Statistical methods for describing temporal order in longitudinal research

Although traditional epidemiological statistical methods (e.g., logistic regression) are useful for describing predictive relationships in longitudinal panel studies in which changes in risk factor levels occur before change in the outcome variable, more sophisticated statistical methods must be used when the temporal order between variables is unknown.

Using national survey data, the current study shows how log-linear models and discrete-time survival analysis can be used to test for temporal order. The relationship between marijuana use and friends' use of marijuana is examined to illustrate these methods.

Using traditional analytic strategies, it appears that friends use and marijuana use are predictive of each other. However, valid tests for temporal order reveal that both variables change concurrently, so there is no temporal order between these variables; rather, these variables tend to change concurrently.

In many current areas of research, temporal order between theoretically important variables is unknown and traditional analytic strategies will yield misleading results. The fundamental problem with prior approaches is that no estimate of concurrent change is made. Without an estimate of concurrent change, estimates of prediction will be biased. The current study illustrates valid methods that can be used to describe temporal orderings.