13.5 Choosing a method
| Method | Use when | Avoid when |
| Direct standardisation | comparing several populations with stable stratum rates, especially for presentation | strata are small or rates unstable |
| Indirect standardisation (SMR) | one small population against reference rates that are known well | comparing SMRs of populations with different structures |
| Inverse-variance pooling | a few large strata; meta-analysis of published estimates | many strata or zero cells |
| Mantel–Haenszel | any number of strata, large or small; a quick, robust summary | effect modification is present (report strata instead) |
| Conditional likelihood | matched or finely stratified data; several exposures | – |
| Regression (maximum likelihood) | several confounders, continuous variables, trends, effect modification | strata are sparse and modelled by separate intercepts |
Whatever the method, three steps come first: look at the stratum-specific estimates, ask whether they are homogeneous enough to be summarised, and compare the adjusted with the crude estimate to judge the confounding. When the stratum estimates agree, the choice of method matters little, as Example 13.2.1 shows. When they disagree, no method rescues a single summary, as Example 13.3.1 shows. The methods differ seriously only with sparse data, where Mantel–Haenszel and conditional methods are the safe choice.
Questions on this section
Stuck on something here? Ask below and it stays attached to this topic.