Applying Multilevel Models In Analysis Of Risk Factors For Peripheral Arterial Disease And Safety And Efficacy Of Drug In Phase Ⅳ Clinical Trails

Objective This application study sought to explore the methods of aplicating multilevel models in analysis of risk factors for peripheral arterial disease and efficacy and safety of drug in phase IV clinical trails, so as to provide the reference for analyzing the hierarchical structure cross-sectional data of clinical epidemiologic studies and repeated measure longitudinal data of phase IV clinical trials.Methods On the basis of the summary of the principles and building steps for multilevel linear model, multilevel logistic regression model and multilevel linear growth model, the PROC MIXED and PROC NLMIXED in SAS software were employed to analyze analyze the mediate effect of group-level variables on the relation of individual-level variables and depend variables in the cross-sectional data, evaluate the variation of safety and efficacy indicators in phase IV clinical trials repeated measure longitudinal data, and the effect of intra-individual correlation and between-individual variation on the interested indicators.Results In the cross-sectional study on the prevalence of abnormal ankle brachial index (ABI) among high risk hypertensive patients, there were statistical significant intra-class correlation and between-group heterogeneity in the data of dependent variables ABI value and abnormal ABI (P<0.05). These data were not adapted to multiple linear regression and fixed effect logistic regression, they would adopt multilevel linear model and multilevel logistic regression model. After adding group-variable research center location to the empty models, the intra-class correlation decreased. The group-variable could explain15%between-group heterogeneity of ABI value and46%between-group heterogeneity of abnormal ABI; it indicated that there were region cluster in the distribution of ABI value and abnormal ABI. The results of multilevel models fitted by group-variable research center location, individual-level variabla aged and other individual background variables indicated that there were across-level interactions between research center location and aged, and the main risk factors for ABI value decrease were sex, slightly elevated serum creatinine, smoked, body mass index (BMI) and pulse pressure; the main risk factors for abnormal ABI were arterial wall thickening, slightly elevated serum creatinine and very high risk hypertension(P <0.05)The data of pitavastatin phase IV clinical trail were incomplete, unbalanced and unequal period repeated measure longitudinal data. These data were not adapt to univariate repeated measure analysis of variance. The information of data were lost if they were analyzed by multivariate repeated measure analysis of variance. Multilevel linear model, however, could take full advantage of data information. The power of multilevel linear model was high. The results of multilevel linear model were as follow:Firstly, during the period of exposure on pitavastatin, the average estimate values of aspartate transaminase (AST), alanine transaminase (ALT) and creatine kinase (CK) increased. The higher baseline values of ALT or CK had smaller increase rate (P<0.05). The main influential factors for AST outcome value were age, drink and baseline ALT value. Those for ALT outcome value were age, sex, baseline BMI value and baseline AST value. Those for CK outcome value were sex, baseline creatinine (Cr) value and baseline AST value. Secondly, during the period of pitavastatin treatment, the average estimate values of total cholesterol (TC), low density lipoprotein cholesterol (LDL-C) and triglyceride (TG) reduced, that of high density lipoprotein cholesterol (HDL-C) increased. The higher baseline values of TC or LDL-C had bigger decrease rate, the higher baseline values of HDL-C had smaller increase rate (P<0.05). The main influential factors for TC outcome value were sex, baseline Cr value and baseline AST value. Those of LDL-C outcome value were age, baseline HDL-C value, baseline TG value and baseline TC value. Those of HDL-C outcome value were sex, baseline TC value, baseline LDL-C valure and baseline TG value. Those of TG outcome value were risk level of abnormal cholesterol, baseline BMI value, baseline Cr value, baseline TC value, baseline LDL-C value and baseline HDL-C value (P<0.05).Conclusion While there are statistical significant intra-class correlation or between-group heterogeneity in the data of clinical epidemio logic studies, multilevel models such as multilevel linear model and multilevel logistic regression model must be adopted. For the incomplete, unbalanced and unequal period repeated measure data collected in phase Ⅳ clinical trails, multilevel models such as multilevel linear growth model must be adopted.