Pre-post data looks simple: measure people before a program, measure them again afterward, and ask whether they improved. The difficult part is deciding what comparison answers that question. A spreadsheet can subtract baseline from follow-up in one click. A regression can adjust the follow-up score for baseline just as quickly. Those analyses are not interchangeable defaults. They use the baseline measure differently, and the choice should happen before anyone sees the result.
For a straightforward randomized study with one continuous baseline and one continuous follow-up outcome, a baseline-adjusted follow-up model is often the efficient starting point. But that is not permission to adjust every pre-post dataset in the same way. The right analysis follows the question, the study design, and the meaning of the measurement. This framework makes that choice visible.
Start with the sentence you need to defend
Write the result in plain language before selecting a formula. “Participants improved” is a within-person statement. “The program improved outcomes more than the comparison condition” is a between-group statement. “People who started lower improved more” is another question again. One dataset may support all three, but they are not answered by the same contrast.
Suppose a workforce program measures interview confidence at intake and six months later. If the question is whether each enrolled participant changed, a change score, follow-up minus baseline, puts that change directly on the table. If the question is whether the program outperformed a comparison group at six months, the key comparison is between groups. Baseline still matters, but it is usually more useful as information that sharpens the follow-up comparison than as a number to subtract automatically.
The three-question fork
Use these questions in order. They do not replace subject-matter judgment; they make the judgment reviewable.
- What is the target outcome? If the decision is about status at follow-up, such as whether students meet a reading benchmark, model the follow-up outcome. If the decision is inherently about change, such as a clinically meaningful improvement threshold, define that change precisely and analyze it as the outcome.
- When was baseline measured relative to the exposure? A true pre-treatment baseline can improve precision in a randomized comparison. A measure collected after participation began can already contain part of the program effect. Adjusting for it may remove some of the very effect the evaluation is meant to estimate.
- Is baseline strongly related to follow-up? In many settings it is. A baseline-adjusted model uses that predictability to reduce unexplained variation at follow-up. When the relationship is weak or the baseline measure is unreliable, the expected precision gain shrinks; inspect the data and measurement plan rather than choosing a method by habit.
What the two common models actually say
A change-score model creates one outcome for each person: follow-up minus baseline. With a treatment indicator, the analysis compares average change across groups. Its main virtue is clarity when change itself is what people care about. It can be useful when the scale has a natural change interpretation, such as dollars of monthly earnings or an established clinical change threshold.
A baseline-adjusted follow-up model keeps follow-up as the outcome and includes baseline as a predictor. In a randomized study, the treatment coefficient compares expected follow-up outcomes between groups that had the same baseline value. Because the model accounts for predictable differences in where people started, it commonly produces a narrower confidence interval than a change-score comparison.
Neither approach turns an observational comparison into a randomized one. If program participants begin with different resources, risks, or referral pathways than nonparticipants, baseline outcome alone does not solve confounding. That planning problem belongs with a clear causal question and a defensible covariate strategy, not an ever-longer list of controls added after the fact.
A small example: the same improvement, different conclusion
Imagine two school districts introduce different tutoring supports. District A begins at an average math score of 55 and reaches 65. District B begins at 70 and reaches 80. Both gained 10 points. For a question about average change inside each district, that is the main result.
But now imagine the program district began at 55 while its comparison district began at 70. A raw follow-up comparison says 65 versus 80 and makes the program look worse. A raw change comparison says 10 versus 10 and makes it look equal. Neither number alone establishes what would have happened to comparable students without the program. The better approach depends on how districts and students were assigned, what other differences existed at baseline, and whether a 10-point gain means the same thing at different portions of the scale.
In a randomized version of this study, a baseline-adjusted follow-up model would generally be a sensible primary analysis. It respects the randomized comparison while using baseline scores to make the six-month comparison more precise. Report the adjusted group difference with its confidence interval, then show descriptive baseline and follow-up means so readers can understand the result without reverse-engineering the model.
Four mistakes that create avoidable confusion
- Testing pre-post change separately in each group. Significant in one group and not significant in another is not a test of a group difference. Compare groups directly.
- Choosing the model after looking at both answers. A change-score analysis and an adjusted model can disagree, especially in small or imbalanced samples. State the primary analysis before outcomes are visible and describe any alternative as a planned check.
- Ignoring baseline measurement quality. A noisy baseline is less helpful for adjustment and can make a change score noisy too. The quality and fit of the measurement belong in the analysis plan.
- Forgetting missing follow-up data. If follow-up is missing differently by group or baseline severity, neither formula repairs the problem by itself. Specify retention procedures and the missing-data approach alongside the primary model.
A pre-analysis worksheet for your protocol
Before analyzing, record the target population; the outcome and time point that drive the decision; whether baseline was measured before exposure; the primary model; the baseline variables included and why; the planned handling of missing follow-up; and one or two sensitivity checks. This is the same discipline that makes a power analysis credible: the calculation should be traceable to the design, not reverse-engineered from a favorable result.
For many randomized pre-post studies, a concise plan might say: “We will compare six-month outcomes between randomized groups using linear regression adjusted for baseline outcome and site. We will report adjusted mean differences and 95% confidence intervals. We will describe unadjusted means at both time points and assess robustness to the prespecified missing-data procedure.” That is short, specific, and easy for a reviewer to evaluate.
The useful question is not “Are change scores bad?” It is “What outcome and comparison would make this decision defensible?” Once that is clear, the model is no longer a software default. If you are setting up a study or sorting through a pre-post dataset that has become complicated, let's discuss the decision the analysis needs to support.