To perform a comparative study of statistical models on datasets from nine large epidemiological studies in the cardiopulmonary field in order to develop tools for identifying appropriate classes of statistical models for use in analyzing longitudinal data.

BACKGROUND: Longitudinal designs are frequently encountered in epidemiologic research, particularly in the cardiopulmonary field. Many different statistical models have been proposed for the analysis of longitudinal data in the statistical literature. These include the general linear model, autoregressive models, random effects models, and simple models based on an analysis of slopes over time. Complex models are not widely used in the epidemiologic literature, due mainly to a lack of understanding of their underlying utility and the types of questions that could be answered with complex models that cannot be addressed using simple models. An additional problem is a lack of software available for fitting complex models. The study has important public health implications, since longitudinal data continue to accumulate rapidly and no guidelines are available as to the appropriate methods of analysis for specific research questions. Furthermore, it is often only through the modelling of longitudinal data that processes pertaining to change can be understood.

DESIGN NARRATIVE: The nine datasets used included: for pulmonary data, the Childhood Respiratory Disease Study, the Netherlands data from Vlagtwedde and Vlaardingen, the Boston Police Study, the Fletcher Study data from England; for cardiopulmonary data, the Veterans Administration Normative Aging Study; for blood pressure data, the Wales Study, the Zinner/Kass Study, the Lee/Zinner Study, and the East Boston Childhood Blood Pressure Study. For each of the nine datasets the following models were fitted and compared for adults, children, and for adults and children combined: autoregressive models both serial-correlation and state-dependence; random-effects models; regression models with intraclass correlation structure; general linear models; models based on fitting slopes to individual persons. New methods for analyzing longitudinal data were developed and included fitting of higher-order autoregressive models with unequally spaced data, nonparametric methods, familial and other clustering effects in the analysis of longitudinal pulmonary function data, robust methods, empirical Bayes methods for estimation of slopes, and hierarchial models based on old and new methods.

The study was renewed in 1996 to perform a comparative study of complex models on datasets from four large epidemiologic studies in the cardiopulmonary field. The models were compared as regards goodness of fit, ease of implementation, and interpretability. In addition, new statistical methods were developed to model phenomena which seemed poorly-fitted by existing models, including adult longitudinal bp and pulmonary function data. The overall goal was to develop tools for identifying appropriate classes of longitudinal statistical models. This has important public health implications, since longitudinal data continue to accumulate rapidly and no guidelines exist as to appropriate methods of analysis. Furthermore, it is often only through modelling of longitudinal data rather than through cross-sectional or separate two time-point analyses that underlying processes pertaining to change can be understood.

The study was renewed in February 2000 to extend and enhance several techniques in the analyses of longitudinal data frequently encountered in epidemiological studies. The techniques include: methods for control for time-dependent confounding in epidemiological studies; developing an incidence model for benign breast disease using the Nurses' Health Study; analysis of incomplete longitudinal data from the Normative Aging Study; extending the penalized likelihood procedures for quantile regression to the repeated measures setting; and development of methods to estimate correlated ROC curves to measure the predictive accuracy of GEE regression models for longitudinal data.

Genders Eligible for Study:  Male

Criteria