Different Statistical Models to Analyze Epidemiological Observational Longitudinal Data: An Example from the Amsterdam Growth and Health Study

With the development of new statistical techniques [such as generalized estimating equations (GEE)] it became possible to analyze longitudinal epidemiological relations, using all available longitudinal data. However, there are different possibilities in modeling longitudinal relations. In this paper four possible models were compared. (1) A simple model in which the actual values of the outcome and predictor variables were related (Y_it = β₀ + β₁X_it...); (2) A model with a time lag between outcome and predictor variables (Yit = (β₀ + β₁X_it-1...); (3) A model in which not the actual values, but changes in values between different time points were related ([Y_it - Y_{it - 1}] = β₀ + β₁, [X_it - X_{it - 1}]...); and (4) A first-order autoregressive model in which the actual value of the outcome variable at time point t is not only related to the actual value of the predictor variable at time point t, but also to the value of the outcome variable at t - 1 (Y_it = β₀ + β₁X_it + β₂Y_{it - 1} +...). In this paper the use of the possible models was discussed by means of an example with data from the Amsterdam Growth and Health Study. In this longitudinal observational study six repeated measurements were carried out over a period of 15 years on subjects with an initial age of 13 years. It can be concluded that each model reflects different parts of the longitudinal relationships and the choice for a particular model must be based on logical considerations. However, in most cases epidemiologists should use the results of different models to obtain a more accurate answer to the particular epidemiological question.

Amsterdam Growth and Health Study - generalized estimating equations - longitudinal studies - statistical models