Application Of GEE And MLM To Case-crossover Study

Case-crossover study is widely accepted as it can control time-invariant confounders from the perspective of design, as well as control time-varying exposures by using suitable referents.Traditionally, data from case-crossover studies were analyzed as matched case- control studies using conditional logistic regression. However, in case-crossover study, some subjects may be also affected in the reference periods and subjects may forget the exact exposures of reference periods, which violate the assumptions of conditional logistic regression. From a statistical point of view, case-crossover designs can be regarded as repeated measures on the individual over time and outcomes are non-independent of course. For the above reasons, the effects of non-independent data, affected in the reference periods and missing data should be fully awared and considered in parameter estimation and hypotheses test in the analysis of case-crossover study data.In the present paper, a serial of numerical simulation studies were performed to compare the statistics features of GEE, MLM and CLR, which can be used to analyze the data from case-crossover studies. Methods are compared across simuated datasets with different correlation among exposures, different sample size, different disease rate, and different OR of expersure. 1000 data sets were generated for each combination of design parameter settings. The main contents of this study are listed as follow:(1) Assessments of three methods with large sample and no missing dataSimulated case-crossover study dataset is used to compare the characteristics of statistics of GEE, MLM, CLR, with sample sizes being 500.(2) Assessments of three methods with small sample and no missing dataSimulated case-crossover study data is used to compare the characteristics of statistics of GEE, MLM, CLR, with sample sizes being 100.(3) Assessments of three methods with large sample and missing dataSimulated case-crossover study data is used to compare the characteristics of statistics GEE, MLM, CLR, with sample sizes being 500.(4) Case studyA case-crossover study was conducted to assesse short-term exposure risk factors of idiopathic ventricular tachycardia (IVT). GEE, MLM, CLR were illustrated with this data.The main results of this study are summarized as follow:(1) When the data from case-crossover study meets the requirement of matched case-control study (disease is rare, the exposure correlation is small, no missing data, large sample), CLR is recommended.(2) When the data of case-crossover study does not meet the requirement of matched case-control study, especially, due to higher disease rate, higher exposure correlation, missing data, or with sevel reference periods for a individual, GEE and MLM are recommended. And the results of GEE are quite similar to ones ofMLM in these situations.(3) When the exposure correlation is very high, parameter estimation and hypotheses test of GEE, MLM and CLR all don't work. Hereby, it is necessary to avoid the high correlation of exposure as far as possible when a period is destined for the reference. Otherwise, we should use case-control study or other designs.