Control Variables in SEM: Theory-Based Selection and Sensitivity Reporting
Age, gender and income are not automatically control variables. Every control needs a place in the causal model and can change the estimand being interpreted.
Choose controls from theory and a causal diagram, not from p-values. Model each variable appropriately, prespecify the main results with and without controls, and conduct sensitivity analysis.
Controls need justification and sensitivity checks
- Theoretical basis
Explain why adjustment is needed
- Coding
Specify type and reference group
- Specification
Include in the relevant equations
- Sensitivity
Compare adjusted and unadjusted results
Conceptual diagram. Adding more controls does not automatically improve inference.
Source: Lê Minh · Conceptual illustration, not software output.1. What does a control variable control?
A control variable is a covariate included to adjust for a source of variation related to the predictor and/or outcome. Adding controls changes the main coefficient into a conditional relationship given the variables held constant.
A variable associated only with the outcome may improve precision without being a confounder; controlling for a mediator or collider may introduce bias. A default list of demographic variables therefore cannot replace causal theory.
A nonsignificant control may still need to remain because of the design; a variable with a small p-value may be inappropriate if it is a consequence of the exposure or a collider.
2. Four questions before including a control
Timing
Does the variable precede X and Y, or result from them?
Causal relationships
Could it cause both X and Y, cause only Y, or act as a collider or mediator?
Measurement quality
Is it reliably measured and coded? Is missingness selective?
Estimand
Does the coefficient conditional on this variable still answer the research question?
3. Modeling and coding
Not every control needs a path to every endogenous variable. Paths should reflect the assumed role. Exogenous controls are usually allowed to covary with other exogenous variables when required by the specification.
| Control | Model entry | Points to check |
|---|---|---|
| Continuous | Direct path to the relevant endogenous variable | Check nonlinearity when justified |
| Binary | 0/1 coding with an explicit reference group | The coefficient is a conditional difference |
| Multicategory | k−1 dummy variables | Do not treat 1/2/3 category codes as metric |
| Latent covariate | Separate measurement model | Account for measurement error |
| Cluster/time | Appropriate multilevel or panel model | Single-level SEM can produce incorrect SEs |
4. Report models with and without controls
A useful practice is to present a baseline model and a prespecified adjusted model. Explain changes in the sign, magnitude or CI of the main coefficient rather than choosing the model that gives the desired result.
Bernerth and Aguinis (2016) show that controls are often used habitually without adequate justification. An appendix recording theoretical sources, coding, missingness and effects on results improves transparency.
| Model | Controls | Main coefficient | CI | Conclusion |
|---|---|---|---|---|
| M0 | None | β = … | […; …] | … |
| M1 | Prespecified | β = … | […; …] | … |
| M2 | Plausible alternative set | β = … | […; …] | … |
5. Reporting template
Reporting example in English
[Variable C] was included as a control because theory/prior research suggests that C may affect [X and Y/Y]. C was coded [ ] and had paths to [ ]. The X → Y coefficient was [ ] before adjustment and [ ] after adjustment. The conclusion was [stable/changed]; sensitivity analysis using [alternative controls] yielded [result].
- Provide a theoretical reason for every control.
- Choose controls before inspecting p-values.
- Avoid unintentionally controlling for mediators or colliders.
- Use appropriate dummy coding for categorical variables.
- Report models with and without controls.
- Describe control-variable missingness and reliability.
Conclusion
A well-chosen control clarifies the estimand; habitual controls can make a model harder to interpret or more biased. Transparent justification and sensitivity analysis matter more than the number of controls.
References
- Bernerth, J. B., & Aguinis, H. (2016). A critical review and best-practice recommendations for control variable usage. Personnel Psychology, 69(1), 229–283.https://doi.org/10.1111/peps.12103
Methodological note: Observational data and the addition of controls do not automatically justify causal interpretation. Design, measurement timing and the assumption of no unmeasured confounding still determine the scope of the conclusions.