Possible Biased Inferences in Tests for Average Partial Association

A set of data relating to the possible effect of cardiovascular disease on driving records is analyzed in several steps.The results of the analysis illustrate how differing conclusions would have been reached if the information about other variables had been less detailed or if the analysis had been stopped before its completion.The importance of proper controlling for covariables, when one is assessing the average partial association between two variables, is emphasized.It is suggested that tests for collapsibility should be performed before the condensing of the original table.Some considerations concerning the selection of relevant covariables are also discussed.