One of the most common criticisms reviewers raise during major revision cycles is this: "Confounding factors have not been controlled for." This criticism calls the study's core finding into question. Multivariable analysis solves this problem, but it must be set up correctly.
What Is a Confounding Factor?
A confounding factor is a variable that is associated with both the independent and the dependent variable and can distort the relationship between them.
A classic example: an association is found between coffee consumption and lung cancer. However, coffee drinkers also smoke more. Smoking is associated with both coffee consumption and cancer, meaning it is a confounding factor. An analysis conducted without controlling for smoking is misleading.
In medical research, age, sex, comorbidity index, BMI, and socioeconomic status are the most common confounding factors.
Which Multivariable Analysis?
The right analysis is chosen according to the type of the dependent variable:
| Dependent Variable | Analysis |
|---|---|
| Continuous (pain score, lab value) | Multivariable linear regression |
| Binary (present/absent, complication) | Logistic regression |
| Time-to-event (survival) | Cox regression |
| Count data (number of recurrences) | Poisson regression |
Which Variables Should Be Included in the Model?
Variable selection should be systematic:
Clinical rationale: Variables known in the literature to be confounding factors should be included in the model. Statistical criteria alone should not be the only consideration.
Univariate criterion: A threshold of p < 0.20 or p < 0.25 is commonly used. Some authors prefer to include all clinically meaningful variables.
Sample size rule: In logistic regression, at least 10 to 15 events are needed for each independent variable (the Events Per Variable rule). Including more than 10 variables in a model with a sample of 50 patients creates overfitting.
Should Stepwise Methods Be Used?
Forward, backward, and stepwise methods perform automatic variable selection in statistical software. However, these methods are problematic:
- Results are sample-dependent and not reproducible
- p-values come out inflated
- Variables without a clinical rationale can enter the model
Recommendation: Perform manual variable selection based on theoretical rationale. Use stepwise methods only in exploratory analyses, and state this explicitly in the paper.
How Should Results Be Reported?
A good multivariable analysis table includes the following:
- For each variable: coefficient (B), standard error, OR/HR (with 95% CI), p-value
- Model fit statistics (Hosmer-Lemeshow, Nagelkerke R², AUC)
- Clearly stated reference categories
Presenting univariate and multivariable results side by side in a table is standard practice; reviewers want to see both.
The Most Common Mistakes
Reporting only the variables that turned out significant: All variables included in the model must be shown in the table, even if they are not significant.
Ignoring missing data: If a variable has missing data, those observations are excluded from the analysis. This can create selection bias. Multiple imputation or complete case analysis should be explained.
Not checking for multicollinearity: If highly similar variables are included in the model at the same time, the coefficients become unreliable. A VIF (Variance Inflation Factor) > 10 signals a multicollinearity problem.
For support with multivariable analysis, request a free consultation.
Where Do People Get Stuck Most in This Analysis?
- You included too many variables in the model and there is a risk of overfitting, but you do not know which ones to remove.
- You performed a VIF check and multicollinearity is present, but it is clinically unclear which variable you should remove.
- A reviewer told you "do not use stepwise methods," but you cannot articulate the rationale for manual variable selection.