Multivariable methods are computationally complex and generally require the use of a statistical computing package. Multivariable methods can be used to assess and adjust for confounding, to determine whether there is effect modification, or to assess the relationships of several exposure or risk factors on an outcome simultaneously. Multivariable analyses are complex, and should always be planned to reflect biologically plausible relationships. While it is relatively easy to consider an additional variable in a multiple linear or multiple logistic regression model, only variables that are clinically meaningful should be included.
It is important to remember that multivariable models can only adjust or account for differences in confounding variables that were measured in the study. In addition, multivariable models should only be used to account for confounding when there is some overlap in the distribution of the confounder each of the risk factor groups.
Stratified analyses are very informative, but if the samples in specific strata are too small, the analyses may lack precision. In planning studies, investigators must pay careful attention to potential effect modifiers. If there is a suspicion that an association between an exposure or risk factor is different in specific groups, then the study must be designed to ensure sufficient numbers of participants in each of those groups. Sample size formulas must be used to determine the numbers of subjects required in each stratum to ensure adequate precision or power in the analysis.