DISEASE DIAGNOSIS - A LONGITUDINAL ANALYSIS By Shengfang (Eleanor) Song A THESIS Submitted to Michigan State University in partial fulfillment of the requirements for the degree of Biostatistics Master of Science 2020 ABSTRACT CHANGES IN BODY COMPOSITION BEFORE AND AFTER DISEASE DIAGNOSIS - A LONGITUDINAL ANALYSIS By Shengfang (Eleanor) Song Background: known when it starts, how it changes as PD progresses, and whether there is a differential loss of lean or fat mass. Methods: In the Health ABC study (n=3075, age range 70 - 79), body composition was assessed by dual - energy X - ray absorptiometry from baseline (year 1) to year 10. For each PD patient at each year, we calculated the difference between body composition measures and the ir expected values had they not developed PD, and then realigned the time scale in reference to the year of PD diagnosis. Using both inverse variance - weighted fixed - effects linear model and linear mixed model with cross random effects, we examined the tren d of change for each body composition measure before and after PD diagnosis. Results: During an average of 7.8 years of follow - up, a total of 81 PD patients were identified. PD patients were more likely to be men (59.3% vs. 48.2%) and white (74.1% vs. 57. 9%) than non - PD participants. Compared with their expected weight, PD case s began to lose weight and fat mass several years before PD diagnosis ( p - values for trend were 0.003 and <0.001 for total body mass and fat mass , respectively ) . Conclusions: In this longitudinal analysis, we found persistent weight loss, predominantly in fat mass, in PD patients starting a few years before diagnosis. Keywords: - mass; Fat - mass; Weight . iii Dedicated to my parents and my husband Jisheng for their love, understanding, and continuing support. iv ACKNOWLEDGEMENTS I would like to first express my deep and sincere gratitude to my thesis advisor, Dr. Honglei Chen, for organizing and planning the ideas and designs of this study and providing invaluable guidance throughout the research. He really opened the door of neur odegenerative disorder and longitudinal data analysis to me and provided me valuable opportunities to learn and work with public health data. I am also very grateful to the support and instruction from my primary advisor Dr. Dorothy Pathak, who has been pr oviding lots of useful advice and encouragement during my graduate study . I would like to thank Dr. Zhehui Luo and Dr. Chenxi Li for providing strong statistical expertise and precious feedback on the methodology applied in my thesis. It was a great privilege and honor to work with and learn from them. I would also like to thank the faculty and staff in the Department of Epidemiology and Biostatistics for their invaluable advice and support during these two years of study . Many t hank s to my lab members Zichun Cao, Aiwen Yang, Yaqun Yuan, Keran Wang, and Pooja Skukla for their support in data analysis, study conducting, and thesis writing. Thank to my cohort of the biostatistics program , I cannot imagine how I can overcome all the diff iculties without working together with them . Finally, I would like to acknowledge my friends and family, I cannot accomplish this without their love and suppor t . v TABLE OF CONTENTS LIST OF TABLES ................................ ................................ ................................ ............ vi LIST OF FIGURES ................................ ................................ ................................ ......... vi i INTRODUCTION ................................ ................................ ................................ ............. 1 METHODS AND MATERIALS ................................ ................................ ......................... 3 Study Design and Study Population ................................ ................................ ..... 3 Measurements ................................ ................................ ................................ ...... 4 Body co mposition ................................ ................................ ........................... 4 PD ascertainment ................................ ................................ ........................... 4 Year s from PD diagnosis ................................ ................................ ................ 7 Covariates ................................ ................................ ................................ ...... 7 STATISTICAL ANALYSIS ................................ ................................ ............................... 8 Summary Statistics ................................ ................................ ............................... 8 Longitudinal A nalys i s on the Timescale of Clinic Visit Year ................................ ... 9 Trend A nalys i s on the Timescale of Years from PD Diagnosis ........................... 12 Inverse variance - weighted fixed - effects linear model ................................ ... 1 3 Mixed model with crossed random effects ................................ .................... 1 7 RESULTS ................................ ................................ ................................ ...................... 2 2 Participant flow chart ................................ ................................ .......................... 2 2 Baseline characteristics ................................ ................................ ...................... 2 5 Longitudinal Analysis on the Timescale of Clinic Visit Year ................................ . 2 7 Trend Analysis on the Timescale of Years from PD Diagnosis ........................... 3 1 Inverse variance - weighted fixed - effects linear model ................................ ... 3 1 Mixed model with crossed random effects ................................ .................... 3 4 DISCUSSION ................................ ................................ ................................ ................ 3 7 REFERENCES ................................ ................................ ................................ .............. 4 2 vi LIST OF TABLES Table 1 Longitudinal m ixed model comparisons by likelihood ratio tests . 1 1 Table 2 Example dataset for the IVW fixed - effect linear regression . 1 6 Table 3 Example dataset for the mixed model with crossed random effects .. 1 7 Table 4 Random effects structure selection for CRE model .. 2 1 Table 5 Baseline Participant Characteristics of PD cases and non - PD participants 2 6 Table 6 Compare main observed outcomes between PD cases and non - PD participants through clinic visit year 1 to 10 2 8 Table 7 Parameter estimates of the longitudinal linear mixed model for the time - trend of changes in relative body composition measures of PD cases compared with non - PD participants . 2 9 Table 8 Parameter estimates of the IVW and CRE model for the time - trend of changes in relative body composition measures of PD cases compared with their expected values 3 3 Table 9 CRE model predicted body composition differences at each timepoint in PD cases in re ference to had they not had PD 3 6 vii LIST OF FIGURES Figure 1 The sample size of timepoint group s for various body composition measures... ................................ ................................ ................................ ................................ ... 1 5 Figure 2 Example hierarchical structure extracted from Table 3 ................................ ..... 1 8 Figure 3 Participant flow chart ................................ ................................ ....................... 2 3 Figure 4 Fitted means (SD) in various body composition measures of PD case s and non - PD participants by clinic visit year 30 Figure 5 Changes in various body composition measures among PD case s before and a fter diagnosis by IVW method ................................ ................................ ...................... 3 2 Figure 6 Changes in various body composition measures of PD cases before and after diagnosis by CRE model ................................ ................................ ................................ 3 5 1 INTRODUCTION neurodegenerative disorder that is clinically diagnosed by the presence of motor dysfunction such as bradykinesia, rigidity, tremor, and postural instability. PD affects approximately 1 million individuals 65 years or older in the US and is projected to in crease 25% by 2030, and thus presents an increasing burden on healthcare systems. 1 Sporadic PD often takes decades to develop and is typically diagnosed when its cardinal motor signs become clinically evident late in life. In the past two decades, we learned that a wide range of nonmotor symptoms may develop many years prior to PD diagnosis. A thorough understanding of these nonmotor symptoms may be critical both to early disease recognition and a better understanding of disease development. To date, the best - studied nonmotor symptoms are olfaction loss, REM sleep behavior disorder, depression, cognitive changes, and constipation 2 . In comparison, ot her symptoms have received less attention. Weight loss is common in PD case s 3 7 and is often associated with severe motor dysfunction, higher comorbidity, poor physical and mental function, frailty, and increased mortality 5,8 12 . Interestingly, two recent prospective studies reported that lower body mass index was associated with a higher risk of PD, suggesting that weight loss may start prior to PD diagnosis 13,14 . To the best of our knowledge, only one previous population - based study has use d repeatedly obtained body weights to assess when and how body weight changes before and after PD diagnosis. In the Health Professionals Follow - up Study, Chen et al. reported that PD case s tend to begin weight loss 2 - 4 years prior to PD diagnosis and the trend persisted after PD diagnosis. On an average, compared to 2 individuals who did not have PD, case s lost 3.86 kg more of body weight throughout the follow - up 4 . However, this study used self - reported body weight, and did not differentiate loss of lean v ersus fat mass which by itself may have implications for the health a nd survival of older adults. We therefore utilized objective measures of body composition, including total, lean, and fat mass, repeatedly assessed in a community - based biracial cohort to examine changes in body composition in PD case s up to a decade befor e and after PD diagnosis. 3 METHODS AND MATERIALS Study Design and Study Population The Health, Aging, and Body Composition (Health ABC) study was designed to study risk factors for functional declines in older adults, especially changes in body composition, behavioral and physiological conditions in the context of aging. The study recruited 3075 well - functioning older adults (age 70 - 79, 51.4% women, 41.6% blacks) in 1997 - 1998 living in Pittsburgh, Pennsylvania, and Memphis, Tennes see. Inclusion criteria included 1) no difficulty in walking 1/4 mile or climbing up 10 steps; 2) no mobility - related difficulty in performing every - day tasks; 3) no intention to move out of the study area in the next three years. Exclusion criteria includ ed 1) active cancer treatment in the past three years; 2) current participation in a lifestyle intervention trial . Participants enrolled in the study by completing the Year 1 baseline clinic visit from April 1997 to June 1998 and their health and survival were monitored for up to 17 years with annual or biennial clinic visit s, semiannual/quarterly phone calls, and hospitalization and death surveillance. Body compositions were measured, using dual - e nergy X - ray absorptiometry (DXA), at clinic visit s on an annual basis from year 1 to year 6, and then biennially from year 6 to year 10, and then again year 16. In this study, we excluded data from the 16th year because of the long gap between year 10 and 16, and few PD case s were alive and participated in the year 16 clinic visit . We followed eligible participants from baseline until the date of death, last contact, or year 10 clinic visit , whichever came first, with an average of 7 .8 years. All participan ts provided written informed consent and the study protocol was approved by the Institutional Review Boards at University of 4 Pittsburgh, University of Tennessee Memphis, NIH 15 . This specific secondary data analysis was IRB exempted as non - human research by the Michigan State University. Measurements Body composition Body weight or t otal body mass , lean mass , and fat mass were acquired from total body scans by using fan - beam dual - energy X - ray absorptiometry (Hologic QDR 4500A version 8.20a, Hologic , Waltham, MA ) with dual - energy X - ray absorptiometry software (Hologic). The percentage of fat mas s was calculated by dividing the total fat mass by the total body mass. The validity and reproducibility of the DXA scanner have been previously reported. 16,17 The reliability was monitored by quality assurance measurements included the use of daily and cross - calibration phantoms at both study sites 18 . It has been determined that for both study sites, Hologic QDR 4500A overestimated fat - free mass by 5.4% compared to criterion methods , 19 and underestimated fat mass and percent fat consequently . Also, t he Pittsburgh DXA scanner overestimated total mass by about 2% relative to scale weight 19 . In this study, DXA data for all years were corrected and recalculated accordingly. Identical scan protocols were employed for all participants from both sites 20 . PD ascertainment We identified potential PD cases from multiple data sources. At the clinic visit s of years 1, 2, 3, 5, and 6 , the study asked participants to show all medications that they used in the past two weeks and, for each medication, to report the name, dose and frequency of use, reason for use, and year of first use. In years 8, 10, and 11, the study asked participants to report medication use in the past thirty days, including name, current use, frequency of use, and duration of use (years 8 and 10). The study also 5 queried about physician diagnosis of PD at enrollment and in a sub - study at year 13. In addition, the stu dy conducted comprehensive hospitalization and death surveillance. For each hospitalization, up to twenty diagnoses were summarized on the discharge form according to the International Classification of Diseases - 9 - Clinical Modification (ICD - 9 - CM). This dis charge summary, along with records of medical history and physical exams, was subsequently reviewed by a local event adjudicator, and diseases presented at the hospitalization were adjudicated. For each death event, the cohort conducted an exit interview w ith a knowledgeable proxy who provided information on physical functioning of the study participant while alive and details of the event. These data, together with other relevant information such as recent hospitalizations, were centrally reviewed by a tea m of experts and the underlying cause of death was adjudicated by consensus. At the time of the present study, hospitalization and death surveillances were complete through August 2012 21 . In 2015, we retrosp ectively adjudicated a total of 81 PD cases by comprehensively reviewing the above - referenced health data. We first identified a total of 156 participants who had at least one of the following: 1) reported the use of antiparkinsonian medications (carbidopa /levodopa, dopamine agonists, monoamine oxidase B inhibitors, amantadine, or anticholinergic drugs) at any of the medication surveys; 2) self - reported PD diagnosis; 3) local adjudication of PD as the cause of hospitalization or ICD - 9 code of PD (332.0) on the discharge form; 4) PD as the centrally adjudicated cause of death or reported on the proxy interview following death. For each potential case, two experienced movement disorder specialists independently reviewed their PD relevant data over the entire f ollow - up, accounting for the number of sources that indicated a PD 6 diagnosis, internal consistency within each source, and evidence against PD diagnosis. Internal consistency was met if the participants reported PD medications in multiple years, reported P D as the reason for taking the medication, or local adjudication of PD as the reason for multiple hospitalizations. Otherwise, if the participants reported neuroleptic use before PD medication, restless leg syndrome as the reason for taking PD medication, use of dementia medication, or hospitalization with dementia prior to or at the first evidence of PD, we treated them as contradictory evidence. The final PD cases were defined as following: 1) at least two independent sources of PD identification without any contradictory evidence (n = 58) or 2) one source of PD identification with clear internal consistency and no contradictory evidence (n = 13). As hospitalization and death became the only sources to identify potential PD cases after the last medication survey in year 11, we also included potential cases whose only source of information was from adjudicated hospitalization (n = 7) or death (n = 3) after the last available medication survey. We further defined the year of diagnosis as the first year that P D medication (n = 48) or diagnosis (n = 1) was reported. If PD was first identified by hospitalization (n = 29) or death (n = 3), we defined year of diagnosis as the middle point of first identification and the previous year of medical survey without repor ts of PD medication use . 21 The current analysis included all PD case s identified in the cohort, defined as a dichotomous variable with 1 for participants who developed PD (cases) and 0 for those who did not develop PD ( non - PD participants ) during the follow - up . For simplicity, we used case s and non - PD participants to refer these two groups. 7 Year s from PD diagnosis In the analysis, we used the year of PD diagnosis as the reference time point. For each PD case at each measurement, we calculated the number of years in reference to the time of PD diagnosis by subtracting the calendar year of PD diagnosis from the year of measurement. For example, if a case got PD diagnosis in 2003 and he had body composition meas ured five times respectively in 1999, 2001, 2003, 2006, 2008, the corresponding time in reference to PD diagnosis will be 4 and 2 years before, the year at, and 3 and 5 years after PD diagnosis, and the corresponding value for the variable Timepoint will b e - 4, - 2, 0, 3, and 5. Covariates We considered the following covariates in this study - age, sex, race, CVY ear ( clinic visit year), study site, and standing height. Age was defined as the age at each clinic visit . Sex, race, and study site were reporte d at baseline. Race is a two - level variable with 1 for white and 2 for black. Study site has two levels - Memphis or Pittsburgh. CVY ear represents the number of years start ing from baseline, which has values from 1 to 10 representing the 1 st (baseline) to the 10 th year of clinic visit . Body height was measured to the nearest 0.1 cm by using a wall - mounted stadiometer at the clinic visit s of years 1, 4, 6, 8, and 10. At each visit, standing height measurements were taken up to 4 times. The final body height was determined as the following: 1) the average if 2 - 3 measurements were obtained; or 2) the average of the last 2 measurements if all 4 measurements were made; or 3) single measurement if only o ne was taken. For study visits that height was not measured, we used measures from the nearest previous visit. 8 STATISTICAL ANALYSIS Summary Statistics We first conducted a multi - variate longitudinal analys i s to explore the yearly change trajectory of the body composition responses ( total body mass, total body lean mass, total body fat mass, and percentage of fat mass ) among PD cases and non - PD participants . To further e xamine how body composition changed in PD case s before and after the diagnosis , w e used a statistical approach similar to what we published previously. 4 The goal is to describe change s in body composition in cases relative to what the changes in body composition would have been had they not developed PD . Briefly, at each clinic visit year, we first fitted among non - PD participants a linear regression model of each body composition meas ure on age, sex, race, study site, and height and obtained the beta - coefficients. W e then applied these coefficients to cases and calculated their expected body composition measures and interpret it as their counterfactual body composition measures had they not developed PD (thereafter referred as expected measures) . We finally calculated residuals which, by definition, represent the difference between their actual body composition me asures in reference to the expected values in that particular calendar year had they not developed PD . After conducting this analysis for each clinic visit that body compositions were measured, we comp il ed all available residuals for PD cases and realigned the time scale using the year of PD diagnosis as reference. We subsequently fitted both an inverse variance - weighted linear regression model and a mixed model with cross ed random effects to examine trends of the above residuals across all years in referen ce to PD diagnosis. For all the analyses in the present study, we assumed m issing completely at random 9 and excluded those with missing responses from the corresponding analysis. The longitudinal analysis , the linear regressions for measures at each clinic visit , and the inverse variance weighted modeling were conducted with SAS version 9.4 (SAS Systems Inc., Cary, NC) and the mixed model with cross ed random effects was conducted with Stata version 16.0 (Sta taCorp, College Station, TX) . A p - value of less than 0.05 was considered statistically significant . Longitudinal Analysis on the Timescale of Clinic Visit Year This is an exploratory analysis for the yearly change trajectories of the body composition s among participants who had (cases) and had not developed PD ( non - PD participants ) during the follow - up . We first fitted a mixed model with a separate mean for each of the 16 combinations ( thereafter referred as unstructured mean) , together with separate unstructured variance - covariance matri ces for cases and non - PD participants ( thereafter referred as separate unstructured variance - covariance matrix ) as : (1) where represents the vector of yearly repeated body composition measurement m ( taking from total body mass, total body lean mass, total body fat mass, and percentage body fat) for participant i ; and stand for the categorical and linear clinic visit year for participant i , respectively ; represents the vector of error components for body composition m in equation (1) where is the corresponding variance - covariance matrix . 10 We t hen tried to further reduc e both the mean and variance - covariance structure to fit a more parsimonious model . Model comparisons were conducted using likelihood ratio tests and the unstructured mean and separate unstructured variance - covariance matrix model was selected to be the fin al model for all the body composition responses ( Table 1). 11 Table 1 Longitudinal mixed model comparisons by likelihood ratio tests Total mass Total lean mass Total fat mass Percentage of fat mass Chi - Square Statistic * D F P - valu es Chi - Square Statistic* D F P - valu es Chi - Square Statistic * D F P - valu es Chi - Square Statistic * D F P - valu es Mean Structure Different quadratic trends 42 10 <.01 66.8 10 <.01 57.3 10 <.01 79.6 10 <.01 Unstructured vs. Different cubic trends 32.3 8 <.01 42.7 8 <.01 56 8 <.01 79 8 <.01 Unstructured vs. Different quartic trends 27.7 6 <.01 23.8 6 <.01 16.8 6 0.01 10.6 6 0.10 Unstructured vs. Different quintic trends 23.5 4 <.01 22.3 4 <.01 11 4 0.03 7.7 4 0.10 Unstructured vs. Different sextic trends 23.4 2 <.01 20.9 2 <.01 8.9 2 0.01 1.9 2 0.39 Variance - Covariance Structure Unstructured vs Heterogeneous Toeplitz 2328.8 42 <.01 1216.1 42 <.01 2471.6 42 <.01 1407.8 42 <.01 variance - covariance 63.8 36 <.01 54.7 36 0.02 62.9 36 <.01 49.9 36 0.06 * Calculated from the difference of the - 2log likelihood for the nested model and the full model. combinations. - covariance structure was the same for cases and non - PD participants. 12 Trend Analysis on the Timescale of Years from PD Diagnosis Step 1: Linear Regression The purpose of this step is to obtain the residuals for PD cas es for the subsequent analys e s in step 2. At each clinic visit year, we first fitted a linear regression in non - PD participants : (2) where represents body composition measurements of non - PD participant i for body composition m (taking from total body mass, total body lean mass, total body fat mass, and percentage body fat) measured at clinic visit year j ; , represents the corresponding error component for body composition m in equation (2) . A ssuming the effects of age, sex , race, height, and site on body composition are the same for PD case s and non - PD participants within each clinic visit year, w e then applied the beta - coefficients obtained from equation (2) to PD case s and calculated the ir predicted values if the y had not developed PD : (3) where denote s the predicted response for PD case i calculated for body composition m at clinic visit year j ; were the corresponding coefficients obtained from equation (2). 13 Residuals were calculated by subtracting the predicted body composition values from the observed measures, which represent the differences in body compositions in PD case s in reference to their expected values if they had not had PD: (4) where denotes the residual of PD case i calculated for body composition m at clinic visit year j ; is the observed body composition m for PD case i measured at clinic visit year j . Step 2: Trend Analysis T o evaluate the change trajectory of body composition s across years from PD diagnosis in PD cases relative to if they had not ha d PD , both an inverse variance - weighted linear regression model and a mixed model with cross ed random effects were fitted on the residuals calculated from step 1 ( from equation (4)) . O nly PD case s were included in the analyses in step 2 . Inverse variance - weighted (IVW) fixed - effect s linear model The i nverse variance - weight ed fixed - effects model is predominantly used in multivariate statistics and meta - analyses for synthesizing data from different sources 22 25 . The idea is that assuming the predictions from larger sample size and smaller variance groups are more precise , and then allocating larger weights to these groups to make the m have greater influence s on the analyses than the groups with smaller sample siz e and/or higher variability 26 . In the present analysis , we were mainly interested in the average of residuals at each timepoint (years from PD diagnosis) and their change trajectory. Therefore, w e fitted a 14 fixed - effects linear model to examine the trajectory of the averages of residuals with timepoint: (5) where is the mean of residuals for body composition m at k th timepoint ; Timepoint represents years from PD diagnosis; , is the error component for body composition m in equation (5). To reduce the bias that might be introduced by the unbalanced sample sizes between timepoint groups in our data ( Figure 1) , we first excluded those timepoints with a sample size less than or equal to 5 from the analyses ( ex cluding timepoint - 13, - 12, 10, 11, 12, 13, and 15). The example dataset after the exclusion is showed in Table 2. W e then borrowed th e idea of inverse variance weighting to allocate a weight for each timepoint group based on the reciprocal of its standard error of the mean to allocate higher weights to those groups with smaller vari ances . The weights for each timepoint group were given by 26,27 : (6) where and denote the weight and the standard error of the mean at k th timepoint for body composition m , respectively . Thus, the weighted least - squares estimates from equation (5) will achieve minimum variance . U nder the assumption that is independent with each other and with timepoint and is normally distributed , the weighted least - squares estimates can be derived as both the maximum - likelihood estimator and the best linear unbiased 15 estimators ( BLUE ) 28 30 . The predicted means and their 95% confidence intervals were then obtained based on thes e estimators. The (7) where is a 3x1 matrix of estimators of the regression parameters in the regression model in equation (5) for body composition m , is a 21x3 matrix of independent variables for body composition m ; is a 21x21 diagonal matrix of w i for body composition m ; is a 21x1 matrix of means of residuals for body composition m , taking the example dataset in table 2 for instance) . Figure 1 The sample size of timepoint group s for various body composition measures 16 Table 2 Example dataset for the IVW fixed - effect linear regression TimePoint Sample Size M ean of Residuals S tandard Error W eight - 11 7 4.17 6.11 0.03 - 10 8 9.29 5.22 0.04 - 9 14 3.11 3.56 0.08 - 8 16 1.92 2.79 0.13 - 7 23 0.67 1.94 0.26 - 6 23 2.41 2.35 0.18 - 5 33 0.51 1.46 0.47 - 4 33 - 2.13 1.83 0.30 - 3 36 0.36 1.73 0.33 - 2 34 - 1.45 1.69 0.35 - 1 32 2.39 2.10 0.23 0 27 - 4.93 1.85 0.29 1 24 - 0.44 2.10 0.23 2 27 - 2.17 2.12 0.22 3 17 0.800 2.61 0.15 4 20 - 4.39 2.77 0.13 5 14 - 5.55 3.14 0.10 6 13 - 0.05 3.62 0.08 7 6 - 8.08 4.79 0.04 8 8 - 5.35 4.54 0.05 9 6 - 5.80 4.19 0.06 Timepoint , years from PD diagnosis. Other assumptions should be satisfied for IVW fixed - effect linear regression including 24,26 : 1) are independent with each other ; 2) are independent with each other and with timepoint ; 3) the conditional variance of is . However, the above assumption s might be violated in our data as the means of residuals at different timepoint s were correlated since each PD case contributed measurements for multiple timepoint s . Thus, were also dependence with each other . Such violations could bias the predicted means and return an incorrect standard error for . Thus, mixed models with crossed random effects were conducted to obtain more precise estimates . 17 Mixed model with cross ed random effects The m ixed model can capture the heterogeneous variances and covariances in the data and is , therefore , more powerful compared to the repeated - measures ANOVA especially when the sphericity and homoscedasticity assumption s are violated 31 . Unlike the IVW fixed - effect linear model, the mixed model with random effects estimated the effects of timepoint on the residuals at an individual level instead of only on their summary statistics. The example dataset for mixed model ing is in Table 3 . Table 3 Example dataset for the mixed model with crossed random effects Obs ID CVYear TimePoint PD Residual 1 1 1 0 1 13.05 2 1 2 1 1 7.38 3 1 3 2 1 3.29 4 1 4 3 1 8.88 5 1 5 4 1 5.55 6 1 6 5 1 1.47 7 2 1 0 1 - 4.84 8 2 2 1 1 - 3.30 9 2 3 2 1 0.59 10 2 4 3 1 0.16 11 2 5 4 1 1.06 12 2 6 5 1 - 3.87 13 2 8 7 1 - 11.04 14 3 1 - 2 1 - 10.60 15 3 2 - 1 1 - 7.81 16 3 3 0 1 - 8.66 17 3 4 1 1 - 12.01 18 3 5 2 1 - 15.85 19 3 6 3 1 - 21.22 20 3 8 5 1 - 25.56 21 3 10 7 1 - 28.23 Timepoint, years from PD diagnosis In traditional hierarchical or multilevel models , random effects are assumed to be nested so that the lower - level units are uncorrelated across higher - level units 31,32 . However, our 18 data did not show a typical hierarchical structure , in other words , none of a unit is nested in another in our data and the levels of units are not fixed. For easy explaining , we chose timepoint as the highest level followed by participant and CVYear as an example to illu strate how our data is different from a traditional hierarchical structure . To be specific, in our da ta, 1) the residuals from the same participant ( lower - level ) were correlated across timepoint ( higher - level ) (Figure 2, solid lines); and 2) the residuals calculated from the same clinic visit year were correlated across the higher levels (participant and timepoint) due to the the same year (Figure 2, orange/blue dashed lines for correlation across participant/timepoint , respectively ) . Thus, we included both random effects of participants and CVYear simultaneously into a single analysis and treated the two random effects as crossed 31 . Figure 2 Example hierarchical structure extracted from Table 3. Correlations are show n by the same colored lines that connected two blocks. The solid lines denote the correlations of multiple measurements from the same participant across timepoint; the or ange and blue dashed lines denote the correlations of the values from the same clinic visit year across participant and timepoint, respectively. 19 The general mixed - effects model is given as: (8) where is a vector of the residual s of PD case i in the analysis for body composition m ; is a vector of the years from PD diagnosis of the i th PD case ; denote s the design matrix for the random effects for the i th PD case in the analysis for body composition m ; represents the corresponding vector of random effects ; represents the corresponding error vector in equation (8) . In the current analysis, the inference was based on the marginal distribution for the outcome variable residual . 33 , (9) where is a vector of the residuals of PD case i in the analysis for body composition m ; is a vector of the fixed effects in model (8) from the i th PD case in the analysis for body composition m ; is the covariance matrix of and is the covariance matrix of . Let be a vector of variance components found in and be estimated by the restricted maximum likelihood (REML) estimation in the current analyses. T he vector of coefficient was estimated by 33 : (10) where N is the total number of PD case s in our dataset. 20 The main assumptions under the present mixed random - effects model are 33 : 1) is multivariate normal with mean vector 0 and with covariance matrix , 2) is multivaria te normal with mean vector 0 and with covariance matrix , 3) all and are independent of each other, and 4) is a positive (semi - )definite matrix . The testing of the need for random effects was conducted by the likelihood ratio test (LRT) and the test statistics are exhibited in Table 4. The structure of random effects was decided based on the LRT results and the criteria of smaller AIC and BIC , and structure 1 was selected as the final random effects structure for all four body composition measures . Thus, the item in equation (8) included the crossed random effects of both by - CVYear and by - participant adjustments to the intercept an d by - participant adjustments to the effect of Timepoint 32,34 . The final model was given by: (11) where is a vector of the residuals of PD case i in the analysis for body composition m ; is a vect or of the years from PD diagnosis of the i th PD case ; represents the CVYear random effects for j th clinic visit year in the analysis for body composition m ; is the participant random effects for i th participant in the analysis for body composition m ; is the error vector for participant i in the analysis for body composition m . 21 Table 4 Random effects structure selection for CRE model AIC BIC Log - likelihoo d Ref struct - ure Chi - Square Statistic* D F P Total mass Structure 1 2401.8 2430.1 - 1193.9 Structure 2 2459.1 2483.4 - 1223.6 1 59.3 1 <.01 Total lean mass Structure 1 1886.3 1915.0 - 936.2 Structure 2 1937.6 1962.2 - 962.8 1 53.3 1 <.01 Total fat mass Structure 1 2182.4 2211.1 - 1084.2 Structure 2 2247.7 2272.3 - 1117.8 1 67.3 1 <.01 Percentage of fat mass Structure 1 1891.1 1919.4 - 938.5 Structure 2 1946.5 1970.7 - 967.2 1 57.4 1 <.01 * Calculated from the difference of the - 2log likelihood for the nested model and the full model . - values. Structure 1: Random intercept for both participants and CVYear; By - participant slope for timepoint . Structure 2: Only random intercept for both participants and CVYear. 22 RESULTS Participant f low chart In all analyses, we excluded participants who were initially identified as possible PD cases but did not get confirmed in the case - adjudication effort. Participants with missing values in body composition measures were excluded from the corresponding analysis. There was no missing value on covariates after this exclusion. The numbers of each exclusion and final sample size for each clinic visit are provided in Figure 3 . 23 Figure 3 Participant flow chart Clinic visit 1 (n=3075) Excluded due to: Uncertain/possible PD(n=58) Missing in total body mass/fat% (n=110) Analyses for total body mass/fat% (n=2907) Excluded due to: Uncertain/possible PD (n=58) Missing in total fat/lean mass (n=20) Analyses for total body fat/lean mass (n=2997) Clinic visit 2 (n=2998) Excluded due to: Uncertain/possible PD(n=58) Missing in total body mass/fat% (n=331) Analyses for total body mass/fat% (n=2609) Excluded due to: Uncertain/possible PD (n=58) Missing in total fat/lean mass (n=228) Analyses for total body fat/lean mass (n=2712) Clinic visit 3 (n=2921) Excluded due to: Uncertain/possible PD(n=58) Missing in total body mass/fat% (n=443) Analyses for total body mass/fat% (n=2420) Excluded due to: Uncertain/possible PD (n=58) Missing in total fat/lean mass (n=329) Analyses for total body fat/lean mass (n=2534) Clinic visit 4 (n=2775) Excluded due to: Uncertain/possible PD(n=56) Missing in total body mass/fat% (n=490) Analyses for total body mass/fat% (n=2229) Excluded due to: Uncertain/possible PD (n=56) Missing in total fat/lean mass (n=370) Analyses for total body fat/lean mass (n=2349) 24 Clinic visit 5 (n=2733) Excluded due to: Uncertain/possible PD(n=56) Missing in total body mass/fat% (n=549) Analyses for total body mass/fat% (n=2128) Excluded due to: Uncertain/possible PD (n=56) Missing in total fat/lean mass (n=428) Analyses for total body fat/lean mass (n=2249) Clinic visit 6 (n=2619) Excluded due to: Uncertain/possible PD(n=55) Missing in total body mass/fat% (n=619) Analyses for total body mass/fat% (n=1945) Excluded due to: Uncertain/possible PD (n=55) Missing in total fat/lean mass (n=507) Analyses for total body fat/lean mass (n=2057) Clinic visit 8 (n=2305) Excluded due to: Uncertain/possible PD(n=44) Missing in total body mass/fat% (n=739) Analyses for total body mass/fat% (n=1522) Excluded due to: Uncertain/possible PD (n=44) Missing in total fat/lean mass (n=617) Analyses for total body fat/lean mass (n=1644) Clinic visit 10 (n=2045) Excluded due to: Uncertain/possible PD(n=35) Missing in total body mass/fat% (n=675) Analyses for total body mass/fat% (n=1335) Excluded due to: Uncertain/possible PD (n=35) Missing in total fat/lean mass (n=538) Analyses for total body fat/lean mass (n=1472) 25 Baseline characteristics We present the baseline population characteristics of participants who had (cases) and had not developed PD ( non - PD participants ) during the follow - up in Table 5 . For simplicity, we used cases and non - PD participants to refer these two groups. Because of the narrow age range of o ur study population, there was no significant age difference between PD cases and non - PD participants (74.0 ± 2. 8 vs. 73.6 ± 2. 9 years, p=0.1 8 ). Compared with non - PD participants , PD case s were more likely to be male (59. 3 % vs. 48. 2 %, p=0.04 9 ) and white (7 4. 1 % vs. 57.9%, p=0.00 4 ). The height and body composition measures did not significantly differ between PD cases and non - PD participants at baseline. 26 Table 5 Baseline Participant Characteristics of PD cases and non - PD participants PD Cases (n=81) Non - PD Participants (n=2936) P - values Continuous Variables, Mean (SD) Age (years) 74.0 (2.8) 73.6 (2.9) 0.18 Standing h eight (m) 1.7 (0.1) 1.7 (0.1) 0.11 Total b ody m ass (Kg) 75.9 (12.4) 75.5 (14.9) 0.62 Total b ody l ean m ass (Kg) 48.4 (9.1) 46.7 (10.0) 0.07 Total b ody f at m ass (Kg) 25.6 (7.1) 26.8 (8.8) 0.21 Percentage of fat mass (%) 33.5 (7.0) 35.0 (7.9) 0.06 Categorical Variables, N (%) Sex 0.049 Female 33 (40.7) 1522 (51.8) Male 48 (59.3) 1414 (48.2) Race 0.004 White 60 (74.1) 1701 (57.9) Black 21 (25.9) 1235 (42.1) Site 0.67 Memphis 39 (48.15) 1483 (50.5) Pittsburgh 42 (51.85) 1453 (49.5) Due to missing data , the exact sample sizes of cases/controls were 77/2830 for total body mass and percentage of fat mass , and 80/2917 for total body lean and fat mass. P - values were calculated through Mann Whitney U test and chi - square test for continuous and categorical v ariables, respectively. 27 Longitudinal Analysis on the Timescale of Clinic Visit Year Table 6 shows the means of each body composition measure by clinic visit year. Overall, all four body composition measures slightly decreased over time in both PD case s and controls. Compare to non - PD participants, PD case s showed a lower total fat mass and percentage of fat mass from clinic visit year 6 to 10 (p<0.05), higher total lean mass at year 6 and 7 (p<0.05), and no difference in total body mass from year 1 to 10. The comparison of fitted means from the linear mixed model ( equation (1) ) was showed in Table 7 and Figure 4 . In this analysis, a significant downtrend in percentage of fat mass was observed among PD case s relative to non - PD participants ( Figure 4 , Table 7, p = 0.03). From baseline to clinic visit year 10, PD case s lose 1.4 kg and 1.8% more total fat mass compared to non - PD participants ( Table 7, p < 0.05 ) . In comparison , the total body mass and lean mass in PD case s were relatively stable over the entire observation period (Figure 4 ) and the change trajectories were not significantly different from non - PD participants ( Table 7, p - values for total body mass and lean mass are 0.58 and 0.86, respectively) . 28 Table 6 Compare main observed outcomes between PD cases and non - PD participants through clinic visit year 1 to 10 Year 1 Year 2 Year 3 Year 4 Year 5 Year 6 Year 8 Year 10 N (PD cases/Controls) Total Mass/Fat% (77/2830) (73/2536) (60/2360) (59/2170) (51/2077) (49/1896) (39/1483) (32/1303) Total Lean/Fat Mass (80/2917) (77/2635) (64/2470) (61/2288) (54/2195) (52/2005) (42/1602) (35/1437) Total Mass (Kg), Mean (SD) PD cases 75.9(12.4) 76.1(11.9) 75.7(13) 75.2(13.4) 75.4(13.1) 74.6(13.6) 74.1(14.1) 73.7(15.1) Non - PD participants 75.5(14.9) 75.3(14.9) 75.1(14.9) 75.0(14.9) 74.5(14.8) 74.4(14.9) 73.6(14.8) 72.5(14.7) Total Lean Mass (Kg), Mean (SD) PD cases 48.4(9.1) 48.3(8.9) 48.4(9.7) 47.9(9.4) 47.9(9.4) 48.4(9.4) * 49.2(9.1) ** 48.0(9.8) Non - PD participants 46.7(10.0) 46.5(9.8) 46.1(9.8) 45.9(9.7) 45.5(9.6) 45.4(9.6) 45.3(9.7) 44.9(9.5) Total Fat Mass (Kg),Mean (SD) PD cases 25.6(7.1) 25.8(7.3) 25.3(6.8) 25.5(7.2) 25.6(7.9) 24.4(7.4) * 23.4(8.1) ** 23.3(9.1) * Non - PD participants 26.8(8.8) 26.6(8.8) 26.9(8.9) 27.0(8.9) 26.9(8.7) 26.8(8.6) 26.3(8.5) 25.6(8.3) Percentage of fat mass (%), Mean (SD) PD cases 33.5(7.0) 33.6(7.5) 33.4(7.0) 33.6(7.0) 33.5(7.6) 32.1(6.9) ** 30.8(7.3) ** 31.8(8.5) * Non - PD participants 35.0(7.9) 34.9(7.8) 35.3(7.7) 35.6(7.7) 35.6(7.6) 35.7(7.5) 35.3(7.4) 34.9(7.4) PD cases and non - PD participants denote those developed PD and had not developed PD during the follow - up, respectively. P - values were calculated throu gh the Kruskal - Wallis test when comparing means between PD cases and non - PD participants within the same year. * P - values < 0.05, ** P - values < 0.01. 29 Table 7 Parameter estimates of the longitudinal linear mixed model for the time - trend of changes in relative body composition measures of PD cases compared with non - PD participants Year 2 Year 3 Year 4 Year 5 Year 6 Year 8 Year 10 P N (PD cases/Controls) Total Mass/Fat% (73/2536) (60/2360) (59/2170) (51/2077) (49/1896) (39/1483) (32/1303) Total Lean/Fat Mass (77/2635) (64/2470) (61/2288) (54/2195) (52/2005) (42/1602) (35/1437) Total mass (Kg), estimate (SE) PD case s - 0.1 (0.3) - 0.3 (0.5) - 0.5 (0.5) - 0.6 (0.6) - 1.0 (0.7) - 1.0 (0.9) - 1.5 (0.8) 0.58 Non - PD participants Ref Ref Ref Ref Ref Ref Ref Total lean mass (Kg), estimate (SE) PD case s - 0.02 (0.2) 0.03 (0.2) - 0.09 (0.2) 0.01 (0.3) 0.09 (0.3) 0.3 (0.4) 0.4 (0.4) 0.86 Non - PD participants Ref Ref Ref Ref Ref Ref Ref Total fat mass (Kg), estimate (SE) PD case s 0.1 (0.2) - 0.2 (0.3) - 0.4 (0.3) - 0.5 (0.4) - 0.8 (0.5) - 1.1 (0.6) - 1.4 (0.6) * 0.15 Non - PD participants Ref Ref Ref Ref Ref Ref Ref Percentage of fat mass (%), estimate (SE) PD case s 0.009 (0.2) - 0.3 (0.3) - 0.4 (0.3) - 0.6 (0.3) - 1.0 (0.4) * - 1.4 (0.5) ** - 1.8 (0.5) ** 0.03 Non - PD participants Ref Ref Ref Ref Ref Ref Ref P - values for trend comparison . * p<0.05, ** p<0.01 . The unstructured main model fully adjusted for age, sex, race, height, clinic site, PD status, and the interaction between PD status, age, height and c linic visit year. Variance - covariance matrixes are unstructured and separate for PD case s and non - PD participants. c linic visit year on body composition, which represents the relative body composition change from baseline for PD case s compared to non - PD participants. 30 Figure 4 Fitted means (SD) in various body composition measures of PD case s and non - PD participants by clinic visit year 31 Trend Analysis on the Timescale of Years from PD Diagnosis I nverse variance - weighted fixed - effect linear model Figure 5 displays changes in body composition measures before and after PD diagnosis by IVW method . T he Y - axis represents the difference between the observed body composition values of PD case s and their expected values if they had not developed PD. A significant linear downtrend was observed for total body mass, lean mass, fat mass , and percentage of fat mass several years before the disease diagnosis (Figure 5 ; Table 8, the p - value for linear trend s are 0.00 3 , 0.00 8 , <0.0001, and <0.0001, respectively) . Nine years after diagnosis, the average cumulative loss was about 8.0 kg for total mass with about 7.0 kg for fat mass compared to 11 years before diagnosis ( Figure 5 ). In comparison, the loss of total lean mass of PD case s was mild over the entire period ( Figure 5 , a loss of about 3.0 kg by year 9). 32 Figure 5 Changes in various b ody composition measures among PD case s before and after diagnosis by IVW method All comparisons were made to their expected values at the given timepoint, which estimated based on data from non - PD participants. The horizontal reference line 0 represents standardized expected changes in body composition measures over years for non - PD c ase s, and for PD case s these refers to the expected changes if they had not developed PD. 33 Table 8 Parameter estimates of the IVW and CRE model for the time - trend of changes in relative body composition measures of PD cases compared with their expected values Intercept Linear Timepoint Quadratic Timepoint Estimate (SE) 95% CI P Estimate (SE) 95% CI P Estimate (SE) 95% CI P Total mass (Kg) IVW - 1.7 (0.6) ( - 3.0, - 0.3) 0.02 - 0.4 (0.1) ( - 0.7, - 0.2) <0.01 0.009 (0.02) ( - 0.03, 0.05) 0.7 CRE - 2.0 (1.3) ( - 2.7, 2.3) 0.9 - 0.4 (0.1) ( - 0.6, - 0.2) <0.01 - 0.03 (0.01) ( - 0.06, - 0.01) <0.01 Total lean mass (Kg) IVW - 0.3 (0.3) ( - 1.0, 0.4) 0.4 - 0.1 (0.05) ( - 0.2, - 0.04) <0.01 0.006 (0.009) ( - 0.01, 0.02) 0.5 CRE 0.6 (0.5) ( - 0.5, 1.6) 0.3 - 0.007 (0.05) ( - 0.1, 0.09) 0.9 - 0.006 (0.005) ( - 0.02, 0.004) 0.3 Total fat mass (Kg) IVW - 1.4 (0.4) ( - 2.3, - 0.6) <0.01 - 0.4 (0.07) ( - 0.5, - 0.2) <0.01 - 0.003 (0.01) ( - 0.03, 0.02) 0.8 CRE - 0.5 (0.8) ( - 2.1, 1.1) 0.6 - 0.3 (0.07) ( - 0.5, - 0.2) <0.01 - 0.03 (0.008) ( - 0.04, - 0.02) <0.01 Percentage of fat mass (%) IVW - 0.9 (0.2) ( - 1.3, - 0.4) <0.01 - 0.3 (0.04) ( - 0.4, - 0.2) <0.01 - 0.02 (0.007) ( - 0.03, - 0.001) 0.04 CRE - 0.7 (0.6) ( - 1.8, 0.4) 0.2 - 0.3 (0.06) ( - 0.4, - 0.2) <0.01 - 0.02 (0.007) ( - 0.04, - 0.01) <0.01 IVW refers to inverse variance - weighted fixed - effects linear regression model. CRE refers to linear mixed model with crossed random effects. Timepoint refers to years from PD diagnosis which was included as a linear term and a quadratic term. Estimate is the fixed effects of timepoint on the difference between actual body composition of PD cases and their expected b ody composition had they not had PD. Crossed random effects in CRE included random intercepts for clinic visit year and for participants, and by - participant random slope on timepoint. CI = confidence interval. P = P - values. 34 Mixed models with cross ed random effects With the inclusion of crossed random effects , the results from the CRE model are believed less biase d than the IVW model and are selected to be our final results . Figure 6 shows the results from the analysis using mixed models with cross ed random effects. The horizontal reference line 0 represents the standardized expected changes in body composition measures of non - PD participants because at any given point their residuals should be 0. Therefore, this line also represents the expected changes in body compo sition of PD case s had they did not develop PD. In this analysis, all body composition measures of PD cases were comparable to their expected values prior to PD diagnosis; however, total body mass, fat mass, and percentage of fat mass began to decrease aro und a few years prior to PD diagnosis which persisted through the all period s after PD diagnosis (Table 8 , p - value for trends were 0.003, <0.001, and <0.001, respectively). By year 5, 3, and 2 after diagnosis, total body mass, fat mass, and percentage of f at mass , respectively, became statistically different from their expected values and the decreasing trend persisted. Nine years after diagnosis, the average cumulative loss was 6.0 kg for total mass with 5. 6 kg for fat mass compared to 11 years before diagnosis (Table 9 ). In comparison, the total lean mass of PD case s was stable over the entire period (Table 8 , p - value for trends=0.26; Table 9 , a gain of 0.09 kg by year 9). 35 Figure 6 Changes in various body composition measures of PD cases before and after diagnosis by CRE model. All comparisons were made to their expected values at the given timepoint, which estimated based on data from non - PD participants. The horizontal reference line 0 represents standardized expected changes in body composition measures over years for non - PD case s, and for PD case s these refers to the expected changes if they had not developed PD. 36 Table 9 CRE model predicted body composition difference s at each timepoint in PD cases in reference to had they not had PD Total mass (Kg) Total lean mass (Kg) Total fat mass (Kg) Percentage of fat mass (%) Time - point Estimates 95% CI Estimates 95% CI Estimates 95% CI Estimates 95% CI - 11 - 0.2 ( - 4.1, 3.6) - 0.1 ( - 1.7, 1.6) - 0.4 ( - 2.9, 2.1) - 0.2 ( - 2.2, 1.8) - 10 0.1 ( - 3.4, 3.6) 0.1 ( - 1.5, 1.6) - 0.1 ( - 2.4, 2.2) 0.0 ( - 1.8, 1.8) - 9 0.4 ( - 2.9, 3.7) 0.2 ( - 1.2, 1.6) 0.1 ( - 2, 2.3) 0.2 ( - 1.5, 1.8) - 8 0.6 ( - 2.5, 3.7) 0.3 ( - 1, 1.6) 0.3 ( - 1.7, 2.3) 0.3 ( - 1.2, 1.8) - 7 0.7 ( - 2.2, 3.6) 0.3 ( - 0.9, 1.6) 0.4 ( - 1.5, 2.3) 0.3 ( - 1.1, 1.7) - 6 0.8 ( - 2, 3.6) 0.4 ( - 0.8, 1.6) 0.5 ( - 1.3, 2.3) 0.3 ( - 1, 1.6) - 5 0.8 ( - 1.9, 3.5) 0.5 ( - 0.7, 1.6) 0.5 ( - 1.3, 2.2) 0.3 ( - 1, 1.5) - 4 0.7 ( - 1.9, 3.3) 0.5 ( - 0.6, 1.6) 0.4 ( - 1.3, 2.1) 0.2 ( - 1, 1.3) - 3 0.6 ( - 2, 3.2) 0.5 ( - 0.5, 1.6) 0.3 ( - 1.4, 1.9) 0.0 ( - 1.1, 1.1) - 2 0.4 ( - 2.1, 2.9) 0.6 ( - 0.5, 1.6) 0.1 ( - 1.5, 1.7) - 0.2 ( - 1.3, 0.9) - 1 0.1 ( - 2.4, 2.6) 0.6 ( - 0.5, 1.6) - 0.2 ( - 1.8, 1.4) - 0.4 ( - 1.5, 0.7) 0 - 0.2 ( - 2.7, 2.3) 0.6 ( - 0.5, 1.6) - 0.5 ( - 2.1, 1.1) - 0.7 ( - 1.8, 0.4) 1 - 0.6 ( - 3.1, 1.9) 0.6 ( - 0.5, 1.6) - 0.9 ( - 2.5, 0.7) - 1.1 ( - 2.2, 0) 2 - 1.1 ( - 3.6, 1.5) 0.5 ( - 0.5, 1.6) - 1.3 ( - 2.9, 0.3) - 1.5* ( - 2.6, - 0.3) 3 - 1.6 ( - 4.2, 1) 0.5 ( - 0.6, 1.6) - 1.8* ( - 3.4, - 0.1) - 1.9** ( - 3.1, - 0.7) 4 - 2.2 ( - 4.9, 0.5) 0.5 ( - 0.7, 1.6) - 2.3** ( - 4.1, - 0.6) - 2.4** ( - 3.6, - 1.2) 5 - 2.9 * ( - 5.7, 0) 0.4 ( - 0.8, 1.6) - 2.9** ( - 4.8, - 1.1) - 3.0** ( - 4.3, - 1.6) 6 - 3.6 * ( - 6.6, - 0.6) 0.3 ( - 1, 1.6) - 3.6** ( - 5.6, - 1.6) - 3.5** ( - 5, - 2.1) 7 - 4.4 ** ( - 7.6, - 1.1) 0.2 ( - 1.2, 1.7) - 4.3** ( - 6.4, - 2.2) - 4.2** ( - 5.8, - 2.6) 8 - 5.3 ** ( - 8.8, - 1.7) 0.1 ( - 1.4, 1.7) - 5.1** ( - 7.4, - 2.8) - 4.9** ( - 6.7, - 3.1) 9 - 6.2 ** ( - 10.1, - 2.3) 0.0 ( - 1.7, 1.8) - 5.9** ( - 8.5, - 3.4) - 5.6** ( - 7.6, - 3.6) Timepoint, years from PD diagnosis. Estimate is the mean estimate of differences between actual and expected body composition of PD case s at each timepoint. CI = confidence interval. * P - values < 0.05, ** P - values < 0.01. 37 DISCUSSION In this longitudinal study with repeated measures of body composition, we found that the body weight of PD case s began to decrease a few years prior to diagnosis which persisted in the years after diagnosis. By year 9, post PD diagnosis, the cumulative los s was on an average of about 6 kg. We further found that on an average 93% of the loss was due to loss of fat mass while the lean mass in PD case s was well preserved. Substantial evidence from case - control and cross - sectional studies have shown that PD case s had lower body weight and body mass index than controls 3,5,7,11,35 39 . Some studies further reported that PD case s lose more weight after diagnosis than individuals without PD. The first report was published in 1976 and found a significant loss of body weight in seven levodopa - treated PD case s compared to healthy controls, with an average loss of 6.28 kg over 1 - 3 years 40 . In 1995, Beyer et al. 3 reported that PD case s were four times more likely to report a significant weight loss of >10 pound s since disease diagnosis. Two later longitudinal studies reported PD case s lost an average of 1.8 kg over 1 year 7 and 4.1 kg over 6.3 years of follow - up 5 . While the clinical and epidemiological data on weight loss in PD is substantial, they are not entirely consistent. Several studies have reported stable or even weight gain in PD case s. For example, Wills et al. found that stable body weight w a s common in relatively young PD case s who participated in a PD clinical trial 8 . Two other studies with small sample sizes also repo rted a stable weight among PD case s over 1 - 3 years of follow - up 41,42 . This inconsistency across studies may in part be explained by differences in study design, case characteristics, sample size, l ength of 38 follow - up, treatment strategies used for case s, and lack of repeated objective assessments of body weight. Several studies further suggest that body weight in PD case s may begin to decline prior to disease diagnosis. In the Health Professionals Follow - up Study (HPFS) and the by 468 PD case s which were compared to participants who had not developed PD using similar methods as in this study. They found that PD case s began to lose weight 2 - 4 years before PD diagnosis despite accompanying increased energy intake 4 and decreased physical activities 43 . These trends persisted after diagnosis. To the best of our knowledge, this is the only longitudinal study in which the body weight s of PD case s were repeatedly assessed prior to disease diagnosis. The finding is supported by another small clinical study of 49 PD case s with body weight prior to disease diagnosis abstracted from medical records. The authors reported PD case s had a mean weight loss of 1.19% in an average period of 2.4 years prior to the time of PD diagnosis 6 . Ther efore, the existing evidence, while limited, suggests that weight loss in PD case s may start in the prodromal stage. Compared to previous studies, the current study has several notable strengths. Like the HPFS and NHS, the Health ABC study repeatedly meas ured body weight of PD case s up to about a decade before and after the diagnosis. Beyond HPFS and NHS, body weight in Health ABC was not self - reported but more accurately assessed using DXA assays along with measures of body fat and lean mass, allowing ana lyses to examine differential changes in body composition compartments. Our study confirmed that a persistent weight loss in PD case s likely starts a few years prior to PD clinical diagnosis. We further found 39 that the loss was almost entirely due to the lo ss of fat mass. This observation of a preferential loss of fat mass in PD case s is supported by some but not all of the earlier small clinical studies. For example, in one study, 19 of 26 PD case s lose weight over a year which was mainly due to loss of fat mass as measured by DXA 7 . In contrast, another study of 58 PD case s reported gain in average body weight and fat mass over 3 years measured by bioelectrical impedance analysis 41 . Compared to this existing l iterature, our study is population - based, had a longer follow - up, repeatedly assessed body composition using DXA, and our finding of a persistent loss of fat mass in PD case s is robust. Although the exact reasons for weight loss in PD are yet to be identified, it may relate to the complex symptoms and signs that develop as PD pathogenesis progresses. For example, in the prodromal stage of PD, poor olfaction may adversely affect the diet a nd nutrition which in turn gradually leads to changes in body composition. In support of this, Purdy et al. reported that poor olfaction was associated with faster weight loss in older adults 44 . Experiments in transgenic mice have also shown poor olfaction to trigger a metabolic response that leads to increased catabolic energy utilization and a subsequent loss of body weight 45 . As PD progresses, its motor signs and complications such as tremor 46 , muscle rigidity 46,47 , and Levodopa - induced dyskinesias 48 may further lead to increased energy expenditure and thus weight loss. Throughout the course of PD, these may further be complicated by other PD symptoms and signs such as depression 49 , gastrointestinal disorders 50 , cognitive impairment 10,51 , dysphagia 52 , and adverse events of dopaminergic treatment 51,53 , all of which may l ead to a progressive and persistent weight loss in PD case s. The preferential loss of fat mass in PD is also intriguing, and may relate to accelerated biological aging 54,55 which is associated with the reduction of subcutaneous 40 fat and the deposition of fat in non - adipose tissues in late adulthood by reduc ing adipogenesis through age - related activation of cellular stress response pathways and increased preadipocyte cytokine generation 56,57 . Further investigations are warranted to investigate the causes of weight loss and pre ferential loss of fat mass in PD case s. This persistent weight loss in PD case s noted above should not be neglected as it may have important adverse health consequences. In PD case s, w eight loss is often associated with higher Hoehn & Yahr stage 9 d isease rating scale 8 , the lower density of nigrostriat al dopaminergic neurons 58 , lower cognitive function 9 11 , declined quality of life 9 , a higher number of comorbidities 9 , and increased risk of dependency and mortality 5 . Moreover, the reduction of fat mass and the subsequent redistribution of adipoc ytes into muscle tissues may result in a higher risk of reduced muscle function and the development of frailty in PD case s. In support of this possibility, several studies reported lower fat mass was associated with more severe motor impairment in PD case s 35,59 . Further, one recent study found that PD case s had a higher MRI measured fat content in the bilateral psoas and thigh muscles than their age - and sex - matched healthy controls, which was associated with disease severity and frailty 12 . Finally, a preferential loss of fat tissue in PD case s may also contribute to PD pathogenesis itself. As a storage tissue for the lipid - soluble neurotoxicants such as organochlorine pesticides, the loss of fat mass may release such neurotoxicants to circulation 60,61 which may in turn contribute to the progressive loss of dopaminergic neurons in PD 62,63 . Therefore, changes in weight and body composition in PD case s may inform disease progression and prognosis, offering an opportunity to improve the health and survival of PD case s. 41 The pres ent study has several limitations. First, although our study identified 81 PD case s, their diagnosis was at different time points of the follow - up, and therefore the actual sample sizes were relatively small and varied at each specific time point in refere nce to PD clinical diagnosis. Nevertheless, our study revealed persistent weight loss in PD case s which was highly statistically significant. Second, the Health ABC study participants were old at enrollment (range 70 - 79). As body composition changes is par t of the aging process, our findings may not be readily generalizable to younger PD case s. Third, both PD diagnostic adjudication and time of diagnosis were retrospectively adjudicated based on cohort data collection, and thus inadvertent errors are likely . However, our findings of persistent weight loss and when it starts are very consistent with those of HPFS and NHS, which prospectively adjudicated PD diagnosis and further conducted rigorous medical record reviews by gathering diagnostic information from study participants and their treating physicians 4,43 . Fourth, we were unable to account for energy intake and expenditure in our analyses due to limited data availability because such information is often difficult to capture in large populations of older adults. In summary, in this longitudinal cohort of older adults with objective and repeated measurement of body composition, we found a per sistent loss of fat mass in PD case s that started a few years before disease diagnosis. Future studies are needed to understand the underlying mechanisms of this change in the body composition of PD case s and its potential impacts on disease progression an d survival. 42 REFERENCES 43 REFERENCES 1. 2020. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6039505/ 2. The Lancet . 2015;386(9996):896 - 912. doi:10.1016/S0140 - 6736(14)61393 - 3 3. Beyer PL, Palarino MY, Michalek D, Busenbark K, Koller WC. Weight Change and Body Co mposition in Case J Am Diet Assoc . 1995;95(9):979 - 983. doi:10.1016/S0002 - 8223(95)00269 - 3 4. disease. Ann Neurol . 2003;53(5):676 - 679. doi:10.1002/ana .10577 5. Cumming K, Macleod AD, Myint PK, Counsell CE. Early weight loss in parkinsonism predicts poor outcomes. Neurology . 2017;89(22):2254 - 2261. doi:10.1212/WNL.0000000000004691 6. Uc EY, Struck LK, Rodnitzky RL, Zimmerman B, Dobson J, Evans WJ. Predictors of Mov Disord Off J Mov Disord Soc . 2006;21(7):930 - 936. doi:10.1002/mds.20837 7. Lorefält B, Toss G, Granérus A - K. Weight loss, body fat mass, and leptin in Mov Disord . 2009;24(6):885 - 890. doi:10.1002/mds.22466 8. Wills A - MA, Pérez A, Wang J, et al. Association Between Change in Body Mass Persons With Parkinson Disease. JAMA Neurol . 2016;73(3):321 - 328. doi:10.1001/jamaneurol.2015.4265 9. Akbar U, He Y, Dai Y, et al. Weight Loss and Impact on Quality of Life in PLoS ONE . 2015;10(5). doi:10.1371/journal.pone.0124541 10. Kim HJ, Oh ES, Lee JH, et al. Relationship between changes of body mass index Arch Gerontol Geriatr . 2012;55(1):70 - 72. doi:10.1016/j.archger.2011.06.022 11. De Lucia N, Peluso S, Esposito M, et al. Frontal def ect contribution to decreasing case s. J Clin Neurosci Off J Neurosurg Soc Australas . 2020;72:229 - 232. doi:10.1016/j.jocn.2019.11.034 12. Chiang P - L, Chen Y - S, Lin AW - C. Altered Body Composition of Psoas and Thigh Int J Environ Res Public Health . 2019;16(19). doi:10.3390/ijerph16193667 44 13. Logroscino G, Sesso HD, Paffenbarger RS, Lee I - M. Body Mass Index and Risk ve Cohort Study. Am J Epidemiol . 2007;166(10):1186 - 1190. doi:10.1093/aje/kwm211 14. Jeong S - M, Han K, Kim D, Rhee SY, Jang W, Shin DW. Body mass index, Mov Disord Off J Mov Disord Soc . 2020;35(2):236 - 244. doi :10.1002/mds.27922 15. National Institute on Aging. Introducing the Health ABC Study: The Dynamics of Health, Aging, and Body Composition. Accessed March 3, 2020. https://healthabc.nia.nih.gov/ 16. Visser M, Fuerst T, Lang T, et al. Validity of fan - beam dual - energy X - ray absorptiometry for measuring fat - free mass and leg muscle mass. J Appl Physiol . 1999;87(4):1513 - 1520. doi:10.1152/jappl.1999.87.4.1513 17. Salamone LM, Fuerst T, Visser M, et al. M easurement of fat mass using DEXA: a validation study in elderly adults. J Appl Physiol . 2000;89(1):345 - 352. doi:10.1152/jappl.2000.89.1.345 18. Murphy RA, Patel KV, Kritchevsky SB, et al. Weight change, body composition, and risk of mobility disability a nd mortality in older adults: a population - based cohort study. J Am Geriatr Soc . 2014;62(8):1476 - 1483. doi:10.1111/jgs.12954 19. Schoeller DA, Tylavsky FA, Baer DJ, et al. QDR 4500A dual - energy X - ray absorptiometer underestimates fat mass in comparison wi th criterion methods in adults. Am J Clin Nutr . 2005;81(5):1018 - 1025. doi:10.1093/ajcn/81.5.1018 20. National Institute on Aging. Year 1 documentation. https://healthabc.nia.nih.gov/sites/default/files/Y1_Documentation.pdf 21. Chen H, Shrestha S, Huang X , et al. Olfaction and incident Parkinson disease in US white and black older adults. Neurology . 2017;89(14):1441 - 1447. doi:10.1212/WNL.0000000000004382 22. - analysis. Stat Methods Med Res . 1993;2(2): 121 - 145. doi:10.1177/096228029300200202 23. Lau J, Ioannidis JPA, Schmid CH. Quantitative Synthesis in Systematic Reviews. Ann Intern Med . 1997;127(9):820 - 826. doi:10.7326/0003 - 4819 - 127 - 9 - 199711010 - 00008 24. Marín - Martínez F, Sánchez - Meca J. Weighting by Inverse Variance or by Sample Size in Random - Effects Meta - Analysis. Educ Psychol Meas . 2010;70(1):56 - 73. doi:10.1177/0013164409344534 45 25. Cantor RM, Lange K, Sinsheimer JS. Prioritizing GWAS results: A review of statistical methods and recommendations fo r their application. Am J Hum Genet . 2010;86(1):6 - 22. doi:10.1016/j.ajhg.2009.11.017 26. Borenstein M, Hedges LV, Higgins JPT, Rothstein HR. A basic introduction to fixed - effect and random - effects models for meta - analysis. Res Synth Methods . 2010;1(2):97 - 111. doi:10.1002/jrsm.12 27. Cochran WG. The Combination of Estimates from Different Experiments. Biometrics . 1954;10(1):101. doi:10.2307/3001666 28. Hedges LV, Vevea JL. Fixed - and random - effects models in meta - analysis. Psychol Methods . 1998;3:486 - 504. 29. Smith VK, Hall TW. A Comparison of Maximum Likelihood Versus Blue Estimators. Rev Econ Stat . 1972;54(2):186 - 190. doi:10.2307/1926280 30. Lee CH, Cook S, Lee JS, Han B. Comparison of Two Meta - Analysis Methods: Inverse - Variance - Weighted Average and We ighted Sum of Z - Scores. Genomics Inform . 2016;14(4):173 - 180. doi:10.5808/GI.2016.14.4.173 31. Quené H, van den Bergh H. Examples of mixed - effects modeling with crossed random effects and with binomial data. J Mem Lang . 2008;59(4):413 - 425. doi:10.1016/j.jm l.2008.02.002 32. Baayen RH, Davidson DJ, Bates DM. Mixed - effects modeling with crossed random effects for subjects and items. J Mem Lang . 2008;59(4):390 - 412. doi:10.1016/j.jml.2007.12.005 33. Verbeke G, Molenberghs G. Linear Mixed Models for Longitudina l Data . Corrected edition. Springer; 2001. 34. Rabe - Hesketh S, Skrondal A. Multilevel and Longitudinal Modeling Using Stata, Third Edition . Stata Press; 2012. 35. Tan AH, Hew YC, Lim S - Y, et al. Altered body composition, sarcopenia, frailty, and their cl inico - Parkinsonism Relat Disord . 2018;56:58 - 64. doi:10.1016/j.parkreldis.2018.06.020 36. Suzuki K, Okuma Y, Uchiyama T, et al. Determinants of Low Body Mass Index in Case ter Case - Control Study. J Park Dis . 2020;10(1):213 - 221. doi:10.3233/JPD - 191741 37. Davies KN, King D, Davies H. A study of the nutritional status of elderly case s Age Ageing . 1994;23(2):142 - . 46 38. Marck MA van der, Dicke HC, Uc E disease: A meta - analysis. Parkinsonism Relat Disord . 2012;18(3):263 - 267. doi:10.1016/j.parkreldis.2011.10.016 39. Barichella M, Cereda E, Cassani E, et al. Dietary habits and neurological sease case s: Implications for practice. Clin Nutr . 2017;36(4):1054 - 1061. doi:10.1016/j.clnu.2016.06.020 40. Vardi J, Oberman Z, Rabey I, Streifler M, Ayalon D, Herzberg M. Weight loss in case s treated long - term with levodopa. Metabolic aspects. J Neurol S ci . 1976;30(1):33 - 40. doi:10.1016/0022 - 510x(76)90253 - 7 41. Vikdahl M, Carlsson M, Linder J, Forsgren L, Håglin L. Weight gain and Clin Nutr . 2014;33(6):1132 - 1139. doi:10.1016/j.clnu.2013 .12.012 42. disease. Nutr Neurosci . 2016;19(1):11 - 20. doi:10.1179/1476830515Y.0000000044 43. Chen H, Zhang SM, Schwarzschild MA, Hernán MA, Ascherio A. Physical activity and t he risk of Parkinson disease. Neurology . 2005;64(4):664 - 669. doi:10.1212/01.WNL.0000151960.28687.93 44. Purdy F, Luo Z, Gardiner JC, et al. Olfaction and changes in body composition in a large cohort of older US adults. J Gerontol A Biol Sci Med Sci . Published online April 8, 2020. doi:10.1093/gerona/glaa085 45. Riera CE, Tsaousidou E, Halloran J, et al. The Sense of Smell Impacts Metabolic Health and Obesity. Cell Metab . 2017;26(1):198 - 211.e5. doi:10.1016/j.cmet.2017.06.015 46. Levi S, Cox M, Lugon M, Hodkinson M, Tomkins A. Increased energy expenditure in Parki BMJ . 1990;301(6763):1256 - 1257. 47. Markus HS, Cox M, Tomkins AM. Raised resting energy expenditure in Clin Sci . 1992;83(2):199 - 204. doi:10.1042/cs0830199 48. Bachmann CG, Zapf A , Brunner E, Trenkwalder C. Dopaminergic treatment is associated with decreased body weight in case dyskinesias. Eur J Neurol . 2009;16(8):895 - 901. doi:10.1111/j.1468 - 1331.2009.02617.x 49. Pilhatsch M, Kroemer NB, Schneider C, et al. Reduced body mass index in J Nerv Ment Dis . 2013;201(1):76 - 79. doi:10.1097/NMD.0b013e31827ab2cc 47 50. Fasano A, Visanji NP, Liu LWC, Lang AE, Pfeiffer RF. Gastrointestinal dysfunction in Park Lancet Neurol . 2015;14(6):625 - 639. doi:10.1016/S1474 - 4422(15)00007 - 1 51. Wills A - M, Li R, Pérez A, Ren X, Boyd J, on behalf of the NINDS NET - PD NET - PD LS - 1 cohort. J Neurol . 2017;264(8):1746 - 1753. doi:10.1007/s00415 - 017 - 8562 - 4 52. Nozaki S, Saito T, Matsumura T, Miyai I, Kang J. [Relationship between weight loss and dysphagia in case Rinsho Shinkeigaku . 1999;39(10):1010 - 1014. 53. Pålhagen S, Lorefält B, Carlsson M, et al. Does L - dopa treatment contribute to reduction in body weight in elderly case Acta Neurol Scand . 2005;111(1):12 - 20. doi:10.1111/j.1600 - 0404.2004.00364.x 54. Hou Y, Dan X, Babbar M, et al. Ageing as a risk factor for neurodegenerative disease. Nat Rev Neurol . 2019;15(10):565 - 581. doi:10.1038/s41582 - 019 - 0244 - 7 55. Horvath S, Ritz BR. Increased epigenetic age and granulocyte counts in the case s. Aging . 201 5;7(12):1130 - 1142. 56. Cartwright MJ, Tchkonia T, Kirkland JL. Aging in adipocytes: Potential impact of inherent, depot - specific mechanisms. Exp Gerontol . 2007;42(6):463 - 471. doi:10.1016/j.exger.2007.03.003 57. Kirkland JL, Tchkonia T, Pirtskhalava T, Ha n J, Karagiannides I. Adipogenesis and aging: does aging make fat go MAD? Exp Gerontol . 2002;37(6):757 - 767. doi:10.1016/S0531 - 5565(02)00014 - 1 58. Lee JJ, Oh JS, Ham JH, et al. Association of body mass index and the depletion of nigrostriatal dopamine in P Neurobiol Aging . 2016;38:197 - 204. doi:10.1016/j.neurobiolaging.2015.11.009 59. Pisciotta MS, Fusco D, Grande G, et al. Untangling the relationship between fat Aging Cli n Exp Res . Published online March 15, 2019. doi:10.1007/s40520 - 019 - 01166 - x 60. Imbeault P, Chevrier J, Dewailly É, et al. Increase in plasma pollutant levels in response to weight loss in humans is related to in vitro subcutaneous adipocyte basal lipolysi s. Int J Obes . 2001;25(11):1585 - 1591. doi:10.1038/sj.ijo.0801817 61. La Merrill Michele, Emond Claude, Kim Min Ji, et al. Toxicological Function of Adipose Tissue: Focus on Persistent Organic Pollutants. Environ Health Perspect . 2013;121(2):162 - 169. doi:1 0.1289/ehp.1205485 48 62. Baltazar MT, Dinis - Oliveira RJ, de Lourdes Bastos M, Tsatsakis AM, Duarte JA, other neurodegenerative diseases -- a mechanistic approach. Toxicol Lett . 2014;230(2):85 - 103. doi:10.1016/j.toxlet.2014.01.039 63. Van Maele - Fabry G, Hoet P, Vilain F, Lison D. Occupational exposure to - analysis of cohort studies. Environ Int . 2012;46:30 - 43. doi:10 .1016/j.envint.2012.05.004