| BackgroundWith the rapid development of the economy and society,and profound changes in the population’s lifestyles,the life expectancy of the global population has been rising,and the aging problem has become increasingly severe,especially in developing countries.The National Aging Development Bulletin 2020 reports that the proportion of China’s elderly population aged 65 and above has increased to 13.50%,indicating that the aging scale in China is expanding and the process is significantly accelerating.It has become an important national strategic plan to actively cope with population aging and promote healthy aging.Aging is the most critical risk factor or the common starting point for most chronic non-communicable diseases such as cardiovascular diseases,diabetes,cancers,and dementia.Clarifying the physiological mechanisms and modifiable risk factors of biological aging and making targeted interventions will help to simultaneously delay and even prevent the occurrence of numerous aging-related diseases or phenotypes and ultimately extend lifespan and healthspan,with more significant social and economic benefits than preventing and treating any single disease.Since behavioral lifestyles are important factors influencing biological aging,and lifestyle interventions have the highest feasibility and accessibility among various potential aging intervention strategies,it is necessary to systematically study and accurately estimate the causal effects of different behavioral lifestyle factors on biological aging and identify critical control points,which can provide scientific evidence for the development of targeted anti-aging interventions or guidelines.Most previous studies examining the association between lifestyle factors and biological aging have applied cross-sectional analyses or traditional cohort study paradigm,which are relatively limited in the strength of the causal evidence and fail to fully consider the time-varying fluctuations of lifestyle exposure and biological aging outcomes.In fact,the longitudinal（repeated measurement）data of the cohort can effectively reflect the covariant relationship of exposure,outcome and covariable in time,and has the structural advantage of self-control,which can provide a good data basis for the above research questions.The fixed effects model（FEM）developed in econometrics is a classical causal inference method dealing with panel data/longitudinal repeated measures data.Its self-control estimation strategy can eliminate all unmeasured time-invariant confounders,thereby significantly improving the validity of the estimates,which is suitable for the study of causal association between time-varying exposure and continuous outcome.FEM has excellent potential for application in longitudinal repeated measures data of cohort,and provides a feasible solution to estimate the causal effect of lifestyle factors on biological aging.However,FEM has two critical identification assumptions,the strict exogeneity（SE）assumption and the common trend（CT）assumption,both might face serious challenges in practical studies,such as the existence of temporal/sequential causal effects of the outcome or the presence of explicit unmeasured time-varying confounders in the studies.At this time,the FEM will fail.Most existing extension methods that loosen the SE or CT assumptions introduce additional parametric assumptions and are invalid when panel waves are very few.At the same time,most of the emerging cohorts are established in a short period,with a small number of waves,especially the biochemical and physical measurements data are only two waves;therefore,a new and robust method is urgently needed to meet the realistic research demands.ObjectivesThis study aims to address two research questions;the first one is a scientific question focusing on the accurate estimation of the causal associations between multiple lifestyle factors and biological aging and the relative contribution of each factor;the second one is a methodological question focusing on the methodological challenges that FEM may face in addressing similar scientific questions as described above.There are four specific research contents:（1）To construct a biological age（BA）index system for multiple organs or systems in the Southwest China to get a preliminary understanding of the status and characteristics of biological aging of the population in this region.（2）To estimate the causal associations and relative contributions of various lifestyle factors with biological aging using FEM and multiple-exposure analysis method to gain a preliminary understanding of the critical intervention targets.（3）To construct the non-parametric fixed effects model（NP-FEM）to relax the SE and CT assumptions,and compare it with the FEM by simulations.（4）To conduct an example application of NP-FEM to analyze the causal associations between lifestyle health levels and biological aging.MethodsThis study was based on collected data from the baseline survey in 2018-2019 and the first repeat survey in 2020 of the China Multi-Ethnic Cohort study,and conducted research and analysis from four aspects mentioned above,with detailed methods for each part as follows:（1）First,all the collected physical measurements and laboratory test biomarkers were screened.A five-step screening process of"evaluation of the reasonableness of markers→evaluation of the missing rate of markers→evaluation of the normality of markers and conversion→evaluation of the correlation between markers and chronological age（CA）→evaluation of the correlation between markers"was carried out to determine the fundamental biological indicators for the construction of BA.Second,the selected indicators was divided into multiple organs or systems according to their functions.The KD method（Klemera and Doubal’s method）was used to construct a series of BA index system by gender,including the overall BA,the cardiopulmonary system BA,the liver system BA,the kidney system BA,the metabolic system BA and the immunohematological system BA.Finally,the overall distribution and correlation between each BA in the population were described.（2）First,the data of five lifestyle factors,namely,smoking,alcohol consumption,dietary,physical activity,and sleep,were dichotomized according to the commonly used health definition criteria,and a comprehensive healthy lifestyle score/index（HLS/HLI）index was constructed.Next,baseline characteristics of the whole baseline population and the part of the population participating in the repeat survey were described and compared,and changes in exposure and outcomes between the two surveys were described.Next,cross-sectional analyses were performed using baseline whole population data,while FEM analyses were performed using data from the two waves to estimate the association between each single lifestyle factor as well as the combined healthy lifestyle level and the overall biological aging acceleration indicator AA（Ageing acceleration,the difference between BA and CA）,then compare the estimation results of the two methods.Finally,the first-differenced data were analyzed using quantile g-computation（QGC）,a multiple-exposure analysis method,to explore the relative contribution of the effects of each single lifestyle factor on biological aging.In addition,sensitivity analyses were further conducted by changing the health definition criteria for each lifestyle factor and the population inclusion and exclusion criteria to confirm the robustness of the results.（3）First,the difference-in-differences（DID）and changes-in-changes model（CIC）were introduced,and the basic idea of NP-FEM was proposed based on its similarity with FEM.Second,the NP-FEM method framework was entirely constructed by fully integrating the research problems and scenarios of interest in this paper and the CIC estimation idea.Then,the theoretical estimation performances obtained by the FEM and NP-FEM methods in four scenarios were justified:i)satisfying the SE and CT assumptions,ii)violating only the SE assumption,iii)violating only the CT assumption,and iv)violating both the SE and CT assumptions,and the sources of bias and expected performances of the effect estimates of the two methods are compared.Finally,under the setting of dichotomous exposure,continuous outcomes,and two-wave balanced longitudinal data,the actual estimated performance of the FEM and NP-FEM methods was compared by simulation studies for the four violation scenarios of SE and/or CT assumptions mentioned above and different violation levels.（4）Using a dichotomized HLS/HLI as an example,a case application of the NP-FEM method was performed to estimate the causal effect of a highly healthy lifestyle compared to a poorly healthy lifestyle on the acceleration of biological aging,with effect point estimates and confidence intervals calculated by a 1000 bootstrap.Results（1）After screening,a total of 15 physical measurements or biochemical markers of systolic blood pressure（SBP）,waist-hip ratio（WHR）,peak expiratory flow（PEF）,γ-glutamyl transpeptidase（GGT）,albumin（ALB）,low-density lipoprotein cholesterol（LDL-CH）,high-density lipoprotein cholesterol（HDL-CH）,triglycerides（TG）,Glutamic oxaloacetic transaminase（AST）,creatinine（Cr）,alkaline phosphatase（ALP）,urea（UREA）,mean corpuscular volume（MCV）,glycosylated hemoglobin（HBA1C）,and platelet count（PLT）were included in the BA construction.The descriptive analysis showed that the mean value of BA of each organ or system at baseline was equal to the mean value of CA,while that at repeated survey was almost greater than the mean value of CA.In addition,the degree of variation of different BA varied widely,depending on its correlation with CA,with the cardiopulmonary system BA having the slightest variation and the liver system BA having the most significant variation;meanwhile,the correlation between BA of each organ or system was significantly different,with the strongest correlation between the overall BA and the cardiopulmonary system BA（r was 0.925 in the baseline and was 0.917 in the repeated survey）and the weakest correlation between the kidney system BA and the liver system BA（r was 0.486 in the baseline and 0.387 in the repeat survey）.（2）The valid sample sizes were 66,990 for the baseline and 5,958 for the two matched waves data,and the distribution of each baseline characteristic of these two populations was similar.Between the two waves,the study population experienced differential changes in each single lifestyle factor and overall lifestyle,with the highest percentage of change in diet at 35.6%and the lowest percentage of change in alcohol consumption at 4.1%.Biological aging acceleration also changed significantly between the two waves,with an approximately normal distribution ofΔAA with a mean of 0.58and a standard deviation of 4.06.Cross-sectional analysis showed a relatively consistent anti-aging acceleration effect for both continuous and dichotomous HLS/HLI index,but keeping healthy in smoking and physical activity showed positive associations with AA.The FEM analysis showed a consistent slowing of biological aging acceleration across different single lifestyle factors and overall health lifestyle level index,but most single factors were not statistically significant.Specifically,the effect estimates for smoking,alcohol consumption,dietary,physical activity,and sleep were-0.052（95%CI:-0.457,0.353）,-0.389（95%CI:-0.896,0.117）,-0.193（95%CI:-0.366,-0.020）,-0.172（95%CI:-0.369,0.025）,and-1.488×10-5（95%CI:-0.183,0.183）,separately.Each one-point increase of HLS/HLI was associated with a mean decrease of 0.130（95%CI:-0.228,-0.031）in the AA.The negative association between dichotomized HLS/HLI and biological aging acceleration was stronger,indicating that a highly healthy lifestyle（HLS/HLI is 4-5）would significantly slow biological aging,causing a mean change in AA of-0.219（95%CI:-0.399,-0.038）.Further relative contribution analysis showed that alcohol consumption（48.3%）,diet（24.0%）,and physical activity（21.3%）were probably the most critical lifestyle components in slowing the acceleration of biological aging,with a combined effect contribution of 93.6%for all three.In addition,different definitions for healthy exposures and different population inclusion and exclusion criteria did not significantly affect the study results and conclusions,suggesting relatively robust results.（3）In the NP-FEM method construction and simulation comparison,the simulation results under each scenario agreed with the theoretical property inference.The main results and conclusions included:（i）In the scenario where the SE and CT assumptions are met,the FEM estimates were constant and unbiased and had slight variance;the NP-FEM estimates were unbiased when the common support assumption was met but had a more considerable variance than FEM,and was biased when it was not met,but the scale of bias was negligible.It is suggested that FEM is the superior analysis method in this scenario,but NP-FEM also performs well unless there is a severe departure from the common support assumption.（ii）Under the scenario of violating only the SE assumption,the FEM estimates were always biased,the bias is large or even lead to reversed effect,while the NP-FEM estimates were unbiased when the common support assumption was met and were biased but with a small scale when it was not met,so the NP-FEM had a very stable estimation performance.It is suggested that NP-FEM can effectively avoid or significantly reduce the bias resulting from violating the SE assumption,thus loosening the SE assumption.（iii）Under the scenario of violating only the CT assumption if there was a heterogeneous time trend associated with the inherent characteristics of individuals,the FEM estimates were always biased and the bias is large or even lead to reversed effect,while the NP-FEM estimates were unbiased when the common support assumption was met and were biased but significantly less than the FEM when it was not.If unmeasured time-varying confounders were present,the FEM and NP-FEM estimates were both biased,and the bias pattern and degree were very similar.It is suggested that NP-FEM is more effective in avoiding or reducing the bias caused by the heterogeneous time trend associated with the inherent characteristics of individuals,but it is not helpful for unmeasured time-varying confounders,so the CT assumption can only be loosened in some scenarios.（iv）In the scenario where both SE and CT assumptions were violated,the final bias was the combined result of the corresponding bias when the SE or CT assumptions were violated separately,with the total bias increasing if they were in the same direction and decreasing if they were in different directions.Overall,NP-FEM estimates performed better than FEM if the violation of the CT assumption was the presence of heterogeneous time trends associated with inherent individual characteristics,and it was difficult to identify and compare the final bias if the presence of unmeasured time-varying confounders caused the violation of the CT assumption.It is suggested that NP-FEM can simultaneously loosen the SE assumption and allow for heterogeneous time trends associated with inherent individual characteristics.（4）In the application analysis of the NP-FEM,the results showed that the transition from a highly healthy lifestyle at baseline to a less healthy lifestyle at follow-up caused an increase in AA（=0.185,95%CI:-0.154,0.541）,while the transition from a less healthy lifestyle at baseline to a highly healthy lifestyle at follow-up caused a decrease in AA（=-0.255,95%CI:-0.570,0.050）,with a combined effect estimate of-0.221（95%CI:-0.453,-0.003）,which was very similar to the corresponding FEM estimate（=-0.219,95%CI:-0.399,-0.038）.ConclusionIn the population of the China Multi-Ethnic Cohort,there showed a mild accelerated biological aging during two surveys,and the biological aging levels of different organs and systems were different.Maintaining health on some single lifestyle factors such as smoking,alcohol consumption,diet,physical activity,and sleep or improving the overall health lifestyle level can help to slow the acceleration of biological aging.Among these,alcohol consumption,diet and physical activity may be the most significant and crucial lifestyle components that contribute.The NP-FEM method allows for significant temporal/sequential causal effects of the outcome variables,and it also allows for heterogeneous time trends associated with inherent individual characteristics,which can,to some extent,loosen the SE and CT assumptions of the FEM method,thus broaden the applicability of the original method,moreover,in scenarios where both SE and CT assumptions are met,the NP-FEM method also has an almost comparable estimation performance with FEM.Therefore,the NP-FEM method may become another option for causal effect estimation based on longitudinal repeated measures data. |
PDF Full Text Request