When should you use logistic regression?
Use it when the outcome has two categories, with predictors named before the tables, and an interpretation that matches the estimand.
Logistic regression is appropriate when the outcome has two categories and the research question is about odds or adjusted associations. It is not a default for every survey.
When the outcome fits
Use it when each unit falls into one of two categories: yes or no, present or absent, survived or died. The model estimates how predictors change the log odds of that outcome. If the research question is about a continuous measure, a count, or a time to event, another family of models is usually the right starting point.
Write the analysis plan before the tables. Name the outcome, the predictors that belong in the model because of the design or the theory, and the confounders you will adjust for. Searching a long list of variables until a p value appears is not a method.
What to check
Check the data against the assumptions the model needs. Events should not be vanishingly rare relative to the number of predictors. Predictors should not be so closely related that the model cannot separate them. Observations should match the sampling design: a clustered survey usually needs a method that respects clusters.
Coefficients on the log odds scale are easy to misread. Report them in a form a reader can use, often as odds ratios with intervals, and say what is held constant. An odds ratio is not a risk ratio. In a common outcome the two can differ a lot. Do not write "more likely" if you mean "higher odds" unless the approximation is justified.
Logistic regression does not prove a cause. Adjustment can remove some confounding. It cannot repair a design that never compared the groups of interest.
When to choose another model
If the outcome has more than two unordered categories, or is ordered, or is the result of a matching design, the binary logit is the wrong instrument. Choose the model that matches the outcome and the sampling.
A methods review asks whether the outcome is truly binary, whether the predictors were specified in advance, whether diagnostics were run, and whether the interpretation matches the estimand. Those questions matter more than whether the software produced stars.
