Research statistics guide
What Pre/Post Correlation Should I Use?
Pre/post correlation describes how strongly participants’ baseline measurements are associated with their follow-up measurements. It is often missing from published papers, but it can materially affect the reconstructed variance and precision of change scores.
What is a pre/post correlation?
The pre/post correlation, usually written as r, is the correlation between repeated measurements of the same outcome in the same participants at baseline and follow-up.
It can range from −1 to +1. In many biological and clinical outcomes it will be positive, but its actual value depends on the outcome, population, measurement reliability, follow-up interval, intervention, and study conditions.
Why does it matter for change scores?
The variance of a paired change score includes the covariance between baseline and follow-up:
When the correlation changes, the SD of change changes. That affects the standard error and confidence interval of mean change and can also affect the sampling variance assigned to a standardized pre/post effect in meta-analysis.
Preferred hierarchy for choosing r
| Preference | Source | Comment |
|---|---|---|
| 1 | Reported study-specific correlation | Best when it corresponds to the same outcome, time points, and analysis sample. |
| 2 | Reconstructed from reported SD of change | Uses the study’s own summary statistics. |
| 3 | Authors or study dataset | Useful when the published report is incomplete. |
| 4 | Comparable external study or dataset | Use only when outcome, population, interval, and measurement conditions are sufficiently similar. |
| 5 | Assumed value plus sensitivity analysis | Appropriate when no defensible empirical estimate is available. |
How to reconstruct r from SD of change
If the SD of change is reported, rearrange the change-score variance equation:
For the full derivation and worked example, see How to calculate SD of change from pre/post data.
Should I just use r = 0.50?
A value such as 0.50 can be useful as a midpoint in a sensitivity analysis, but it should not be presented as a universal default supported for every outcome and population. A single assumed correlation can create a false impression of precision when the true correlation is unknown.
A better approach is to report the assumption transparently and test a plausible range — for example, several values chosen from prior evidence or a broad exploratory range when little information is available.
How to perform a correlation sensitivity analysis
Recalculate the SD of change, standard error, and confidence interval across multiple plausible values of r. Then examine whether substantive conclusions about precision are sensitive to the assumption.
Useful outputs include:
- SD of change at each assumed correlation;
- SE of mean change;
- confidence-interval width;
- sampling variance of a standardized effect when the chosen method depends on r.
One correlation or separate correlations for two groups?
If empirical evidence supports different correlations in two independent groups, group-specific values can be appropriate for reconstructing each group’s change-score variance. However, some standardized-effect variance estimators assume a common correlation structure. Do not silently substitute a common value solely to force a preferred estimator.
How should the assumption be reported?
A concise Methods statement might read:
If the correlation was reconstructed instead:
Test different correlations
The Blazon Scientific Pre/Post Change Calculator lets you enter a known or assumed correlation, reconstruct it from an available SD of change, and examine how alternative values affect precision.
Open the Pre/Post Change CalculatorReferences and further reading
- Clifton L, Clifton DA. The correlation between baseline score and post-intervention score, and its implications for statistical analysis. Trials. 2019. PubMed
- Cochrane Handbook. Analysing data and undertaking meta-analyses: change scores. Cochrane
Frequently asked questions
Is there a standard pre/post correlation?
No. The correlation is outcome-, population-, interval-, and study-specific. A commonly used assumed value should not be mistaken for a universal constant.
Can I use a correlation from another study?
Yes, if it is sufficiently comparable in outcome, population, measurement method, and follow-up interval. The choice should be justified and sensitivity analysis is still useful.
Can I calculate r from an SD of change?
Yes. If baseline SD, follow-up SD, and SD of change are all available for the same paired sample, the correlation can be reconstructed algebraically.
Does r change Hedges’ g?
For the baseline-standardized point estimates used by the Blazon calculator, changing the assumed correlation does not change Hedges’ g itself. It can change change-score precision and, depending on the variance estimator, the sampling variance of the standardized effect.
What if my reconstructed correlation is above 1 or below −1?
The reported summary statistics are not mutually compatible with the standard paired-change formula. Check rounding, sample sizes, analysis populations, and whether adjusted or transformed statistics have been mixed.