“What should we control for?” is one of the most common regression questions, and one of the easiest to answer badly. A spreadsheet can contain dozens of plausible variables: age, baseline score, education, site, income, diagnosis, prior service use, and whatever else happened to be collected. Adding more can feel cautious. But covariate adjustment is not a cleanup step that automatically makes an estimate more credible. Each variable changes the question the model answers.
A useful rule is to choose covariates from the study story before looking at the outcome results. Do not start with a long list and keep variables that make the p-value look better. Start with the claim you want to support, then use a short, recorded set of questions to decide what belongs in the model. This guide is for that planning conversation—not a substitute for subject-matter knowledge or a causal design, but a way to make the reasoning visible.
First, name the estimate you want
Regression coefficients do not arrive with one universal meaning. Imagine an evaluation asks whether participation in a job-training program is associated with earnings six months later. An unadjusted comparison asks about the observed difference between participants and nonparticipants. A model adjusted for prior earnings, education, and local labor market asks about a difference between people similar on those factors. Both may be useful. They are not interchangeable.
Write a one-sentence target before choosing variables: “We want to estimate the association between exposure and outcome, comparing people who were similar on factors that existed before exposure began.” If the project is a randomized trial, the primary estimate may be the assigned-group difference, with a baseline outcome included mainly to improve precision. If it is observational, the sentence should be even more specific about the comparison being attempted. This step prevents a model from becoming a grab bag of available columns.
The five questions for every candidate covariate
For each variable, ask these questions in order. The point is not to turn causal thinking into a mechanical checklist. It is to notice the consequences of adding a variable before software makes the decision for you.
- Did it exist before the exposure or intervention? A pre-existing factor can plausibly explain why people received different exposures and why they have different outcomes. Age at enrollment, a pre-intervention test score, or clinic location may fit this description. A post-program confidence score does not.
- Could it affect both the exposure and the outcome? That is the practical confounding question. In the job-training example, prior work experience could influence who enrolls and later earnings, so it is a strong candidate. A variable related only to the outcome may still improve precision, but it has a different justification.
- Is it on the pathway the intervention is meant to change? If training increases credential completion and credentials increase earnings, adjusting for credential completion would remove part of the program’s effect. It answers a narrower question: the difference not operating through credentials. That can be a valid secondary analysis, but it should not quietly replace the main one.
- Could conditioning on it create a misleading comparison? Variables caused by two other factors can create a false link when you adjust for them. You do not need to memorize every causal-graph term to use the warning: do not control for a variable simply because it predicts something. Ask how it came to be measured and what may have caused it.
- Can we measure and explain it well enough? A variable with extensive missingness, vague definitions, or inconsistent timing can add noise and ambiguity. Sometimes the right answer is still to include it and handle its missingness carefully. Sometimes it is to acknowledge that the needed confounder was not measured rather than pretending a poor proxy solved the problem.
These questions yield three useful buckets: pre-exposure confounders for the primary model; predictors of the outcome that may improve precision; and variables that belong in a secondary, explicitly different analysis—or not in the regression at all. A small table in the analysis plan with the variable, timing, rationale, and planned role is often more valuable than a longer model formula.
Baseline outcomes are often special
When an outcome is measured before and after an intervention, the baseline value deserves separate attention. Including it in a follow-up outcome model can improve precision because it explains part of where participants started. It also respects the fact that people with very different starting points rarely have identical follow-up expectations.
But the model should match the study question. A common approach is to model follow-up as the outcome and include baseline as a covariate. Modeling change scores can also be appropriate in some designs, but it is not automatically more intuitive or more valid. The key is to decide before examining results, state the primary approach, and use alternatives as sensitivity checks rather than selecting whichever version produces the stronger claim. The same foresight matters when you calculate sample size: the planned analysis must align with the assumptions in your power analysis.
More controls are not always safer
There are real costs to an overfilled model. Extra covariates consume information, especially in a small sample. Closely related variables can make coefficients unstable: the model may have trouble separating education, income, and occupation when they all tell nearly the same story. Missing data can shrink the analytic sample if every added covariate requires a complete record. And a long, post-hoc list invites a reader to wonder how many other reasonable lists were tried.
Suppose a school district evaluates a tutoring program. It adjusts for baseline math score, grade, school, attendance before tutoring, and English-learner status because each was available before students started. That is a defensible starting set if those factors inform participation and achievement. Adding midyear attendance, number of tutoring sessions completed, and a spring motivation survey changes the question. Those may reflect the program’s operation or students’ response to it. They are useful descriptively, and perhaps in a carefully labeled mechanism analysis, but they should not be mixed into the primary effectiveness estimate by default.
Plan a primary model and a small set of checks
A defensible analysis rarely depends on one magical formula. It does need a declared primary model. Record the exposure, outcome, covariates, timing, coding choices, and how clustered or repeated observations will be handled. If students sit within schools or patients within clinics, the dependence may require the kind of modeling decision covered in this multilevel-model diagnostic; it is not fixed merely by putting site into a covariate column.
Then pre-specify only the sensitivity checks that answer a real concern. You might compare results with and without a plausible precision covariate, use an alternative sensible coding for a skewed baseline measure, or assess the influence of a reasonable missing-data approach. Each check should have a purpose stated in advance. A dozen exploratory versions of the model do not make the primary result more robust; they make interpretation less clear.
A short covariate-planning worksheet
Before analysis begins, ask the team to complete this worksheet in plain language:
- What exact exposure and outcome define the primary question?
- Which variables were measured before exposure, and why could each influence both exposure and outcome?
- Which candidate variables might be consequences of exposure or selection into the observed sample?
- Which variables improve precision only, and do we have enough data to include them?
- What is the one primary model, and what limited sensitivity checks will be reported regardless of their results?
That worksheet creates a record reviewers and collaborators can follow. It also connects a regression decision to the wider methods story: the design, measurement, missing-data plan, and reporting choices all need to describe the same study. The habits that make a methods section credible make a regression model credible too.
There is no universal covariate list for a topic area. The right set comes from timing, subject-matter knowledge, and the exact estimate you intend to report. If your team has a model with a long list of “controls,” or an observational study where the comparison needs to be made explicit, get in touch before the results dictate the plan.