A statistically correct result can still be explained badly. Select an estimate to see what it means, how to report it and what not to claim.
ESTIMATE · UNCERTAINTY · SUBSTANTIVE MEANINGINTERPRETATION LAB
The examples use hypothetical values. Good interpretation also requires knowledge of coding, units, reference groups and study design.
LINEAR REGRESSION
After adjustment for the variables in the model, the exposed group had a mean outcome 2.4 units lower than the unexposed group (95% CI: 1.7 to 3.1 units lower).
Exposure reduced the outcome by 2.4 units—unless the research design supports a causal interpretation.
Check: outcome units, reference category, functional form and whether adjustment variables were pre-specified.
LOGISTIC REGRESSION
After adjustment, the odds of the outcome were 65% higher in Group B than in Group A (OR 1.65, 95% CI: 1.21–2.25).
Group B was 65% more likely to experience the outcome. Odds and probability are not interchangeable.
Check: outcome prevalence, reference group and adjusted predicted probabilities.
COUNT OR RATE MODEL
The adjusted event rate in Group B was 22% lower than the rate in Group A (IRR 0.78, 95% CI: 0.66–0.92).
Group B had 22% fewer people with the outcome. An IRR concerns a count or rate, not necessarily the proportion of people affected.
Check: exposure time, overdispersion, excess zeroes and whether the model estimates counts or rates.
EFFECT MODIFICATION
The adjusted difference associated with exposure was 4 percentage points in Group A and 11 percentage points in Group B—a 7-point larger difference in Group B.
The interaction coefficient proves that Group B is more affected. The scale, uncertainty and underlying predicted values all matter.
Useful STATA step: estimate adjusted predictions with margins, followed by contrasts or marginal effects on a meaningful scale.
STATISTICAL UNCERTAINTY
The estimated difference was −0.8 units, with a 95% confidence interval from −1.3 to −0.3. The interval is compatible with effects ranging from a small to a moderately lower outcome.
There is a 95% probability that the true value lies inside this particular interval.
Check: precision and substantive importance, not only whether the interval crosses the null value.
RETURN TO THE ANALYSIS