b) for the lOO-day data, during the post-installation period, from day 982 to the end; and (c) for the 365-day data, during the post-installation period, from day 1246 to the end. 88 HYPOTHESIS 10: The correlation between the means on the two split halves will be significant and will approach one for each of these three sets of reliability corre- lations. Results for Hypothesis 10 have been summarized in Chapter V, Table 11. Note: A test of the frequency-severity contro- versy could also be done by inspection of the overall set of time-series correlations. This leads to Hypothesis lO-a. HYPOTHESIS lO-a: The time-series reliability' estimates of daily mean values for both frequency and severity indi- ces will both be highly reliable when these are tested by correlation with their split-half equiv- alents over time, but neither will have uniformly demonstrated a significantly higher reliability. Results are presented in Chapter V, Table 11. Under special conditions, a number of the reliability values for espe- cially small samples of time-use behavior could be anoma- lously low. The list below summarizes various conditions which could bring about such anomalously low reliabilities for the time-series correlation procedure. See Appendix C for a full explanation. 89 Conditions Conducive to Anomalously Low Reliability in the Time-Use Measures 1. The sample size can be too small, including too few employees. 2. Very few employees may participate in the respective time use. 3. The time-use itself may rarely occur, even for those who did choose it. 4. The time-use may be so carefully regulated or may other- wise have occurred with such consistency that the amount of observed variation underlying the smdit-half reli- ability estimate was too tiny. 5. The amount of time-use in one split half or the other may be constant over time causing a standard deviation of zero and making calculation of a correlation impossi- ble. 6. The subdivision of overall categories of measurement into more detailed ones (e.g., three detailed absence categories are components of All Time Absent) may intro- duce too many distinctions for the amount of behavior in the sample and so cause the subsample of behavior assigned to one or more of the detailed time uses to become too small for effective study. 7. The length of the time-series may be too short to effec- tively establish reliability for the series. 9O 8. The length of the observation period for collecting each day's values on each measure may be too short to smooth out the sporadic daily fluctuations. D: Time-Series Correlations to Test the Substitution Hypotheses and Investigate the Impact of the Absence Control Program Over Time Both the 100-day measures and the 365-day measures had potential for examining substitution effects induced by the program. Five specific subseries were created using appro- priate portions of the timeéseries. From these two types of measures the new datasets were constructed by selecting out the first 99 days or 364 days respectively, for the time segments ‘when part of the apparent variation. was there merely due to the construction of the measures and the accu- mulation of daily values into incompletely developed lOO-day or 365-day measures. To rely on such data would have incor- rectly inflated the time-series correlations. The following datasets were then subjected to the SAS Proc Corr procedure to generate rectangular matrices of time-series correlations and associated significance levels. On the lOO-day dataset: From day 281 to day 546 ===> baseline period From day 982 to day 1277 ===> first program year ... after first 100 days of data collection From day 1278 to the end (1680) ===> last 14 months From day 982 to day 1680 ===> entire post installation per- iod ... after first 100 days of data collection 91 And, on the 365-day dataset: From day 1246 to day 1680 ===> the last 14 months The lOO-day datasets would show effects due to whatever happened to be occurring ... seasonal, holiday, annual renewal, substitution, or long-term change effects ... with only a limited capability to distinguish annual cycles from other co-occurrent patterns. However, the 365-day dataset is relatively insensitive to annual cycles by construction and could, therefore, be used to examine the purely inter- annual, non-cyclical effects. The use of these respective sets of data is described below. D-l. Correlational Tests on the lOO-Day Dataset: These tests were done to investigate four general trends in time-use behavior over time. D-la. Group-level progression effects: The first trend to be examined was a hypothetical pro- gression from less severe to more severe forms of absence and withdrawal from work, for all employees in the workforce in general. This progression effect was investigated by correlation of the group-level time-series measures for short-term legitimate and long-term legitimate absence. A significant negative correlation would have supported the hypothesis that an inverse relationship between short and long-term. absence existed and that a progression. effect occurred. 92 HYPOTHESIS 11: The time-series correlations between the short- term legitimate and long-term legitimate absence indices in the lOO-day dataset within the span of each year will show that significant and negative correlations occurred. Short-term legitimate absences were replaced by long-term legitimate absence each year. I Results for Hypothesis 11 have been summarized in Chapter V, Table 12. D—lb. Legitimacy effects: The argument has been made that under the impetus of an absence control program, employees will shift the form of their absence instead of eliminating absence from their behavior. Non-legitimate absence under such conditions will tend to be replaced with legitimate absence. To test for this effect, the following hypothesis was tested. HYPOTHESIS 12: Time-series correlations between non-legitimate and legitimate short—term absence in the lOO-day dataset over all four datasets will be negative and significant, indicating a legitimacy effect on the short-term absence. 93 Results for Hypothesis 12 have been summarized in Chapter V, Table 12. D-lc. Combined progression and legitimacy effects: If both these effects were present and working together, then the combined effect should have been stronger than either progression or legitimacy when the two operated in tandem. To check for this combined effect, the following hypothesis was tested. HYPOTHESIS 13: Time-series correlations between non—legitimate short-term absence and legitimate long-term absence will be significant and negative, indicat- ing a combined progression and legitimacy effect. Results for Hypothesis 13 have been summarized in Chapter V, Table 12. D—ld. Shifts toward work and away from non-legitimate short-term absence: The intended result of the program was to induce a shift directly from short-term non-legitimate absence into work itself rather than various other time-uses. To verify this impact. of the absence control program, the following hypothesis was tested. HYPOTHESIS 14: Time-series correlations between time worked on the job and the short-term non-legitimate absence measures will be negative and significant. 94 Results for Hypothesis 14 have been summarized in Chapter V, Table 12. Each of the above hypotheses (ll, 12, 13 and 14) was tested with respect to each of the four time periods (see page 90). across which correlations among the lOO-day meas- ures had been calculated. D—2. Time-Series Correlations Between Measures in the 365-Day Dataset: These correlations were calculated for data from day 1246 onward to examine the strength of long-term inter-an— nual effects, exclusive of amnually repeating seasonality, holidays, and renewal effects. As was done for the lOO-day data, the 365-day measures were analyzed for four patterns of change: (a) progression, (b) legitimacy, (c) combined progression and legitimacy, and (d) shifts between work and absence. Results are presented in Table 12. For all of the above hypotheses tests, tables which summarize the results have Ibeen. explained” presented. and then discussed in Chapter V. The sequence of presentation in Chapter V for results of the procedures described above has been described in Figure 6. Step-by-step, it follows the same general order of presentation used here in Chapter IV to outline the methodology and operational hypotheses. CHAPTER V Results and Discussion Introduction: The results presented here in Chapter V are organized to parallel the discussion of methodology presented in Chap- ter IV. Sections A, B, C, and D in the methodology have described the procedures which produced corresponding Sec- tions A, B, C, and D in these results. See Chapter IV, Figure 6 for an outline of these procedures. Results for all hypotheses which rely upon a specific analytical proce- dure have been presented together, either in a single table or in a cluster of consecutive tables or time-series graphs. A selective summary of the results relevant to each hypothe- sis has been reported here. A more detailed presentation of the results can be found in Appendix D. For reader conven- ience, a brief descriptive guideline for interpreting the various results and summary statistics presented in, the tables precedes each table or cluster of tables. Accompany- ing each respective cluster of tables, the results have been discussed and interpreted relative to the related hypotheses presented in Chapter IV. The extent of support for each hypothesis was assessed and examined briefly in the light of other co-occurrent findings. Results of the (A) normality tests and statistics describing distributional form have been presented first; followed by (B) the T-tests; (C) the time-series split-half 95 96 reliability correlations; and (D) the time—series correla- tional tests of substitution effects. Note: For completeness, the database was constructed to allow for the occurrence of termination time. But for all results presented here, no findings on time-use indices for termination time have been reported because in the ran- domly selected sample of 164 cases, which is the basis of the present report, no occurrences of termination time were reported in the raw data. Consequently, that time-use cat- egory is neither needed nor available to help account for the impact of the program on the system of employee time uses. A: Normality and Empirical Data Function Statistics for the T-Test Data The selected EDF statistics described below have been tabulated here to illustrate the distributional forms typi- cal of data used in the T-test analyses. Since there was an overwhelming repetitiveness in these EDF statistics, only statistics for All-Time-Absent have been reported here, and only for the dates selected for T-tests of year-to-year overall change. A more detailed set of tables in Appendix D allows comparison of the absence distributions with those for all-time-worked and for all-time scheduled off. EDF statistics for the 30-day, lOO-day and 365-day datasets have been presented in Table 1 below, following a brief explana- tion of the meaning and usage of each type of statistic in the table. 97 Below is an annotated list describing EDF statistics in Table 1 below. Relative Date: Refers to the relative date to which the daily cross-section. of 'T-test. data jpertain. Day 182 means July 1, 1978, day 183 means July 2, 1978, etc. Refers to the number of cases for' which data existed on the respective day. Variation in N is indicative of the extent of. turnover across the time duration of this study. Of the 164 cases in the sample, 126 were there on day 546, others hired on later, and still later a few employees apparently quit. Mean and Med: Refer to the group mean value of the selected time-use index on a particular day, or to the group median value, . respectively. The units of measurement involved are: frequency of absence episodes per time period, total severity of absence episodes added across the time period, and mean severity (which is a ratio calculated each day as severity divided by frequency). The size and direction of differences between mean and 98 median are crude ibut ‘useful indicators of the direction and magnitude for departures from nor- mality. Where the mean is greater than the median, the distribution will generally be posi- tively skewed. Std. Dev.: Refers to the standard deviation for data in a particular daily cross-section. Skew: Refers to the skewness of the data distribution, ideally, in a normal distribution, skew = 0. Any negative skewness in time-use data will tend to be rather small because such a result is limited by minimum values at zero. On the other hand, posi- tive skewness in time-use data can get quite large, depending on the number and size of unusu- ally large outliers above the mean. Kurtosis: The "peakedness" of the distribution relative to a normal distribution, ideally = O. Negative values mean a flat distribution, positive values mean a high modal concentration. Table 1: 30-Day, 99 Normality and EDF Statistics for Selected loo-Day and 365-Day Data Distributions Used for Cross-Sectional Tests of All-Time-Absent Relative Std. Kur- Prdb. Index Date N Mean Med. Dev. Skew tosis D > D 30-Day 211 129 1.64 2.0 1.14 0.61 0.68 0.19 (0.01 Frequency 912 151 1.19 1.0 1.05 0.90 0.98 0.23 (0.01 30-Day 211 129 103.30 120.0 58.12 -0.69 -0.84 0.15 (0.01 Severity 912 151 125.58 152.0 61.59 -1.10 -0.02 0.19 (0.01 30-Day Mean 211 129 67.25 56.0 53.31 0.65 ~0.60 0.16 (0.01 Sev./Occur. 912 151 73.77 74.5 65.06 0.38 -1.14 0.17 (0.01 lOO-Day 546 126 3.44 3.0 2.09 1.04 1.64 0.17 (0.01 Frequency 1277 151 1.74 2.0 1.51 1.07 2.55 0.17 (0.01 1642 141 3.77 3.0 2.44 0.51 0.08 0.14 (0.01 lOO-Day 546 126 419.30 477.0 156.33 -l.59 1.53 0.25 (0.01 Severity 1277 151 281.83 336.0 151.65 -0.96 -0.44 0.20 (0.01 1642 141 412.06 504.0 197.96 -1.13 ~0.05 0.23 (0.01 lOO-Day Mean 546 126 144.95 124.0 108.28 1.94 5.45 0.16 (0.01 Sev./Occur. 1272 151 134.43 114.3 127.11 1.03 0.49 0.15 (0.01 1642 141 130.59 103.2 122.12 2.12 5.53 0.17 (0.01 Annual 546 126 11.17 11.0 5.44 0.53 0.35 0.12 (0.01 Frequency 1277 151 9.34 9.0 5.29 0.23 -0.23 0.08 (0.04 1642 141 12.21 12.0 6.80 0.39 0.27 0.06 (0.15 Annual 546 126 1383.33 1594.0 517.67 -O.95 -0.17 0.16 (0.01 Severity 1277 151 1322.63 1583.0 591.38 -1.34 0.31 0.27 (0.01 1642 141 1490.70 1727.0 591.42 ~l.42 1.10 0.20 (0.01 Annual 546 126 151.20 128.0 108.38 2.68 9.79 0.19 (0.01 Sev./Occur. 1272 151 155.07 135.4 106.95 1.59 4.46 0.11 (0.01 1642 141 147.33 122.1 127.49 3.30 17.74 0.13 (0.01 100 D: Refers to the Kblmogorov "D'I statistic described in Chapter IV. The D statistic varies between 0 and l, with lower values indicating less normal distribution. In a normal distribution, D = 1.0. Prob > D: Means "the probability of this distribution pro- ducing a higher D statistic." The D and Prob > D statistics occur as pairs. Discussion of Results for the Normality and EDF Statistics In general, the D statistics are all less than 0.5 and most are less than 0.4, indicating that the samples were definitely non-normal, while the associated probabilities of the underlying distribution producing a rugher D statistic were without exception less than 0.01 (D was highly signifi- cant). Skew and kurtosis were generally positive and a clear relationship exists between the sizes of the mean and standard deviation. This set of distributions is definitely non-normal and positively skewed. The relatively large standard deviations in the values of 30-day severity per occurrence (compared to the respective means) are a conse- quence of compounding two sources of variation in measure- ment. The severity per occurrence (mean severity) was cal- culated as a function of 30-day severity divided by 30-day frequency. 101 The T-tested data distributions are clearly not normal. Hypothesis 1 is definitely supported. Changes in the means remain helpful as indices of the absolute levels of time- use, but estimates of iconfidence levels associated. with those results are run: at all accurate. These results are entirely consistent with prior findings about cross-sections of data in the absenteeism literature. On the other hand, these findings with respect to the individual values in daily cross-sections relate to the T-tests only. The data distributions for time-series of the group mean values that were used imi later correlational analyses are zni entirely different matter, and as means based upon the daily cross- sections, those will be less affected by extreme values and should be distributed more normally. B: Results for the T-Tests of Cross-Sectional Comparisons T-tests were implemented on differences between the means of 17 different, time-specific daily datasets to exam- ine five sets of hypotheses (see page 57 and Figure 6). An explanation of the reasons for choosing these particular dates has been presented in Chapter IV. Results of the nor- mality and EDF analyses for these data distributions indi- cate 'very serious non-normality, just as prior research would imply. Nonetheless, for consistency with prior research, the T-test results are presented below. In most cases the variances of the two distributions were different enough (F-tested) that the T-test for distributions with 102 unequal variance was used. This decision was made sepa- rately for each T-statistic, however, since the F-test val- ues and probabilities were all available. For the T-tests, the table or tables of selected sta- tistics associated with each hypothesis have been presented, one hypothesis at a time, followed immediately by a tuief interpretation of the results and a discussion of support for the hypothesis. More detailed T-test results are included 1J1 Appendix IL Values jpresented ixi the 'T-test tables include: Var/Index: the variable and time-use index tested - Relative Date 1: the first of the two dates for which the T-test was done - Mean 1: the mean value of the index variable on Date 1 - Std.Dev. 1: the standard deviation for the data distribution at Date 1 - Relative Date 2: the second of the two relative dates for which the T-test was done - Mean 2: the mean value of the index variable on Date 2 - Std.Dev. 2: the standard deviation for the data distribution at Date 2 - F-Prob: the probability of equal variance in the two data distributions _ 103 - T: the T-test statistic, in most cases calcu- lated for unequal variance - T-Prob: the probability‘ that the difference between the means occurred by chance - Units of measurement were as follows: (a) fre- quency was reported as number of episodes per period of observation; (b) severity was reported as number of hours across all episodes in the period of observation; (c) mean severity was reported as the daily ratio of severity over frequency (or mean number of hours per time-use episode). Note: The number of missing values reported here and else- where in this research is a direct result of excluding every questionable raw value from further analysis and reassigning them as missing values. Missing values, as a result, do reduce sample size and may absorb undocumented changes in other time-use indices but are not otherwise included in calculation of the other time-use indices. 104 B-1: Discussion of T-Tests on Overall Pre-Program Shifts (See Table 2) Contrary to the null statement in Hypothesis 2, some highly significant shifts in time-use behavior were appar- ently recorded between the start of the baseline year and the month prior to program commencement. Anticipatory effects due to pre-program training apparently did have an impact as the discussion below indicates. Across the pre-program interval, between July 30, 1978, and June 30, 1980, the severity and frequency indices for all time worked both changed by about 25%, but in reverse direction, indicating that more time was worked with fewer interruptions. Mean "severity" (or duration) of time worked was up from about 67 to 73 hours per occurrence. The same was true for routine time worked which is the largest single contributing element in all time worked. Significance lev- els of P = 0.002 were reported for changes in both of these indices. At the same time, scheduled time-off dropped to 69 hours from 104 hours per worker, with a frequency that did not change much at all, and a drop in mean severity from 52 hours per occurrence to 23 hours. Absence levels increased from 31 hours per employee in 30 days to 43 hours per employee. An improvement in quality of data recorded is evident, between the baseline period where the data came from prior payroll records, and the later period 105 T-Test Results for the Hypothesis that Pro-Program Occurred - 30-Day Data Table 2: Shifts in Time-Use Behaviors Relative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 211 912 Frequency 1.64 1.14 1.19 1.05 0.37 3.47 0.00 Severity 103.30 58.12 125.58 61.59 0.50 ~3.11 0.00 Mean Sev. 67.25 53.31 73.77 65.06 0.02 -0.92 0.35 Routine Work 211 912 Frequency 1.65 1.16 1.15 1.04 0.19 3.80 0.00 Severity 103.09 58.07 125.11 61.52 0.50 -3.01 0.00 Mean Sev. 67.60 54.01 72.91 65.16 0.03 -0.75 0.45 All Scheduled Off 211 912 Frequency 4.31 2.11 4.49 1.49 0.00 -0.77 0.43 Severity 103.60 60.25 69.48 41.83 0.00 5.14 0.00 Mean Sev. 52.02 78.11 23.24 47.85 0.00 3.64 0.00 Missing Values 211 912 Frequency 0.26 0.50 0.09 0.29 0.00 3.24 0.00 Severity 43.53 86.20 10.65 46.34 0.00 3.88 0.00 Mean Sev. 40.50 82.19 10.65 46.34 0.00 3.66 0.00 Routine Off 211 912 Frequency 4.10 2.45 4.39 1.61 0.00 -1.15 0.25 Severity 59.26 34.29 56.23 21.71 0.00 0.87 0.38 Mean Sev. 12.02 6.26 11.84 4.46 0.00 0.26 0.79 All Time Absent 211 912 Frequency 1.23 1.07 1.08 0.96 0.18 1.25 0.21 Severity 31.73 45.96 43.78 61.31 0.00 ~1.88 0.06 Mean Sev. 22.50 41.03 36.59 60.21 0.00 -2.32 0.02 Sht. Leg. Absence 211 912 Frequency 0.89 0.97 0.66 0.86 0.17 2.07 0.03 Severity 22.36 35.17 20.30 33.51 0.57 0.50 0.61 Mean Sev. 16.94 31.33 17.86 32.23 0.74 -0.24 0.80 Long Leg. Absence 211 912 Frequency 0.02 0.12 0.11 0.32 0.00 -3.46 0.00 Severity 3.72 29.77 18.75 58.13 0.00 -2.78 0.00 Mean Sev. 3.72 29.77 18.75 58.13 0.00 -2.78 0.00 Sht. Non—Leg. Ab. 211 912 Frequency 0.38 0.66 0.36 0.70 0.58 0.27 0.78 Severity 5.64 13.63 4.72 10.91 0.01 0.62 0.53 Mean Sev. 4.20 9.14 3.47 7.86 0.08 0.71 0.47 NOTE: Sample sizes for the daily cross-sections ranged from 145 to 164 case 106 where observers were trained to record time-uses with greater accuracy. The mean number of hours of missing val- ues dropped from 43 hours to 10 hours per employee per month. These changes were probably due to the extensive par- ticipation of employees and supervisors in group meetings where absenteeism was discussed. In this case, a positive "Hawthorne Effect" seems to have been induced. There were significant increases in infiji time worked (P == .002) and time absent (P = .06), with a corresponding significant decrease in time scheduled off (P = .0001). It is also notable that frequency indices and severity indices do not always vary, either in the same directions or by correspond- ing amounts. Indeed, they may in fact have some tendency to do the reverse for these 30-day measures, a possibility that will be investigated and discussed later, using the time-se- ries correlations. While such a tendency would imply a need to stop using frequency and severity as alternative but par- allel measures for the same "absence," the relationship is quite congruent with theories of substitution and progres- sion effects. T-test results indicate that. prior to full program installation, substantial anticipatory shifts in relevant patterns of time-use behavior had already occurred. This finding indicates that the results of later T-tests of pro- 107 gram impact which incorporate these anticipatory shifts must be interpreted as the result of both the pre-program activi- ties and the absence control program itself. It should be emphasized once again that these pre-pro- gram data, although randomly selected, come from less than a 7% sample of the workforce, and for the assessment of pre- program shifts, these were collected in each of the two cross sections across only 30 days of observation. As such, this was a relatively small sample of time-use behavior. D-statistics have indicated.a1 serious non-normality in the respective distributions. Accordingly, although difference in the means has practical significance and is, in itself, a helpful statistic, the T-test, jprobability statistics reported here are at best tenuous. It should also be noted that for time-use reported in Other Work, Punishment Time, and Late Time, the index values are based upon scarce behav- ior from a much smaller sample of employees, so that results for those time-use variables are even more questionable. For this reason, the results for those indices have not been reported here, although they are included in Appendix D. Note also the extremely low levels of long-term legitimate absence reported prior to the program installation. 108 A Note on Internal Consistency and Relations Among Different Time Uses There is a deliberate but noteworthy internal consis- tency that should be present, by construction, among the changes of severity values especially, one of which demon- strates the integral construction of the set of measures. Consider the following example. There is a general "fund" of eight hours available in each employee's day, seven days a week. For the 30-day values reported in Figure 2, this fund must total to 8 hours times 30 days, or 240 hours. Since the combined measures must account for how all those hours get used, an internal total of 240 hours per employee, (8 hours per day for 30 days), provides a common base or account that all measures of time-use basically draw against. Because the moving values include rounding errors and are calculated against a variable number of employee "cases" due to turnover, the total of (all-time-worked) plus (all-time-scheduled-off) plus (all-time-absent) will tend to vary to some extent above and below this target level. But, a rough equivalence should exist and indeed it does, a veri- fication and testimonial to the integrity of the measures (total = 239 hours instead of 240). This same type of internal relationship holds for the lOO-day and. 365-day measures also, although the "fund" of hours involved is greater. 109 B-2: T-Tests for Pre-Program Intra-Annual Change B-2a: Discussion of T-Test Results on Pre-Program Seasonal Shifts During the baseline year, on the 100-day data, there was no evidence of any significant seasonal change in absence behavior itself whatsoever. This is important in that prior to the program, highly significant changes occurred in both the amount of time worked (up from 342 hours per employee to 409 hours, T-Prob = 0.003), and in the amount of time scheduled off (down from 355 hours per employee to 287 hours, T-Prob = 0.002), and in that during the following year the 30-day data have indicated a shift in absence toward lower short-term absence and more time worked. This suggests that the shifts in absence behavior itself did not begin until after the baseline year when the installation period had started. The pre-program seasonal differences above presumably reflect summer vacation time. It is interesting to note that short-term non-legitimate absence became a much more variable time-use behavior in the winter months than it was during the summer. Furthermore, there was a very large decrease (from 174 to 90 hours) in the amount of time accounted for by ndssing values across the period (see Table 3 below). Some seasonal shifts in non-absence time-uses were occurring prior to program installation, and Hypothesis 3 is supported for time worked and for time scheduled off Table 3: 110 T-Test Results on Hypotheses of General Seasonal Shifts in the Pro-Program loo-Day Measures: October 20, 1978 to April 14, 1979 Relative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 294 469 Frequency 3.65 2.44 3.23 2.08 0.07 1.48 0.13 Severity 342.92 192.61 409.95 165.98 0.09 -3.00 0.00 Mean Sev. 103.83 102.19 156.40 136.01 0.00 -3.51 0.00 Routine Work 294 469 Frequency 3.70 2.54 3.14 1.94 0.00 1.98 0.04 Severity 340.27 191.79 409.51 166.05 0.10 -3.10 0.00 Mean Sev. 101.53 96.80 157.66 135.75 0.00 -3.82 0.00 A11 Scheduled Off 294 469 Frequency 12.58 6.49 13.65 5.29 0.02 -l.45 0.14 Severity 355.48 203.25 287.87 139.84 0.00 3.11 0.00 Mean Sev. 98.66 206.04 48.59 138.99 0.00 2.29 0.02 Missing Values 294 469 Frequency 0.67 1.23 0.40 0.76 0.00 0.21 0.03 Severity 175.69 287.29 90.23 190.11 0.00 2.81 0.00 Mean Sev. 115.12 225.31 71.03 170.35 0.00 1.77 0.07 Routine Off 294 469 ‘ Frequency 12.12 6.99 13.45 5.46 0.01 -1.71 0.08 Severity 177.74 '4.66 194.97 66.93 0.00 ~1.69 0.09 Mean Sev. 13.92 4.86 14.15 3.70 0.00 -0.42 0.67 All Time Absent 294 469 Frequency 3.32 2.28 2.87 2.16 0.52 1.63 0.10 Severity 106.61 126.05 107.23 154.33 0.02 -0.04 0.97 Mean Sev. 41.64 83.67 56.70 140.03 0.00 -1.05 0.29 Sht. Leg. Absence 294 469 Frequency 2.51 1.94 1.77 1.64 0.06 3.33 0.00 Severity 67.45 83.39 52.60 97.88 0.07 1.31 0.19 Mean Sev. 30.99 70.53 33.28 91.39 0}00 0.22 0.82 Long Leg. Absence 294 469 Frequency 0.03 0.21 0.02 0.12 0.00 0.71 0.47 Severity 9.61 79.78 8.56 75.57 0.54 0.11 0.91 Mean Sev. 6.51 51.63 8.56 75.57 0.00 -0.25 0.79 Sht. Non-Leg. Ab. 294 469 Frequency 1.09 1.55 1.22 1.36 0.15 -0.68 0.49 Severity 29.55 64.91 46.08 109.13 0.00 -1.48 0.14 Mean Sev. 19.48 56.43 28.21 93.21 0.08 -0.91 0.36 111 (changes at least partly controlled by the emmdoyer), but not supported at all for absence behavior which occurs at the discretion of the employee. This result is clearly dif- ferent from conventional findings in absence research. It also means that effects upon absenteeism induced by the pro- gram itself may be free from spurious and large cyclical effects. However, once again note that the EDF statistics show little evidence of normality in the distributions compared here, and that these T-test statistics should be considered with caution, especially since missing values accounted for about 40% as much employee time as does routine time worked during the winter of the baseline year prior to the entire intervention. Although the data about time worked or time paid and absent had to be accurately recorded because these were originally legally audited payroll data, no such con- straint was placed upon other unpaid categories or upon the detailed record of reasons for absence, time off, etc., until the installation of the program began. To minimize contamination where error could have otherwise been intro- duced by miscoded data, these were identified and corrected, or where correction wasn't possible, converted to missing data. 112 B-Zb: Discussion of the 'T-Tests on Pre-Program. Holiday Shifts Observed in the lOO-day Measures The lOO-day moving values for the measures were aggre- gated across a short enough period to be sensitive to holi— day and seasonal effects which would be smoothed out of the 365-day data. Nonetheless, the Christmas Holidays had no effect at all on long-term legitimate absence, nor any sig- nificant effect on any of the absence measures except the frequency of short-term legitimate absence, which decreased, as expected, from the holiday to non-holiday period from 2.30 to 1.77, T-Pnflo = 0.01. It appears very likely that the co-occurrent non-significant decrease in the severity index for short-term legitimate absence of five hours per employee was the result of many employees taking a legiti- mate personal leave day during the Christmas period. Except for this minor fluctuation, the holiday had little effect on absence per se prior to the program itself (see Table 4 below). However, as 1J1 the 30-day cross-sections, significant shifts did occur in time worked and in time scheduled off. Time worked per 100 days went up from 354 hours per employee to 409 hours, T-Prob = 0.01, and time scheduled off went down after the holiday period by a corresponding amount from 355 hours to 287 hours ‘per employee (T-Prob 2= 0.002). Apparently both the employees and the employer cooperated to arrange legitimate time off during the holiday period. Table 4: Pro-Program Time-Use Measures: 113 T-Test Results of Holiday Shifts in the loo-Day January 5, 1979 to April 14, 1979 Relative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 370 469 Frequency 3.45 2.30 3.23 2.08 0.24 0.79 0.42 Severity 354.70 201.76 409.95 165.98 0.03 -2.40 0.01 Mean Sev. 113.94 107.86 156.40 136.01 0.01 -2.78 0.00 Routine Work 370 469 Frequency 3.38 2.19 3.14 1.94 0.17 0.93 0.35 Severity 353.93 201.73 409.51 166.05 0.03 -2.42 0.01 Mean Sev. 115.53 111.97 157.66 135.75 0.03 -2.72 0.00 All Scheduled Off 370 469 Frequency 12.63 6.68 13.65 5.29 0.01 -1.35 0.17 Severity 355.79 203.72 287.87 139.84 0.00 3.12 0.00 Mean Sev. 109.67 228.24 48.59 138.99 0.00 2.60 0.01 Missing Values 370 469 Frequency 0.71 1.35 0.40 0.76 0.00 2.32 0.02 Severity 174.57 292.18 90.23 190.11 0.00 2.75 0.00 Mean Sev. 112.70 234.37 71.03 170.35 0.00 1.63 0.10 Routine Off 370 469 Frequency 12.26 7.15 13.45 5.46 0.00 -l.50 0.13 Severity 177.67 100.00 194.97 66.93 0.00 -1.44 0.15 Mean Sev. 12.99 5.80 14.15 3.70 0.00 -1.91 0.05 All Time Absent 370 469 Frequency 3.12 2.20 2.87 2.16 0.80 0.94 0.34 Severity 95.71 136.53 107.23 154.33 0.17 -0.64 0.52 Mean Sev. 48.81 127.11 56.70 140.03 0.27 -0.47 0.63 Sht. Leg. Absence 370 469 Frequency 2.30 1.74 1.77 1.64 0.48 2.54 0.01 Severity 57.79 97.35 52.60 97.88 0.95 0.43 0.66 Mean Sev. 29.92 89.02 33.28 91.39 0.77 0.30 0.76 Long Leg. Absence 370 469 Frequency 0.02 0.12 0.02 0.12 1.00 0.00 1.00 Severity 6.33 71.14 8.56 75.57 0.49 -0.24 0.80 Mean Sev. 6.33 71.14 8.56 75.57 0.49 -0.24 0.80 Sht. Non-Leg. Ab. 370 469 Frequency 1.07 1.56 1.22 1.36 0.12 -0.81 0.42 Severity 31.59 85.68 46.08 109.13 0.01 -l.19 0.23 Mean Sev. 18.65 66.03 28.21 93.21 0.00 -0.95 0.34 114 Hypothesis 4 is supported for shifts in time-use that were at least partly subject to control by the employer, but not for absence in general. Contamination of the program evaluation by a pre-existing Christmas holiday effect would not appear to have been a problem, although this may not generalize from Christmas holidays to the summer July holi- day period. 115 B-2c: Discussbma of the T-Tests on Combined Seasonal and Holiday Effects in the Pre-Program lOO-Day Measures Between Christmas 1978 and June 15, 1979, there was very little change in absence behavior. Once again there was an increase in time worked, from 354 hours per employee to 423 hours, T-Prob = 0.003. There was also a significant decrease in time scheduled off, from 355 to 255 hours per employee, T-Prob = 0.0001; and a very large decrease in time accounted for by missing values, 174 hours to 48 hours, T-Prob = 0.0001. Again, the change in time absent was non- significant except for the frequency of short-term legiti- mate absence (see Table 5). Hypothesis 5 is therefore supported to the extent that combined seasonal and holiday shifts are evident in time use behavior that is at least partly under the control of the employer. However, no significant evidence exists that a significant Christmas and winter to late spring holiday/sea- sonal effect CH1 discretionary! non-legitimate absence occurred within the baseline year and before the program went into effect. Table 5: 116 Effects on the loo—Day Pro-Program Measures: T-Test Results for the Combined Seasonal and Holiday January 5, 1979 to June 15, 1979 Relative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 370 531 Frequency 3.45 2.30 3.40 2.20 0.60 0.17 0.86 Severity 354.70 201.76 423.01 159.02 0.01 -3.02 0.00 Mean Sev. 113.94 107.86 150.06 123.72 0.12 -2.50 0.01 Routine Work 370 531 Frequency 3.38 2.19 3.24 2.03 0.40 0.53 0.59 Severity 353.93 201.73 422.51 158.70 0.01 -3.03 0.00 Mean Sev. 115.53 111.97 157.04 124.26 0.24 -2.82 0.00 A11 Scheduled Off 370 531 Frequency 12.64 6.69 13.65 4.51 0.00 -1.43 0.15 Severity 355.79 203.72 255.44 104.66 0.00 4.98 0.00 Mean Sev. 109.67 228.24 28.59 76.95 0.00 3.82 0.00 Missing Values 370 531 Frequency 0.71 1.35 0.23 0.54 0.00 3.75 0.00 Severity 174.57 292.18 48.25 131.87 0.00 4.47 0.00 Mean Sev. 112.70 234.37 39.94 112.14 0.00 3.18 0.00 Routine Off 370 531 Frequency 12.26 7.15 13.57 4.61 0.00 -1.74 0.08 Severity 177.67 100.00 204.84 53.32 0.00 -2.52 0.01 Mean Sev. 12.99 5.80 15.15 3.03 0.00 -3.76 0.00 All Time Absent 370 531 Frequency 3.12 2.20 2.85 2.07 0.48 1.02 0.30 Severity 95.71 136.53 126.46 174.34 0.01 -1.58 0.11 Mean Sev. 48.81 127.11 71.03 154.34 0.03 -1.26 0.20 Sht. Leg. Absence 370 531 . Frequency 2.30 1.74 1.63 1.69 0.73 3.15 0.01 Severity 57.79 97.35 61.29 109.88 0.17 -0.27 0.78 Mean Sev. 29.92 89.02 39.82 88.34 0.93 -0.90 0.37 Long Leg. Absence 370 531 Frequency 0.02 0.12 0.02 0.12 1.00 0.00 1.00 Severity 6.33 71.14 12.16 97.28 0.00 -0.55 0.58 Mean Sev. 6.33 71.14 12.16 97.28 0.00 -0.55 0.58 Sht. Non-Leg. Ab. 370 531 Frequency 1.07 1.56 1.28 1.46 0.41 -1.11 0.26 Severity 31.59 85.68 53.02 122.59 0.00 -1.63 0.10 Mean Sev. 18.65 66.03 34.44 106.97 0.00 -1.43 0.15 117 B-3: T-Tests of Long-Term Change During the Time that the Program Was Fully in Force B-3a: Discussion of the T-Tests for Year-to-Year Change in the 100-Day Frequency and Severity Measures During the Pro- gram Period (See Table 6 and Table 7 below) Between the summer of 1980 and the summer of 1981, both the frequency indices and severity indices show significant shifts in time worked (down from 376 hours to 264 hours), in time scheduled off (up from 250 hours to 394 hours), and in missing values (up from 38 hours to 246 hours). This increase in missing values was startling. 0n the average, an increase of only one occurrence in the frequency measure co-occurred with this anomaly, which indicates a severe one-episode time-use occurred for each employee. Comparison between day 1347 and day 1390 suggests that this event must have occurred sometime between July 1 and August 12, 1981, since the event was not included in the lOO-day base for day 1390. The pattern here suggested a strike or massive shutdown during the summer of 1981. The frequency indices also indicate significant year-to-year increases for other periods in overall frequency of legitimate absences and a corresponding overall increase in the frequency of work intervals, and these two findings are both consistent with the this work interruption hypothesis. An anomaly so pervasive and unusual could neither be an accident nor be ignored, so the DOT Personnel Director was telephoned for Table 6: 118 Change in the loo-Day Measures, Frequency Indices Only During the Program Period T-Test Results for Hypotheses about Long-Term Year-to-Year Relative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 1166 2.95 2.22 1531 4.02 2.52 0.12 -3.83 0.00 982 3.37 2.04 1347 2.62 1.87 0.31 3.29 0.00 1025 3.45 2.38 1390 4.06 2.69 0.13 -2.06 0.04 1210 3.01 2.17 1575 4.02 2.50 0.09 3.69 0.00 Routine Work 1166 2.89 2.28 1531 3.93 2.43 0.43 —3.75 0.00 982 3.34 2.02 1347 2.60 1.82 0.22 3.27 0.00 1025 3.42 2.28 1390 3.99 2.59 0.12 -l.96 0.05 1210 2.97 2.25 1575 4.00 2.52 0.17 -3.68 0.00 A11 Sch. Off 1166 13.56 6.67 1531 13.38 5.15 0.00 0.26 0.79 982 12.91 4.30 1347 9.30 3.82 0.16 7.61 0.00 1025 13.64 4.74 1390 12.28 5.04 0.46 2.37 0.01 1210 12.75 6.65 1575 13.54 5.25 0.01 1.12 0.26 Missing Values 1166 0.15 0.43 1531 0.28 1.76 0.00 -0.82 0.41 982 0.17 0.41 1347 1.10 0.94 0.00 -10.85 0.00 1025 0.13 0.37 1390 0.21 1.48 0.00 -0.68 0.49 1210 0.15 0.37 1575 0.30 1.74 0.00 -l.06 0.28 Routine Off 1166 13.42 6.87 1531 13.31 5.29 0.00 0.15 0.88 982 12.77 4.41 1347 9.39 4.26 0.69 6.67 0.00 1025 13.54 4.87 1390 12.20 5.19 0.44 2.27 0.02 1210 12.58 6.87 1575 13.47 5.39 0.00 -1.24 0.21 All Time Absent 1166 2.47 2.04 1531 3.27 2.31 0.14 -3.12 0.00 982 3.18 1.97 1347 2.54 1.81 0.32 2.89 0.00 1025 3.40 2.32 1390 3.59 2.59 0.20 -2.66 0.50 1210 2.48 1.97 1575 3.45 2.38 0.02 -3.80 0.00 Sht. Leg. Ab. 1166 1.81 1.69 1531 2.51 2.09 0.01 -3.12 0.00 982 2.36 1.69 1347 2.26 1.78 0.56 0.50 0.61 1025 2.70 2.00 1390 3.30 2.63 0.00 -2.16 0.03 1210 1.73 1.75 1575 2.67 2.13 0.02 -4.10 0.00 Long Leg. Ab. 1166 0.22 0.54 1531 0.33 0.78 0.00 -1.36 0.17 982 0.60 0.62 1347 0.33 0.59 0.57 3.78 0.00 1025 0.59 0.64 1390 0.34 0.65 0.74 3.30 0.00 1210 0.21 0.48 1575 0.37 0.76 0.00 -2.18 0.03 Sht Non-Leg. Ab. 1166 0.68 1.23 1531 0.73 1.17 0.58 -0.39 0.69 982 0.45 0.96 1347 0.20 0.45 0.00 2.89 0.00 1025 0.30 0.93 1390 0.17 0.43 0.00 1.52 0.12 1210 0.75 1.19 1575 0.74 1.17 0.82 0.03 0.97 Table 7: 119 Change in the loo-Day Measures, Severity Indices Only, During the Program Period T-Test Results for Hypotheses about Long-Term Year-to-Vear Relative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 1166 412.55 204.59 1531 436.37 184.62 0.22 -1.05 0.29 982 376.40 151.03 1347 264.89 126.99 0.04 6.84 0.00 1025 415.21 172.11 1390 403.56 182.65 0.47 0.56 0.57 1210 418.51 214.19 1575 435.45 191.03 0.17 -0.71 0.47 Routine Work 1166 411.81 204.68 1531 434.87 184.84 0.22 -1.01 0.31 982 375.34 150.62 1347 264.43 126.71 0.04 6.83 0.00 1025 414.52 171.82 1390 403.33 182.38 0.47 0.54 0.59 1210 417.77 213.92 1575 434.43 190.77 0.17 0.70 0.48 All Sch. Off 1166 282.32 197.86 1531 257.06 161.11 0.01 1.20 0.23 982 250.52 135.30 1347 394.01 105.51 0.00 -10.14 0.00 1025 261.13 146.87 1390 222.48 146.30 0.96 2.26 0.02 1210 270.39 203.50 1575 255.13 168.09 0.02 0.70 0.48 Missing Values 1166 92.03 256.02 1531 56.06 193.95 0.00 1.36 0.17 982 38.78 157.57 1347 246.07 85.02 0.00 -14.11 0.00 1025 59.60 186.22 1390 30.92 137.55 0.00 1.50 0.13 1210 95.74 256.74 1575 63.06 206.91 0.01 1.20 0.23 Routine Off 1166 186.75 89.00 1531 193.50 72.37 0.01 -0.71 0.47 982 207.50 79.67 1347 137.74 65.84 0.02 8.18 0.00 1025 200.09 71.03 1390 178.97 79.69 0.17 2.38 0.01 1210 168.17 81.55 1575 186.91 72.52 0.16 -2.08 0.03 All Time Absent 1166 110.01 190.81 1531 110.21 164.91 0.08 -0.01 0.99 982 168.99 173.15 1347 140.04 170.53 0.86 1.44 0.15 1025 126.99 172.50 1390 178.52 218.35 0.00 -2.23 0.0 1210 115.05 195.41 1575 112.53 165.16 0.05 0.60 0.54 Sht. Leg. Ab. 1166 44.68 95.66 1531 50.62 78.71 0.02 -0.58 0.56 982 65.06 76.12 1347 60.10 82.12 0.36 0.53 0.59 1025 68.29 80.62 1390 77.89 102.19 0.00 -0.89 0.37 1210 46.20 103.29 1575 42.50 58.71 0.00 0.38 0.70 Long Leg. Ab. 1166 55.63 170.79 1531 49.73 150.92 0.14 0.31 0.75 982 97.33 172.49 1347 75.74 166.34 0.66 1.09 0.27 1025 53.74 168.58 1390 97.41 216.84 0.00 -1.91 0.05 1210 58.01 170.66 1575 68.35 161.32 0.50 -0.02 0.98 Sht Non-Leg. Ab. 1166 9.70 25.08 1531 9.85 17.38 0.00 -0.06 0.95 982 6.54 19.36 1347 4.19 13.35 0.00 1.21 0.22 1025 4.97 18.08 1390 3.22 11.13 0.00 1.00 0.31 1210 10.83 23.32 1575 11.69 27.13 0.03 -0.29 0.77 120 comment on unusual events during the period. The explana- tion fits very well. Apparently, the mayor of Detroit gave those employees an extra 80 hours of legitimate paid vaca— tion in a concession bargaining decision at that time. This was not appropriately coded as short-term legitimate absence "missing values." In addition, and therefore became part of a 70 hour reduction in routine time off and a 30 hour reduc- tion in all time absent contributed their effect to account for the rest of these missing hours. Except for the anomaly during the summer of 1981, the only significant change in overall absence for these tests was in long-term legitimate absence which increased from 53 to 97 hours per employee immediately after the summer ano- maly occurred. Hypothesis 6 was supported for the frequency of both absence and time worked. Both increased. On the other hand, this did not significantly affect severity of either index which suggests that the employees were developing a long-term habit of using their casual leave days in several occurrences instead of just a few. Overall, it appears that except for some increase in long-term absence, the lOO-day measures do not show evidence of major shifts in absence within the time span of the fully implemented absence con- trol program. 121 These findings are highly vulnerable to short—term intra-annual effects, however, and despite the comparison of corresponding year-to-year dates, they should be considered with caution. B-3b: Discussion of T-Test Results for Hypotheses about Holiday Effects in the 100-Day Measures During the Program Period Across the period when the program itself was fully implemented, both the frequency and severity indices were T-tested to detect significant holiday effects. But, the results were essentially the same for both so that a detailed presentation of only the severity shifts is included in Table 8 below. In brief, two significant shifts have occurred across major holiday periods within the span of the program. These occurred for tflma time-use measures during the summers of 1981 and of 1980. One large shift (an increase in time worked, decrease in time scheduled off, and in missing values) occurred just after the interval for which the anomaly which affected Hypothesis 6 showed up (Day 1347 to Day 1390). Apparently things went back to normal after the sudden increase of paid time off had been used by employees. A smaller effect, a decrease in long term legit- imate absence was also associated with a summer (July 4th) holiday period in 1980 and was also noticeable and signifi- cant. 122 Table 8: T-Test Results for Holiday Effects During the Program Period on the loo-Day Data, Using Severity Indices Only Relative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 1166 412.55 204.59 1210 418.51 214.19 0.58 -0.25 0.80 1531 436.37 184.62 1575 435.41 191.03 0.69 0.04 0.96 982 376.40 151.03 1025 415.21 172.11 0.11 -2.08 0.03 1347 264.89 126.99 1390 403.56 182.65 0.00 —7.40 0.00 Routine Work 1166 411.81 204.68 1210 417.77 213.92 0.59 -0.25 0.80 1531 434.87 184.84 1575 434.43 190.77 0.71 0.02 0.98 982 375.41 150.52 1025 414.52 171.82 0.11 -2.11 0.03 1347 264.43 126.71 1390 403.33 182.38 0.00 7.43 0.00 A11 Sch. Off 1166 282.32 197.86 1210 270.39 203.50 0.73 0.52 0.60 1531 257.06 161.11 1575 255.13 168.09 0.62 0.10 0.92 982 250.52 135.30 1025 261.23 146.87 0.32 -0.66 0.51 1347 394.01 105.51 1390 222.48 146.30 0.00 11.29 0.00 Missing Values 1166 92.03 256.02 1210 95.74 256.74 0.97 -0.13 0.90 1531 56.06 193.95 1575 63.06 206.91 0.45 -0.29 0.76 982 38.78 157.57 1025 59.60 186.22 0.04 -l.05 0.29 1347 246.07 85.02 1390 30.92 137.55 0.00 15.80 0.00 Routine Off 1166 186.75 89.00 1210 168.17 81.55 0.29 1.89 0.05 1531 193.50 72.47 1575 186.91 72.52 0.99 0.76 0.44 982 207.50 79.67 1025 200.09 71.03 0.16 0.85 0.39 1347 137.74 65.84 1390 178.97 79.69 0.02 -4.74 0.00 All Time Absent 1166 110.01 190.81 1210 115.05 195.41 0.77 -0.23 0.82 1531 110.21 164.91 1575 112.53 165.16 0.99 -0.12 0.90 982 168.99 173.15 1025 126.99 175.50 0.96 2.11 0.03 1347 140.04 170.53 1390 178.52 218.35 0.00 -1.65 0.10 Sht. Leg. Ab. 1166 44.68 95.66 1210 46.20 103.29 0.35 -0.13 0.89 1531 50.62 78.71 1575 42.50 58.71 0.00 0.98 0.32 982 65.06 76.12 1025 68.29 80.62 0.48 -0.36 0.72 1347 60.10 82.12 1390 77.89 102.19 0.01 -1.61 0.10 Long Leg. Ab. 1166 55.63 170.79 1210 58.01 170.66 0.99 -0.12 0.90 1531 49.73 150.92 1575 58.35 161.32 0.43 -0.46 0.64 982 97.38 172.49 1025 53.74 188.58 0.78 2.22 0.02 1347 75.74 166.34 1390 97.41 216.84 0.00 —0.94 0.34 Sht Non-Leg. Ab. 1166 9.70 25.08 1210 10.83 23.32 0.37 -0.41 0.68 1531 9.86 17.38 1575 11.69 27.13 0.00 -0.67 0.50 982 6.54 19.36 1025 4.97 18.08 0.40 0.73 0.46 1347 4.19 13.35 1390 3.22 11.13 0.03 0.66 0.50 123 No holiday effects were apparent across the Christmas periods however, and while this one finding does offer weak support for Hypothesis 7 as it relates to the summer holi- days, this holiday effect may have been at least partly the effect of annual renewal or vacation. There was, for exam- ple, no significant effect across the winter (Christmas) holiday season when renewal would not be expected to have much effect and holidays might. Once again, some rather anomalous shifts and reductions in time-worked as well as time scheduled off occurred during the summer of 1981. 124 B-3c: Discussion of Results for the T-Tests of Combined Seasonal, Holiday and Annual Renewal Effects During the Pro- gram Period on the lOO-Day Measures Across the span of the program period itself, the 100-day measures were also examined for combined cyclical effects. Several T-test comparisons for the lOO-day meas- ures offer general support for Hypothesis 8. Combined sea- sonal, holiday and renewal effects on absence do appear to take pdace across surrounding periods not contaminated by the anomaly in the early summer of 1981, producing signifi- cant decreases in levels of absence and increases in time worked from summer to winter. Since the same summer of 1981 anomaly (Date 1 = 1347) which was evident in earlier T-tests was also presumably contributing to combination effects in that year, the 1981 results (using Day 1347) really don't reflect cyclical change (see Table 9 below). Comment on the Measures Used to Examine Intra-Annual T-Test Results In general, the frequency and mean severity indices co-vary with the severity indices and tend to provide sig- nificant results across the same intervals, although in this there were some exceptions. For convenience, in the present analysis the results on the frequency indices have not been included. Detailed examination of the relationship between severity and frequency of absence is definitely an interest- ing problem. Indeed, the parallelism in the 100-day indices between severity and frequency is in the reverse direction l25 Table 9: T-Test Results for Hypotheses of Combined Holiday, Seasonal, and Renewal Effects During the Program Period on the loo-Day Measures, Severity Indices Only Relative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 982 376.40 151.03 1166 412.55 204.59 0.00 -.175 0.08 1347 264.89 126.99 1531 436.37 184.62 0.00 -9.09 0.00 1025 415.21 172.11 1210 418.51 214.19 0.01 -0.15 0.88 1390 403.56 182.65 1575 435.45 191.03 0.60 -1.43 0.15 Routine Work 982 375.34 150.52 1166 411.81 204.68 0.00 -1.76 0.07 1347 264.43 126.71 1531 434.87 184.84 0.00 -9.03 0.00 1025 414.52 171.82 1210 417.77 213.92 0.01 -0.15 0.16 1390 403.33 182.38 1575 434.43 190.77 0.60 -l.40 0.16 All Sch. Off 982 250.52 135.30 1166 282.32 197.86 0.00 -1.63 0.10 1347 394.01 105.57 1531 257.06 161.11 0.00 -8.44 0.00 1025 261.23 146.87 1210 270.39 203.50 0.00 -0.45 0.65 1390 222.48 146.30 1575 255.13 168.09 0.10 -1.74 0.08 Missing Values 982 38.78 157.57 1166 92.03 256.02 0.00 -2.18 0.03 1347 246.07 85.02 1531 56.01 193.95 0.00 10.65 0.00 1025 59.60 186.22 1210 95.74 256.74 0.00 -1.40 0.162 1390 30.92 137.55 1575 63.06 206.91 0.00 -1.54 0.12 Routine Off 982 207.50 79.67 1166 186.75 89.00 0.18 2.13 0.03 1347 137.74 65.84 1531 193.50 72.47 0.26 -6.76 0.00 1025 200.09 71.03 1210 168.17 81.55 0.09 3.63 0.00 1390 178.97 79.69 1575 186.91 72.52 0.27 -0.88 0.38 All Time Absent 982 168.99 173.15 1166 110.01 190.81 0.24 2.81 0.00 1347 140.04 170.53 1531 110.21 164.91 0.69 1.49 0.13 1025 126.99 172.50 1210 115.05 195.41 0.13 0.56 0.57 1390 178.52 218.35 1575 112.53 165.16 0.00 2.86 0.00 Sht. Leg. Ab. 982 65.06 76.12 1166 44.68 95.66 0.00 2.05 0.04 1347 60.10 82.12 1531 50.62 78.71 0.62 0.98 0.32 1025 68.29 80.62 1210 46.20 103.29 0.00 2.07 0.03 1390 77.89 102.19 1575 42.50 58.71 0.00 3.57 0.00 Long Leg. Ab. 982 97.38 172.49 1166 55.63 170.79 0.90 2.11 0.03 1347 75.74 166.34 1531 49.73 150.93 0.25 1.38 0.17 1025 53.74 168.58 1210 58.01 170.66 0.88 -0.22 0.82 1390 97.41 216.84 1575 58.35 161.32 0.00 1.72 0.08 Sht Non-Leg. Ab. 982 6.54 19.36 1166 9.70 25.08 0.00 -1.23 0.22 1347 4.19 13.35 1531 9.85 17.38 0.00 -3.07 0.00 1025 4.97 18.08 1210 10.83 23.32 0.00 -2.44 0.01 1390 3.22 11.13 1575 11.69 27.13 0.00 -3.43 0.00 126 to an apparent co-variation suggested by the 30—day meas- ures, where frequency went down as severity went up. It is possible that this latter effect is made evident by the use of a shorter time period as the base for these measures, and that both patterns can and do operate at the same time so that they covary positively over long periods and negatively over short periods. Further insight might be forthcoming fnmn the use of time—series correlations between measures across relatively short time periods such as one week, but no support for the inverse relationship appears when the lOO-day measures are correlated (see Table 13). 127 B-4: Summary and Discussion of the T-Tests on Overall Pro- gram-Induced Change in the loo-Day Measures The program induced changes between the baseline year and later periods were not entirely as predicted for the lOO—day data, but were unquestionably strong and apparently significant. Time Scheduled Off, Time Worked, and All Time Absent decreased significantly from the end of the baseline year to the end of the first program year. Then all three significantly increased. again. back toward. their original levels by the end of the second program year. In the mean time, missing values were of course affected by the anomaly in the late spring or early summer of 1981 and increased noticeably between the end of the baseline year and the end of the first program year, then decreased dramatically back to much less than their original level by the end of the second program year. Overall, between the last 100 days of the baseline year and the last 100 days of the second pro— gram year, time worked apparently showed no significant change, nor did time scheduled off or overall time absent. However, these overall year-to—year tests of program induced changes in the lOO-day measures were likely to be affected both by the anomaly during the summer of 1981 and by other short-term fluctuations for the lOO-day measures, and in any case are only comparisons of the fourth quarter values in each respective year. The 365-day measures below were used instead to assess overall program impact on annual time- 128 detailed tabular presentation of evidence of long-term inter-annual change in the 100-day measures has been omitted here. B-5: Discussion of the T-Tests on Overall Long-Term Program Induced Changes Indicated by the 365-Day Measures, A Direct Evaluation of the Program Figure 7 below offers a rough graphical summary of overall changes in the major annual time-use indices. Full details have been presented in Table 10. Overall, time worked per employee per year went up by 107 hours to 1490 hours per year. Time scheduled off was down by 200 hours to 915 hours. And an overall increase in time absent of 110 hours per employee per year took place, all of which was in long-term legitimate absence since short-term non-legitimate absence dropped 112 hours to 27 hours per employee per year from 139 hours, and short-term legitimate absence also dropped 40 hours (non significant). Most of the observed change in absence behavior had occurred by the end of the first complete year of the program. Also worthy of note is a substantial decrease in the level of missing values present in the data, over the second year of the program especially. It appears likely that many of the 80 hours given away in concession bargaining during the first year were recorded as missing values and that a certain amount of time must have been needed after program start-up to streamline the administration of data collection by supervisors, data-entry clericals, etc. Day Day 546 1277 All Time Worked '05 Missing Values MS I "4» 129 Frequency of time worked up 1 Severity of time worked up 107 hrs Mean severity time worked - same Mean Sev. Time Sch. Off up 57 hrs Freq. Time Scheduled Off - same Severity Time Sch. Off down 200 hrs Severity All Absence up 110 hrs Mean Severity A11 Absence up 42 hrs Frequency of All Absence up 1 Long-Term Legitimate up 223 hrs Short-Term Legitimate down 1 hr Short-Term Non-Legitimate down 112 Frequency Missing Values up 0.5 Severity Missing Values down 218 Mean Sev. Missing Values down 114 NOTE: The relative scales of sketches above are merely roughly "eyeballed" pictures of trends in each line, clustered for convenience. Figure 7: Approximate Graphic Profiles of Long-Term Impact on the Overall 365-Day Time-Use Measures (Relative Scales) 130 Table 10: T-Test Results of Program Induced Change in the 365-Day Measures Relative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 546 1277 Frequency 11.17 5.44 9.34 5.29 0.74 2.81 0.00 Severity 1383.33 517.67 1322.63 591.38 0.12 0.91 0.36 Mean Sev. 151.10 108.38 155.07 106.95 0.87 -0.30 0.76 Routine Work 546 1277 Frequency 10.90 5.33 9.20 5.16 0.69 2.69 0.00 Severity 1379.21 517.23 1320.09 590.91 0.12 0.89 0.37 Mean Sev. 154.12 109.31 156.97 107.91 0.87 -0.22 0.82 A11 Scheduled Off 546 1277 Frequency 47.79 18.37 44.21 18.61 0.88 1.61 0.10 Severity 1130.93 501.61 1160.86 588.36 0.07 -0.46 0.64 Mean Sev. 40.09 70.25 164.61 573.57 0.00 -2.64 0.00 Missing Values 546 1277 Frequency 1.29 2.27 1.09 0.36 0.00 1.02 0.31 Severity 419.68 703.82 506.07 753.95 0.43 -0.98 0.32 Mean Sev. 227.95 541.41 471.54 710.64 0.00 -3.23 0.00 Routine Off 546 1277 Frequency 47.28 18.85 43.70 18.79 0.97 1.57 0.11 Severity 701.85 245.18 641.01 269.04 0.28 1.97 0.50 Mean Sev. 15.09 2.47 14.09 5.02 0.00 2.14 0.03 All Time Absent 546 1277 Frequency 10.76 5.50 8.92 5.11 0.39 2.87 0.00 Severity 396.29 459.44 431.03 560.36 0.02 -0.57 0.57 Mean Sev. 52.92 141.63 103.63 338.48 0.00 -1.67 0.09 Sht. Leg. Absence 546 1277 Frequency 7.44 3.90 6.95 3.99 0.80 1.02 0.31 Severity 222.06 277.14 180.44 256.47 0.36 1.29 0.19 Mean Sev. 39.41 105.95 32.11 79.58 0.00 0.64 0.52 Long Leg. Absence 546 1277 Frequency 0.04 0.23 0.95 1.05 0.00 -10.35 0.00 Severity 34.60 274.80 221.22 509.37 0.00 -3.88 0.00 Mean Sev. 23.05 159.09 127.80 333.81 0.00 -3.42 0.00 Sht. Non-Leg. Ab. 546 1277 Frequency 3.94 3.80 1.75 2.55 0.00 5.49 0.00 Severity 139.63 275.64 29.36 61.49 0.00 4.40 0.00 Mean Sev. 35.86 90.91 8.33 13.94 0.00 3.37 0.00 131 Table 10 Continued RElative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 546 1642 Frequency 11.17 5.44 12.21 6.80 0.01 -l.39 0.16 Severity 1383.33 517.67 1490.70 591.42 0.13 -1.58 0.11 Mean Sev. 151.10 108.38 147.33 127.49 0.06 0.27 0.78 Routine Work 546 1642 Frequency 10.90 5.33 11.94 6.56 0.02 -1.43 0.15 Severity 1379.21 517.23 1487.47 591.21 0.13 -1.60 0.11 Mean Sev. 154.12 109.31 148.20 125.65 0.11 0.41 0.68 A11 Scheduled Off 546 1642 Frequency 47.79 18.37 48.18 16.45 0.20 -0.18 0.85 Severity 1130.93 501.61 915.97 510.73 0.84 3.47 0.00 Mean Sev. 40.09 70.25 92.32 387.79 0.00 -1.57 0.11 Missing Values 546 1642 Frequency 1.29 2.27 1.70 5.14 0.00 -0.84 0.40 Severity 419.68 703.82 201.50 597.66 0.06 2.71 0.00 Mean Sev. 227.95 541.41 113.30 396.57 0.00 1.95 0.05 Routine Off 546 1642 , Frequency 47.28 18.85 47.81 16.62 0.15 -0.24 0.81 Severity 701.85 245.18 683.84 234.47 0.61 0.61 0.54 Mean Sev. 15.09 2.47 13.75 3.87 0.00 3.41 0.00 All Time Absent 546 1642 Frequency 10.76 5.50 11.74 6.58 0.04 —l.33 0.18 Severity 396.29 459.44 506.01 564.39 0.02 -1.75 0.08 Mean Sev. 52.92 141.63 94.15 350.52 0.00 -1.28 0.20 Sht. Leg. Absence 546 1642 Frequency 7.44 3.90 10.01 5.82 0.00 -4.29 0.00 Severity 222.06 277.14 221.06 224.37 0.02 0.03 0.97 Mean Sev. 39.41 105.95 34.91 123.56 0.08 0.32 0.74 Long Leg. Absence 546 1642 Frequency 0.04 0.23 1.03 1.61 0.00 7.22 0.00 Severity 34.60 274.80 257.50 534.78 0.00 —4.35 0.00 Mean Sev. 23.05 159.09 107.46 286.28 0.00 -3.02 0.00 Sht. Non-Leg. Ab. 546 1642 Frequency 3.94 3.80 1.75 2.25 0.00 5.62 0.00 Severity 139.63 275.64 27.44 54.52 0.00 4.49 0.00 Mean Sev. 35.86 90.91 7.80 11.44 0.00 3.44 0.00 132 Table 10 Continued Relative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 1277 1642 Frequency 9.34 5.28 12.21 6.80 0.00 -4.00 0.00 Severity 1322.63 591.38 1490.70 591.42 1.00 -2.43 0.21 Mean Sev. 155.07 106.95 147.33 127.49 0.03 0.56 0.57 Routine Work 1277 1642 Frequency 9.20 5.16 11.94 6.56 0.00 -3.96 0.00 Severity 1320.09 510.91 1487.47 591.21 0.99 -2.42 0.01 Mean Sev. 156.97 107.91 148.20 125.65 0.07 0.64 0.52 All Sch. Off 1277 1642 Frequency 44.27 18.61 48.18 16.45 0.14 -1.93 0.05 Severity 1160.86 588.36 915.97 510.73 0.09 3.80 0.00 Mean Sev. 164.61 573.57 92.32 387.79 0.00 1.27 0.20 Missing Values 1277 1642 Frequency 1.09 0.36 1.70 5.14 0.00 -1.40 0.16 Severity 506.07 753.95 201.50 597.66 0.00 3.84 0.00 Mean Sev. 471.54 710.64 113.30 396.57 0.00 5.36 0.00 Routine Off 1277 1642 Frequency 43.70 18.79 47.81 16.62 0.14 -1.98 0.04 Severity 641.01 269.04 683.84 234.47 0.10 -1.45 0.14 Mean Sev. 14.09 5.02 13.75 3.87 0.00 0.66 0.50 All Time Absent 1277 1642 Frequency 8.92 5.11 11.74 6.58 0.00 -4.08 0.00 Severity 431.03 560.36 506.01 564.39 0.93 -1.14 0.25 Mean Sev. 103.63 338.48 94.15 350.52 0.67 0.23 0.81 Sht. Leg. Ab. 1277 1642 Frequency 6.95 3.99 10.01 5.82 0.00 -5.21 0.00 Severity 180.44 256.47 221.06 224.37 0.11 -1.44 0.15 Mean Sev. 32.11 79.58 34.91 123.56 0.00 -0.23 0.81 Long Leg. Ab. 1277 1642 Frequency 0.95 1.05 1.03 1.61 0.00 -0.47 0.64 Severity 221.22 509.37 257.50 534.78 0.58 -0.59 0.55 Mean Sev. 127.80 333.81 107.46 286.28 0.07 0.56 0.57 Sht Non-Leg. Ab. 1277 1642 Frequency 1.75 2.55 1.75 2.25 0.14 0.01 0.99 Severity 29.36 61.48 27.44 54.52 0.15 0.28 0.77 Mean Sev. 8.33 13.94 7.80 11.44 0.02 0.36 0.72 133 Overall, it appears from mere comparison of the means that the program was clearly successful although the T-test probability statistics are of questionable value. Once again, a caveat is critical here. These T-test values are based on distinctly non-normal data distributions and come from a 7% sample of the workforce, a sampling procedure necessitated by the sheer volume of computer time and memory space requirements needed to prepare the derived data for detailed analysis. Hypothesis 9 is supported. See the over-time plots in Figure 7 for a rough graphical picture of the program impact over time. Aside from the anomaly in the early summer of 1981, several issues become apparent from analysis of the T-test data. 1. There was an apparent overall decrease in short-term non-legitimate absence over the long term as intended. 2. This was absorbed by an increase in long-term legitimate absence instead of producing a larger increase in the number of hours worked, but was accompanied by an increase in hours worked which was large enough for practical if not statistical significance (an estimated 7% increase). 134 There was a small but significant effect induced by annual renewal of the number of casual leave days granted to every employee, such that in early July a surge of short-term legitimate absence occurred. No significant evidence of this effect occurred during the Christmas holiday period, which suggests that the holiday effect only happened in associa- tion with the annual renewal, basically a cyclical version of the legitimacy effect. Seasonal shifts in time worked and time sched- uled off do occur surrounding holiday periods and the annual renewal period in July, but seasonal cycles seem to have little effect on absence itself, a finding that was counter to prior research. As in prior research, the T-tested data dis- tributions are non-normal to the point where probability estimates are of questionable merit, and although findings in the present work are generally consistent with much of the prior work, this leaves the problem of how to improve 13x31 the cross-sectional comparisons unresolved. 135 C. The Split-Half Time-Series Reliability Correlations The procedure for calculating split-half reliability correlations has been presented and analyzed in detail in Appendix C. The results have been described below. Table 11 presents reliabilities for those measures that were cen- tral to the substantive questions to be investigated, and identifies the conditions which brought about a few anoma- lously low reliabilities. A list of such conditions, any of which could cause low reliabilities, follows the results. The Reliability 9f Measures for the Substantive Hypotheses The results below have been reported for three sets of time-series correlations between split-halves of the time- series data for specific measures central to analysis of the substantive hypotheses. The three time-series. data sets were: (1) using the lOO-day data, an interval in the base- line year from day 281 to'day 546; (2) also using the 100-day data, an interval during the posteinstallation period, from day 982 to day 1680; and (3) using the 365—day data, an interval starting 365 days after data collection began for the post-installation period and continuing until day 1680. For each of the three sets, reliabilities for 27 different time-use indices have been presented. In general, these reliability correlations were unusually good, for data related to absenteeism. See Table 11 and the following dis- cussion for details. 136 Table 11: Time Series Correlations for Split-Half Reliability Variable Pre-Prog. or Index rleO-Day All Time Worked Frequency 0.75 Severity 0.98 Mean Sev. 0.88 Routine Work Frequency 0.76 Severity 0.97 Mean Sev. 0.91 All Scheduled Off Frequency 0.93 Severity 0.99 Mean Sev. 0.97 Missing Values Frequency 0.89 Severity 0.99 Mean Sev. 0.98 Routine Off Frequency 0.95 Severity 0.97 Mean Sev. 0.86 All Time Absent Frequency 0.82 Severity 0.74 Mean Sev. 0.58 Sht. Leg. Absence Frequency 0.91 Severity Sample too Mean Sev. small Long Leg. Absence Frequency No Values in Severity 1/2 Prior to Mean Sev. Installation Sht. Non-Leg. Ab. Frequency Variation too small Severity 0.96 Mean Sev. 0.68 Post-Inst r:100-Day GOO GOO GOO 000 000 GOO 000 000 .88 .92 .50 .89 .91 .57 .99 .96 .82 .88 .96 .96 .99 .98 .92 .89 .91 .73 .92 .57 .21 .84 .94 .89 .76 .36 .56 Post-Inst r:365-Day .99 .63 .50 000 .98 .61 .64 000 .98 .87 .92 000 .46 .87 .84 000 0.97 0.99 Variation too small 0.98 0.74 0.64 0.97 Variation too small 0.88 0.87 0.69 Sample became too small 137 As Table 11 shows, reliabilities for most of the time- use measures were high compared with conventional absence measures, typically ranging between 0.5 and 0.99 with most values between 0.8 anui 0.98. Correlations for indices of All Time Worked and for Routine Work (which was the largest single component of "All Time Worked") were both very strong and both very similar, almost identical when compared across the indices of Frequency, Severity and Mean Severity. The same strong results were also found for All Time Scheduled Off and for All Time Absent. Not surprisingly, Mean Sever- ity uniformly had somewhat lower reliabilities than Fre- quency or Severity. The procedure for calculation of values for mean severity computes a ratio each day for Severity divided by Frequency, thereby allowing any measurement error for either Severity or Frequency to affect the daily value of Mean Severity. "Missing values" were an unusual but apparently relia- ble measure. As Tables 2 through 10 above have shown, the frequency of missing periods across the measurement inter- vals for DOT employees was quite low (less than one per employee) while the severity of missing value occurrences tended to be high, especially in the baseline period and early in the program when the incidence of undocumented periods of employee time-use Ibehavior' was much. greater. These conditions could very easily have brought a sporadic distribution of the time-use across the two split halves and 138 led to low reliabilitis for the missing values. But, appar- ently this time-use was common to many employees and was adequately represented in both split halves. The reliabil- ity estimates of the missing values shown in Table 11 are quite strong and consistent, although they do show slightly smaller reliabilities for frequency than severity. Anomalous Values in the Split-Half Reliabilities Among the 78 split half reliabilities on measures for the substantive hypotheses were fifteen unusually low or anomalous values (see Table 11). Conditions under which these could be expected to occur have been described in detail in Appendix C. A brief outline of those conditions follows. Conditions Conducive to Anomalously Low Reliability in the Time-Use Measures 1. The sample size can be too small, including too few employees. 2. Very few employees may participate in the respective time use. 3. The time-use itself may rarely occur, even for those who did choose it. 4. The time-use may be so carefully regulated or may other- wise have occurred with such consistency that the amount of observed variation underlying the split-half reli- ability estimate was too tiny. 139 5. The amount of timeduse in one Split half or the other may be constant over time causing a standard deviation of zero and making calculation of a correlation impossi- ble. 6. The subdivision of overall categories of measurement into more detailed ones (e.g., three detailed absence categories are components of All Time Absent) may intro- duce too many distinctions for the amount of behavior in the sample and so cause the subsample of behavior assigned to one or more of the detailed time uses to become too small for effective study. 7. The length of the time-series may be too short to effec- tively establish reliability for the series. 8. The length of the observation period for collecting each day's values on each measure may be too short to smooth out the sporadic daily fluctuations. All eight of the conditions discussed above were poten- tial threats to the reliability of the time-use measures as the detailed discussion in Appendix C has shown. In examin- ing the reliabilities reported in Table 11, some of these conditions became evident from cross-referencing related statistics that correspond to each anomaly. The reasons behind each anomaly were quite clear and these have been noted in Table 11. In general they may be reduced to one or both of two problems: (1) small sample size, or (2) very little variation in the underlying behaviors being measured. 140 For the measures with a large enough sample of variation in the underlying behavior, the reliabilities were quite high. The anomalous reliabilities all occurred in the detailed categories of behavior where sample sizes of behavior were bound to be much smaller, and generally occurred at times when the behavior was unusually scarce. Both Hypothesis 10 and Hypothesis 10(a) were generally supported for the broad. overall measures. The detailed small sample measures do not all lend themselves well to assessment of the reliability hypotheses because of sample size. But, where the reliability estimates for both fre- quency and severity were useable, they were generally quite strong and neither was clearly superior. D-1 & D-2: Discussion of Time-Series Correlations Testing the Substantive Hypothesis, Using Severity Indices Time-series correlations were calculated within selected time intervals for both the lOO-day and 365-day datasets to examine various substantive hypotheses as out- lined in Chapter IV and in Appendix C. Results for Hypotheses 11, 12, 13, and 14 were examined in that order for each of the five selected combinations of: lOO-day or 365-day measure, and time interval of analysis. The results reported to show association between time inter- vals, correlations and associated significance levels, data- sets, and related hypotheses are summarized in Table 12. 141 Table 12: Time-Series Correlations Testing Substantive Hypotheses for Severity Indices Only Related Variable 1 Variable 2 Interval r r:Prob Dataset Hypothesis 281- 546 * 0.45 0.0001 loo—day Short-Term Long-Term 982-1277 0.36 0.0001 loo-day Hypothesis Legitimate Legitimate 1278-1680 0.56 0.0001 lOO-day 11 Absence Absence 982-1680 0.52 0.0001 lOO-day Progression 1247—1680 * 0.53 0.0001 365-day 281- 546 *-0.17 0.0058 lOO-day Short-Term Short-Term 982-1277 -0.89 0.0001 100-day Hypothesis Legitimate Non-Leg. 1278—1680 -0.93 0.0001 loo-day 12 Absence Absence 982-1680 -0.92 0.0001 loo-day Legitimacy 1247-1680 *-0.67 0.0001 365-day 281- 546 * 0.69 0.0058 100-day Short-Term Long-Term 982-1277 -0.44 0.0001 loo-day Hypothesis Non-Leg. Legitimate 1278-1680 -0.55 0.0001 loo-day l3 Absence Absence 982-1680 -0.53 0.0001 loo-day Progression 1247-1680 *-0.56 0.0001 365-day Legitimacy 281- 546 0.97 0.0058 lOO-day Short—Term Time 982-1277 -0.02 0.0001 lOO-day Hypothesis Non-Leg. Worked 1278-1680 0.25 0.0001 lOO-day 14 Absence 982-1680 0.06 0.0001 lOO-day Program 1247-1680 *-0.10 0.0001 365-day Impact 281- 546 —0.98 0.0058 loo-day Missing Time 982-1277 -0.91 0.0001 lOO-day As Values Worked 1278-1680 -0.75 0.0001 lOO-day Above 982-1680 -0.96 0.0001 lOO-day 1247-1680 -0.97 0.0001 365-day 281- 546 *-0.71 0.0058 loo-day Missing Long-Term 982-1277 -0.22 0.0001 lOO-day As Values Legitimate 1278-1680 -0.45 0.0001 loo-day Above Absence 982-1680 -0.18 0.0001 loo-day 1247-1680 -0.40 0.0001 365-day 281- 546 * 0.08 0.0058 lOO-day Missing Short-Term 982-1277 -0.38 0.0001 lOO-day Hypothesis Values Non-Leg. 1278-1680 -0.63 0.0001 lOO-day 14 Absence 982-1680 -0.32 0.0001 lOO-day Program 1247-1680 *-0.47 0.0001 365-day Impact Note: * indicates a finding in which one or both of the measures correlated showed low reliability for the split-half samples. Since the overall sample size is twice that in the split-halves, such findings may or or may not be based upon a reliable series of values (see Appendix II 142 There is evidence CM? a very strong legitimacy effect after the program installation occurred, but not before. In fact, before installation, short-term non-legitimate absence correlated +0.97 with time worked, while after installation it correlated -0.92 with short-term legitimate absence, and was uncorrelated with time worked. Hypothesis 12 is defi- nitely supported (see Table 12). Ancedotal evidence from the DOT Personnel Director indicated that before installa- tion a great part of absence was informally excused by the former Personnel Director as a distribution of these per- sonal favors. The more an employee was there at work, the greater the chance of personal favors, which in any case were unpunished absences at that time. After the program, the legitimate portion of this was given formal recognition (and so was the non-legitimate absence), so that after the program became routine the record includes fewer missing values. The group-level progression hypothesis (Hypothesis 11) is clearly not supported, however. Short-term and long-term forms of legitimate absence are positively correlated, both on the 100-day and 365-day measures, although the reliabil- ity of the 365-day measure is questionable due to sample size on the short-term legitimate absence measures. A combined progression and legitimacy effect (Hypothe- sis 13) is supported, but very probably the apparent effect 143 is largely induced by legitimacy alone. As noted above, there is no support for any independent progression effect at the group level, and in fact, the evidence even indicates the reverse -- that an increase in more severe forms co-oc- curs with more of the less severe forms of absence. Hypothesis 14 is not supported, presumably because the overall impact of the program was allowed to dissipate in long-term legitimate absence which. had. been. deliberately placed outside the purview of the program. 144 A Comment on Frequency-Severity Relationships For the sake of curiosity, the relationship between frequency and severity of the three major time-use catego- ries was examined briefly. Table 13 below summarizes the results. Time scheduled off shows a strong negative corre- lation between frequency and severity. But, for the other two time uses, the relationship is strong and positive. As the extremely low variability of the indices for Time Sched- uled Off indicated, the DOT regulated Scheduled Time Off very thoroughly. The negative frequency-severity correla- tion presumable just reflects careful scheduling of a fixed total amount of scheduled time off per' employee. This inverse relationship might also obtain for shorter-span-of— observation measures (i.e., a one week or one day base for the moving values). But, to reliably test that would require a larger sample than. the jpresent one since the smoothing effect of the longer moving averages would be lost. Table 13: Time-Series Correlations To Inspect Relations Between Frequency Indices and Severity Indices Variable 1 Variable 2 Interval r rzProb Dataset Time-Worked Time-Worked 982-1680 0.70 Severity Frequency 1246-1680 0.60 .0001 100-day .0001 365-day Time Sch Off Time Sch off 982-1680 -0.78 Severity Frequency 1246-1680 -0.74 .0001 lOO-day .0001 365-day Time Absent Time Absent 982-1680 0.51 Severity Frequency 1246-1680 0.73 .0001 lOO-day .0001 365-day OO 00 OO 145 Chapter VI Conclusions and Implications General Conclusions: The findings presented in Chapter V lead to seven gen- eral conclusions. 1. The dissertation set fundamental methodological goals of reconceptualizing the definition of absenteeism to overcome problems of theoretical inconsistency and low reli- ability encountered in prior research and of establishing an appropriate procedure to test reliability for those meas- ures. Absence has been redefined as one among many forms of time-use, and time-series reliability tests have demon- strated a high and generally consistent reliability for these measures where sample size and variability were suffi- cient for an informative test. On the whole, the time-series reliability correlations show a smebility in these measures that is about twice as high as findings in prior research, especially for severity (see Breaugh, 1981; Hammer and Landau, 1981). Part of this is probably due to the care taken in design, installation and administration of the absence data collection itself after the program started. Part is due to the use of a time-series approach instead of the cross-sectional proce- dures used in prior research. In particular, the time-se- ries correlations, although very simple, are a necessary 146 change in procedure for such work. They provide a reliable base for hypothesis testing, while the cross-sectional T-tests did not, despite their utility for comparing the means and providing a handy test of program impact. 2. The findings also shed light on the question of whether measures of absence frequency or severity (duration) are more reliable for absence research. The controversial use of frequency measures in replacement of severity meas- ures is run: necessary. Both indices are hdghly reliable, but while absence frequency can tell us what starts to hap- pen and how often, absence (or time-use) severity can tell us with far greater precision exactly what does happen, at least with respect to overall impact on work and the flow of time expenditure from one time-use to another. 3. Hill and Trist and others have hypothesized that absence as a form of withdrawal may, over time, occur with increasing severity. Although the analysis implemented after derivation of the time-use measures is not suitable for examination of this effect at an individual level, there was no evidence of such a progression effect at a group level. 4. There was some minimal evidence of seasonal or hol- iday shifts in absence occurring together with strong renewal effects, and in the baseline year, evidence of other shifts occurring in seasonal patterns Ibut not affecting 147 absence. In general, the T-test data support seasonality, but not for absence itself, only for scheduled work and time off. 5. The legitimacy hypothesis, that employees will sub- stitute legitimate absence for non-legitimate absence, is very strongly supported, both as an intra-annual and a long- term trend. Employees emit absence behavior in patterns that are allowed by the rules and the absence control pol- icy. This appears to have dominated the impact of the pro- gram far beyond the level of a mere side effect. It was the main effect anml in this respect, somewhat surprising. An hypothesis has been introduced in the present work that annually renewing the bank of legitimate days of discretion- ary absence an employee may "take" would result in an annual renewal of substitution effects. This was clearly supported by the T-test results. There is some evidence for seasonal fluctuation, a very clear renewal effect, and a minor holi- day effect associated with the renewal effect. The renewal effects strongly illustrate the impact of collectively bar- gained time-use allowance upon employee behavior, a sort of special condition legitimacy effect. 6. The dissertation was also intended to identify the overall impact of the absence control program, if any occurred. There was a significant decrease in short-term non-legitimate absence, and some increase in the overall 148 amount of time worked (about 7%, a return on the investment that. was jpractically' but not statistically significant), plus an overall increase in long-term legitimate absence (significant). Prior to the program the archival data showed long-term absence to be minimal so the program delib- erately included a decision not to focus on long-term absence until later. As a result, due to the legitimacy effect, mudh of the program impact was apparently absorbed in the long-term absence shift. The 7% gain in time-worked was a very clear practical success in that the success cri- terion set by DOT before the program started was a 1% gain. Further comment on the program impact follows below. 7. On a methodological level, the evidence brought to bear on procedural hypotheses clearly supported the use of group level time-series analysis, and in particular sup- ported the time series correlation technique employed here. The evidence also supported the contention that daily cross- sections of individual absence data are distributed highly non-normally. Despite this finding, T-tests comparing daily cross-sections of individual data over time still showed results that are consistent with the results of group-level time series analyses. The comparismn of daily cross sec- tions has obvious practical merit since it allows estimation of costs and benefits. But for inferential tests to analyze the significance of trends in time-use behavior, conven- tional cross-sectional procedures are at best tenuous and 149 where possible these should probably be set aside in favor of the time-series procedure. A Detailed Evaluation of Long-Term Program Impact On the whole, the absence control program had several long-term impacts on employee time-use. l. The program was, by contract, supposed to decrease short-term non-legitimate absence significantly. A signifi- cant decrease of 112 hours per employee per year did occur. Allowing for turnover and assuming an average size for the DOT workforce of 2000 employees, this would correspond to 224,000 hours per year. At an average of $20 per hour for wages and benefits, this constitutes a savings to the public of about $4.5 million per year. 2. The program was not intended to change the amount of time scheduled off, but during the program a significant 200 hours per employee per year decrease in scheduled time off did take place. This would correspond to about 400,000 employee hours per year, or $8 million per year at $20 per hour, presumably an unexpected side effect of the more pre- cise knowledge the program information system provided for scheduling. This discovery would not have surfaced without the detailed statistics about time-use shifts that were pos- sible due to reconceptualizing both absence and attendance and other employee activities as alternative forms of time- use. 150 3. A massive increase of 223 hours per employee per year in long-term legitimate absence occurred, mostly during the first year following installation. 4. There was also: (a) a strong negative correlation between short-term non-legitimate absence and long-term legitimate absence; (b) a moderate positive correlation between short-term legitimate absence and long-term legiti- mate absence (which refutes the progression hypothesis at a group level); and (c) no correlation worthy of note between time worked and short-term non-legitimate absence after the program began. 5. In order for the program to pay for itself, it was expected to move 1% of employee time-use from other catego- ries into time worked. Number of hours worked should have increased; and they apparently did increase by 107 hours per employee per year, an amount which closely corresponds to the saving of hours generated by reducing short-term non-le- gitimate absence, and a 700% return on the investment within two years. 6. The number of hours hidden in missing values was reduced by 218 hours per employee per year. These went into both work and absence. Overall, these six events in concert indicate that the program did four types of things to time-use behavior. 151 a. It moved about 110 hours per employee per year from. short-term non-legitimate absence into legitimate time-use behaviors, primarily long- term legitimate absence. b. It moved a large amount of time formerly hid- den in the records as missing values into time worked, and some time into long-term legiti- mate absence. c. It increased the amount of time worked over- all. d. The program also induced a very strong effect attributable 11) annual renewal. Immediately after the renewal date itself on July lst, an annual cycle occurs in which time worked and short-term non-legitimate absence decreases while short-term legitimate absence increases. This impact of the program offers further sup- port for the legitimacy hypothesis. On the face of these results, the project worked well. But, between the concession bargaining giveaway of 80 hours per employee per year and the lack of any effective restraint. on long-term. legitimate absence, much. of' this impact was dissipated in substitution toward forms of absence that remained legitimate. DOT employees are appar- ently even more resistant to giving up their absence time than the program planners anticipated. 152 The positive results obtained are largely attributable to the thorough groundwork prepared by the consultants and the enthusiastic and widespread support from the DOT top management and general supervisors. That support began in the planning phase and carried throughout group workshops and supervisory training in how to operate the program, com- puterization, support of the attendance incentive and day- to—day program administration, all in full cooperation with the consultants. Implications of the Present Research for Future Work There are 21 number of implications which fo110w from limitations and insights discovered through the present work. The time interval between the baseline year and the start of the program presents an unfortunate unknown period, and although the 30-day sample was small, there were signif- icant changes iii absence evident in time 30-day cross-sec- tional comparisons, as well as other changes. For a fully detailed analysis, this period of time should also be sam- pled. The data required are apparently not as clean as data recorded later, but are all still available in the DOT pay- roll records and this task could very conveniently be com- pleted. Furthermore, data between August 8, 1982, and the present are now available and these will have captured the long-term impact of the" program in fuller detail. Those data should also be incorporated into the database before further analysis is carried out. 153 The absence literature has recognized. variations in absence across demographic and occupational subgroups. This work still needs to be done. Very possibly the correlations reported here would be even higher if other trends were not confounded by demographic variations. Also very possible is an accompanying increase in normality if non-absence prone- ness varies by demographics of employees and this would strengthen the T-tests. Given additional computer resources, the» rest. of the *workforce, excluded from. the present work, could be included so that the problems intro- duced by small sample sizes would be overcome. This would also permit a comparison of the sample results with results for the entire workforce, an intriguing methodological curi— osity. Furthermore, the .measurement technique used in this work should be introduced and tested in other settings, both for methodological and substantive insights. There may be variation due to cultural differences or industrial sectors and these have policy and measurement implications concern- ing the validity of both the absence control program itself and the derived measures. Shorter time spans of observation such as one day or one week measures or 30 day time-series{measures should be examined using a time—series correlation procedure. There is need to check for an inverse relationship between fre— 154 quency and severity of absence time-uses that could become evident across the narrow time interval. Using the same methods applied here, a more detailed analysis of substitution across additional pairs of time-use variables could be done, supplemented by cross-log panel correlations to assess causality. Only major trends were examined here. Causal relationships and induced changes attributable to the program could be inferred to a greater extent if a formal time-series analysis were conducted to establish functional forms of change and interdependence. These pro- cedures use exploratory research tools to isolate underlying patterns that are of theoretical interest and are consistent in mathematical form. From that point, it is a short step to synthesize the separated patterns described by each iso- lated component of absence into a general model. A more detailed examination of the distributional forms of time-use behaviors should be completed. The present work substantiates prior findings in that absence behavior is distinctly non-normal in its distribution. The implications of this for applying transformation procedures to make absence data more tractable for use in inferential statis- tics have yet to be clarified. It is not clear, for exam- ple, just what the mean of a transformed distribution of absence data actually means. Does it have to be transformed 155 back to the original units before its practical significance may be interpreted? An empirical check on the result of doing that, or of comparing the results with those for the untransformed data from the entire workforcce would be instructive. The 'use of techniques ;presented. here combined. with time-series researdh on attitude changes should definitely be carried out in respect to absenteeism. The hypothetical relationship between attitude change and behavioral change has yet to be established, both in general and specifically in respect to absence behavior. The time-series correlation procedure has promise, par- ticularly for measurement and analysis of events in the per- sonnel field. That possibility should be explored. In par- ticular, the associations between time-use, monetary use and other resource use, and performance on the job has sweeping implications, both for modeling the performance system man- agement tool and for fundamental research on human work behavior. A more technical insight from the present research con- tributes in a direct and pragmatic way. The type of work presented here creates from merely large amounts of raw data enormous amounts of derived data and these data require com- puterization to make them manageable. Anyone planning to use the technique should be prepared to resolve the associ- 156 ated problems of database construction and information man- agement. Despite these problems, using a detailed time-se- ries procedure for program evaluation of absence control is clearly so much superior to the cross-sectional approach in scope, precision and reliability that the effort is war- ranted, especially since many of the technical problems involved have already had to be resolved for completion of the present work. Implications for Personnel Practice 1. The phenomena of employee attendance, absence or other time uses constitute a complex and interdependent set of events which covary over time in measurable and predicta- ble ways. This interdependence has been reduced to a trac- table form. Four requirements were met to bring this about. (A) Program evaluation to measure change in those events was based upon a set of time-use definitions that are custom tailored and fundamentally anchored in the absence control policy. (B) The administrative precision and consistency of data collection was sufficient to provide a thorough record of the employee time uses. This was one of the functions served by pre-program supervisory training. (C) The entire information set and procedures to create summary statistics from it were computerized so that the plethora of interde- pendent facts became a coherent, interpretable system. (D) 157 There was extensive administrative and supervisory support for the program throughout the work system. Even the unions cooperated. It is probable that in the absence of any one of these four elements, both the absence control program itself and the information system which supported it would have failed. 2. The dominant influence on employee absence and attendance is the implementation of control practices. Absence or attendance is apparently a highly calculated decision. If allowed to be absent with no associated cost then employees will be absent. The legitimacy effect is very clear and very strong. For absence control these two issues mean that every loophole in the absence control policy and its implementa- tion in practice must be plugged. These behaviors covary as a system. To suppress one aspect of absence will induce an increase in some other aspect unless the system specifically prevents the substitution. APPENDIX A THE DETAILED DESCRIPTION OF THE WORKFORCE DEMOGRAPHICS Appendix A The Detailed Description of the Workforce Demographics Unpacking: The initial unprocessed raw data tapes received from the research site contained, by construction, some coding idiosyncracies. To save space in computer mem- ory, the data for each employee had been packed into 6-month records within which the position of any character indicated what it represented, or what date it referred to. This cre- ated, in effect, ab’out four hundred different variables, before any new derived variables could even be created. Yet most of those variables were really the same variable taking on a value that had occurred on a different day. Before any further analysis could be done, all of the raw absence data were split off from the demographics and both were rewritten onto new tapes. In the process, a unique (and confidential) case number‘ was introduced. to replace each employee social security number. This case number provides a convenient link between the demographics for each employee and the absence data for that employee. The raw absence data were changed in the procedure from 6-month records to daily strings containing the employee case number, the relative date of the time-use (absence) behavior (starting with day 1 on January 1, 1978), the time- use (absence) code, and the duration of the time-use in 159 160 hours. The demographic data were (selectively) written out to a separate file each time they appeared on the raw data tapes (about 5 to 8 times per employee). These were later edited to create one clean, nonredundant file of all 2336 employee demographic records. The statistics below summarize the demographic charac- teristics of the workforce. They were generated by submit- ting the file of demographics to analysis using the Biomedi- cal Dataprocessing Package (1982), subroutine BMDP2D. The first such run generated an overall description of the entire workforce and those statistics are presented below in detail. That set of statistics provided numerical values for the composition of the entire workforce by demographic subgroup. A fortran subroutine was then written to randomly select a sample group representing all employees in the DOT workforce. Demographic statistics for the sample were then generated using BMDP2D. On the following page, in Figure 8, is a list of the demographic variables analyzed. 161 Variable Case Number Bargaining Unit Code Year First Month Year Hired Month Hired Birth Year Birth Month Birth Day Job Class Cost Center Marital Status Description 9f the Data Item A 4-digit integer case label used to link the absence data tapes and the employee demographics. A 4-digit alphanumeric code which may allow later data analysis comparing results for different collective bar- gaining constituencies. 1 2-digit integer for the last two digits in the year the data represent. A 2-digit integer for the first month the data represent ... 1 through 12. A 2-digit integer value. 2-digit integer value. 2—digit integer value. 2-digit integer value. > 3’ ID > 2-digit integer value. A 6-digit integer code representing the payroll job category of each employee. A 4-digit integer code identifying the organizational profit-center associated with each employee. May be used in later work for comparison across occupational subgroups. M for married, S for single. A 1- digit alphabetic code. Dependents A 2-digit integer variable. Race A l-digit integer code. Sex M or F, a l-digit alphabetic code. L Figure 8: The Demographic Data Variables Continued on next page L_____________.______—_—___———_____—______._._._.___._______.______._*J 162 Figure 8 Continued Location Group A 2-digit integer code which identi- fies the work site at which each employee is stationed. Could be used in validation across locations if the coding documentation for it was available, but at present, it was not. Job Class A 2-digit brief occupational Group Category used for Affirmative Action. The coding documentation for these values is also unavailable. Quit Month A 2-digit integer code for the month Quit Day and the day of termination. In most cases this field is blank. r———————*—’—‘”—'—"——“—————‘———-1 L_._.____.__..._ ..- __..____..__________._._..____J Detailed Statistics for the DOT Workforce Demographics Case Number: These ‘values were generated to jprovide a unique link between the demographic data and the time-use data for each respective employee, a link which intrinsi- cally protects the confidentiality of data for each employee. One-by-one they range from 1001 up to 2336. Bargaining Unit: The bargaining unit code for each employee represents the subdivision of each given union which holds representation rights for the employee. There are 1342 dif- ferent subdivisions of this kind represented in the DOT demographics, plus 60 employees for whom no code was recorded (massing values). These results have been clus- tered. to provide the following summary statistics. The 163 largest employee group, Transportation Equipment Operators (bus drivers) were represented by the Analgamated Transit Union Local 26. A substantial segment of the DOT workforce was represented by various AFSCME Locals. Two of these, AFSCME Local 312, representing 533 mechanics and general transportation employees over the research period, and AFSCME Local 214, representing 251 Clerical Transportation employees, together encompassed 34.4% of the DOT employees. A further 21.8% of the workforce (479 employees) was repre- sented by locals of the Detroit Building Trades Council. 11.3% of the employees were members of a variety of unions, including Teamsters Local 214, all non-supervisory. 3.0% of the workforce (68 employees) were members of supervisors' and technical/supervisory employees' unions. The remaining 72 employees are non-unionized. Year and Month 9f Data Collection: The variables were a vital component of the raw absence data, but in the demo- graphics they are merely coincidental and therefore uninfor- mative. Year and Month and Day Hired (Seniority Data): The median and modal values for each of these is reported in the table below. For the month and day the mean is also included. 164 Table 14: Year and Month and Day Hired r 1 I l I Variable | Year N | Month N I Day N I I --------- I ------------- I ------------- I ------------ I I Mean I not reported I June 139 | 15th 74 I I Median I 1974 196 I July 210 I 15th 74 I I Mode | 1979 319 I August 311 I 6th 114 I I N I 2240 I 2241 I 2241 I I Range | 1917-82 I 1-12 I 1-31 I I Missing I 96 I 95 I 95 I I I From these statistics, several patterns in hiring at DOT are indicated. 1979 was the biggest hiring year for DOT (and 1974 was second). This is consistent with a sizable pur- chase of new busses at about that time. The strongest hir- ing month appears to be August; September was second, with 214 hires. Day of hire during the month is fairly variable, but quite evenly distributed in general with a slight modal peak on the 6th and a fluctuation on approximately weekly intervals that suggests a weekly pattern, perhaps attributa- ble to hiring practices. The median seniority was about 8 years (starting in June 1974) at the end of the program year in June, 1982. Year, Month and Day 9_f Birth (Age): Birth year for DOT employees ranged from 1914 to 1962, so that at the end of the study in 1982, the oldest employees were about 68 years of age and the youngest employees about 20 years of age. The median age was about 38 years and the mode (n of 107) was 32 years. These figures again reflect a strong upsurge 165 in recruitment levels in recent years, in which the new recruits tended to be somewhat younger. Day and Month of birth are both quite evenly distributed across the year for the DOT workforce. Summary statistics are presented in the table below. These are included here for comparison with the sample characteristics. Table 15: Birth Dates I I Variable I Birth Year I Birth Month I Birth Day I I """""" I """"""""" I """""""" I """"""" I I Mean I not reported I June I 15th I I Median I 1945 I July | 16th I I Mode I 1950 I August I 23rd I I Range I 1914-1962 I Jan. - Dec. | 1 - 31 I I N I 2335 I 2336 I 2335 I I Missing I 1 I O I 1 I I I L J Job Class (Occupational Payroll Category): Two types of occupational data were recorded for the two types of DOT workforce. The first type described here was recorded as a six-digit, detailed payroll code assigned to each specific job. The second type is described later. There were 115 different detailed occupational categories for the DOT employees. These range from the Director and Deputy Direc- tor of the department to clerical employees and office work- ers, transportation terminal workers, transportation equip- ment operators (TEO's 1335 of these), mechanics, and other craft workers. For convenience of interpretation, 166 these detailed occupational categories have been summarized into the seven clusters below. Table 16: Job Class I I I No. of No. of % of I I Job Class Cluster Employees Jobs Total I I I I I I Managerial - Executive, I I Supervisory 105 23 4.5 I I Administrative & Techni- I I cal/Professional 28 18 1.2 I I Craft Workers 401 ‘ 3 17.2 I | Transportation Equipment I l Operators 1335 1 57.1 I I Guards 42 3 1.8 I I Clerical Employees 182 23 7.8 I I Terminal Service Workers 234 16 10.0 I I Miscellaneous Others 9 8 0.4 I I -— — I I TOTALS 2336 107 100.0 I I I Cost Center (Geographic/Functional Distributbma of Employ- ees): 40 distinct cost centers were used to identify the specific work location and type of work of DOT employees. These are listed with summary statistics showing: the num- ber of employees in each geographic location from all job categories and the number of salaried and non-salaried employees from all locations and job categories. The codes used to represent the "miscellaneous" category seem to sug- gest that the jobs involved are all salaried, but since the coding manuals available from DOT were not specific in this 167 regard, no use can be made of those particular coding cat- egories yet. The other 4-digit cost-center codes, nonethe- less, offer an excellent sampling-frame for later subsam- pling from the database. # Sal- aried Cost # of Employ- Cost Center Center N Employees ees per Code by Location Cost Center 10 Gilbert Garage Salaried 1166 10 - Gilbert Terminal Workers 1230 41 I Gilbert Gilbert TEO's 5420 420 I 539 Gilbert Garage Workers 6390 89 - 9 Shoemaker Garage Salaried 1167 9 - Shoemaker Terminal Workers 1200 28 I Shoemaker Shoemaker TEO's 5460 461 I 572 Shoemaker Garage Workers 6380 100 - 13 Coolidge Garage Salaried 1168 13 - Coolidge Terminal Workers 1220 36 I Coolidge Coolidge TEO's 5480 452 I 593 Coolidge Garage Workers 6370 119 - 16 Administration Salaried 1100 16 - 55 Personnel Salaried 1110 55 I 15 Auditing Salaried 1120 15 I 3 Stationery Stores Salaried 1121 3 I 30 Cashiers Salaried 1140 30 I 41 Technical Services Salaried 1164 41 I 23 Rolling Stock Office Salar. 1165 23 I 8 Shops Salaried 1161 8 | - 23 Plant Maintenance Salaried 1160 23 I I Plant Maintenance Trades 6361 38 I I Plant Mainten. Electricians 6362 7 I I Plant Maintenance Drivers 6363 22 I I Main Heating Plant Maintenance 6360 21 I | Office Technical Services, Maint. 6345 5 | I 477 Unit Repair, Maintenance 6300 35 I | Heavy Repair, Maintenance 6330 56 I I Body Shop, Maintenance 6340 20 - I Security 6350 39 ----- Loc. TEO's on L.T. Disability 5499 2 I I Indet- Miscellaneous (1118, 6367, 6377, | I ermin. 6387, 6397, 7370) 41 ----- 80 6 Charter Services Salaried 1170 6 - I 20 Transp. Planning & Schedul. 1180 20 I - 32 Transport. Office Salaried 1190 32 - TOTALS: 304 A11 Salaried Employees 1991 Non-Salaried Employees 41 Miscellaneous Employees 2336 All Employees in the DOT Figure 9: Distribution of A11 Employees in the DOT Workforce by Geographic & Functional Cost Centers 169 Marital Status: Marital status of DOT employees was only recorded as either married or single for payroll purposes. Out of 2336 cases read, there were 1940 single employees, 392 married employees, and 4 cases with missing values. In short, 83% of the workforce is single. Dependents: The number of dependents for members of the DOT workforce ranged from 0 to 10 and 12, with a mean of 0.75 and both median and mode of 0. 1592 employees had no depen- dents (so presumably about 352 employees had dependents but were not married). For those 744 employees who did have dependents, the mean number of dependents was 2.34, the mode was 1 dependent, and the median was 2. There were 5 missing values. Race: Racial characteristics of the DOT workforce are defined as follows. "Black" means persons have origins in any of the black racial groups of Africa and does not include those of Hispanic origin. "White" means all persons have origins in any of the original peoples of Europe, North Africa, or the Middle East and does not include those of Hispanic origin. "Asian or Pacific Islanders" means all persons having origins in any of the peoples of the Far East, Southeast Asia, the Indian Subcontinent, or the Pacific Islands. This category includes, for example, Chi- nese, Japanese, Korean, the Phillipine Islands and Samoa. "American Indian or Alaskan native" includes all persons 170 having origins in any of the original peoples of North Amer- ica and maintaining a cultural identification through tribal affiliation or community recognition. The "Hispanic" cat- egory means all persons of Mexican, Puerto Rican, Cuban, Central or South American, or other Spanish culture or ori- gin, regardless of race. The category "other" was included to catch any misfits or unknowns. The distribution of DOT employees by category was as follows. Table 17: Race r I 1 I I I I Category # % of DOT I Category # % of DOT I I I ' I I Black 1856 79.5% I Hispanic 3 0.1% I I White 442 18.9% I Am. Indian 1 0.0% I I Asiatic 7 0.3% I Other 3 0.1% I I I I L l There were 23 missing values. Sex: Out of 2336 employees on the DOT Demographic File, 1981, or 84.8% were male. 354 of the employees were female, or 15.2%. There was 1 missing value. Location Group and Job Class Group: Two of the most poten- tially useful categories of demographic data were flawlessly recorded throughout the entire measurement period. Unfortu- nately, the coding ‘manuals are not available for these groups at this time. But, on the other hand, the constructs of theoretical interest described by those data are also described by combinations of the cost center codes above, 171 and those 4-digit codes can be used to provide the necessary sampling frame. Each value is a 2-digit code ranging from 0 to 14 and 99. Year and Month of Ouit: 400 employees quit between July 1, 1978 and June 30, 1982; but this figure should be taken more as an indicator for June 1, 1980 through August 10, 1982 since no values appear in the demographic record from 1978 or 1979. This amounts to about 17.1% of all employees on the payroll over the period of study, and calculated across 26 months, it approximates an annual rate of 7.9%. 1981 was the largest year for quits (180), but the number of termina- tions in 1982 would probably have exceeded those for 1981 (182) by December since the 1982 quits recorded only repre- sent 7 months of 12 in 1982. On that basis, the rate of terminations was rising. A look at the monthly quits distribution shows a defi- nite pattern. The monthly distribution of the 400 quits on record by percentage is noted on the following page. Winter and Spring (January to July) are about 3.4 times as likely a quit period as Summer and Fall (August to December). On the basis of progressive levels of alienation as a theory explaining absence and related turnover, January would therefore be a likely month in which to watch for increases in long—term absence prior to termination. 172 Table 18: Year and Month of Quit I %0f I Month Count Total Histogram of % by Month I I ---------------------------------------------- I I Jan 157 39.3% IXXXXXXXXXXXXXXXXXXXX - I | Feb 25 6.3 IXXX I Winter I I Mar 54 13.5 IXXXXXX 312 I & I I Apr 33 8.3 IXXX or I Spring I I May 27 6.8 Ixxx 78% I I I June 16 4.0 IXX - I I July 17 4.3 IXX - I I Aug 11 2.8 IX I I I Sep 33 8.3 IXXXX 88 I Summer I I Oct 11 2.8 IX or | & I I Nov 4 1.0 IX 2 % I Fall I I Dec 12 3.0 IXX - I I I Brief Demographics for the Subsample Group Since the computer time required to create the derived data for one case alone is about 30 seconds, it was quite impossible to carry out an analysis of all 2336 cases in the DOT database. Instead, the statistics that were generated by analyzing demographic data for the entire DOT workforce were used to calculate a group-specific "sampling fraction." This in turn was incorporated into a fortran program, which generated a random selection of cases representing the entire employee population. The target sample size for this group was 160 cases. Due to the way in which the random selection program operated, the actual sample generated includes 164 cases. This sample was quite adequate in size to accommodate both attrition and replacement over time due 173 to turnover. It was also randomly selected and just small enough to accommodate the computer' processing time con- straints. Statistics for the most salient demographic char- acteristics of each subpopulation sample were generated using BMDP2D again, and are summarized below. Note that values for events in 1982 represent only the first seven months, while employees terminating before the study began are pp: represented at all. Demographic figures which were merely incidental have been omitted from the subcroup sta- tistics. Demographics for the Sample from All DOT Employees There were 2336 employees in the entire DOT database. The computer resources available could. handle about 160 cases at a time, enough to cover sample size decreases due to turnover, etc. Selecting about 160 cases required a sam- pling fraction of 160/2336 or 0.0684932. This limit was used in the random case selection program and generated a set of 164 cases randomly selected from the data for all DOT employees. The deviation from 160 cases to 164 is presumably a result of rounding error. The median year of hire was 1974 and by 1974 56.1% of the employees in this sample were already on the payroll with 9% hired in 1974. By 1979 (the modal year) 95.5% of the sample had been hired and 16.1% were hired in 1979 alone. The seniority year of these employees ranged from 174 1942 to 1982. The modal month of hire (13. % of employees) was August, with a mean in June and median in July. Hiring also seemed to have peaked from these employees in the first and last few days of each month. The median year of birth for employees in this sample was 1944, with 51.8% of the employees born in or before 1944. The mode was 1950 and the range was from 1916 to 1960. In general, the age distribution of this sample is moderately skewed ixi favor of younger employees. Day and month of birth are evenly distributed. According to the Cost-Center and Job-Class codes, 55.5% of the sample for all DOT employees were Transportation Equipment Operators (TEO's), i.e., 91 bus drivers. 83.5%; or 137 employees in. the sample were single. 15.9% or 26 employees were married. 73% or 119 of the employees had no dependents. For those 27% who had depen- dents, the modal number of dependents was one, the median was two, and the range was from one to ten. Most employees in the overall sample were black (78% or 126 employees). The remainder (22% or 36 employees) were white. 87.2% or 143 employees were male. The remaining 12.8% or 21 employees were female. During the three years spanned by the study, 28 out of the 164 employees in the sample quit, or about 17%, an 175 annual turnover of about 2.5% in 1980 and about 7.3% in both 1981 and 1982. Most of these terminations occurred in Janu- ary (42.9%, July (14.3%), and September (14.3%), and there was a definite tendency to quit in the first half of the year, not in the late summer or fall. APPENDIX B DERIVATION OF THE RAW TIME-USE DATA AND CREATION OF THE DERIVED DATA STRUCTURE 177 Appendix B Derivation of the Raw Time-Use Data and Creation of the Derived Data Structure The raw absence data tape as described in Figure 10 below contains daily values for the case number, relative date, absence type, anui "absence" severity 1J1 hours. The first two values are assigned to the second two and identify to which employee and on what day those last two data values relate. The meanings of these absence data ‘values are dependent on several things. 1. Of primary importance is the reason for absence (or worked time) which was originally coded for the specific employee on a given day. 2. Second, the duration of the occurrence of the time- use over time may persist until is forces a recoding of the entire occurrence or at least of part of it. Indeed, the duration of an absence episode may initially be understood to be short and yet later may turn out to last for weeks, or even months. Reassignment of such time-use occurrences to corrected categories from their onset is time first of two main functions served by the Fortran program which generates the derived database. 3. Third in importance but critical to the integrity of the derived data are a number of systematic coding errors 177 178 which must be removed from the data before they can interpreted and used. be Variable Case Number Relative Date Absence Type Absence Severity __————.—_._——_—_————_—_———e___.__.r __._.—_————._—_—__——— Description of the Data Item A unique 4-digit integer case- label used also on the demo- graphics tape. A 4-digit integer for the rela- tive date on which respective absences occur. January 1, 1978 was Day 1. January 2, 1978 was Day 2, etc. The data extend from Day 182 to Day 1680. A 2-digit integer code as list- ed in Figure 1. Each code re- presents how the time was spent on a given day by the employee. A 1-digit integer for the num- ber of hours expended in the type of absence behavior coded under absence type above. -4-.._..______———_—_ l________._._____..____.______._____.___ Figure 10: The Raw Absence Data Tape Variables The logical structure of relations between absence rea- sons has been used to create a pmogram that converts each primitive "absence reason" and duration coded in the raw data into derived measures for one or more of the 13 time- use indices. Figure 11 is a flowchart of the recategoriza- 179 tion procedures. The recategorization rules used have been explained following the flowchart. Read the flowchart from right to left, following the arrows. Note that where the condition associated with a given arrow in the flowchart was satisfied, the data values in the raw data categories (on the right) will have been processed as contributing elements and added into the associated higher-level "resultant" cat- egory of time use (on the left). "Resultant" categories are each marked with an asterisk in the respective box of the flowchart. 180 Not Rehired for 365 Days or More Termination Time T eLaid or: 19 Termination eRcsigned 26 Without Not Resigned, Nor Retired. sDischarged 27 I. Rehire Nor Discharged. eRetired 28 M Rehired in 5 Days or Less. sDeceased 29 s Transferred 3O eScheduled or: Time 3 All 00ft Day - Trade 4 00:: Day - Adjustments Scheduled I eSwing Holiday 9 0 Time 0“ eaoliday 10 Not F oStrike 42 Fired and Rehired Discharged a Rehired in 5 in 30 Days or Less Days to Punishment 365 DayeI Time E Suspended Workeer L Day Not Scheduled for This WorkerZZ . G 13:-13:19 y' or Kissing Value 1 eSchool/Training (Work Related) 13 All 0Court Time (Witness or Work RelatedEB Time Other I OJury Duty 16 AWorked Work eUnion Business 17 C OOonventions for Work 18 eCivil Service BusinessZO It Over 5 L 5 Days or Less and Less ; Hilitary LeavelS Than 366 Over 365 Days Days 3 eScheduled Workday l " Routine Catt-Day, Worked 2 Time at as s A with Injury 32 Work Time eLongterm Absent - Injured“ 3 Worked eLongterm - Disabled 24 OSick - 6 Days or More 25 eLeave of Absence 14 eSingle Miss 40 eDouble Miss 41 sTardy Employee 43 Late Time Lost to Time _. Lateness L Short Non- All Time Absent 0 Legit. K Absence eUnexcused Absence 45 esick Employee 33 ODeath - Not Family 35 *"" 0 Moving 37 eOther Unpaid/Non-Legitimate Short Absence 44 eAWOL 39 oSick Family Member 34 e Wedding 36 00!: with Permission 38 J Longterm Legitimate Absence Short I Legitimate Absence Retired & Rehired eCasual Leave Day 5 eDeath in Family 11 in 5 Days or “3‘ eCasual Leave Time 7 eOccupational InjuryIZ 3 0 Vacation 3 ‘ 4* I Figure 11: Logical Relations Among Different Uses of Employee Time I. .s .I. .00-.. o I... 'IV..t |s| 181 r I I Type 9: Time Worked pp the gee I I Computer Code # Code Definition I I 00 Created for generating the de- I rived database as a code for Day I Present and Worked. I 04 Off Day - Worked I I I Type 1: Scheduled/Contractually Provided Absenteeism I I Computer Code # Code Definition I I 01 Regularly Scheduled Day Off I 02 Off Day - Trade I 03 Off Day - Adjustment I 11 Casual Leave Day I 12 Casual Leave Time I 13 Vacation I 14 Swing Holiday I 15 Holiday I 23 Death in Family I 25 Occupational Injury I 30 School/Training (DOT Related) I 32 Leave of Absence I 34 Military Leave I 35 Jury Duty I 36 Union Business I 37 Conventions (DOT approved) I 38 Laid Off I 39 Civil Service Business I 42 Suspended I 62 Not Scheduled I 33 Court Time (Witness or DOT | related) I Figure 12: Summary of the Raw Absence Data as Recorded in the Department of Transport Absence Management Database or as Modified and then Used to Generate the Derived Data Continued on next page l_—_—_____.__——_——.—_——_——____._.___________________—_J 182 Figure 12 Continued Type 2: Long-Term/Non-Recurring Absenteeism Computer Code # Code Definition 19 Long-Term Disability 20 Sick Employee (6 days or more) (These codes may be used to handle catastrophic illnesses approved by the Personnel Division without the 6 day restriction.) 1*._______________._______.J r I I I I I I I I I I I I I 51 Resigned I 52 Discharged I 53 Retired I I 54 Deceased I I 55 Transferred I I 18 Long-Term with Injury I I 17 S & A with Injury I I I I Type 3: Excessive/Recurring Absenteeism I I I I Computer Code # Code Definition I I 21 Sick Employee I I 22 Sick Family Member I I 24 Death - Not Family I I 26 Wedding I I 27 Moving I I 31 Off with Permission I I 41 Absent without Permission I I 43 Single Miss (TEO's only) I I 44 Double Miss (TEO's only) I I 46 Strike I I 47 Tardy I I 61 Other I I 40 Unexcused Absence I I I Notes: 1. According to the Attendance Control Policy, employee absenteeism has been grouped into three categories, with time worked on the job there are four: Type 0 - Time Worked on the Job Type 1 - Scheduled/Contractually Provided Absenteeism Type 2 - Long-Term/Non-Recurring Absenteeism Type 3 - Excessive/Recurring Absenteeism 2. These types are reworked as shown in Figure __ to create the various levels and categories of derived data. 183 Derivation of the Time-Use Data Values The programs which generate the derived time-use data calculate them in two separate datasets from the raw absence types and severities and from the respective relative dates. This is done in seven stages for each dataset. 1. First, all the data for a single individual employee is read into an array and the size of that array is determined in the process. 2. Second, the program goes through the array of raw values and discovers and recodes each of the several types of systematically miscoded data values wherever they occur in the array. 3. Third, the data are scanned to detect any scheduled days off intervening' within longer span time-use occur- rences. The "second" portion of the respective occurrence long-span time-use is then flagged to allow calculations of frequency to be corrected (return from days off is not counted as a second occurrence). This "flag" is a value that tells the computer how many days back to go looking for the "first" portion whenever it encounters a flagged "sec- ond" portion. Instead of incrementing the counter by one for the second portion of the occurrence, the program just works around the days off for treatment of that specific time-use. 184 4. Long term absences for reasons of termination are checked. to see if the employee was afterwards rehired. Rehires within 365 days or less are allocated to the long- term absence time-use category. Military leave anxi long- term sickness are similarly treated. For fired and rehired employees, the first 30 days are recoded as punishment time, the remainder as long-term legitimate absence unless they are off for over one year. Employees coded as deceased and later as rehired are recoded for that interval as missing values. 5. Once all the raw absence codes have been checked and recoded or cleaned as necessary, the program sorts and aggregates occurrences from the original 45 different absence-types into 13 time-use categories. Seven of these categories are intermediate time-use categories, and when clustered together in groups of one, two, or three, these in turn contribute their effect to create the six final or overall, outside-level time-use categories. The pattern of this aggregation and reasons for each decision are both indicated in Figure 11, and the 13 derived time-use indices are listed in Figure 13. 6. Two different computer runs are then executed. For each time-use index in each day of the study period for each individual employee, two sets of three derived values (lOO-day frequency, 100-day severity, mean 100-day severity, 185 and annual frequency, annual severity, annual mean severity) are then generated as the derived values in two arrays of three types of measures across the 13 time-use indices for the respective employee. Estimates for the two sets of three types of measures, on all 13 indices, were generated on a daily basis for every day of the program. Numeric codes for the 13 time use indices are listed below in Figure 13, and these correspond to the numbers on boxes in Figure 11. Note: All codes are 2-digit integer values. I I I Code for I I Time Use Time Use Category/Index I I I I 10 All time worked on or off the job I I 11 Time worked on the job | I 12 Time worked off the job I I 20 All scheduled time off I I 22 Time off due to punishment I I 23 Routine normal time scheduled off I I or strike I I I 21 Missing value or unscheduled day I I 30 Time absent for all reasons I I 31 Short-term legitimate absence I I 32 Long-term legitimate absence I I 33 Short-term non-legitimate absence I I (except late) I I 34 Late time I I 40 Unended termination time (no rehire) I I I I I I L J Figure 13: Derived Data Types 7. For each of the two sets of three measures, all 13 time-use indices for every day are written onto tape in a. 186 comprehensive array which constitutes the track record of time-use for that employee throughout the employee's partic- ipation in the study. These two arrays are created daily, a piece at a time, as the new values are derived. When the calculations are completed, each entire array of daily val- ues for the employee is written out onto tape in daily records 219 characters long. The two derived arrays are described in Figure 14 and Figure 15 below. The selective use of these individual derived data values to generate sta- tistics for each subgroup in later methodological and subs- tantive analyses is described in Chapter IV. Selection of subsample groups of cases from the overall 2336 cases in the DOT database is done automatically during derivation of the time-use categories. An early step in the process is a for- tran subroutine that rejects all cases whose case numbers don't match those in the respective subgroup demographics. 187 Variable Case Number Relative Date Absence Type Absence Severity Day Type A Day Type B Day Type 1 Day Type 2 l"—"—‘_—————————__——_———‘—_—___'_"_—_"’_'fi—'——_’—'—‘_—“‘_'—_———fi Description A 4-digit code linking the indivi- duals' derived statistics for time usage with the demographics tape. Used for sampling purposes. A 4-digit number indicating number of days from the starting date of the program. Range is from 182 to about 1686. A 2-digit code for the raw absence type used to sort the data in creating the derived statistics. A 1-digit code for the number of hours consumed in the time usage indicated by the absence type above on a given day. A 2-digit code (see Figure __) for the highest level of derived category into which the sorted absence types above become categorized. A 2-digit code (see Figure __) for the alternate data type on a given day. For example: if Day Type A is "late" then Day Type B would be "worked time". A 2-digit code (see Figure __) for am intermediate data type also de- rived by categorically sorting the absence types above according to the logical structure noted in Figure __ A 2-digit code which alternates for Day Type 1 ... like Day type B. Figure 14: The Individual "Derived" lOO-Day Data Types Continued on next page |___________________—_—__..________—.__.__________--____________—_-J ____.__.____.___._____.__._._——_1 r__________________.____._________.._____.__________.—_____________ . .-- 188 Figure 14 Continued Variable Day Severity Day Severity Day Severity Day Severity lOO-Day 100-Day 100-Day 100-Day lOO-Day 100-Day 100-Day 100-Day 100-Day 100-Day lOO-Day 100-Day lOO-Day 100-Day 100-Day 100-Day lOO-Day 100-Day 100-Day 100-Day 100-Day lOO-Day lOO-Day 100-Day lOO-Day lOO-Day 100-Day 100-Day lOO-Day lOO-Day lOO-Day 100-Day 100-Day lOO-Day 100-Day 100-Day lOO-Day 100-Day lOO-Day Frequency Frequency Frequency Frequency Frequency Frequency Frequency Frequency Frequency Frequency Frequency Frequency Frequency Severity [vi—‘03:” 10 11 12 20 21 22 23 3O 31 32 33 34 4O 10 Severity Severity Severity Severity Severity Severity Severity Severity Severity Severity Severity Severity 11 12 20 21 22 23 3O 31 32 33 34 40 Mean Mean Mean Mean Mean Mean Mean Mean Mean Mean Mean Severity Severity Severity Severity Severity Severity Severity Severity Severity Severity Severity Severity 34 Severity 40 Description 1-digit codes for the number of hours spent in the respective time usage on a given day. 4-digit values for the frequency of occurrence derived for each respec- tive form (see Figure __ and Figure __) of time usage in the past 100 days, as calculated for the specific day indicated by relative date. 4-digit integer values for the cumu- lative severity of each respective form (see Figure __ and Figure __) of time usage in the past 100 days, as calculated for the specific day indicated by relative date. 10 6-digit (plus a 7th for 11 decimal point) real values 12 calculated for each respec- 20 tive type of time usage each 21 day by dividing 100-Day Sev- 22 erity by annual frequency. 23 There are 2 digits to the 30 right of the decimal place. 31 32 33 1 l I I I I I I I I I I I I I I I | I I | I I I I I I I I I I I I I I I I I I I I | I I I | I I I I I J 189 Variable Case Number Relative Date Absence Type Absence Severity Day Type A Day Type B Day Type 1 Day Type 2 r—_—_.——__———————_—.—_—_——.———__—_—_——__—__—__——————————.—_____———__——__fi Description A 4-digit code linking the indivi- duals' derived statistics for time usage with the demographics tape. Used for sampling purposes. A 4-digit number indicating number of days from the starting date of the program. Range is from 182 to about 1686. A 2-digit code for the raw absence type used to sort the data in creating the derived statistics. A 1-digit code for the number of hours consumed in the time usage indicated by the absence type above on a given day. A 2-digit code (see Figure __) for the highest level of derived category into which the sorted absence types above become categorized. A 2-digit code (see Figure __) for the alternate data type on a given day. For example: if Day Type A is "late" then Day Type B would be "worked time". A Z-digit code (see Figure __) for am intermediate data type also de- rived by categorically sorting the absence types above according to the logical structure noted in Figure __ A 2-digit code which alternates for Day Type 1 ... like Day type B. Figure 15: The Individual "Derived" Annual Data Types Continued on next page 190 Figure 15 Continued r——'_——————_—_——_——_——_—'_——'—_——_——-—_~_——'"_7""‘_-——’-___-——__———fi Variable Description Day Severity A Day Severity B Day Severity 1 Day Severity 2 1-digit codes for the number of hours spent in the respective time usage on a given day. 1 I I I I | I I I Annual Frequency 10 4-digit values for the frequency of I Annual Frequency 11 occurrence derived for each respec- I Annual Frequency 12 tive form (see Figure __ and Figure I Annual Frequency 20 __) of time usage in the past 365 I Annual Frequency 21 days, as calculated for the i Annual Frequency 22 specific day indicated by relative I Annual Frequency 23 date. I Annual Frequency 30 I Annual Frequency 31 I Annual Frequency 32 I Annual Frequency 33 I Annual Frequency 34 I Annual Frequency 40 I I Annual Severity 10 4-digit integer values for the cumu-I Annual Severity 11 lative severity of each respective I Annual Severity 12 form (see Figure __ and Figure __) I Annual Severity 20 of time usage in the past 365 days, I Annual Severity 21 as calculated for the specific day I Annual Severity 22 indicated by relative date. I Annual Severity 23 I Annual Severity 30 I Annual Severity 31 I Annual Severity 32 I Annual Severity 33 | Annual Severity 34 I Annual Severity 40 I | Ann. Mean Severity 10 6-digit (plus a 7th for decimal I Ann. Mean Severity 11 point) real values calculated I Ann. Mean Severity 12 for each respective type of time I Ann. Mean Severity 20 usage each day by dividing Annual I Ann. Mean Severity 21 Severity by annual frequency. I Ann. Mean Severity 22 There are 2 digits to the right I Ann. Mean Severity 23 of the decimal place. These I Ann. Mean Severity 30 digits are significant due to the I Ann. Mean Severity 31 way raw values for absence fre- I Ann. Mean Severity 32 quency and severity were record- I Ann. Mean Severity 33 ed. I Ann. Mean Severity 34 I Ann. Mean Severity 40 I I J APPENDIX C METHODOLOGICAL NOTES Appendix C Methodological Notes (A): Procedures for Assessing Normality and Distributional Form Various statistics offer insight into the relative nor- mality of data distributions. In the use of inferential statistics, a tacit assumption is made that there is a con- sistency between the shape of empirical data distribution and the shape of some theoretical distribution ... such as the normal curve. The merit of T-tests and like procedures depends upon the correctness of that assumption. In the present work, T-tests on both the lOO-day and 365-day data are used to examine substitution and long-term effects by comparing daily cross-sections. The assumption of an under- lying normal curve is accordingly worthy of verification, especially since the absenteeism literature has reported a history of asymetric data distributions in cross sectional studies of employee absence behavior. Various EDF (empirical data function) statistics have been developed to assess departure from normality and to assess the direction of its impact. For a first approxima- tion, conventional estimates for the mean, median and mode, when compared, may be quite different, and this difference would indicate departure from normality. The SAS software 192 193 which was used here to generate EDF statistics also calcu- lates customary measures for skewness and kurtosis. How- ever, SAS also offers two other statistics that provide a complimentary set of overall indices of normality: the Sha- piro-Wilk W statistic and the Kolomogorov D statistic. Each of these two statistics is a number created by making a ratio of a theoretical variance over the observed variance for the data, assuming a theoretical distribution that is uniform and normal with the mean and standard deviation observed in the empirical sample. Values for the W and D statistics range between 0 and 1 with. smaller ‘values indicating' less normality. Unfortu- nately, interpreting the values of D is not yet a very con- venient procedure. For sample sizes as large as 126 cases (the smallest case T-tested here), a D-statistic of less than about 0.8 would indicate a significant departure from normality. The degree to which this departure affects the meaning of the T-test statistic will tend to vary with the type of departure from normality (kurtosis, skewness, bimo- dality, etc.). Tables for comparison of these D values with the results of monte—carlo studies (M.A. Stephens, 1974) can be used to estimate the probability that the population dis- tribution underlying the observed distribution departs from normality with a .05 significance level. Low probabilities mean a low probability of drawing a sample with a better W or D-statistic from the respective distribution. The W sta- 194 tistic is a theoretically better estimate than D, and can be used for samples of 50 observation or less, while the D sta- tistic is still calculable and interpretable for larger sam- ples where W becomes inconvenient. In use, the normality indices become helpful in two ways. (1) Low values constitute a warning not to depend heavily on inferential statistics (like T-tests) from the relevant dataset. (2) Increasingly higher values may be generated by trial and error, transforming the data toward symmetry in successive approximations. Given enough com- puter time, this may be continued until a transformation is discovered that most effectively normalizes the distribu- tion. However, in the present research situation there are limitations on time and problems associated with interpret- ing the transformed data relative to its external reference frame. The present researCh will not go beyond assessing normality of the untransformed data to check the degree to which assumptions of normality underlying the T-tests may be correct. The data for each date involved in direct cross- sectional comparisons were accordingly subjected to SAS pro- cedure UNIVARIATE and W or D statistics for each date, plus the associated probability of getting a higher test statis- tic (significance level), were generated as appropriate. Selected results for these are listed with other EDF statis- tics in Table 19, Table 20 and Table 21. 195 While it might be feasible to create a set of meaning- ful comparison data for a Monte-Carlo type study of normal- ity for a distribution of the means used in the correla- tions, the computer resources and time this requires are not available at present. It should be noted, however, that since the standard error of the means is by construction = 1/n (the standard deviation of the raw data), the means used in. time—series correlations should. be substantially less affected than the individual data by extreme values and, as group means they will rarely have zero values so that the truncation problem is bypassed by construction. To provide a frame of reference with more traditional measures of the empirical data function and insight into the nature of departures from normality, the respective means, medians, ranges, skewness, and kurtosis are also reported in .Tables 19 through 21. (B): A Methodological Note on the Time-Series Correlation Procedures Two sets of correlations were calculated using the SAS "CORR” procedure. The first. set. of these ‘were done to establish the split-half reliability of the 39 derived time- series measures (time-use indices) for three sets of data. The two sets of daily means for corresponding indices for both pre-program and post-installation data on the 100-day data, and for data from day 1247 onward (May 21, 1981) for the 365-day data, were correlated between the split-halves 196 over time generating three sets of 39 time-series correla- tions. The second set of time-series correlations was done to examine the substantive relationships between changes in the various types of employee time-use behaviors. These inter- nal relations between time—series of means for the two over- all datasets (164 cases for each measure involved) were also compared. Frequencies, severities and mean severities for each of the two datasets were analyzed separately, creating essentially six sets of time-series correlations comparing the 13 different time-use variables. For each set of 13 time-use indices, a 13—by-l3 matrix of time-series correla- tions was created. By time-series correlation between selected pairs of the 13 different indices across selected intervals of time, both intra-annual and inter-annual evi- dence of progression effects, substitution effects, and gen- eral shifts between work and absence were examined. Introduction to the Time-Series Correlation Procedure Split-Half Reliabilities: Traditional (procedures to assess the reliability of absence measures have been compli- cated by contamination from the unreliability of absence behavior itself. Prior work has characteristically compared a cross-section of data from a group of employees at one point in time with another cross-secthmn of data from the same group of employees at some other time. The fundamental 197 procedure used in the present work does not compare cross- sections at all. Instead, a mean for each index is calcu- lated for each day across all employees in each split-half of the overall sample. This is done for every day of the program and the exact same procedure is applied to the other split-half. This generates two series of daily means for each time-use index for every day of the program (one for each split half), and these two subsets of means ought to be very nearly congruent if the two subsamples are each approx- imately representative of the overall sample. If these two sets of means are then correlated, day 1 with day 1, day 2 paired with. day 2, etc., the result containes two basic sources of incongruence: error of measurement, and any var- iation attributable to unintended sampling effects of group composition. As sample size increases, such sampling effects should become tiny by comparison with measurement error as the law of large numbers enhances representativity of the means. This incongruence or unreliability should then appear as a reduction for the value of the correlation. Perfectly reliable measures should generate a split-half reliability of 1. Lower values indicate less reliability. Any value of 0.7 to 0.8 or larger is particularly worthy of note. Tradi- tional absence research has very seldom produced reliability correlations greater than 0.35 for severity data and 0.65 for frequency data. It is possible that by using the time- 198 series reliability approach such findings could be raised for the traditional studies even with traditional measures. On the other hand, this would require a choice to sacrifice the insights possible from the detailed time-use data. While the conventional measures would probably be reliable if appropriately tested, the question which is of bottom line interest is whether they would be informative. Without the level of detail in the data toward which the careful observation and recategorization. procedures canI move the measurement process, such detail seems doubtful at the stage of analysis. It was necessary tx> prevent tautological inflation of the reliabilities due to variation while the moving values were still aggregating from a base period of 1 day up to the full lOO-day or 365-day starting point. For this reason, the calculation of time-series‘split-half reliability corre- lations excludes data from "early" periods in which moving values for these measures had not yet aggregated across a period equal to the full number of days for which the meas- ures were defined. Since the correlation is defined as covariance divided by the product of the standard devia- tions, any similarity between the overall absolute size of the means across the two samples would not affect the size of the correlations, while day-to-day changes can and will. 199 Substantive Correlational Tests: The same basic proce- dure that was ‘used for' split-half reliabilities can be applied to compare changes in two different types of time- use over time. In this way, for example, short-term non-le- gitimate absence can be correlated with long-term legitimate absence or with routine time worked to look for long-term program impacts. All that is needful is to calculate daily means for each index of the two measures each day of the program and then correlate the sets of means over time. Underlying the calculation of time-series reliabilities was a simple plan to create daily summary statistics and compare those instead of comparing the unsummarized raw data directly. The resulting correlations are not nearly as sen- sitive to two fundamental problems that have plagued conven- tional cross-sectional reliabilities. Conventional cross-sectional correlations compare two sets of values from the same individual's behavior at two points in time. Differences in those values between time 1 and time 2 can occur either (1) because of a fundamental unreliability in the measurement procedure, or (2) because the observed behavior of those individuals itself tends to change over time. Absence levels for individuals do change over time, however, which means that prior absence research has in fact found low reliabilities because the two series of values compared weren't from comparable samples of 200 observed behavior, even though they were in all probability collected using the same measurement procedure. It might even be as close to the truth to call those findings esti- mates of the reliability of individual absence behavior than estimates of reliability for the measures used. To overcome this limitation in the conventional para- digm, the time-series correlation procedure had to begin with two series of values that were: (a) representative of the absence behavior of all the workers over time, and (b) series from comparable samples of behavior over time. The second problem was resolved first -- by splitting the over- all random sample of DOT employees into two subsamples of randomly selected individuals. The time-series of values for those two subsample groups should then have been compa- rable samples of behavior over time. Each series, however, was still made up of values from many employees each day and to compared these values as correlated series, a single value to represent the values for all employees each day had to be found for every day in each of the two series. The group mean value was selected. Then these daily group means for each series were sorted into ascending order by date to create two time-series of comparable group means. Those two time-series were then correlated by date. The procedure was followed for each of the three indices on each measure for each of the three datasets tested. 201 Two theoretically comparable series were thereby estab- lished for every index in each of the datasets. Calculation of reliability for the time—use measures over time should have been straightforward. Thee empirical results, however, showed consistently high reliabilities for the majority of the measures and erratically low reliabilities for the oth- ers (see Table 31). Such anomalies could have resulted from the research procedure if any one or more of the following conditions occurred. 1. The workforce or the sample drawn from it could be too small to provide an adequate sample of time-use behavior for study or to prOperly represent the occurrence of behavior for the workforce over time. In effect, this would mean that one or both series of means came from too small a group of employees and would be overly affected by the behavior of any individual employee, a behavior past reserch has shown to be generally unrelia- ble. The result would be low reliability estimates for all time-uses. 2. Most of the employees may have spent very little or no time in the respective time-use behavior so that for them the time-use value was at or very close to zero most of the time. A mean would then still be calculable for the entire sample of employees, but only the idio- syncratic time-use behavior of relatively few employees 202 would have contributed to the time-series of means underlying the estimates of variance and covariance for each split half. Given a small enough number of con- tributing employees, the pattern of means in the time- series would then start to become as idiosyncratically variable as the time use behavior of the individual employee, while this variation in the total amount of the time use would then be divided by the number of employees in the total sample in calculating the mean level of time use for the entire group. If these conditions occurred, it could cause anomalous reliabilities in 'two ‘ways. First, the idiosyncratic pattern of means in one split half could then be quite different than that in the other split half. To the extent that this occurred, the split half reliability would have been lower. Second, if the total amount of time-use in a given category were to be small enough, then after calculation of the group mean for each day, the level of measurement error or even of rounding error in calculation could be large enough to mask this real variation in the time use. This, too, would dilute the actual observed pattern and cause low estimates of reli- ability. In effect, the cause of such low reliabilities boils down to another result of too small a smaple of the actual time-use behavior. 203 The amount of variation in either one or both series of means might be so low that the standard deviation of one or both series was computed as zero. Such a zero varia- tion for some time-use could occur either because the extremely regular occurrence of that time-use behavior so that the value of the mean never varied, or because there was none of the behavior occurring in one or both of the split halves. In either situation, the procedure to calculate a correlation would require division by zero, an operation that has no defined meaning. The SAS software used in the present research will simply report a zero value for the correlation in such a case. In a less extreme but potentially more confusing case, the sample of variation observed in the mean level of a time-use for the group could be quite tiny but non-zero. This would occur if very little of the behavior occurred at all, or if it was regulated to an extreme degree so that the level of the mean scarcely varied. Contrary to the result in condition 3 above, this would result in a calculation of the correlation as a ratio of two very small non-zero values, the covariance divided by the product of standard deviations. Those very small values would be distorted severely by surrounding measurement error and their covariance could very well even be neg- ative, tiny, and quite definitely meaningless but still negative. Negative values for the covariance would 204 result in a negative estimate for the reliability corre- lation. There may be a need to subdivide overall categories of time-use in order to discover‘ patterns in. the more detailed but less inclusive time-uses that are contribu- tory to the overall category. This also subdivides the sample of time-use behavior in the overall category into smaller samples of behavior in each of the more detailed categories. This, in effect, can bring into play one or more of conditions 2, 3 or 4 above and induce anomalous reliability estimates. The length of the time-series across time could be too short so that the number of means in each series was too small to capture the day-to-day variation in group—level time-use and too small to produce a reliable correla- tion. The length of the observation period used to accumulate the moving values can also cause the reliability corre- lations to be larger or smaller. On any given day the time-use of an individual employee may vary from one category to another. If the period of observation were limited to one day only, this would result in zero val- ues for every time-use category the employee did not use on that given day. Quite apart from giving no indica- tion of time-use rates across longer periods of special interest (such as annual absence), using these one day 205 values would also cause the truncation effect in the data distribution for each day (Hammer and Landau, 1981) to become very severe for all of the time-use measures. In fact, the more detailed the number of time-use cat- egories become, the more severely the truncation effect would distort the values of the daily group means for every time-use. Instead of permitting the effect of daily differences to be smoothly integrated with daily fluctuations from other employees by averaging the mov- ing values over time, the direct use of such daily fluc- tuations to calculate the means would transfer the trun- cation effect into the series of means and make the means more easily affected by sporadic individual employee behavior. In effect, using a period of time longer than one day initiates a moving average procedure that is later completed by calculating the daily group means. The series of means is then less affected by idiosyncratic local fluctuations from individual employ- ees, although sensitivity to daily fluctuations at a group level is preserved. The choice of a length for the time period of observa- tion was made as 100 days and 365 days for the present research. A seven day period and 30-day period were also considered because the business use of weekly and monthly scheduling periods would make them convenient for program evaluation. The longer time spans chosen 206 here are equivalent to the length of the quarterly plan- ning period and the business year. 8. Another time related condition can also affect the vari- ance in the series of means and result in anomalous reliability values. As the length of the time period of observation approaches 365 days, recurring cyclical effects that are locked into an annual cycle will tend to make the annual moving values settle around a total annual value which from then on would only show varia- tion from year to year. In effect, this would mean that the 365 day measures, which by their construction damp out intra-annual cycles, may not retain enough of a sam- ple of inter-annual variation to produce a meaningful reliability estimate. Because annual absence level and long-term inter-annual changes are of particular inter- est in the present work, the 365-day interval was selected anyway on a gamble that the other conditions listed above would not combine to make the sample size too small and the measures unreliable. Finally, it should be noted that the split-half proce- dure itself requires splitting the overall sample in half prior to any other steps. Measurement error for the origi- nal sample can be shown to approximate that of the split half multiplied by the square root of two. In effect, the split-half reliability is a value that relates to a repre- sentative half of the overall sample. This means that while 207 a high split-half reliability indicates reliability for the measure in the overall sample (where sample size would be twice as large), a low value merely indicates that no such reliability has been demonstrated with the smaller sample size. APPENDIX D A REVIEW AND DISCUSSION OF THE DETAILED RESULTS Appendix D A Review and Discussion of the Detailed Results The results presented here and in Chapter V are both organized to parallel the discussion of methodology pre- sented in Chapter IV and Appendix II. Sections A, B, C, D, and E in the methodology described the procedures which have produced corresponding Sections A, B, C, D, and E in the results. See Chapter IV, Figure 6 for an outline of these procedures. To minimize the number of tables necessary (and sur- rounding explanatory material), results ikn: all hypotheses which rely upon a specific analytical procedure are pre- sented together, either in a single table or in a cluster of consecutive tables or time-series graphs. Where results are extremely repetitive, only a selective summary was reported here. Procedures for deriving each one of the three forms of each of the 13 time-use indices have been presented in Appendix III. A brief description and guidelines for inter- preting the various results and summary statistics presented in the tables are described briefly preceding each table or set of tables. Following each respective block of tables, results are discussed and interpreted relative to the spe- cific related hypotheses presented in Chapter IV. The extent of support for each hypothesis is assessed and dis- cussed briefly in the light of other co-occurrent findings. 210 Results of the (A) normality tests and statistics describing distributional form are presented first; followed by (B) the T-tests; (C) the time-series split-half reliabil- ity correlations; (D) the time-series correlational tests of substitution effects; and (E) the graphical time-series analysis. Note: For all results presented here, no findings on indexes for termination time have been reported because in the randomly selected sample of 164 cases, which is the basis of the present report, no occurrences of termination were reported in the raw data. A: Normality and EDF Statistics for the T-Test Data The selected EDF statistics described below are tabu- lated here to illustrate the distributional forms typical of the T-test data. Since there was EH1 overwhelming repeti- tiveness in these EDF statistics, only sample statistics for the three overall indices of All-Time-Worked, All-Time-Ab- sent, and All-Time-Scheduled-Off have been reported here, and only for the dates involved in T-tests of year-to-year overall change. Those particular dates were selected because they provided a basis for comparing the 100-day and 365-day data distributions. EDF statistics for the 30-day, lOO-day and 365-day datasets respectively have been pre- sented in Table 19, Table 20 and Table 21. Date: Mean Std . 211 Refers to the relative date to which the daily cross-section of T-test data pertain. Day 182 means July 1, 1978, day 183 means July 2, 1978, etc. Refers to the number of cases for which data existed on the respective day. Variation in N is indicative of the extent of turnover across the time duration of this study. and Med: Refer to the group mean value of the selected time-use index on a particular day, or to the group median value, respectively. The size and direction of differences between mean and median are crude but useful indicators of the direction and magnitude for departures from normality. Where the mean is greater than the median, the distribution will generally be positively skewed. DeV.: Refers to the standard deviation for data in a particular daily cross-section. Skew: 212 Refers to the skewness of the data distribution, ideally, in a normal distribution, skew = 0. Any negative skewness in time-use data will tend to be rather small because such a result is limited by minimum values at zero. On the other hand, posi- tive skewness in time-use data can get quite large, depending on the number and size of unusu- ally large outliers above the mean. Kurtosis: Prob The "peakedness" of the distribution relative to a normal distribution, ideally = 0. Negative values mean a flat distribution, positive values mean a high modal concentration. Refers to the Kolmogorov "D" statistic described in Chapter IV. The D statistic varies between 0 and 1, with lower values indicating less normal distribution. In a normal distribution, D = 1.0. > D: Means "the probability of this distribution pro- ducing a higher D statistic." The D and Prob > D statistics occur as pairs. Table 19: Normality and EDP Statistics for 21J3 Selected 30-Day T-Test Data Distributions Std. Kur- Pbe. Index Date M Mean Med. Dev. Skew tosis D > D ALL TIME WORKED 30-Day 211 129 1.64 2.0 1.14 0.61 0.68 0.19 <0.01 Frequency 912 151 1.19 1.0 1.05 0.90 0.98 0.23 <0.01 30-Day 211 129 103.30 120.0 58.12 -0.69 -0.84 0.15 <0.01 Severity 912 151 125.58 152.0 61.59 —1.10 -0.02 0.19 <0.01 30-Day Mean 211 129 67.25 56.0 53.31 0.65 -0.60 0.16 <0.01 Sev./Occur. 912 151 73.77 74.5 65.06 0.38 -l.14 0.17 <0.01 ALL TIME ABSENT 30-Day 211 129 1.23 1.0 1.07 0.95 1.28 0.23 <0.01 Frequency 912 151 1.08 1.0 0.96 1.09 1.76 0.28 <0.01 30-Day 211 129 31.73 16.0 45.96 2.63 8.14 0.24 <0.01 Severity 912 151 43.78 20.0 61.31 1.90 2.95 0.24 <0.01 30—Day Mean 211 129 22.50 8.0 41.03 3.85 16.69 0.29 <0.01 Sev./Occur. 912 151 36.59 8.0 60.20 2.25 4.27 0.29 <0.01 ALL TIME SCHEDULED OFF 30-Day 211 129 4.32 5.0 2.11 -0.30 -0.33 0.26 <0.01 Frequency 912 151 4.49 5.0 1.50 -1.21 3.14 0.32 <0.01 30-Day 211 129 103.61 80.0 60.25 1.42 0.95 0.30 <0.01 Severity 912 151 69.48 64.0 41.83 2.89 10.40 0.31 <0.01 30-Day Mean 211 129 52.02 16.0 78.11 1.82 1.59 0.43 <0.01 Sev./Occur. 912 151 23.24 14.4 47.85 4.31 16.98 0.48 <0.01 1214 Table 20: Normality and ED! Statistics for Selected loo-Day T-Test Data Distributions Std. Kur- 55 Index Date M Mean Med. Dev. Skew tosis D > D ALL TIME WORKED loo-Day 546 126 3.44 3.0 2.09 1.04 1.64 0.17 <0.0 Frequency 1277 151 1.74 2.0 1.51 1.07 2.55 0.17 <0.0 1642 141 3.77 3.0 2.44 0.51 0.08 0.14 <0.0 100~Day 546 126 419.30 477.0 156.33 -l.59 1.53 0.25 <0.0 Severity 1277 151 281.83 336.0 151.65 -0.96 -0.44 0.20 <0.0 1642 141 412.06 504.0 197.96 -1.13 -0.05 0.23 <0.0 loo-Day Mean 546 126 144.95 124.0 108.28 1.94 5.45 0.16 <0.0 Sev./Occur. 1272 151 134.43 114.3 127.11 1.03 0.49 0.15 <0.0 1642 141 130.59 103.2 122.12 2.12 5.53 0.17 <0.0 ALL TIME ABSENT loo-Day 546 126 2.92 3.0 1.93 1.33 2.45 0.18 <0.0 Frequency 1277 151 1.86 2.0 1.53 0.97 2.10 0.20 <0.0 1642 141 3.33 3.0 2.29 0.68 0.09 0.16 <0.0 loo-Day 546 126 131.35 64.0 173.33 2.38 5.34 0.29 <0.0 Severity 1277 151 88.77 40.0 143.65 2.51 5.69 0.28 <0.0 1642 141 136.25 72.0 169.46 2.29 5.26 0.25 <0.0 loo-Day Mean 546 126 75.75 21.3 158.64 3.36 11.26 0.36 <0.0 Sev./Occur. 1277 151 54.36 16.0 123.21 3.69 13.17 0.34 <0.0 1642 141 59.24 22.7 135.67 4.43 20.07 0.33 <0.0 ALL TIME SCHEDULED OFF loo-Day 546 126 14.31 15.0 4.27 -0.50 5.26 0.31 <0.0 Frequency 1277 151 9.72 11.0 4.99 -0.22 0.21 0.20 <0.0 1642 141 13.06 14.0 5.22 -0.78 2.55 0.33 <0.0 lOO‘Day 546 126 254.37 236.0 96.04 2.77 13.95 0.24 <0.0 Severity 1277 151 433.97 400.0 154.71 1.54 1.71 0.28 <0.0 1642 141 255.19 220.0 176.60 2.46 5.32 0.39 <0.0 loo-Day Mean 546 126 26.50 16.0 73.30 9.96 105.58 0.43 <0.0 Sev./Occur. 1277 151 141.16 37.1 256.78 2.18 2.91 0.42 <0.0 1642 141 74.45 15.4 205.26 3.29 9.11 0.50 <0.0 Table 21: 215 Normality and EDF Statistics for Selected 365-Day T-Test Data Distributions Std: Kur- Prob. Index Date N Mean Med. Dev. Skew tosis D > D ALL TIME WORKED Annual 546 126 11.17 11.0 5.44 0.53 0.35 0.12 <0.01 Frequency 1277 151 9.34 9.0 5.29 0.23 -0.23 0.08 <0.04 1642 141 12.21 12.0 6.80 0.39 0.27 0.06 <0.15 Annual 546 126 1383.33 1594.0 517.67 -0.95 -0.17 0.16 <0.01 Severity 1277 151 1322.63 1583.0 591.38 -1.34 0.31 0.27 <0.01 1642 141 1490.70 1727.0 591.42 -1.42 1.10 0.20 <0.01 Annual 546 126 151.20 128.0 108.38 2.68 9.79 0.19 <0.01 Sev./Occur. 1272 151 155.07 135.4 106.95 1.59 4.46 0.11 <0.01 1642 141 147.33 122.1 127.49 3.30 17.74 0.13 <0.01 ALL TIME ABSENT Annual 546 126 10.76 10.5 5.50 0.49 0.24 0.09 <0.02 Frequency 1277 151 8.92 9.0 5.11 0.37 0.16 0.07 <0.04 1642 141 3.33 3.0 2.29 0.68 0.09 0.16 <0.01 Annual 546 126 396.29 264.0 459.44 3.08 11.04 0.26 <0.01 Severity 1277 151 431.03 240.0 560.3 2.78 7.54 0.25 <0.01 1642 141 506.01 289.0 564.40 2.70 7.89 0.24 <0.01 Annual 546 126 52.91 24.3 141.63 8.28 78.74 0.36 <0.01 Sev./Occur. 1277 151 103.63 27.4 388.48 6.14 41.40 0.42 <0.01 1642 141 59.24 22.7 135.67 4.43 20.07 0.33 <0.01 ALL TIME SCHEDULED 01"? Annual 546 126 47.79 53.0 18.37 -0.01 1.21 0.22 (0.01 Frequency 1277 151 44.21 48.0 18.61 -0.98 0.95 0.27 <0.01 1642 141 48.18 52.0 16.45 -1.14 3.57 0.24 <0.01 Annual 546 126 1130.93 918.0 501.61 1.44 1.73 0.27 <0.01 Severity 1277 151 1160.86 996.0 588.36 1.96 3.23 0.32 <0.01 1642 141 915.97 835.0 510.73 2.52 7.10 0.33 <0.01 Annual 546 126 40.09 16.6 70.25 4.90 28.14 0.33 <0.01 Sev./Occur. 1277 151 164.61 21.0 573.57 4.54 19.40 0.43 <0.01 1642 141 92.32 15.8 387.79 6.23 41.31 0.46 <0.01 216 Summary 9f Results for the Normality and EDF Statistics Between July 30, 1978, and June 30, 1980, there was a 25% increase in the size of the means for all time worked, a 30% increase in the mean for all time absent, and a corre- sponding 35% decrease in all time scheduled off. The D sta- tistics are uniformly less than 0.5 and most are less than 0.4, indicating that the samples were non-normal, while the associated probabilities of the underlying distribution pro- ducing a higher D statistic were without exception less than 0.01 (D was highly significant). Skew and kurtosis were generally positive and.aa clear relationship exists between the sizes of the mean and standard deviation. This set of distributions is definitely non-normal and positively skewed. The relatively large standard deviations in the values of 30-day severity per occurrence (compared to the respective means) are a consequence of compounding two sources of variation in measurement. The severity per occurrence (mean severity) was calculated as a function of 30-day severity divided by 30-day frequency. For the lOO-day data, three sets of data were examined: June 30, 1979, June 30, 1981, and June 30, 1982. There was a 30% drop in the mean amount of time worked per hundred days from 1979 to 1981, followed by a return to the 1978 level by 1982. This was accompanied by a 30% reduction in all time absent from 1979 1x3 1981, which also returned to 217 the 1979 level by 1982. These two fluctuations were possi- bly due to a 40% increase in the mean of time scheduled off between June 30, 1979, and June 30, 1981, which also settled back down to the 1979 level by June 30, 1982. The differ- ences between lOO-day and 30-day values suggest that in the last month of the pre-program period a lot of scheduled time off may have been formally granted. The D statistics for the lOO-day data are uniformly less than 0.50 with associ- ated probabilities of a higher D value that are all less than 0.01. Skew and kurtosis are similar to the 30-day data. The mean and standard deviation are strongly and positively correlated. For the 365-day data, there is a fluctuation in the mean that follows the changes for corresponding dates in the short-term lOO-day data, but is much less severe. There is also a generalized long-term 8% increase in the number of hours worked per year and a 20% increase in number of hours absent, with a corresponding 20% decrease in the number of hours scheduled off. Skewness for frequency of time sched- uled off was small and negative, as was skewness for sever- ity of all time worked. The D statistics are uniformly less than 0.50 with accompanying probabilities for getting a higher D value of less than 0.01. In general, these distributions are non-normal, posi- tively skewed, and indicate some sizeable changes from year- 218 to-year. Those differences have been T-tested despite the non-normality of their associated distributions and the results are reported below. Hypothesis 1 is definitely not supported. The T-test results should be considered, but estimates of confidence levels associated with those results are not.an: all accu- rate. These results. are (entirely' consistent with. prior findings in the absenteeism literature. B: Results for the T-Tests of Cross-Sectional Comparisons T-tests ‘were implemented. on 48 hours, T-Prob == 0.0001. Again, the change 1J1 time absent was non-significant, except for the frequency of short-term legitimate absence. Hypothesis 5 is therefore supported to the extent that it is evident in time use behavior that is at least partly under the control of the employer. However, no significant evidence exists that 21 significant holiday/seasonal effect on discretionary non-legitimate absence occurred. Table 26: 23C) T-Test Results for Hypotheses about Long-Term Year-to-Year Change in the loo-Day Measures, Frequency Indices Only ATRelative RElative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F—Prob T T-Pr All Time Worked 1166 2.95 2.22 1531 4.02 2.52 0.12 ~3.83 0.00 982 3.37 2.04 1347 2.62 1.87 0.31 3.29 0.00 1025 3.45 2.38 1390 4.06 2.69 0.13 -2.06 0.04 1210 3.01 2.17 1575 4.02 2.50 0.09 3.69 0.00 Routine Work 1166 2.89 2.28 1531 3.93 2.43 0.43 -3.75 0.00 982 3.34 2.02 1347 2.60 1.82 0.22 3.27 0.00 1025 3.42 2.28 1390 3.99 2.59 0.12 -1.96 0.05 1210 2.97 2.25 1575 4.00 2.52 0.17 -3.68 0.00 Other Work 1166 0.05 0.24 1531 0.06 0.39 0.00 -0.27 0.78 982 0.07 0.32 1347 0.04 0.24 0.00 0.72 0.46 1025 0.05 0.24 1390 0.03 0.24 0.88 0.64 0.52 1210 0.07 0.30 1575 0.07 0.41 0.00 -0.11 0.91 All Sch. Off 1166 13.56 6.67 1531 13.38 5.15 0.00 0.26 0.79 982 12.91 4.30 1347 9.30 3.82 0.16 7.61 0.00 1025 13.64 4.74 1390 12.28 5.04 0.46 2.37 0.01 1210 12.75 6.65 1575 13.54 5.25 0.01 1.12 0.26 Missing Values 1166 0.15 0.43 1531 0.28 1.76 0.00 -0.82 0.41 982 0.17 0.41 1347 1.10 0.94 0.00 -10.85 0.00 1025 0.13 0.37 1390 0.21 1.48 0.00 -0.68 0.49 1210 0.15 0.37 1575 0.30 1.74 0.00 -1.06 0.28 Punishment Time 1166 0.11 0.37 1531 0.13 0.50 0.00 -0.42 0.67 982 0.15 0.46 1347 0.11 0.35 0.00 0.96 0.33 1025 0.09 0.33 1390 0.12 0.42 0.00 -0.62 0.53 1210 0.12 0.40 1575 0.11 0.36 0.23 0.13 0.89 Routine Off 1166 13.42 6.87 1531 13.31 5.29 0.00 0.15 0.88 982 12.77 4.41 1347 9.39 4.26 0.69 6.67 0.00 1025 13.54 4.87 1390 12.20 5.19 0.44 2.27 0.02 1210 12.58 6.87 1575 13.47 5.39 0.00 -1.24 0.21 All Time Absent 1166 2.47 2.04 1531 3.27 2.31 0.14 -3.12 0.00 982 3.18 1.97 1347 2.54 1.81 0.32 2.89 0.00 1025 3.40 2.32 1390 3.59 2.59 0.20 -2.66 0.50 1210 2.48 1.97 1575 3.45 2.38 0.02 -3.80 0.00 Sht. Leg. Ab. 1166 1.81 1.69 1531 2.51 2.09 0.01 -3.12 0.00 982 2.36 1.69 1347 2.26 1.78 0.56 0.50 0.61 1025 2.70 2.00 1390 3.30 2.63 0.00 -2.16 0.03 1210 1.73 1.75 1575 2.67 2.13 0.02 -4.10 0.00 Long Leg. Ab. 1166 0.22 0.54 1531 0.33 0.78 0.00 -l.36 0.17 982 0.60 0.62 1347 0.33 0.59 0.57 3.78 0.00 1025 0.59 0.64 1390 0.34 0.65 0.74 3.30 0.00 1210 0.21 0.48 1575 0.37 0.76 0.00 -2.18 0.03 Sht Non-Leg. Ab. 1166 0.68 1.23 1531 0.73 1.17 0.58 -0.39 0.69 982 0.45 0.96 1347 0.20 0.45 0.00 2.89 0.00 1025 0.30 0.93 1390 0.17 0.43 0.00 1.52 0.12 1210 0.75 1.19 1575 0.74 1.17 0.82 0.03 0.97 Late Time 1166 0.37 1.43 1531 0.58 1.16 0.01 -l.39 O.l6 982 0.35 1.05 1347 0.43 0.94 0.21 -0.70 0.48 1025 0.33 0.96 1390 0.50 1.07 0.19 -l.45 0.14 1210 0.33 0.88 1575 0.52 0.98 0.20 -1.71 0.08 231 Table 27: T-Test Results for Hypotheses about Long-Tern Year-to-Year Change in the loo-Day Measures, Severity Indices Only Relative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 1166 412.55 204.59 1531 436.37 184.62 0.22 —1.05 0.29 982 376.40 151.03 1347 264.89 126.99 0.04 6.84 0.00 1025 415.21 172.11 1390 403.56 182.65 0.47 0.56 0.57 1210 418.51 214.19 1575 435.45 191.03 0.17 -0.71 0.47 Routine Work 1166 411.81 204.68 1531 434.87 184.84 0.22 -1.01 0.31 982 375.34 150.62 1347 264.43 126.71 0.04 6.83 0.00 1025 414.52 171.82 1390 403.33 182.38 0.47 0.54 0.59 1210 417.77 213.92 1575 434.43 190.77 0.17 0.70 0.48 Other Work 1166 0.74 4.56 1531 1.50 12.31 0.00 -0.69 0.48 982 1.06 6.38 1347 0.45 2.49 0.00 1.08 0.28 1025 0.69 4.13 1390 10.23 1.90 0.00 1.24 0.21 1210 0.74 3.38 1575 1.02 7.15 0.00 -0.42 0.67 A11 Sch. Off 1166 282.32 197.86 1531 257.06 161.11 0.01 1.20 0.23 982 250.52 135.30 1347 394.01 105.51 0.00 -10.14 0.00 1025 261.13 146.87 1390 222.48 146.30 0.96 2.26 0.02 1210 270.39 203.50 1575 255.13 168.09 0.02 0.70 0.48 Missing Values 1166 92.03 256.02 1531 56.06 193.95 0.00 1.36 0.17 982 38.78 157.57 1347 246.07 85.02 0.00 -14.11 0.00 1025 59.60 186.22 1390 30.92 137.55 0.00 1.50 0.13 1210 95.74 256.74 1575 63.06 206.91 0.01 1.20 0.23 Punishment Time 1166 3.54 17.63 1531 7.50 38.41 0.00 -1.12 0.26 982 4.24 22.28 1347 10.20 56.25 0.00 -1.18 0.24 1025 1.54 7.71 1390 12.58 74.47 0.00 -1.75 0.08 1210 6.49 30.73 1575 5.16 25.32 0.02 0.40 0.68 Routine Off 1166 186.75 89.00 1531 193.50 72.37 0.01 -0.71 0.47 982 207.50 79.67 1347 137.74 65.84 0.02 8.18 0.00 1025 200.09 71.03 1390 178.97 79.69 0.17 2.38 0.01 1210 168.17 81.55 1575 186.91 72.52 0.16 -2.08 0.03 All Time Absent 1166 110.01 190.81 1531 110.21 164.91 0.08 —0.01 0.99 982 168.99 173.15 1347 140.04 170.53 0.86 1.44 0.15 1025 126.99 172.50 1390 178.52 218.35 0.00 -2.23 0.02 1210 115.05 195.41 1575 112.53 165.16 0.05 0.60 0.54 Sht. Leg. Ab. 1166 44.68 95.66 1531 50.62 78.71 0.02 -0.58 0.56 982 65.06 76.12 1347 60.10 82.12 0.36 0.53 0.59 1025 68.29 80.62 1390 77.89 102.19 0.00 -0.89 0.37 1210 46.20 103.29 1575 42.50 58.71. 0.00 0.38 0.70 Long Leg. Ab. 1166 55.63 170.79 1531 49.73 150.92 0.14 0.31 0.75 982 97.38 172.49 1347 75.74 166.34 0.66 1.09 0.27 1025 53.74 168.58 1390 97.41 216.84 0.00 -1.91 0.05 1210 58.01 170.66 1575 68.35 161.32 0.50 -0.02 0.98 Sht Non-Leg. Ab. 1166 9.70 25.08 1531 9.85 17.38 0.00 -0.06 0.95 982 6.54 19.36 1347 4.19 13.35 0.00 1.21 0.22 1025 4.97 18.08 1390 3.22 11.13 0.00 1.00 0.31 1210 10.83 23.32 1575 11.69 27.13 0.03 -O.29 0.77 Late Time 1166 0.79 2.94 1531 1.19 3.28 0.18 -1.09 0.27 982 1.32 4.74 1347 0.72 2.22 0.00 1.39 0.16 1025 1.25 3.69 1390 1.18 3.77 0.79 0.15 0.87 1210 0.60 2.25 1575 1.18 3.18 0.00 -l.79 0.07 232 B-3a: Results of the T-Tests for Year-to-Year Change in the lOO-Day Measures Both the frequency indices and severity indices show significant shifts in time worked (down from 376 hours to 264 hours), in time scheduled off (up from 250 hours to 394 hours), and in missing values (up from 38 hours to 246 hours between summer of 1980 and summer of 1981). The pattern here suggests a strike or plant shutdown during the summer of 1981. Comparison between day 1347 and day 1390 suggests that this event must have occurred between July 1 and August 12, 1981, since the event was not included in the lOO-day base for day 1390. The frequency indices however, also indicate significant year-to-year increases for other peri- ods in overall frequency of legitimate absences and a corre- sponding overall increase in the frequency of work inter- vals, and these two findings are both internally consistent with the progression hypothesis, although no corresponding increase in time used in absence was observed. Furthermore, except for the anomaly during the summer of 1981, the only significant change in overall absence for these tests was in long-term legitimate absence which increased from 53 to 97 hours per employee just after the summer anomaly occurred and which suggests that some workers may have taken a long-term leave during those summer months. 233 Hypothesis 6 is only supported for the frequency of both absence and time worked, workers increased both. On .the other hand, this did not significantly affect severity of either index: which suggests that the employees were developing a long-term habit of using their casual leave days in several occurrences instead of just a few. Some weak support for an hypothesis on long-term absence increase is evident. These findings are highly vulnerable to short-term intra-annual effects, however, and despite the comparison to corresponding year-to-year dates, they should be considered with caution. 2134 Table 28: T-Test Results for Holiday Effects on the loo-Day Data, Severity Indices Only Relative ’Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 1166 412.55 204.59 1210 418.51 214.19 0.58 —0.25 0.80 1531 436.37 184.62 1575 435.41 191.03 0.69 0.04 0.96 982 376.40 151.03 1025 415.21 172.11 0.11 -2.08 0.03 1347 264.89 126.99 1390 403.56 182.65 0.00 -7.40 0.00 Routine Work 1166 411.81 204.68 1210 417.77 213.92 0.59 -0.25 0.80 1531 434.87 184.84 1575 434.43 190.77 0.71 0.02 0.98 982 375.41 150.52 1025 414.52 171.82 0.11 -2.11 0.03 1347 264.43 126.71 1390 403.33 182.38 0.00 7.43 0.00 Other Work 1166 0.74 4.56 1210 0.74 3.38 0.00 0.00 1.00 1531 1.50 12.31 1575 1.02 7.15 0.00 0.40 0.68 982 1.06 6.38 1025 0.69 4.13 0.00 0.60 0.54 1347 0.45 2.49 1390 0.23 1.90 0.00 0.86 0.39 All Sch. Off 1166 282.32 197.86 1210 270.39 203.50 0.73 0.52 0.60 1531 257.06 161.11 1575 255.13 168.09 0.62 0.10 0.92 982 250.52 135.30 1025 261.23 146.87 0.32 —0.66 0.51 1347 394.01 105.51 1390 222.48 146.30 0.00 11.29 0.00 Missing Values 1166 92.03 256.02 1210 95.74 256.74 0.97 -0.13 0.90 1531 56.06 193.95 1575 63.06 206.91 0.45 -0.29 0.76 982 38.78 157.57 1025 59.60 186.22 0.04 -1.05 0.29 1347 246.07 85.02 1390 30.92 137.55 0.00 15.80 0.00 Punishment Time 1166 3.54 17.63 1210 6.49 30.73 0.00 -1.02 0.30 1531 7.50 38.41 1575 5.16 25.32 0.00 0.60 0.54 982 4.24 22.28 1025 1.54 7.71 0.00 1.41 0.16 1347 10.20 56.25 1390 12.58 74.47 0.00 -0.30 0.76 Routine Off 1166 186.75 89.00 1210 168.17 81.55 0.29 1.89 0.05 1531 193.50 72.47 1575 186.91 72.52 0.99 0.76 0.44 982 207.50 79.67 1025 200.09 71.03 0.16 0.85 0.39 1347 137.74 65.84 1390 178.97 79.69 0.02 -4.74 0.00 All Time Absent 1166 110.01 190.81 1210 115.05 195.41 0.77 —0.23 0.82 1531 110.21 164.91 1575 112.53 165.16 0.99 -0.12 0.90 982 168.99 173.15 1025 126.99 175.50 0.96 2.11 0.03 1347 140.04 170.53 1390 178.52 218.35 0.00 -1.65 0.10 Sht. Leg. Ab. 1166 44.68 95.66 1210 46.20 103.29 0.35 -0.13 0.89 1531 50.62 78.71 1575 42.50 58.71 0.00 0.98 0.32 _ 982 65.06 76.12 1025 68.29 80.62 0.48 -0.36 0.72 1347 60.10 82.12 1390 77.89 102.19 0.01 -1.61 0.10 Long Leg. Ab. 1166 55.63 170.79 1210 58.01 170.66 0.99 -0.12 0.90 1531 49.73 150.92 1575 58.35 161.32 0.43 -0.46 0.64 982 97.38 172.49 1025 53.74 188.58 0.78 2.22 0.02 1347 75.74 166.34 1390 97.41 216.84 0.00 -0.94 0.34 Sht Non-Leg. Ab. 1166 9.70 25.08 1210 10.83 23.32 0.37 -0.41 0.68 1531 9.86 17.38 1575 11.69 27.13 0.00 -0.67 0.50 982 6.54 19.36 1025 4.97 18.08 0.40 0.73 0.46 1347 4.19 13.35 1390 3.22 11.13 0.03 0.66 0.50 Late Time 1166 0.79 2.94 1210 0.60 2.25 0.00 0.66 0.51 - 1531 1.19 3.28 1575 1.17 3.18 0.70 0.04 0.97 982 1.32 4.74 1025 1.25 3.69 0.00 0.14 0.89 1347 0.72 2.22 1390 1.18 3.77 0.00 -1.25 0.21 235 B-3b: T-Test Results for Hypotheses about Holiday Effects in the lOO-Day Measures Both the frequency and severity indices were T-tested to detect significant holiday effects, but the results were essentially the same for both so that a detailed presenta- tion of only the severity shifts is included in Table 28. In brief, two significant shifts have occurred across major holiday periods. These occurred for the time-use measures during the summers of 1980 and of 1981. A large shift occurredd across the same interval for which an anomaly was discovered which affected Hypothesis 6 and showed up as a year-to-year overall change. A smaller shift also associ- ated with a summer (July 4th) holiday period in 1980 is also apparent and significant. No holiday effects were apparent across the Christmas periods. Hypothesis 7 is supported for summer holidays, but this holiday effect may be at least in part the effect of annual renewal or vacation. There is no significant effect across the winter (Christmas) holiday season when renewal would not be expected to have much effect and holidays might. Once again, some rather anomalous shifts and reductions in time- worked as well as time scheduled off occurred during the summer of 1981. Table 29: 1236 T-Test Results for Hypotheses of Combined Holiday, Seasonal, and Renewal Effects on the loo-Day Measures: Severity Indices Only gfielative Relative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 982 376.40 151.03 1166 412.55 204.59 0.00 -.175 0.08 1347 264.89 126.99 1531 436.37 184.62 0.00 -9.09 0.00 1025 415.21 172.11 1210 418.51 214.19 0.01 -0.15 0.88 1390 403.56 182.65 1575 435.45 191.03 0.60 -1.43 0.15 Routine Work 982 375.34 150.52 1166 411.81 204.68 0.00 -1.76 0.07 1347 264.43 126.71 1531 434.87 184.84 0.00 -9.03 0.00 1025 414.52 171.82 1210 417.77 213.92 0.01 -0.15 0.16 1390 403.33 182.38 1575 434.43 190.77 0.60 -1.40 0.16 Other Work 982 1.06 6.38 1166 0.74 4.56 0.00 0.50 0.61 1347 0.45 2.49 1531 1.50 12.31 0.00 -O.99 0.32 1025 0.69 4.13 1210 0.74 3.38 0.01 -0.12 0.90 1390 0.23 1.90 1575 1.02 7.15 0.00 -1.28 0.20 All Sch. Off 982 250.52 135.30 1166 282.32 197.86 0.00 -1.63 0.10 1347 394.01 105.57 1531 257.06 161.11 0.00 -8.44 0.00 1025 261.23 146.87 1210 270.39 203.50 0.00 -0.45 0.65 1390 222.48 146.30 1575 255.13 168.09 0.10 -1.74 0.08 Missing Values 982 38.78 157.57 1166 92.03 256.02 0.00 -2.18 0.03 1347 246.07 85.02 1531 56.01 193.95 0.00 10.65 0.00 1025 59.60 186.22 1210 95.74 256.74 0.00 -1.40 0.16 1390 30.92 137.55 1575 63.06 206.91 0.00 -l.54 0.12 Punishment Time 982 4.24 22.28 1166 3.54 17.63 0.00 0.30 0.76 1347 10.20 56.25 1531 7.50 38.41 0.00 0.47 0.63 1025 1.54 7.71 1210 6.49 30.73 0.00 -1.92 0.05 1390 12.58 74.47 1575 5.16 25.32 0.00 1.12 0.26 Routine Off 982 207.50 79.67 1166 186.75 89.00 0.18 2.13 0.03 1347 137.74 65.84 1531 193.50 72.47 0.26 -6.76 0.00 1025 200.09 71.03 1210 168.17 81.55 0.09 3.63 0.00 1390 178.97 79.69 1575 186.91 72.52 0.27 -0.88 0.38 All Time Absent 982 168.99 173.15 1166 110.01 190.81 0.24 2.81 0.00 1347 140.04 170.53 1531 110.21 164.91 0.69 1.49 0.13 1025 126.99 172.50 1210 115.05 195.41 0.13 0.56 0.57 1390 178.52 218.35 1575 112.53 165.16 0.00 2.86 0.00 Sht. Leg. Ab. 982 65.06 76.12 1166 44.68 95.66 0.00 2.05 0.04 1347 60.10 82.12 1531 50.62 78.71 0.62 0.98 0.32 1025 68.29 80.62 1210 46.20 103.29 0.00 2.07 0.03 1390 77.89 102.19 1575 42.50 58.71 0.00 3.57 0.00 Long Leg. Ab. 982 97.38 172.49 1166 55.63 170.79 0.90 2.11 0.03 1347 75.74 166.34 1531 49.73 150.93 0.25 1.38 0.17 1025 53.74 168.58 1210 58.01 170.66 0.88 -0.22 0.82 1390 97.41 216.84 1575 58.35 161.32 0.00 1.72 0.08 Sht Non-Leg. Ab. 982 6.54 19.36 1166 9.70 25.08 0.00 -1.23 0.22 1347 4.19 13.35 1531 9.85 17.38 0.00 -3.07 0.00 1025 4.97 18.08 1210 10.83 23.32 0.00 -2.44 0.01 1390 3.22 11.13 1575 11.69 27.13 0.00 -3.43 0.00 Late Time 982 1.32 4.74 1166 0.79 2.93 0.00 1.15 0.25 1347 0.72 2.22 1531 1.19 3.28 0.00 -1.31 0.19 1025 1.25 3.69 1210 0.60 2.25 0.00 1.86 0.06 1390 1.18 3.77 1575 1.18 3.18 0.04 0.02 0.98 237 B-3c: Summary of Results for the T-Tests of Combined Sea- sonal and Annual Renewal Effects on the lOO-Day Measures Several T-test comparisons for the lOO-day' measures offer general support for Hypothesis 8. Combined seasonal and renewal effects on absence do appear to take place, pro- ducing significant decreases in levels of absence and increases in time worked from summer to winter. However, the same summer of 1981 anomally (Date 1 = 1347) which was evident in earlier T-tests would also presumably be contrib- uting to this pattern, so to some extent these results too are suspect. Short-tenn absence does account for part of the change, both legitimate and non-legitimate short-term absence. Long-term absence was also affected, but not to the same extent. Comment on the Measures Used to Examine Intra-Annual T-Test Results In general, the frequency indices and mean severity co-vary with the severity indices and tend to provide sig- nificant results across the same intervals, although in this there were some exceptions. For convenience, in the present analysis the results on the frequency indices have not been included. Detailed examination of the relationship between severity and frequency of absence is definitely an interest- ing problem. Indeed, the parallelim in the lOO-day indices between severity and frequency is in the reverse direction to an apparent co-variation suggested by the 30-day meas- 238 ures, where frequency went down as severity went up. It is likely that this latter effect is present by construction of the measures and that both patterns can and do operate at the same time. The internal consistency' and. detail built into the integral set time-use measures should allow such an analysis to be carried out. The respective mathematical forms should be distinguishable and any relationships between them may be as well. Formal procedures for detailed time-series analy- sis need to be applied. But, that problem is beyond the scope of this present inquiry due to limitations on time and computer resources. Some insight may be forthcoming from the time-series correlations and graphical analysis. B-4: Results of the T-Tests on Overall Program—Induced Change in the lOO-Day Measures The so-called program induced changes predicted in the lOO-day data unquestionably appear significant. Time sched- uled off, time worked, and time absent all decreased signif- icantly from the end of the baseline year to the end of the first. program year. Then. all three again significantly increased back to their original level by the end of the second. program year. Meantime, missing ‘values increased dramatically between the end of the baseline year and the taxi of the first program year, then decreased back toward their original level by the end of the second program year. 239 Overall, between the end of the baseline year and the end. of the second program. year, time worked apparently showed no significant change, nor did time scheduled off or time absent. These overall year-to-year tests of program induced changes were likely to be affected by short-term fluctuations for the 100-day measures, however. Rather than focusing on them, the 365-day measures were used instead to assess overall program impact. 240 Table 30: T-Test Results of Program Induced Change in the 365-Day Measures Relative RelaEiVe Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 546 1277 Frequency 11.17 5.44 9.34 5.29 0.74 2.81 0.00 Severity 1383.33 517.67 1322.63 591.38 0.12 0.91 0.36 Mean Sev. 151.10 108.38 155.07 106.95 0.87 -0.30 0.76 Routine Work 546 1277 Frequency 10.90 5.33 9.20 5.16 0.69 2.69 0.00 Severity 1379.21 517.23 1320.09 590.91 0.12 0.89 0.37 Mean Sev. 154.12 109.31 156.97 107.91 0.87 —0.22 0.82 Other Work 546 1277 Frequency 0.34 1.67 0.17 0.53 0.00 1.13 0.25 Severity 4.11 19.09 2.54 8.33 0.00 0.86 0.39 Mean Sev. 1.63 7.87 2.04 7.02 0.18 -0.46 0.64 All Scheduled Off 546 1277 Frequency 47.79 18.37 44.21 18.61 0.88 1.61 0.10 Severity 1130.93 501.61 1160.86 588.36 0.07 -O.46 0.64 Mean Sev. 40.09 70.25 164.61 573.57 0.00 -2.64 0.00 Missing Values 546 1277 Frequency 1.29 2.27 1.09 0.36 0.00 1.02 0.31 Severity 419.68 703.82 506.07 753.95 0.43 -0.98 0.32 Mean Sev. 227.95 541.41 471.54 710.64 0.00 -3.23 0.00 Punishment Time 546 1277 Frequency 0.45 1.07 0.36 0.82 0.00 0.81 0.41 Severity 9.40 27.64 13.79 48.23 0.00 -0.95 0.34 Mean Sev. 4.17 8.85 6.88 17.61 0.00 -1.65 0.09 Routine Off 546 1277 Frequency 47.28 18.85 43.70 18.79 0.97 1.57 0.11 Severity 701.85 245.18 641.01 269.04 0.28 1.97 0.50 Mean Sev. 15.09 2.47 14.09 5.02 0.00 2.14 0.03 All Time Absent 546 1277 Frequency 10.76 5.50 8.92 5.11 0.39 2.87 0.00 Severity 396.29 459.44 431.03 560.36 0.02 -0.57 0.57 Mean Sev. 52.92 141.63 103.63 338.48 0.00 -1.67 0.09 Sht. Leg. Absence 546 1277 Frequency 7.44 3.90 6.95 3.99 0.80 1.02 0.31 Severity 222.06 277.14 180.44 256.47 0.36 1.29 0.19 Mean Sev. 39.41 105.95 32.11 79.58 0.00 0.64 0.52 Long Leg. Absence 546 1277 Frequency 0.04 0.23 0.95 1.05 0.00 -10.35 0.00 Severity 34.60 274.80 221.22 509.37 0.00 -3.88 0.00 Mean Sev. 23.05 159.09 127.80 333.81 0.00 -3.42 0.00 Sht. Non-Leg. Ab. 546 1277 Frequency 3.94 3.80 1.75 2.55 0.00 5.49 0.00 Severity 139.63 275.64 29.36 61.49 0.00 4.40 0.00 Mean Sev. 35.86 90.91 8.33 13.94 0.00 3.37 0.00 Late Time 546 1277 Frequency 1.17 3.04 1.05 2.81 0.37 0.32 0.74 Severity 3.44 9.80 2.95 7.79 0.01 0.46 0.64 Mean Sev. 1.17 2.76 1.04 2.19 0.01 0.44 0.66 2411 Table 30 Continued Relative 4Re1ative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 546 1642 Frequency 11.17 5.44 12.21 6.80 0.01 -1.39 0.16 Severity 1383.33 517.67 1490.70 591.42 0.13 -1.58 0.11 Mean Sev. 151.10 108.38 147.33 127.49 0.06 0.27 0.78 Routine Work 546 1642 Frequency 10.90 5.33 11.94 6.56 0.02 -1.43 0.15 Severity 1379.21 517.23 1487.47 591.21 0.13 -1.60 0.11 Mean Sev. 154.12 109.31 148.20 125.65 0.11 0.41 0.68 Other Work 546 1642 Frequency 0.34 1.67 0.18 0.90 0.00 0.98 0.32 Severity 4.11 19.09 3.23 19.73 0.71 0.37 0.71 Mean Sev. 1.63 7.87 1.24 4.97 0.00 0.48 0.63 A11 Scheduled Off 546 1642 Frequency 47.79 18.37 48.18 16.45 0.20 -0.18 0.85 Severity 1130.93 501.61 915.97 510.73 0.84 3.47 0.00 Mean Sev. 40.09 70.25 92.32 387.79 0.00 -1.57 0.11 Missing Values 546 1642 Frequency 1.29 2.27 1.70 5.14 0.00 -0.84 0.40 Severity 419.68 703.82 201.50 597.66 0.06 2.71 0.00 Mean Sev. 227.95 541.41 113.30 396.57 0.00 1.95 0.05 Punishment Time 546 1642 Frequency 0.45 1.07 0.34 0.86 0.01 0.93 0.35 Severity 9.40 27.64 30.63 142.68 0.00 -1.73 0.08 Mean Sev. 4.17 8.85 21.57 133.91 0.00 -1.54 0.12 Routine Off 546 1 1642 Frequency 47.28 18.85 47.81 16.62 0.15 -0.24 0.81 Severity 701.85 245.18 683.84 234.47 0.61 0.61 0.54 Mean Sev. 15.09 2.47 13.75 3.87 0.00 3.41 0.00 All Time Absent 546 1642 Frequency 10.76 5.50 11.74 6.58 0.04 -1.33 0.18 Severity 396.29 459.44 506.01 564.39 0.02 -1.75 0.08 Sht. Leg. Absence 546 1642 Frequency 7.44 3.90 10.01 5.82 0.00 —4.29 0.00 Severity 222.06 277.14 221.06 224.37 0.02 0.03 0.97 Mean Sev. 39.41 105.95 34.91 123.56 0.08 0.32 0.74 Long Leg. Absence 546 1642 Frequency 0.04 0.23 1.03 1.61 0.00 7.22 0.00 Severity 34.60 274.80 257.50 534.78 0.00 -4.35 0.00 Mean Sev. 23.05 159.09 107.46 286.28 0.00 -3.02 0.00 Sht. Non-Leg. Ab. 546 1642 Frequency 3.94 3.80 1.75 2.25 0.00 5.62 0.00 Severity 139.63 275.64 27.44 54.52 0.00 4.49 0.00 Mean Sev. 35.86 90.91 7.80 11.44 0.00 3.44 0.00 Late Time 546 1642 Frequency 1.17 3.04 1.79 3.03 0.99 -1.67 0.09 Severity 3.44 9.80 4.22 10.50 0.43 -0.62 0.53 Mean Sev. 1.17 2.76 0.99 2.09 0.00 0.59 0.55 1242 Table 30 Continued RElative 4ARElative Variable Date Std. Date Std. Index 1 Mean 1 Dev.1 2 Mean 2 Dev.2 F-Prob T T-Pr All Time Worked 1277 1642 Frequency 9.34 5.28 12.21 6.80 0.00 -4.00 0.00 Severity 1322.63 591.38 1490.70 591.42 1.00 -2.43 0.21 Mean Sev. 155.07 106.95 147.33 127.49 0.03 0.56 0.57 Routine Work 1277 1642 Frequency 9.20 5.16 11.94 6.56 0.00 -3.96 0.00 Severity 1320.09 510.91 1487.47 591.21 0.99 -2.42 0.01 Mean Sev. 156.97 107.91 148.20 125.65 0.07 0.64 0.52 Other Work 1277 1642 Frequency 0.17 0.54 0.18 0.90 0.00 -0.13 0.89 Severity 2.54 8.33 3.23 19.73 0.00 -0.39 0.70 Mean Sev. 2.04 7.02 1.24 4.97 0.00 1.14 0.25 All Sch. Off 1277 1642 Frequency 44.27 18.61 48.18 16.45 0.14 -1.93 0.05 Severity 1160.86 588.36 915.97 510.73 0.09 3.80 0.00 Mean Sev. 164.61 573.57 92.32 387.79 0.00 1.27 0.20 Missing Values 1277 1642 Frequency 1.09 0.36 1.70 5.14 0.00 -1.40 0.16 Severity 506.07 753.95 201.50 597.66 0.00 3.84 0.00 Mean Sev. 471.54 710.64 113.30 396.57 0.00 5.36 0.00 Punishment Time 1277 1642 Frequency 0.36 0.82 0.34 0.86 0.55 0.17 0.86 Severity 13.79 48.23 30.63 142.68 0.00 -1.33 0.18 Mean Sev. 6.87 17.61 21.57 133.91 0.00 -1.29 0.19 Routine Off 1277 1642 Frequency 43.70 18.79 47.81 16.62 0.14 -1.98 0.04 Severity 641.01 269.04 683.84 234.47 0.10 -1.45 0.14 Mean Sev. 14.09 5.02 13.75 3.87 0.00 0.66 0.50 All Time Absent 1277 1642 Frequency 8.92 5.11 11.74 6.58 0.00 -4.08 0.00 Severity 431.03 560.36 506.01 564.39 0.93 -1.14 0.25 Mean Sev. 103.63 338.48 94.15 350.52 0.67 0.23 0.81 Sht. Leg. AD. 1277 1642 Frequency 6.95 3.99 10.01 5.82 0.00 -5.21 0.00 Severity 180.44 256.47 221.06 224.37 0.11 -1.44 0.15 Mean Sev. 32.11 79.58 34.91 123.56 0.00 -0.23 0.81 Long Leg. Ab. 1277 1642 Frequency 0.95 1.05 1.03 1.61 0.00 -0.47 0.64 Severity 221.22 509.37 257.50 534.78 0.58 -0.59 0.55 Mean Sev. 127.80 333.81 107.46 286.28 0.07 0.56 0.57 Sht Non-Leg. Ab. 1277 1642 Frequency 1.75 2.55 1.75 2.25 0.14 0.01 0.99 Severity 29.36 61.48 27.44 54.52 0.15 0.28 0.77 Mean Sev. 8.33 13.94 7.80 11.44 0.02 0.36 0.72 Late Time 1277 1642 Frequency 1.05 2.81 1.79 3.03 0.36 -2.14 0.03 Severity 2.95 7.79 4.22 10.50 0.00 -1.17 0.24 Mean Sev. 1.04 2.19 0.99 2.09 0.57 0.18 0.85 243 Day Day Day 546 1277 1642 All Time Worked '93 Frequency of time worked up 1 Severity of time worked up 107 hrs Mean severity time worked - same Mean Sev. Time Sch. Off up 57 hrs Freq. Time Scheduled Off - same Severity Time Sch. Off down 200 hrs Severity All Absence up 110 hrs Mean Severity A11 Absence up 42 hrs Frequency of All Absence up 1 Long-Term Legitimate up 223 hrs Short-Term Legitimate down 1 hr Short-Term Non-Legitimate down 112 Frequency Missing Values up 0.5 Severity Missing Values down 218 Mean Sev. Missing Values down 114 NOTE: The relative scales of sketches above are merely roughly "eyeballed" pictures of trends in each line, clustered for convenience. Figure 16: Approximate Graphic Profiles of Long-Term Program Impact on the Overall 365-Day Time-Use Measures: Relative Scales 244 B-S: Results of the T-Tests on Overall Program Induced Changes Indicated by the 365-Day Measures Figure 16 offers a rough graphical summary of overall changes in the major annual time-use indices. Full details have been presented in Table 30. Overall, time worked per employee per year is up by 107 hours to 1490 hours per year. Time scheduled off was down by 200 hours to 915 hours. And an overall increase in time absent of 110 hours per employee per year took place, all of which was in long-term legiti- mate absence since short-term non-legitimate absence dropped 112 hours to 27 hours per employee per year from 139 hours. Most of the observed change in absence behavior occurred before the end of the first complete year of the program. Also worthy of note is a substantial decrease in the level of missing values present in the data, over the second year of the program especially. It appears likely that a certain amount of time must have been needed after program start-up to streamline the administration of data collection by supervisors, data—entry clericals, etc. Overall, it appears that the program was clearly suc- cessful if the T-test statistics are to be believed. Once again, a caveat is critical here. These T-test values are based on distinctly abnormal data distributions. Hypothesis 9 is supported. See the time-series plots (Figure 16 above) for a more detailed picture of the program impact over time. 245 C. The Split-Half Time-Series Reliability Correlations Results for three sets of time-series correlations between split-halves of the time-use datasets are summarized in Table 31. A reliability correlation and the associated probability of that correlation occurring by mere coinci- dence (significance level) has been listedd for each of the 36 time-use indices (time due to termination of employees was dropped because sample size was equal to 0). The three datasets for which reliability correlations were generated are: (1) during the baseline year from day 281 to day 546, using the 100-day data; (2) during the post-installation period from day 982 to day 1680, using the lOO-day data; and (3) during the post-installation period from day 1247 to day 1680, using the 365-day data. The time-series split-half reliability correlations have been discussed above in Chapter IV. An interpretation and discussion of the results follows. Table 31: 246 Time Series Correlations for Split-Half Reliability Variable Index Pre-Prog r lOO-Day Post-Inst r IOO-Day Post-Inst r 365-Day All Time Worked Frequency Severity Mean Sev. Routine Work Frequency Severity Mean Sev. Other Work Frequency Severity Mean Sev. A11 Scheduled Off Frequency Severity Mean Sev. Missing Values Frequency Severity Mean Sev. Punishment Time Frequency Severity Mean Sev. Routine Off Frequency Severity Mean Sev. All Time Absent Frequency Severity Mean Sev. Sht . Leg. Absence Frequency Severity Mean Sev. Long Leg. Absence Frequency Severity Mean Sev. Sht. Non-Leg. Ab. Frequency Severity Mean Sev. Late Time Frequency Severity Mean Sev. 000 \D ‘1 small Sample too small Sample too small Sample too small 000 \D \1 000 \J p 0.91 Sample too small Sample too small Sample too small No values in 112 during baseline 0.39 Sample too small Sample too small .88 .92 .50 00° .89 .91 .57 000 Sample too Sample too Sample too 'Sample too Sample too Sample too 0.99 0.98 0.92 0.89 0.91 0.73 0.92 0.57 0.21 .84 .96 0.89 00 .76 .36 .56 OOO Sample too Sample too 0.57 small small Sample Sample small 0. small small small Sample Sample Sample 0 O Var. 00° 00° GOO 00° .99 .50 .98 too small too small 70 too small too small too small .97 .99 too small 0.97 Var. Var. too small too small 0.88 0.87 0 Sample Sample Sample small 0 small Sample 0 .69 too small too small too small .51 too small .57 247 Summary of Results of the Time-Series Split-Half Reliability Correlations Except for two jpeculiarities, the reliabilities are extremely high for all measures (see Table 31). The construction of 13mg time-series split-half reli- ability correlations is such that reliability is based upon the actual sample size of time-use behaviors underlying the correlated series of means as noted in Appendix III. Where the time-use behavior involved becomes an idiosyncratic and sporadic event associated with relatively few employees, then the size of the respective mean will start to vary unpredictably instead of systematically, and the two "halves" of the overall sample may be anomalously (but mean- inglessly) correlated as has occurred for small sample behaviors in a few of the present measures. Five types of the 12 reported time-use measures were apparently affected by this type of problem. These were: other' work, late time, disciplinary time-off (punishment time), short-term absence (both legitimate and non-legiti- mate after program installation), and long-term legitimate absence before program installation. Where the amount of available behavior (and sample size) became larger (even for these five exceptions), the anomalous correlations disap- peared. 248 The second anomaly in the split-half reliability corre- lations was an unusual value for reliability of the routine time-off mean severity index in the 365-day measures. Exam- ination of the covariance and prodduct of standard devia- tions which contributed to that result showed that both had extremely tiny values (less than .0001 of the size of the mean). The same period for the 365-day measures were high and consistent. Routine scheduled time off is controlled by the employer' and. is otherwise probatdjr the least likely time-use to sum»! a low reliability. For consistency with frequency and severity, it should be about +0.94. Setting these anomalies aside, the results of the split-half reliability correlations are overwhelmingly exactly what they should be by construction, very high and all positive. Hypothesis If) is unquestionably supported, and this is nothing remarkable but merely the result of measurement design. Comparison of the reliabilities for the 100-day meas- ures and 365-day measures separately to examine the frequen- cy-severity controversy shows an ixmxiguing result. When the "small sample" behaviors were excluded.i11 the 100-day measures, severity was marginally more reliable than fre- quency in 9 out of 12 cases. When only the "small sample" behaviors were examined, severity was more reliable than frequency in 7 out of 12 cases. For the 365-day measures 249 neither measure was more reliable. Frequency was a more reliable measure than severity in four out of six cases for the "large sample" behaviors, and more reliable than sever- ity for three out of six of the "small sample" behaviors. Hypothesis 10(a) is very well supported although it does require an adequate sample size, andboth the frequency and severity of the time-use behaviors can be highly relia- ble. Overall, neither was preferable on the basis of reli- ability. Results of the Time-Series Correlations on Substitution Effects on the Severity Indices Time-series correlations were calculated within selected time intervals for both the lOO-day and 365-day datasets to examine various substantive hypotheses as out- lined in Chapter IV. Results for Hypotheses 11, 12, 13, and 14 were examined in that order for each of the five selected combinations of: lOO-day or 365-day measure, and time interval of analysis. Intervals, correlations and associated significance levels, datasets, and related hypotheses are summarized in Table 32. 250 Table 32: Time-Series Correlations Testing Substantive Hypotheses for Severity Indices Only Related Variable 1 Variable 2 Interval r r:Prob Dataset Hypothesis 281- 546 * 0.45 0.0001 loo-day Short-Term Long-Term 982-1277 0.36 0.0001 loo-day Hypothesis Legitimate Legitimate 1278-1680 0.56 0.0001 loo-day 11 Absence Absence 982-1680 0.52 0.0001 loo-day Progression 1247-1680 * 0.53 0.0001 365-day 281- 546 *-0.17 0.0058 loo-day Short-Term Short-Term 982-1277 -0.89 0.0001 loo-day Hypothesis Legitimate Non-Leg. 1278-1680 -0.93 0.0001 loo-day 12 Absence Absence 982-1680 -0.92 0.0001 loo-day Legitimacy 1247-1680 *-0.67 0.0001 365-day 281- 546 * 0.69 0.0058 lOO-day Short-Term Long-Term 982-1277 -0.44 0.0001 loo-day Hypothesis Non-Leg. Legitimate 1278-1680 -0.55 0.0001 loo-day l3 Absence Absence 982-1680 -0.53 0.0001 lOO-day Progression 1247—1680 *-0.56 0.0001 365-day Legitimacy 281- 546 0.97 0.0058 loo-day Short-Term Time 982-1277 -0.02 0.0001 lOO-day Hypothesis Non-Leg. Worked 1278-1680 0.25 0.0001 lOO-day 14 Absence 982-1680 0.06 0.0001 lOO-day Program 1247-1680 *-0.10 0.0001 365-day Impact 281- 546 -0.98 0.0058 loo-day Missing Time 982-1277 -0.91 0.0001 lOO—day As Values Worked 1278-1680 -0.75 0.0001 lOO-day Above 982-1680 -0.96 0.0001 loo-day 1247-1680 40.97 0.0001 365-day 281- 546 *-0.71 0.0058 loo—day Missing Long-Term 982-1277 -0.22 0.0001 loo-day As Values Legitimate 1278-1680 -0.45 0.0001 lOO-day Above Absence 982-1680 -0.18 0.0001 lOO-day 1247-1680 -0.40 0.0001 365-day 281- 546 * 0.08 0.0058 loo-day Missing Short-Term 982-1277 -0.38 0.0001 lOO-day Hypothesis Values Non-Leg. 1278-1680 -0.63 0.0001 loo-day 14 Absence 982-1680 -0.32 0.0001 lOO-day Program 1247-1680 *-0.47 0.0001 365~day Impact Note: * indicates a finding in which one or both of the measures correlated showed low reliability for the split-half samples. Since the overall sample size is twice that in the split-halves, such findings may or or may not be based upon a reliable series of values (see Appendix II 251 Summary of Time-Series Correlations Testing the Substantive Hypothesis, Using Severity Indices There is evidence of a very strong legitimacy effect after the program installation occurred, but not before. In fact, before installation, short-term non-legitimate absence correlated +0.97 iwth time worked, while after installation it correlated -O.92 with short-term legitimate absence, and was uncorrelated with time worked. Hypothesis 12 is defi- nitely supported. The progression hypothesis (Hypothesis 11) is clearly not supported, however. Short- and long-term forms of legitimate absence are positively correlated, both on the lOO-day and 365-day measures. A combined progression and legitimacy effect (Hypothe- sis 13) is supported, but very probably the apparent effect is largely induced by legitimacy alone. Hypothesis 14 is not supported, presumably because the overall impact of the program was allowed to dissipate in long-term legitimate absence. A Comment on Frequency-Severity Relationships For the sake of curiosity, the relationship between frequency and severity of the three major time-use catego- ries was examined tudefly. Table 33 below summarizes the results. Time scheduled off shows a strong negative corre- 252 lation between frequency and severity. But, for the other two time uses, the relationship is equally strong and posi- tive. Table 33: Time-Series Correlations To Inspect Relations Between Frequency Indices and Severity Indices Variable 1 Variable 2 Interval r r:Prob Dataset Time-Worked Time-Worked 982-1680 0.70 0.0001 lOO-day Severity Frequency 1246-1680 0.60 0.0001 365-day Time Sch Off Time Sch off 982-1680 -0.78 0.0001 lOO-day Severity Frequency 1246-1680 -O.74 0.0001 365-day Time Absent Time Absent 982-1680 0.51 0.0001 lOO-day Severity Frequency 1246-1680 0.73 0.0001 365-day BIBLIOGRAPHY 254 BIBLIOGRAPHY Allen, Steven (3. "Compensation, Safety, and Absenteeism: Evidence From The Paper Industry," Industrial and Labor Relations Review, Vol. 34, No. 2 (January 1981). Baum, John F. "Effectiveness of an Attendance Control Policy in Reducing Chronic Absenteeism," Personnel Psychology, 1978, Vol. 31. Beatty, Richard W. & Beatty, James R. "Longitudinal Study of Absenteeism of Hard-Core Unemployed," Psychological Reports, 1975, Vol. 36, 395-406. Bloomfield, Peter. Fourier .Analysis (n: Time-Series: An Introduction, John Wiley & Sons, 1976. Breagh, James A. "Predicting Absenteeism from Prior Absen- teeism and Work Attitudes," Journal 9f Applied Psychol- ogy, 1981, Vol. 66, NO. 5, 555-560. Behrend, Hilde & Pocock, Stuart, "Absence and the Individ- ual: A Six-Year Study in One Organization," Industrial and Labor Review, Vol. 114, No. 3, Nov-Dec, 1976, pp. 311-327. Campbell, Edward J. "Sick Leave Abuse and What To Do About It: A.lmmflt at Government Employees," Personnel, Vol. 57, Nov-Dec 1970. Carlson, John G. & Hill, Kenneth D. "The Effect of Gaming - Attendance and Amtitude," Personnel Psychology, 1982, Vol. 35. Chadwick-Jones, J.K., Nicholson, Nigel, and Brown, Colin. Social Psychology 9f Absenteeism, New York: Praeger Publishers, 1982. Cheloha, Randall S. & Farr, James L. "Absenteeism, Job Involvement, and Job Satisfaction in an Organizational Setting," Journal 9f Applied Psychology, 1980, Vol. 65, No. 4. pp. 467-473. Elkouri, Frank and Elkouri, Edna Asper. How Arbitration Works, Third Edition, Washington, D.C.: The Bureau of National Affairs, Inc., 1960. Garrison, Kathleen R. & Muchinsky, Paul M. "Evaluating the Concept of Absentee-Proneness with Two Measures of Absence," Personnel Psychology, 1977, Vol. 30. 255 Glass, Gene V., Willson, Victor L. & Gottman, John M. Design and Analysis pf Time-Series Experiments, Colo- rado Associated University' Press, Boulder Colorado, 1975. Glasser, Jay H. "A Stochastic Model for Industrial Illness Absenteeism," American Journal pf Public Health, Vol. 60, No. 10, 1970. Hamblin, Robert L. "Social Attitudes: Magnitude Measure- ment and Theory," Chapter' 3 in Measurement .ip the Social Sciences, H. M. Blalock, Jr., ed., 1974. Hammer, Tove Helland & lenday, Jaqueline. "Methodological Issues in the Use of Absence Data," Journal pf Applied Psychology, 1981, Vol. 66, No. 5. pp. 574-581. Hedges, Janice Niepert. "Special Labor Force Report: Absence from Work - Measuring the Hours Lost," Monthly Labor Review, Oct. 1977, Vol. 100, No. 10. Ilgen, Daniel R. "Attendance Behavior: A Reevaluation of Latham and Pursell's Conclusions," Journal pf Applied Psychology, 1977, Vol. 62, No. 2, pp. 230-233. Ilgen, Daniel R. and Hollenback, John H. "The Role of Job Satisfaction in Absence Behavior, " Organizational Behavior and Human Performance, Vol. 19, pp. 148-161 (1977). Ivancevich, John M. & McMahon, J. Timothy. "Black-White Differences in a Goal-Setting Program," Organizational Behavior and Human Performance, Vol. 20, pp. 287-300, (1977). Johns, Gary. "A Multivariate Study of Absence from WOrk," Academy pf Management Proceedings, 1978, pp. 69-73. Kim, Jay S. & Campagna, Anthony F. "Effects of Flextime on Employee Attendance and Performange: A Field Experi- ment," Academy pf Management Journal, 1981, Vol. 24, No. 4, 729-741. Kuzmito, Frank E. "How Much is Absenteeism Costing Your Organization," The Personnel Administrator, June 1979. Latham, Gary P. & Pusell, Elliott D. "Measuring Absenteeism From the Opposite Side of the Coin," Journal pf Applied Psychology, 1975, Vol. 60, No. 3, 369-471. Latham, Gary P. & Pursell, Elliott D. "Measuring Atten- dance: A Reply to Ilgen," Journal pf Applied Psychol- ogy, 1977, Vol. 62, No. 2, 234-236. 256 Liu, Lon-Mu. "Box-Jenkins Time -n- Series Analysis," BMDP Statistical Software, 1978 Edition, Ch. 20.2. Luthans, Fred & Mario, Terry L. "Evaluating Personnel Pro- grams Through the Reversal Tchnnique," Personnel Jour- nal, October 1979. Macy, Barry A. & Mirvis, Phillip H. "A Methodology for Assessment of Quality of WOrk Life and Organizational Effectiveness in Behavioral - Economic Terms," Adminis- trative Science Quarterly, June 1976, Vol. 21. Markham, Steven E.; Dansereau, Fred Jr. & Alutto, Joseph A. "Group Size and Absenteeism Rates: A Longitudinal Analysis," Academy pf Management Journal, 1982, Vol. 25, No. 4, 921-927. Markham, Steven E.; Dansereau, Fred & Alutto, Joseph A. "On the Use of Shift as an Independent Variable in Absen— teeism Research," Journal pf CKtupational Psyghology, 1982, Vol. 55, 225-231. Markham, Steven E.; Dansereau, Fred & Alutto, Joseph A. "Female vs. Male Absence Rates: A Temporal Analysis," Personnel Psychology, 1982, Vol. 35. Mikalachki, A. & Chapple, D. C. "Absenteeism and Overtime: Double Jeopardy," Relations Industrielles, 1977, Vol. 32. Miner, Mary Green. "Job Absence and Turnover: A New Source of Data," Monthly Labor Review, 1977. Mirvis, Philip H. & Lawler, Edward E., III. "Measuring the Financial Impact. of' Employee .Attitudes," Journal _gf Applied Psychology, 1977, Vol. 62, No. 1, 1 - 8. Mobley, William H. "Some Unanswered Questions in Turnover and Withdrawal Research," Academy pf Management Review, 1982, Vol. 7, NO. 1, 111 - 116. Montgomery, Douglas C. & Johnson, Lynwood A. Forecasting and Time-Series Analysis, McGraw-Hill, Inc., 1976. Mowday, Richard T. & Spencer, Daniel G. "The Influence of Task and Personality Characteristics on Employee Turn- over and Absenteeism Incidents," Academy pf Management Journal, 1981, Vol. 24, No. 3, 634 - 642. Nicholson Nibel; Brown, Colin A. & Chadwick-Jones, J.K. "Absence from Work and Personal Characteristics," Jour- nal pf Applied Psychology, 1977, Vol. 62, No. 3, 319 - 327. 257 Nicholson, Nigel; Brown, Colin A. & Chadwick-Jones J.K. "Absence from Work and Job Satisfaction," Journal pf Applied Psychology, 1976, Vol. 61, No. 6, 727-737. Nicholson, Nigel &: Goodge, Peter' M. "The Influence of Social, Organizational and Biographical Factors on Female Absence," Journal pf Management Studies, October 1976, Vol. 13. Nicholson, Nigel; Jackson, Paul & Howes Gillian. "Shiftwork and Absence: An Analysis of Temporal Trends," Journal pf Applied Psychology, 1978, Vol. 51, 127 - 137. Nicholson, Nigel; Wall, Toby & Idscheron, Joe. "The Pre- dictability of Absence and Propensity to Leave from Employees' Job Satisfaction and Attitudes Toward Influ- ence in Decision-Making," Hpmap Relations, Vol. 30, No. 6, 1977, 499-514. Parasuraman, Saroj. "Predicting' Turnover Intentions and Turnover Behavior: A Multivariate Analysis," Journal pf Vocational Behavior, Vol. 21, 111 - 121, 1982. Popp, Paul O. & Behlohlav, James A. "Absenteeism in a Low Status Work Environment," Academy pf anagement Jour- nal, 1982, Vol. 25, No. 3, 677 - 683. Rousseau, Denise M. "Relationship> of’ Work to INonwork," Journal pf Applied Psychology, 1978, Vol. 63, No 4, 513 - 517. SAS Institute Inc., SAS User's Guide: Basics, 1982 Edition, Cary, NC: SAS Institute Inc., 1982. Scott, Dow & Markham, Steve. "Absenteeism Control Methods: A Survey of Practices and Results," Personnel Adminis- trator, June 1982. Shapiro, 8.8. and Wilk, M.B. "An Analysis of Variance Test for Normality (Complete Samples)," Biometrika, 52, 1965, pp. 591-611. Smulders, Peter G. W. "Comments on Employee Absence/Atten- dance as a Dependent Variable in Organizational Research," Journal pf Applied Psychology, 1980, Vol. 65, No. 3, 368 - 371. Steers, Richard M. "Employee Turover and Absenteeism: A Future Research Agenda," Invited address, Michigan State University, April 24, 1981. Steers, Richard M. & Rhodes, Susan R. "Major Influences on Employee Attendance: A Process Model," Journal pf Applied Psychology, 1978, Vol. 63, No. 4, 391 - 407. 258 Steers, Richard M. & Rhodes, Susan R. "A New Look at Absen- teeism," Personnel, November - December 1980. Stephens, M.A. "EDF Statistics for Goodness of Fit and Some Comparisons," Journal pf the American Statistical Asso- ciation, 69, pp. 730-737. Stevens, S.S. "Mathematics, Measurement and Psychophysics," Handbook pf Experimental Psychology, 1951. Taylor, Daniel E. "Absences from Work Among Full-Time Employees," Monthly Labor Review, Vol. 104, 1981. Terborg, James R.; Thomas W. Lee; Frank J. Smith; Gregory A. Davis & Mark S. Burbin. "Extension of the Schmidt and Hunter Validity Generalization Procedure to the Pre- diction of Absenteeism Behavior from Knowledge of Job Satisfaction and Organizational Commitment," Journal pf Applied Psychology, 1982, Vol. 67, No. 4, 440 - 449. Thrall, Tony & Engelman, Laszlo. '"Univariate and Bivariate Spectral Analysis," BMDP Statistical Software, 1981 Edition, Ch. 20.1. Tukey, John W. Exploratory Data Analysis, Addison - Wesley, 1977.