— = .——= —— _— _——' — _— _— —- (DOC) (HES‘S 3 1293 This is to certify that the thesis entitled An Analysis of Interview Frequency and Reference Period in Rural Consumption Expenditure Surveys: A Case Study From Sierra Leone presented by Sarah Gibbons Lynch has been accepted towards fulfillment of the requirements for Economics /:i2;zéZ/Zo£§4./5Kafiflgfgk Major professor Date (WALé/fi']? 0-7639 91w: 25¢ per day per item RETURNING LIBRARY MATERIALS: Place in book return to remve charge from circulation records . f y ~15 , an: " I“ flaw“; .' ‘93]. ‘ i I“ v: I 4: ~'.T:I. 5" ~“‘\ {55% 1113' ' .-.;.i g 1". "HEM“. AN ANALYSIS OF INTERVIEW FREQUENCY AND REFERENCE PERIOD IN RURAL CONSUMPTION EXPENDITURE SURVEYS: A CASE STUDY FROM SIERRA LEONE By Sarah Gibbons Lynch A THESIS Submitted to Michigan State University in partial fulfillment of the requirements for the degree of MASTER OF SCIENCE Department of Agricultural Economics 1979 ABSTRACT AN ANALYSIS OF INTERVIEW FREQUENCY AND REFERENCE PERIOD IN RURAL CONSUMPTION EXPENDITURE SURVEYS: A CASE STUDY FROM SIERRA LEONE By Sarah Gibbons Lynch Interview frequency and length of reference period are two facets of survey design crucial to the collection of reliable and cost-efficient consumption expenditure data. The influence of these two factors on con- sumption expenditure estimates was analyzed using parametric and non- parametric techniques. The data base used in this study was a comprehen- sive rural consumption expenditure survey conducted in Sierra Leone in 1974-1975. The objectives of this study were: 1) to analyze the differences in household expenditure estimates based on data collected during two inter- views per month with data collected during one interview per month; 2) to compare expenditure estimates derived from each of the four individual days of recall contained in one interview; and 3) to determine if the expenditure estimates based on a first or sum of the second and third day of recall differed depending on whether the data were drawn from the first or second interview. Results of this analysis provided some, though not conclusive, evi- dence that expenditure estimates based on one interview per month were not statistically different from two interviews per month. Expenditure esti- mates fran the first day of recall were statistically different from and Sarah Gibbons Lynch consistently higher than those from the other three days of recall. Expen- diture estimates of days of recall from the first interview were higher than those from the second interview in a month. Problems of memory decay, respondent fatigue, and some telescoping of expenditures were cited as explanations for the results. TABLE OF CONTENTS List of Tables List of Figures Chapter 1 Introduction 1.1 Consumption Expenditure Survey Methodology in Low Income Countries 1.2 Focus of the Study 1.3 Outline of Remaining Chapters The Research Problem 2.1 Factors in Survey Design 2.2 Factors in Determining Interview Frequency 2.3 Factors in Determining the Length of Interview Reference Period Survey Methodology Used in Sierra Leone Rural Consumption Expenditure Survey 3.1 Sample Selection 3.2 Description of Questionnaires and Interview Schedule 3.3 Description of Interview Categories Non-Parametric Analysis of the Influence of Interview Frequency on Expenditure Estimates 4.1 Non-Parametric Tests and Their Application to This Analysis 4.2 Comparison of the Two-Interview Set with the One-Interview Subset 4.2.1 Comparison of Total Mean Expenditures for All Comnodities 4.3 Comparison of the Two-Interview Set and the One-Interview Subset with the One-Interview Independent Set Analysis of Aggregated Comnodity Groups 5.1 Data Preparation 5.2 Comparison of Mean Expenditure Estimates 5.3 Comparison of Total Annual Expenditures ii 15 15 18 22 25 25 27 3O 31 38 38 45 Chapter 6 Analysis of Reference Period 6.1 Introduction 6.2 Sample Description 6.3 Comparison of Mean Expenditure Estimates from Individual Days of Recall 6.4 Differences in Expenditure Estimates Between the First and Second Interview 6.5 Discussion of Results 7 Conclusion 7.1 Summary of Research Findings 7.2 Research Implications Appendices A Disaggregated Commodity List 3 Assignment of Interviews Which Overlapped Two Months C Procedure to "Puff Up" the Data D Indexing Procedure Selected Bibliography Page 47 47 48 48 55 58 63 63 64 68 71 73 77 Table 3.1 4.1 LIST OF TABLES Number of Household-Month Observations Results of Non-Parametric Tests Comparing the Two-Interview Set with the One-Interview Subset Comparison of Variance of Estimates From the Two-Interview Set and the One-Interview Subset Comparison of Total Mean Expenditures Comparison of the Two-Interview Set with the One-Interview Independent Set Comparison of Variances of Estimates From the Two-Interview Set and the One-Interview Independent Set Comparison of the One-Interview Subset with the One-Interview Independent Set Comparison of Variance of Estimates From the One-Interview Subset and the One-Interview Independent Set Aggregated Commodity Groups Total Number of Two-Interview Household-Month Observations Results of Comparison of Mean Annual Estimates Distribution in Probability Results of Comparison of Total Mean Annual Conmodity Estimates Test Statistics for Equality of Means of the Four Individual Days of Recall From the First Interview Comparison of Mean Expenditures of Each Day of Recall Comparison of First Day of Recall with the Average of the Second, Third and Fourth Day of Recall iv 24 28 29 3O 32 33 34 34 38 4O 42 43 45 50 51 52 Table 6.4 6.5 6.6 Comparison of Mean Expenditures From the Second, Third and Fourth Day of Recall Results of Comparison of Annual Expenditure Estimates From the First and Second Interview Based on the First Day of Recall Results of Comparison of Annual Expenditure Estimates From the First and Second Interview Based on the Average of the Second and Third Day of Recall 54 57 59 Figure 3.1 3.2 LIST OF FIGURES Sierra Leone Rural Resource Regions Example Interview Schedule vi 16 20 CHAPTER 1 INTRODUCTION 1.1 Consumption Expenditure Survey Methodology in Low Income Countries Knowledge of consumption patterns derived from rural household expen- diture surveys is an important input into policy analysis and economic planning in many low income countries. Besides providing useful informa- tion on the general state of health and nutrition in rural areas, household budget surveys can help identify the trends in consumption expenditure patterns of different income groups, and the distribution of food within and among different groups. These surveys can also help identify potential consumption-based linkages with local small-scale industries. From these surveys it may be possible to estimate elasticities of demand for goods and services, information which is crucial in both short- and long-run econom- ic planning. In many low income nations the paucity of reliable information on rural consumer behavior represents a serious constraint on development planning. Lacking country specific consumer data many of these nations have been forced to use general income elasticities of demand provided by the FAO in order to project consumer demand for some types of commodities. The lack of information also thwarts the efforts of international agencies to develop and implement strategies designed to reach the rural poor. While the need for information on consumption patterns is clear, there is no consensus on the optimum survey methodology to obtain it. The numerous consumption expenditure surveys that have been conducted in developing countries reflect a wide range of objectives and methods. Exam- ples of some of these studies are found in Massell and Heyer (1967). Ikhtiar Ul Mulk (1966), Jamei (1966), Houyouk (1973), and King (1977). There are several reasons for the lack of consensus on methodology. Firstly, more is generally known about the interpretation of results than about the methodology used to obtain those results. Often methodological mistakes are buried, barring others from learning from them. Also, the purpose of the survey is seldom the investigation of methodological issues, and thus improvements in survey design are not field tested and evaluated systematically. This is understandable, though not desirable, given the high costs that complicated replications of different survey techniques under similar conditions would entail. In the profession's uncertainty over what is essential in the collec- tion of comprehensive rural consumption expenditure data. hi low income countries there has been a tendency to implement the frequent visit survey methodology. This survey methodology is based on an interview schedule that includes repeated visits to participating households during a month and extending over a relevant period such as one crop season or calendar .year. The advantage of the frequent visit methodology when compared with other survey types is that less reliance is placed on a respondent's ability to remember events. With frequent interviewing events are record- ed as they occur. It is hypothesized that this improves the quality of the data by reducing measurement error. Given the heterogeneity of popula- tions in rural areas of low income countries it is often believed that this methodology is essential in order to generate accurate expenditure esti- mates for different regions, income groups and seasons. The problem with this methodological approach is that it is generally costly and time-consuming. Its comprehensive nature generates higher costs in every phase of the data collection process. It generally requires significant administrative capacity to supervise the implementation of the survey and the interpretation of results. Usually, the sheer physical quantity of data collected cannot be absorbed and analyzed by local pro- cessing facilities and personnel. Often the SOphistication of the data obtained goes far beyond what Collinson (1979) describes as the "bread and butter" needs of the host government. There is an important trade-off to be considered between the reduc- tion of measurement error resulting from the intensive interview schedule and the increased costs of .obtaining that improvement in accuracy. Improvements in accuracy can always be achieved, but at a diminishing rate. At some point the marginal utility of an increase in accuracy is exceeded by ‘the Inarginal cost. of obtaining ‘it. This happens either because resources are limited or because the increase in accuracy is not necessary given the objectives of the study. The imperative need for knowledge of rural consumption patterns for planning purposes, and the lack of available resources and capital in many low income countries, make it essential that the most cost-efficient survey'methodology be adopted. Efforts must be made to develop a methodol- ogy which can quickly generate the kind of "bread and butter" information needed by governments with some minimmn criterion of reliability. It should also be compatible with the nation's human and physical capacity to collect, process and absorb information if it is to have an impact on the developmental process. It is important, therefore, that survey'methodolo- gies be developed which strike a balance between theory, necessity and cost. 1.2 Focus of the Study The purpose of this paper is to explore two issues relevant to the cost of collecting, processing and using information as well as its relia- bility. First is interview frequency, that is the number of times during a month a household is visited. The frequent visit methodology assumes that a more intensive interview schedule improves the reliability of the expen- diture estimates by reducing the measurement error in the sample. A more intensive interview frequency, however, requires a greater commitment of resources which are generally in scarce supply. The second is the reference period used in an interview. The refer- ence period is the length of time over which a respondent is requested to report purchases during one interview. ‘This period can range anywhere from a twenty-four hour recall to a month, three-month, six-month, or year recall. The reference period is extremely important because it influences both the measurement and sampling error in the survey. A central issue in determining its length is the ability of a respondent to remember purchases over time. It is presumed that memory decays over time and, therefore, a direct relationship exists between the length of the reference period and the degree of measurement error. An empirical assessment was made of these two issues using data col- lected in a 1974-75 comprehensive frequent visit micro-level study con- ducted in rural Sierra Leone. Parametric and non-parametric tests were used to examine the differences in mean expenditure estimates derived from one interview per month versus two interviews per month. This was done on a monthly and annual basis for both a very disaggregated list of commodi- ties and a consolidated list of corrmodity groups. The reference period used for an interview in the Sierra Leone study was four days. An assessment was made of the differences between the mean expenditure estimates derived from each of the four different days of recall obtained from one interview in order to determine if the problem of memory decay was more evident in a particular day of recall. Since the purpose of this paper is to explore methodological issues, an effort has been made to describe in detail the steps taken in conducting this analysis. Wherever appropriate, tables giving the statistical results are included to allow readers to assess the data for themselves. 1.3 Outline of Remaining Chapters In Chapter 2, the issues involved in determining interview frequency and reference period are discussed in greater depth. The concepts of measurement error and sample error are described and their relationship to interview frequency and reference period is explored. Chapter 3 of this paper describes the methodology used in the micro- level survey conducted in Sierra Leone, one component of which was the consumption expenditure study which provides the data base for this paper. Detailed information is given on sample selection, the household interview schedule, and the length of reference period. Also contained in this chapter is a description of the data preparation carried out for this analysis. Particular attention is given to describing the three catego- ries of interview frequency used in this analysis. The procedures and results of non-parametric tests performed on 257 disaggregated commodity groups using monthly expenditure estimates are presented in Chapter 4. This approach compared three different data sets representing expenditure estimates based on one and two interviews per month. This will be followed in Chapter 5 by a description of the procedures and results obtained when using the correlated t-test to determine whether the differences between annual commodity expenditure estimates based on two interviews are significantly different from those based on one inter- view per month. For this analysis, 16 conmodity groups were created representing food items, beverages and some frequently purchased items. The four days of recall obtained during one interview are examined individually in Chapter 6. An analysis of the differences in expenditure estimates generated by the four different days of recall is tested using Hotelling's T2 test. In this chapter a comparison is made of first and second interview expenditure estimates derived from particular days of recall. Finally, Chapter 7 provides a summary of the research findings and the conclusions of this analysis. CHAPTER 2 THE RESEARCH PROBLEM 2.1 Factors in Survey Design There are numerous methodological factors involved in survey design which contribute to the cost per unit of information and data turn-around time. Some of the factors relevant to survey design include sample size, sample selection procedure, collection technique (e.g., interview, ques- tionnaire, group interview), and duration of survey. Critical to the choices made are the objectives of the intended research. The survey design implemented should generate the type of information and level of accuracy needed to test the desired hypotheses. The attempt should be made, therefore, to minimize the relevant threats to validity, which vary depending on the objectives of the study. While many of the factors mentioned above represent important and sometimes controversial issues in survey design, they are beyond the scope of this paper. It is recognized, however, that there is a great deal of interdependence between the decisions made with respect to interview fre- quency and reference period and other variables involved in survey design. The trade-offs between these variables should be given serious considera- tion in designing a survey methodology. Central to the issues of interview frequency and reference period are the concepts of sample and measurement error. The validity of the infer- ences drawn from the data depends to a great extent on the degree to which these two types of errors exist in the data. Boruch (1972) defines measurement or response error as the difference between the recorded response to the inquiry and a potentially measurable, true condition associated with that inquiry. Sources of measurement error in survey questionnaires are identified as faulty recall, a deliberate or accidental distortion of responses, structural weakness or ambiguity in the item, lapses in the quality of data reporting, and errors in proceSsing and maintaining the data. Moser and Kalton (1972) also identify interviewer bias as a source of measurement error. Another source of measurement error arises when the panel method is used in survey design. The panel method, which was incorporated into the design of the Sierra Leone study, specifies the collection of data from the same sample on more than one occasion. Two specific problems arise when using this method which Moser and Kalton (1972) identify as sample mortal- ity and conditioning. The former occurs when over the course of the survey participants drop out, move or die. Sample mortality does not necessarily result in biased results if the exit of participants is random. Problems, however, could arise if the participants' discontinued participation could be correlated with particular characteristics such as income, education, ethnic group and/or religion. The other problem associated with the panel method, also discussed by Neter and Waksberg (1964), is conditioning. With repeated visits to par- ticular households there is a risk that they will in some way become untypical. If this happens the panel or sample of households may become, as Moser and Kalton (1972) point out "...untypical--not in composition but in its characteristics--of the population it was selected to represent." This may affect the accuracy of the expenditure records obtained from these households. The repeated visits can sensitize the participants, making them more aware of their expenditures, thereby improving the expenditure records. Alternatively, repeated visits to households can result in respondent fatigue which can cause a decrease in the accuracy of expendi- ture records. Measurement error is a critical factor in data reliability. Its presence can introduce significant bias in expenditure estimates. This is especially serious if the bias introduced is large and in an unknown direction. The problem is made more difficult because there is no method for statistically measuring the extent or direction of the bias from the data themselves. The other type of error influencing sample reliability is the class of errors described as sample errors. As Moser and Kalton (1972) describe it, sample errors lead to fluctuations of the sample or population estimates around their true or expected values. The standard error is the measure of this fluctuation. Two factors influencing the degree of sample error present are sample size and the variability in the population. The smaller the sample size and/or the greater the variance in population characteris- tics the greater the standard error; What this intuitively implies is that a wide variation in population characteristics makes the estimation of the population mean from one sample less reliable. The size of the standard error also influences the ability to use certain types of statistical tests. A large standard error widens the confidence intervals within which the population's expected value is found. Conversely a smaller standard error tightens these boundaries, improving the reliability of statistical tests. There are several factors involved in determining the interview fre- quency and reference period. The trade-off between increased accuracy and 10 increased cost has been discussed earlier. The possibility of obtaining more accurate data always exists. However, at some point diminishing returns to accuracy set in. Thus, it is important to be able to judge when the marginal cost of improved accuracy is greater than its marginal utility for the particular objectives of the survey. Determining this point helps 'to determine the survey's tolerance to measurement and sample error. 2.2 Factors in Determining Interview Frequency A trade-off between the two types of errors is inherent in the choice of frequency of interview. A large sample size results in a smaller standard error; A large sample size and/or an intensive interview schedule result, in general, in a smaller standard error. However, the costs of collecting data from a large sample or from repeated visits to households can be quite high. The implementation of surveys reflecting these two types of design generally requires significant administrative and super— visory capacity. It is also necessary to have the facility to handle and process the extensive amount of data being collected. If these capabili- ties are not available, significant measurement error could be introduced into the data. A balance must be struck between sample error and bias. Considerations important in assessing this relationship are the extent of variations in household expenditures due to income, household size, and cultural or regional preferences. Rey (1976) suggests that an important concern in determining inter- view frequencies is that they cover the span of time during which consump- tion expenditures follow a certain pattern. They should include at least one buying cycle for each interval into which the year is divided. Know- ledge of the population characteristics, and production and marketing cycles will give the first indication of what the necessary frequency 11 pattern might be. It is essential that the influence of marketing cycles on household expenditures not be overlooked given the dominance of period- ic markets in many low income countries. Also, in many low income coun- tries where the majority of the population is involved in subsistence agricultural production, seasons will have great impact on expenditure patterns. It is essential, therefore, that the influence of seasons be accounted for in intramonth interview scheduling. Other factors involved in determining interview frequency are the availability of administrative capacity and trained personnel to partici- pate in the study. Poorly trained and/or supervised enumerators can intro- duce significant bias in the data collection process which could threaten the validity of the results. An increase in interview frequency per household also puts a greater strain on respondents. This could possibly generate fatigue on the part of respondents and the potential for decreas- ing reliability in response. Non-response on the part of participating households due to absenteeism requires callbacks which can be costly both in terms of travel expenses and enumerator's time. Supervision of data collection and processing procedures in multi-visit surveys can also be demanding of scarce administrative capacity. 2.3 Factors in Determining the Length of Interview Reference Period Directly related to the intermonth interview schedule is the length of the reference period chosen. Choices concerning the length of reference . period reflect trade-offs between accuracy and cost, and sample and mea- surement error, similar to those involved in determining the interview frequency. A longer reference period per interview reduces the cost per unit of information. This is because a long interview reference period 12 permits the collection of more data points during the one interview at little extra cost. Alternatively, this information could be obtained in separate interviews but the costs would be significantly higher. At the same time, however, a long reference period increases the possibility of response error due to memory decay which threatens the reliability of the data. Thus, in this case there is a trade-off between decreasing the marginal cost by lengthening the reference period and reducing the margin- al utility of the data by introducing significant measurement error. The reference period chosen also influences the size of the standard error. A longer reference period decreases the sampling error in that more data points are collected which capture more of the variation in a population's expenditures thereby reducing the standard error. However, as mentioned previously, memory decay which increases over time can introduce a poten-3 tially significant bias in expenditure estimates. A decision must thus be made as to the point at which the benefits brought about by the reduction in standard error are swamped by the increase in measurement error due to memory loss. Moser and Kalton (1972) identify two primary factors which influence a respondent's ability to remember expenditures. The first is the length of time since the event took place. There is a greater probability of forgetting a purchase the longer the period for which it must be remem- bered. The importance of the purchase to the respondent is the second factor which influences how well the expenditure is remembered. The less significant the item the easier it is to forget. To avoid this type of bias some studies have used reference periods of different lengths depend- ing on the type of purchase (Hussain, 1966; King, 1977). A shorter refer- ence period was used for items with a shorter recall, i.e., those items 13 frequently purchased and less significant to the respondent. A longer reference period was used to collect information on those items which are purchased less frequently but are major or more significant purchases. Twounajor issues in determining the length of the reference period are identified in the literature (Neter, 1965; Moser and Kalton, 1972; Prais and Houthakker, 1971). One concern is what is referred to by Prais and Houthakker (1971) as recall loss. This has been described in the preceding paragraphs and refers to the respondent's failure to report an activity because of memory failure. Neter. notes that the probability of this occurring increases as time passes and is a more important influence on the ability to recall frequent and less significant purchases. The second issue is the end period or telescoping effect. This describes the tendency to include expenditures incurred just before the beginning of the inquiry. The telescoping effect is believed to have greater influence (Ml the reporting of exceptional expenditures such as those made on major durables (Prais and Houthakker, 1971). There is also some evidence to suggest that there is a greater general telesc0ping effect for shorter reference periods. This has been suggested as a potential explanation for the relatively higher expenditure levels associated with short recall periods commonly found in survey results (Moser and Kalton, 1972). Another factor which can influence the magnitude of the telesc0ping effect is whether the recall period is bounded or unbounded. Unbounded recall occurs when respondents are asked to report expenditures made since a given date but where no control is exercised over the possibility that expenditures from the previous period are repeated. Bounded recall tech- niques attempt to reduce the telescoping effect through repetition of past 14 purchases to prevent duplication in subsequent interviews (Moser and Kalton, 1972). Empirical tests have been conducted to analyze the influence of tele- sc0ping using bounded-and unbounded recall periods. In a study done by Neter and Waksberg (1964) it was found that expenditure estimates derived from a one-month unbounded recall period were significantly higher than the expenditure estimates obtained from a bounded one-month recall period. The issues discussed in the preceding paragraphs must be considered when determining the interview frequency and reference period used in a particular study. The accuracy of the data and the cost per unit of information are heavily influenced by these decisions. Unfortunately very little is known about the magnitude of the trade-offs involved in choosing among the alternative frequency and recall patterns. While theory and common sense suggest these factors have significant influence on reducing measurement errors, there is little existing empirical evidence to tell us either how much or at what cost the improved accuracy is obtained. CHAPTER 3 SURVEY METHODOLOGY USED IN SIERRA LEONE RURAL CONSUMPTION EXPENDITURE SURVEY 3.1 Sample Selection The data used in this analysis were collected in a comprehensive rural household budget survey conducted in Sierra Leone from March 1974 through May 1975. A frequent visit or cost route survey methodology was used to collect 14 months of cross-sectional data covering a wide spectrum of rural activities. The integrated survey was designed primarily to collect micro-level information on farm production and non-farm activities for an entire cr0p year. A secondary objective of the survey was the collection of data on migration and consumption expenditures. The following descrip- tion of the Sierra Leone study relies heavily on the information provided in Spencer, et al. (1976); Spencer and Byerlee (1977); King (1977); and Rural Employment Research Project (1974). In the Sierra Leone survey the enumeration areas as well as the participating households were selected using a stratified sampling proce- dure. Based on available secondary data Sierra Leone was divided into eight resource regions reflecting different physical and climatic factors. Each of the eight resource regions shown in Figure 3.1 was subdivided into enumeration areas. Each of these areas was approximately ten squarelniles. Roughly 130 farm families located in one to ten villages were contained in each enumeration area. As the purpose of the survey was to obtain information on rural households, enumeration areas falling into or containing urban areas were 15 If I? "a l I 1 ‘L r-~_----.“ o _‘ ,.—.‘ ’ ’.l ‘ ." - ‘ I! . M .. \"I c f \ I F: \\ “ onu- , [I 1 7 11— z s ‘ c0 If ‘I \"-“~ A, \\ 4 hn-h : ., \ \. .' . .- ‘ oar—u /. f,’ (I , .~/‘ 1..- _ Ila-Ma . a. . ’-'. hail... ------x. A: (_ r4“ 1.- f “‘2‘ /° .1 K ' ('V” I .1 ~- - —I.C 2am“; /' . i /' ‘ . - nonal- . j l \- "n L.“ (#4133... 13/.) I ‘.‘ ”In” I-“ 'l ”at? ) «on: I) \u g .l"._ 1.\ . : I" r- ram 3 \\mm~z~ m4¢zcmucm H N m e zmw>cmucH acoumm zoo ppaumm H N m c 3mp>cmucH “mew; moo Ppmumm NH 0H ma «H mfi NH HH mpoo heaved chmczgp amomwcumz anemone amuse: xmuczm mmnczumm xmmz do zoo 21 commodity expenditure estimates based on the more intensive two-interview- per-month data set to be compared with expenditure estimates obtained from the one—interview subset. The fact that the households included in each sample are identical reduces the possibility that factors other than the experimental variable of interview frequency are responsible for any observed variation in expenditure estimates between the two sets.* The C-2 questionnaire asked respondents to report purchases made on durable and less frequently purchased goods. This questionnaire was administered once a month, theoretically at the end of the month and had a reference period of one month. Checks were made in the data processing to ensure that purchases reported on one form were not also included on the other. Both questionnaires allowed respondents to report purchases on a highly disaggregated set of commodities (see Appendix A). Very specific information was requested on each purchase. The type and/or brand, if known, of each item was recorded. The total expenditure on each item was recorded in Leonian cents. Special codes were used to reflect the specific unit measurement of the item and the quantity of units purchased. Detailed information was collected on where the item was purchased, e.g., in the village market, a store, from a trader. Where possible names were obtained. The last category of information collected on each expenditure was the origin of the item, or where it was produced. Respondents could choose between four general categories: 1) rural areas (population less *1" this part of the analysis only information on expenditures obtained from the C-1 or short reference period questionnaire is being included. The C-2 or long reference questionnaire administered once a month would not be relevant in an intermonth comparison of different inter- view or recall patterns. 22 than 2,000); 2) large urban areas (population greater than 100,000); 3) small urban areas (population greater than 2,000 but less than 100,000); and 4) imported. On the C-1 or short reference questionnaire this information was recorded for each purchase made during the four-day reference period. The C-2 questionnaire recorded all this information for purchases made during an entire month. 3.3 Description of Interview Categories Problems with the data were encountered in attempting to test our research hypotheses. While each household was to have been interviewed twice obtaining seven days of information per month, this was not always the case. Households were often over- or under-interviewed. As a result, complete monthly data for some households were missing. This is not the same thing as reporting that no expenditures were made. The latter was considered a valid expression of an expenditure pattern. What is being referred to here is that for some reason a household was not interviewed during a given month and, therefore, has zero days of information. At the other extreme some households had information for more than seven days per month. Presumably there are numerous reasons for the wide variation in the amount of monthly data collected for each household. A household might have an inconsistent interview pattern because the family moved during the survey period, deaths, and/or absenteeism at the time of interview. Alter- natively, enumerators could miss the first, second or even both interviews in a particular month for any number of reasons. Incomplete information could be collected during an interview. Over-interviewing a particular household could reflect an attempt to compensate for other missed 23 households. Finally, some of the missing data might be explained by coding and processing errors. In order to test the research hypotheses we had to identify for each household those months for which there were at least seven days of informa- tion recorded. A household could have more than seven days of information in a given month but only seven were used for purposes of analysis. Fur- ther for a seven-day set of information to be included in our sample the following had to hold: 1) the seven days had to represent two interviews; 2) the days had to be seven consecutive calendar days; 3) the sequence of the recall pattern had to be 4-3-2-1-3-2-1 or, though rarely observed, 3-2- 1-4-3-2-1. After identifying and making a separate computer tape which consisted of only those months for which a household had seven days of information, there remained a quantity of household month observations for which there were four days or more of information but less than seven. If in this residual data there existed information for a particular household on four consecutive calendar days, with a recall pattern of 4-3-2-1, for a month for which a seven-day record did not exist, then it was included on a tape containing four-day or, in this paper's terminology, the one-interview independent sets. If, for a particular month, a household had both a seven-day set and four-day independent set, priority was always given to including the seven-day set. If the data collection process overlapped two months, the overlap data set was assigned arbitrarily with the guiding principle being to include as many seven-day sets as possible. Details of this procedure are given in Appendix B. Table 3.1 shows the number of household observations contained in each month. Estimates for the two- interview set and one-interview independent set are given separately. 24 TABLE 3.1 NUMBER OF HOUSEHOLD-MONTH OBSERVATIONS Month and Two Interviews One Interview Year Per Month Per Month Total May 1974 88 32 120 June 1974 118 33 151 July 1974 142 32 174 August 1974 167 30 197 September 1974 152 44 196 October 1974 136 57 193 November 1974 160 42 202 December 1974 156 38 194 January 1975 146 45 191 February 1975 120 36 156 March 1975 159 37 196 April 1975 149 33 182 In order to make the seven-day and four-day expenditures representa- tive for the same period of time, they were expanded to reflect one month's purchases. This was accomplished by multiplying each estimated expendi- ture by the number of days in the month divided by the number of days of information. The details of this procedure can be found in Appendix C. CHAPTER 4 NON-PARAMETRIC ANALYSIS OF THE INFLUENCE OF INTERVIEW FREQUENCY ON EXPENDITURE ESTIMATES 4.1 Non-Parametric Tests and Their Application to This Analysis Several approaches were used to examine the influence of interview frequency on expenditure estimates. In order to compare the data in its most disaggregated form non-parametric tests were used. This statistical procedure allowed the comparison of each of the original 257 commodities listed in the C-1 questionnaire on a monthly basis. Using this highly disaggregated list of commodities parametric tests could not be used because of their restrictive assumption that the popula- tion sample have a normal distribution. The assumption of normalcy is clearly not the case when dealing with expenditures data where purchases of zero represent a large proportion of the observations for a particular conmodity. The zero observations cannot be eliminated as they are not a reflection of non-reSponse, but rather non-purchase. The latter is a valid expression of a household's demand and should not be automatically exclud- ed from the sample. In light of the inability to assume a normal distribution in monthly commodity estimates, non-parametric tests, which do not depend on assump- tions concerning the form of the underlying distribution, were used. Non- parametric methods provide statistical tests in which no hypotheses are made about specific values of parameters. These methods are useful in many situations where ordinal data are being examined. Ihi this analysis the 25 26 non-parametric sign test was employed. This test is based on the signs generated by the differences between pairs of observations. It uses plus or minus signs as data rather than qualitative measures. Thus it does not take into consideration magnitudes of the differences between the paired observations. The non-parametric sign test is particularly useful when dealing with two samples that are not independent. To conduct the sign test mean monthly expenditure estimates and vari- ances were calculated for each of the 257 commodities and services (see Appendix A for listing of these) using data obtained from the two—interview set, the one-interview subset, and the one-interview independent set. The differences between the means of these three samples were calculated using paired data. The number of times that the difference was greater than or less than zero was counted. Similarly, a ratio of variances was construct- ed for each pair. The number of times the ratio was greater than or less than one was counted.* Assuming for the moment that the three samples were drawn randomly from the same population it would be expected that their estimated mean expenditures would be equal. In comparing any pair of monthly expenditure estimates there would presumably be a 50-50 chance that one sample's expen- diture estimate would be larger than the other sample's estimate. Thus, *A non-parametric comparison took place only in those cases where the two-interview set contained some positive observation for a particular commodity. This restriction was implemented because of the number of zero observations. In any given month there were a number of commodities which were not purchased by any household. In this case, expenditure estimates based on either interview frequency would have means and variances of zero. These were, therefore, not calculated. Given the way these data were prepared for analysis, if the mean derived from two interviews per month equaled zero, then by definition the means of the one-interview subset equaled zero. Basing our decision rule on the value of the two-interview set seemed to be the most efficient way of handling this problem. 27 the probability on any comparison of means between two samples is p = .5 that one would be larger than the other and vice versa. If the sample size is large, the binomial probability distribution approaches the normal dis- tribution, permitting the computation of test statistics with which to test the research hypothesis. The research hypothesis tested here was that there was no difference in the probability distribution of the means and variances when comparing the two-interview set with the one-interview subset, the two-interview set with the one-interview independent set, and the one-interview subset with the onelinterview independent set. Put in another way, the hypothesis tested was that the probability of one sample's commodity'mean and variance being larger than the other sample's equaled p = .5. 4.2 Comparison of the Two-Interview Set with the One-Interview Subset The first testing of the research hypothesis compared the two- interview set with the one-interview subset. As shown in Table 4.1, the means from the two-interview set were larger in 509 instances while the opposite was true in 617 cases. In computing the standardized binomial variable a Z value of -3.22 was obtained. This statistic has a two-tailed significance level of .0014. Thus, at the .05 level of significance the research hypothesis of no difference between the means cannot be accepted on the basis of these sets of data. The inability to accept the null hypothesis based on this outcome suggests that the frequency of interview does influence expenditure esti- mates, at least in statistical terms. In practical terms, however, the numbers are not extremely dissimilar. They indicate that 5/11 of the time XTjk > XSjk and that 6/11 of the time the opposite is true. This suggests 28 TABLE 4.1 RESULTS OF NON-PARAMETRIC TESTS COMPARING THE TWO-INTERVIEW SET WITH THE ONE-INTERVIEW SUBSET Ho: p = .5 where p = probability that (R . >"R . ) Ha: p f .5 TJk SJk where: x t“ _T'k = mean monthly expenditure on the j commodity (1,...,257) 3 in the k month (1,...,14) based on two interviews per month. iS'k mean montnjy expenditure on the jth commodity (1,...,257) J in the k month (1,...,14)_based on one interview per month which is a subset of ijk' n = 1126 From the estimates for ijk and iSjk the following were calculated: ijk - XSjk:> O in 509 cases and XTjk - XSjk‘< O in 617 cases. These are standard binomial random variables with a standardized normal distribution = N(O,1). z = 509-.5(l126) = _3.22 7l126(.5)(l-.5) 29 that there is on average a tendency for expenditure estimates based on one interview to be larger than the expenditure estimates based on two inter- views per month. As might be expected in the analysis of variances using the non- parametric sign test, the variances of the two-interview expenditure esti- mates were smaller than those of the one-interview subset. As shown in Table 4.2, the variances of the two-interview set were smaller than the variances of the one-interview independent set in 407 cases; the opposite was true in 721 cases. This occurs because in interviewing twice a month expenditure variations are averaged out over a greater number of days. This results in a smaller variance. TABLE 4.2 COMPARISON OF VARIANCE OF ESTIMATES FROM THE TWO-INTERVIEW SET AND THE ONE-INTERVIEW SUBSET OZT’k variance of monthly1§xpenditure estimate for the jth commodity J (1,...,257) in the k month (1,...,14) based on data collected in two interviews per month. OZS'k ; variance of monthly Efipenditure estimate for the jth commodity J (1,...,257) in the k month (1,...,14) based on one interview per month which is a subset of the two-interview set. n = 1128 In calculating the ratio of variances, it was observed that: 2 O 2 O - ° .2115. > 1 in 407 cases, while gzl%£'< 1 in 721 C6565- 0 Sjk SJk 30 4.2.1 Comparison of Total Mean Expenditures for All Commodities While these non-parametric tests indicate that there is a tendency for the one-interview subset expenditure estimates to be greater than estimates based on two interviews, the figures do not tell what the magni- tude of this difference is. To obtain some rough indication of this magnitude, all available mean monthly expenditure estimates were totaled using both the two-interview set and one-interview subset. The research hypothesis that the two means were equal was tested. 'The hypothesis tested and the derivation of these figures is shown in Table 4.3. TABLE 4.3 COMPARISON OF TOTAL MEAN EXPENDITURES Ho: X. = X -TE ~SE ”3 XTE I XSE where: RTE = total mean expenditure for all commodities for all months based on two interviews per month. 755 = total mean expenditure for all commodities for all months based on the one-interview subset. n = 1126 and where: ._ 257 lz4 (Leones) X = /n - .25 TE j- _1 k- fiijk 257 lz4 '7 = In - .27 SE j= -l k=l leSjk j = commodity (1,...,257) k = month (1,...,14) t = 25' 27 = -3.135 [[OTE+SE o2 --2((1)V)]:,I 31 The total mean expenditure estimate for the two-interview data set for fourteen months of information is 25.095. The total mean expenditure estimate for the one-interview subset is 26.920. Using the correlated T- test procedure to test the difference between the two means, the test statistic derived was -3.135. From a statistical point of view the differ- ence between these two means is significant at the .05 level. Therefore, the research hypothesis that the total mean expenditure estimate based on two interviews per month is equal to the>mean expenditure estimate obtained from a one-interview subset cannot be accepted. These figures support the results obtained earlier that the expenditure estimates based on one interview have a tendency to be slightly larger than those based on two interviews per month. Again, while these figures are different from a statistical point of view, they are in practical terms very similar. The one-interview subset estimate is only 7 percent larger than the expenditure estimate generated by the two-interview set. Depending on the purpose of the survey, and the level of accuracy needed, these differences could be viewed as very slight. If so, the additional cost of a second monthly interview might not be deemed necessary. 4.3 Comparison of the Two-Interview Set and the One-Interview Subset with the One-Interview Independent Set The same research hypothesis of no difference in the probability distribution of the means and variances of the paired data was tested by comparing the two-interview set and the one-interview subset with the one- interview independent set. The results present an interesting contrast to those obtained from the first tests. Mean monthly expenditure estimates based on the two-interview set are larger than those derived from the 32 one-interview independent set in 973 cases. The reverse situation prevails in only 425 cases. TABLE 4.4 COMPARISON OF THE TWO-INTERVIEW SET WITH THE ONE-INTERVIEW INDEPENDENT SET Ho: p = .5 where p = probability that (-'. >'7 . ) Ha: p f .5 . Tjk Ijk where: iT‘k = mean monthly expeqfljture on the' jth commodity J (1,...,257) in the k month (1,...,14) based on two interviews per month. YI'k = mean monthly expe iture on the jth commodity J (1,...,257) in the k month (1,...,14) based on the one-interview independent set. n = 1398 From the estimates for iTjk and iSjk the following were calculated: XTjk - XSjk > O in 973 cases and ~ Tjk - ijk < 0 in 425 cases. These are standard binomial random variables with a standardized normal distribution = N(O,1). _ _973-.5(l398) _ Z ‘ 398 .5 -.5 ' 14-556 These results are the reverse of those obtained in the previous test comparing the two-interview set with the one-interview subset. In that test the one—interview means tended on average to be larger than the two- interview means. In this test not only are the means of the two-interview set larger on average than the one-interview independent set but the fre- quency with which one is larger than the other is much greater as evidenced by the larger Z statistic of 14.655. 33 The variances of the two-interview set estimates also are consistent- ly higher than those for the one-interview independent set as shown in Table 4.5. TABLE 4.5 COMPARISON OF VARIANCES OF ESTIMATES FROM THE TWO-INTERVIEW SET AND THE ONE-INTERVIEW INDEPENDENT SET Oszk = variance of monthly exp‘enditure estimate for the jth corrmodity (1,...,257) in the k month (1,...,14) based on the two- interview set. UZIjk - variance of monthly expenditure estimate for the jth commodity (1,...,257) in the k month (1,...,14) based on the one- interview independent set. n = 1398 In calculating the ratio of variances, it was observed that: 02 2 O . ' 32115., I in 998 cases, while Ezflf < 1 in 401 cases. Ijk j In comparing the one-interview subset with the one-interview indepen- dent set, similar results are obtained. As shown in Table 4.6, the mean expenditure estimates generated by the one-interview subset are higher than the one-interview independent set in 767 cases. The opposite occurs 429 times. This difference has a Z value of 9.774 using the normal approximation. The variances for the one-interview subset are higher than those of the four-day independent set by a margin of 794 to 403. 34 TABLE 4.6 COMPARISON OF THE ONE-INTERVIEW SUBSET WITH THE ONE-INTERVIEW INDEPENDENT SET Ho: Ha: DU where: ‘I'kll X Sjk Ijk n 3 where p = probability that (ijk> CXIjk) mean expenditure for the jth commodity (1,...,257) for the k month (1,...,14) based on the one-interview subset. mean expenditure for the jth commodity (1,...,257) for the k month (1,...,14) based on the one-interview independent set. 1198 From the estimates for?Sjk and?Ijk the following were calculated: X Sjk’x -Y Sjk Ijk Ijk > O in 767 cases and < O in 429 cases. Z = 767-:5tll96) = 9.774 TABLE 4.7 COMPARISON OF VARIANCE OF ESTIMATES FROM THE ONE-INTERVIEW SUBSET AND THE ONE-INTERVIEW INDEPENDENT SET 2 ° Sjk 2 ° Ijk n = variance (1,...,2 of expenditure estimates for the jth commodity 57) in the k month (1,...,14) based on the one- interview subset. = variance (1,...,2 th commodity of expenditure estimates for the j k month (1,...,14) based on the one- 57) in the interview independent set. = 1197 In calculating the ratio of variances, it was observed that: 0 Sjk 0’2 . Ijk > 1 in 794 ° Sjk . cases, while EIIER' < l in 403 cases. 35 The results presented in Tables 4.4-4.7 present a potentially impor- tant contrast. In the first test of the research hypothesis comparing the two-interview set with the one-interview subset the only difference between the two samples was frequency of interview. Since the one- interview subset was taken from the two-interview data set, the households contained in each sample were the same. This significantly reduced the possibility of other factors such as income, household size, and education having any influence on the results. Thus, to the extent possible the impact of interview frequency on expenditure estimates at the monthly level was isolated. The data suggested that the isolated effect of the difference in interview frequency was for one-interview mean expenditures to be on average somewhat larger than those based on two interviews per month. In contrast, when comparing the one-interview independent set with the two-interview set and its subset, the expenditure estimates of the former were smaller than those of the other two sets. The inaccessibility of information on the characteristics of the households contained in the two sets prohibits a conclusive explanation of these observed differences. However, several hypotheses can be offered to explain these results. The first deals with an issue concerning the internal validity of the study. One could hypothesize that the households visited in the specified manner (two interviews in a month) went through a conditioning process such as that discussed briefly in Chapter 2. Because these households were visited consistently during the survey period they became more sensitive to the survey process. Thus, they had a greater tendency to remember more accurately the purchases made during subsequent recall periods. Households visited inconsistently and not in the speci- fied manner might report fewer expenditures because they had been 36 interviewed infrequently and were not necessarily anticipating further interviews. Another hypothesis with far more serious implications is that the two samples were not drawn randomly'from the same population. This would imply that the two samples reflect different population characteristics. This might occur for two reasons. One deals with the respondent's willingness to participate or the sample's morbidity rate while the other deals with an enumerator's interviewing techniques. In the former case a respondent's willingness or unwillingness to participate in a survey might be reflected in whether or not the household was interviewed in the correct manner. A household's receptiveness to the survey, their availability during inter- view sessions, and general interest in the survey could influence the number of times per month and year the household was visited by enumera- tors. What can cause serious problems in the reliability of the data is if this difference in receptivity is not random but based on specific popula- tion characteristics such as income, education, type of employment, or ethnic group. In survey design this is known as the problem of self- selection. This same type of difference in population characteristics mentioned above could also influence the number of times an enumerator visited a particular household. Enumerators could be less rigorous in their attempts to interview households of a particular ethnic group, income bracket, or level of education. This could explain the results obtained when comparing the two- interview and one-interview subset with the one-interview independent set. The latter might reflect a greater pr0portion of households with a lower income, more removed from urban areas and thus less involved in a market 37 economy and/or more difficult to travel to. If this were the case, the lower means might reflect fewer purchases, a smaller variety in purchases and/or less total income spent on commodity purchases. This would also explain why the variance of the one-interview independent set is charac— teristically smaller than those of either the two-interview or one- interview subset. If this hypothesis is valid, then a potentially significant distor- tion has been introduced into the data. Failure to obtain data from this genre of households could result in biased expenditure estimates and eco- nomic policies which might have undesired consequences. Assuming for the moment that this hypothesis is true the results reveal how essential well-trained enumerators and adequate field supervi- sion are in the collection of reliable data. If the complexity of the survey design goes beyond the capacities of enumerators and administra- tors, then serious problems might arise. CHAPTER 5 ANALYSIS OF AGGREGATED COMMODITY GROUPS 5.1 Data Preparation The analysis using non-parametric tests compared mean monthly expen- diture estimates of different interview frequency for a highly disaggre- gated set of commodities. However, for many research and planning purposes annual commodity expenditure estimates are required. These estimates are essential in deriving elasticities of demand and in the formulation of eco- nomic policy. In order to compare the annual expenditure estimates derived from the two-interview set and the one-interview subset the original commodity list was aggregated into 16 groups. An attempt was made to aggregate individual commodities with sensitivity to the demand, origin and nutritional charac- teristics of that item. This particular aggregation, shown in Table 5.1, TABLE 5.1 AGGREGATED COMMODITY GROUPS 1. Rice 9. Sugar 2. Grains 10. Fresh Fish 3. Cassava and Other Root Creps 11. Dried Fish 4. Vegetables, Leguminous 12. Bakery Items Products and Fruit 13. Other Processed Foods 5. Groundnuts 14. Alcoholic and Non-Alcoholic 6. Palm and Other Oils Beverages 7. Meat and Other Livestock 15. Tobacco and Kola Nuts 8. Salt and Other Condiments 16. Fuel and Light 38 39 contains all the possible food items listed on the original survey code along with all beverages, tobacco and kola nuts, and fuel and light. All other types of durables, home, and personal goods were excluded. For the most part these purchases are recorded on the C-2 questionnaire. As mentioned in Chapter 2, this questionnaire had a reference period of one month and was used to collect information on durables and other less frequently purchased goods. Since this analysis involved making compari- sons between expenditure estimates based on one and two interviews per month, the C-2 questionnaire was not relevant. The research hypothesis to be tested was that the annual mean expendi- ture estimates based on two interviews per month were equal to those based on the one-interview subset. The alternative hypothesis was that the means were not equal. In estimating annual mean commodity expenditures based on this data several issues were encountered. The first matter of concern was the households to be included in the sample. As discussed earlier in Chapter 2, very few households were interviewed for all 12 months. Table 5.2 shows how many households have data based on two interviews per month and for how many months they have it. The cumulative frequency is also given. What this table shows is that only three households included in the survey have 12 complete months of data. Eleven households have 11 months of data making the cumulative frequency of households with greater than 11 months of data equal to 14. The least restrictive criterion, that a household have at least one month of data generates a cumulative frequency of 247 households. 40 TABLE 5.2 TOTAL NUMBER OF TWO-INTERVIEW HOUSEHOLD-MONTH OBSERVATIONS No. of Months for No. of Households in Cumulative Which Household Has Data Two-Interview Sample Frequency Based on Two Interviews with X Months of Data 12 months 3 3 11 11 14 10 24 38 9 30 68 8 36 104 7 42 146 6 26 172 5 25 197 4 26 223 3 15 238 2 244 1 3 247 The number of months for which a household possesses valid data is an important concern in this analysis because of the lack of independence between the two samples.. It cannot be assumed that purchases made and recorded in the second interview are independent from the purchases made in the first interview. Nor, for that matter are purchases made in January independent of expenditures made in December or February. This lack of independence between samples can be corrected for through the use of the correlated t-test. Unlike the»more common Student's t-test, the correlated t-test does not assume that the two samples share a common variance. Nor does the correlated t-test assume that the covariance between the two samples is zero. In using the correlated t-test the variance of each sample is computed individually and then the covariance 41 between the two samples is computed and subtracted out of the denominator. This removes any double-counting in the pooled variance arising from the non-independence of the samples. Analyzing the difference in mean annual commodity expenditure esti- mates with the correlated t-test requires using households with 12 months of data. This is necessary in order to compute the individual variances of each sample from which the covariance between the two samples can be calculated. As Table 5.2 indicates, few households have 12 months of data. In order to overcome this problem,lnonthly indices for the 16 commodity groups were computed using the procedure described in Appendix 0. Separate monthly indices were calculated for both the two-interview set and the one- interview subset. Missing expenditure information was imputed for only those households that had eight months or more of data. Households with less than eight months were excluded from the sample. Taking households with eight or more months of data generated a sample of 104 households and held the maximum number of months to be imputed for any given household to only one-third of the total. 5.2 Comparison of Mean Expenditure Estimates These indexed data were then used to test the research hypothesis that the means of the two samples were equal. This hypothesis was tested for each of the 16 commodity groups using the correlated t-test. The alterna- tive hypothesis was that the means were not equal. Table 5.3 summarizes the results of this analysis. For 14 out of the 16 commodity groups the difference between the means proved insignificant at the .05 level. Rice and Palm and Other Oils were the two commodity groups where the difference between the Ineans was determined to be 42 _a>m_ mo. as» am p=a8_escmsm u m Peso. mo. men “a s=a8_c_=mwm so: u m2. mz 054. m“. Po.m_ mo.o_ B;m_s new szd .0. mz mam. No. - mm.m_ Nm.mF mpaz apex ecu cadence .m. mz mph. hm. - mfi.m mo.m mamacm>mm ._< .4, mz moo. NG.P- mm._ mo.. coed gammauoca gasps .mp mz 0mm. km. om.N em.N mEBBH acmxam .N. mz map. mm..- ms.om N¢.¢m gang ua_ca ._P mz omm. om. om.~ ¢P.m saw; gamed .op mz eme. a“. NN.N N¢.N camam .m mz ¢m_. em.F- mo._P mm.op mpcmewueou Lasso ace open .w mz Fem. .o. Nm.m mm.m muuauoca cucumm>23 cmguo use new: .5 m coo. em.m om.mN mm.cm m_.o Logno ecu 2.”; .m mz Noe. «N. - mm. mm. muzcuczocc .m mz amp. mm..- No.m .p.m 3.32; new meamm .ma_aapmmm> .e mz omm. mm. - NN.m mF.m maocu boom cmggo ecu a>ammau .m mz mpm. Po.P- mm.N mN.P menace Lasso .N m Noe. mp.m- am.om om.mm muse .P Amocomnv Amocomdv mucmu_c_=mwm ABPPPQmaoLa m=_a>-h (mx «em sumuoeeou Amp onmwhwmmmoo - fl so. u e . . u m “(A . mcaficcmaxm SSS-=8 3:55 some 33% 33335.8“. u gmucp ozu n m_ mo. on» no poou_$wco_m no: u mzs m ooo. No.o Nu.oN ooo_o new .832 .op m ooo. mo.o No.mN muoz apex one ouuooo» .m. mz «Po. NP. Nm. mmoogm>mo PP< .ep mz now. no. oo.N wooed ommmmooeo L93o .m— mz mom. oo._ Pp.m mama“ oemxoo .Np m ooo. we.o om.mN good amoeo .pp m “mo. eo.N mm.o good some; .op mz Noo. mN.N «N.o Looam .o m ooo. Ne.o e¢.mN muomswocoo L55o coo u_om .o mz oeN. om.~ ¢_.¢ muooooco xuoumm>oo coguo use How: .N mz ooN. oN.P o~.m mppo L23o one spam .o mz Noe. No. ~¢.N muocooooeo .m mz mom. mo.. o_.m muons; coo mcomo .mwpnoumow> .e mz Noe. oo. om.N mooeo poem eoooo ooo o>ammoo .m mz mpm. op.P om.m m=_oeo L85o .N mz NNN. oN._ no.m muwm ._ eeooau_cweo_m ooo_ooeooeo o_em_ooom o e sowooesoo Aop......v sooooeeoo u o 3mp>emucw awe?» mgu ease Ppoome we zoo gassed mop co ommoo meoumccmoxm some n ox 3mw>cmucw posed mg» Eocm Ppoomg .1 mo moo news» oz» co ommoo weouwocmoxm some n aw comp u : 3mp>emucw ume_$ may see; .Poume mo zoo ocouom on» so comma oeouwocmoxm some n mm Ponce moo m.x use ppm no: no: zow>emucm “meow on“ soee P—oume o o.1 o o co zoo “move use so ommon me=u_ocwoxm cows u «m .ox u .on u .mw u .mmhz~ hmmmu mzh zoom 44~oz~ mood mzh do mzh~4<30m mom mufihmHHe_ mo. men we uceewwweowm moceewwweowm we _m>e— mo. esp ee weeewwweowm we: mm 2 -k m ooo. op.e woo. woo. “sow; ece Feed .op m ooo. oo. Nmo. ooo. muzz e_ex eee eeoeeew .m. mz omm. Po. ooo. opo. meoeee>eo ._< .ep mz ooo. om.p moo. ooo. meeew eemmeeeee L55o .mp mz NNP. e... ooo. ooo. maeeo zeexeo .N_ m ooo. mo.e moo. oNF. new; neweo .__ m opo. oo.N NNo. Nmo. now; some; .op m oNo. Po._ ooo. epo. eeoom .o m ooo. oo.m oNo. moo. meeesweeeo Lexpo eoe “Fem .o m omo. ww.p Ppo. Nmo. meeeneeo xeepme>wo Lefio eee.ueez .w m Poo. ow.~ Noo. oNP. mpwo eezuo eee Ewe; .o mz Poo. mm._ Noo. ooo. meoeeeoeeo .m mz woo. om._ ooo. Npo. peace eee mceeo .mepeeeeoe> .e mz mop. oo. . Ppo. ooo. moeeo pee: L55o eee e>emmeo .m mz ooP. «N.P Noo. mpo. meweeo Lefio .N mz ooo. eF. mew. mew. eewm .F Ameceeov Ameeeeov teeeeo_weeo_m oe___oeooeo eo.e>-w we ex seweoeeoo Ao_....._v se_eoEEoe u o how. u e Ppeeee we see n n oeeeew eee .eewee .eceeem on» :e eemee sees meeeweeeexe u kw PM A «M "e: o o ppeoee we xee emeww use :e cemee sees eeepweoeexe u «M w u < mzh zhmz 44ew mo. one pe aceewwweowm we: u mze mz oNN. oe._ wo.N poowo ece week .o_ mz owo. ow. oN. meez epex wee eeeeeew .mp mz omo. oo. mw. meoeee>em pw< .ew mz woo. Ne. em. meeew eemmeeeeo Lefo .mw mz owe. ow. oe.p mseum meexem .N_ mz mmo. op. wm. new; eewLo .pp mz opN. om._ mp.m sewn some; .ow mz mNN. oe.~ oo.N eeoom .o mz ooo. mo. om.m mueesweceo L65o eee upem .o mz oFN. em. oo. meoeeeeo xeeume>w4 L55o eee pee: .w mz ooo. mo. .p. wwwo L55o eoe Ewen .o mz ooo. Po. No. meoeeeoeco .m mz moo. me. me.” “weed eoe mceeo .mepeeueoe> .e mz mpm. ow. _m.N meeeo ueem cezuo eee e>emmeo .m mz oom. em. oo.w mcweeo eezao .N mz ooN. om.F o_.m eewm .P 82.8.... to .5 a: _ 288.. 2. NH 3 £2.28 .peoec we met gueeew we oweuweeeexe gees u ow soup u : Ppeeee we mee eewze we eeoeweeeoxe gees u MM Peace wee mom on» PPe we: "e: Ppeeee we mee eceeem we egauweemoxe :ees u mm on: u mox u nox "e: 44ew mo. one we Hneewwwnowm wen u mze mz __m. No._- oN.PN NN.wP enowo one Peed .o_ mz moo. oo. . Ne.w_ «N.o_ muoz epex one eeeeeew .op mz oom. so. oo.m oo.o memene>eo Fp< .ew mz Noo. —e. mo. NN._ moeen oemmeeenn nonuo .mp mz New. mm. . oo.m ow.N mseeo xnmxem .NP mz wwo. om. . ow.wv o_.oe new; oewno .P. mz pmo. PN. . mo.op eo.o new; nmeno .ow mz PNm. oo._- wo.o wo.N neoem .o mz owe. Nu. Nm.e_ No.ow meneswoneo nonuo one u_em .o mz mpo. po. . mo.N_ oo.o muezoenn xeeume>wo noneo one new: .w mz woN. NF._ Po.oN eo.om wwwo nonuo one Ewen .o mz oNo. eo. No. oo. meenonoeno .o mz ooo. No. op.e oo.o ewenn one mnemo .mmweeueoe> .o mz mpN. oN._ _o.w we.N weeno poem nonuo one e>emmeo .m mz mNm. om. so. oo.w mnweno nonuo .N mz opp. oo.~ ww.oe wo.NN eewm ._ Ameneeov Ameneeov eeoneewwwnowm xuwwweeeenn enpe>iw mm ow ouwerEeo Aop.....pv ouwoeesee u o cow u n mueswamm no o< enouwonenxm Peenne news wweeen xeo one zew>nmunw oneoem u ox . x e . x ”e: eueewume .I o.| o.| enouwoneexe .eonne news Fweemn xeo one zew>neunw umnww u «M .o u .< ”e: 44mmhz~ ozoumm oz< wmmwu mxh zoom muhe_ mo. one ee eeeeenwoowm _e>o_ mo. eoe ee eoeewwwnowm eon mz owo. mN. nm.ep ow.e_ unowo one Poem .ow m Nmo. w_.N nN.o_ om.Nw meez epex one eeeeeew .op mz ooo. mo. ee.m oe.e momeno>oo pw< .ew mz mop. Ne.w no.. o~.P moeeo oommooenn noneo .m_ mz ope. Po. . ne.m vo.N mEouH anoxeo .N_ m woo. Po.m N~.mN mo.wm new; oowno .Fp mz on. om. . —w.w om.w new; noon; .ow mz wow. mN. Po._ mo.N neonm .o mz oN_. no.P no.5 oe.o monoewoneo noneo one opem .o mz PNo. no. oe.e N_.o mueooenn xeepwo>wo nonoo one ueoz .w mz woo. on. op.eN mo.NN mpwo nonuo one Ewen .o mz woo. mo. oe. No. meononeeno .o mz wwo. ow._ oo._ mo.N “wand one mneoo .moweeeooo> .e mz eNe. ow. oo.m oe.o moeno ueem noneo one e>emmeo .m mz oom. Po. ow. oo.~ mnweno nonuo .N m epo. oo.N oN.mo o¢.NN oewm .w Amomeoov Amoxeoov woeneowwwnowm zewpweeeeno oowe>iw .M .M zewoeeseo Aop.....pv zuwoessee u o wow u n oneswumo oneuwonoexo Fennne neos pweeon we o o n meo onwn» one oneuom one we ooeno>e .zow>noenw oneeom u .ox .ox w ox “e: oneswamo onouwonoexo wennne neos ppeoon we o.| o.| o.| xeo onwne one oneeom one we ooeno>e .3ow>noenw amnww u .ok .o u .o "e: 44< mxh zo ommthz~ ozoumm oz< hmmaw m2» zoom muh