12.5 Using models well

• The model is an assumption. A regression coefficient adjusts for confounding only as well as the model describes the dependence of risk on the confounders. A continuous confounder entered as a straight line when its effect is curved leaves residual confounding; categorising it, or using splines, is safer.
• Choose confounders by causal reasoning, as in Section 7.5: adjust for common causes, not for mediators or colliders. Automatic selection by significance tests is not a substitute.
• Check for effect modification with product terms and report stratum-specific estimates when it is present.
• Sparse data distort maximum likelihood estimates away from the null when there are few events per parameter; with many matched sets or strata, conditional methods are needed.
• Trends. For an exposure with ordered levels, entering it as a single score tests for a dose–response trend with one degree of freedom, more powerful against a monotone effect than the test with one indicator per level.

Questions on this section

Stuck on something here? Ask below and it stays attached to this topic.