Free scoping read. A written assessment of your manuscript in two business days, at no cost.

Start one
Contents

Academy · Research Data, Analytics & AI · SD of the Change Score

Estimating the Standard Deviation of a Change Score from Baseline and Post-Intervention Values

The standard deviation (SD) of a change score is calculated from the variance sum law by combining the baseline and final SDs with their correlation coefficient. The 2019 analysis by Clifton and Clifton shows that the chosen correlation value has a direct, and sometimes decisive, effect on the result.

Pubrica AcademyWritten by a biostatistics writer Reviewed bya meta-analysis methodologist Published10 October 2026 Last checked10 October 2026 Reading time8 minutes Sources10 peer-reviewed

When a study reports only the standard deviation (SD) of a baseline measurement and the SD of the corresponding post-intervention measurement, but not the SD of the change between them, that missing change-score SD can be estimated rather than left as a gap in a meta-analysis's data extraction. Clifton and Clifton's 2019 analysis in the journal Trials sets out the underlying formula: because a change score is the post-intervention value minus the baseline value, the variance sum law gives its variance as the baseline variance plus the final variance minus twice the correlation between the two, multiplied by both SDs, and taking the square root of that combined variance gives the change score's SD1. The only information this formula requires beyond what a study typically reports directly is the correlation coefficient between the baseline and post-intervention scores, and the sourcing of that value accounts for most of the practical difficulty in applying the method.

Learning objectives

  • The formula for estimating the SD of a change score from baseline and final SDs and a correlation coefficient
  • Why the correlation coefficient has a disproportionately large effect on the resulting estimate
  • The alternative methods available when no usable correlation coefficient can be found or assumed

Common misconceptions

  • “If the correlation between baseline and final scores is unknown, the change-score SD cannot be estimated at all.” Weir and colleagues’ 2018 systematic review of methods for handling missing SDs in meta-analysis describes several approaches that allow a usable change-score SD to be estimated even when the correlation is not directly reported: Bayesian estimation of the correlation with an uninformative prior, meta-regression across the trial’s own baseline and follow-up SDs, and borrowing a plausible correlation from a similar, previously published study.

Formula for estimating the SD of a change score

The formula Clifton and Clifton describe follows directly from the statistical variance sum law applied to a difference between two correlated measurements. If a change score is defined as the post-intervention value minus the baseline value, its variance equals the baseline variance plus the final variance, minus two times the correlation coefficient between the two measurements, multiplied by the baseline SD and the final SD1. Written out, this is SD(change) equals the square root of SD(baseline) squared plus SD(final) squared minus two times the correlation coefficient times SD(baseline) times SD(final). Because a study that reports both baseline and final SDs almost always presents those two values in its results tables, the only missing input for this calculation is usually the correlation coefficient itself, which captures how closely an individual participant's baseline score predicts their own final score.

The same relationship, rearranged to derive the correlation rather than the change-score SD, appears in related methodological work by Clifton and Clifton, which the authors describe as a companion analysis showing how a correlation coefficient can be derived from a study that reports baseline, final, and change-score SDs together, using the identical variance sum law rearranged to solve for the correlation1. This is of practical importance because a meta-analyst extracting data across many trials may find one or two studies in the literature that report all three SDs together, calculate an implied correlation from those studies using this rearranged formula, and then apply that borrowed correlation to other, similar trials that report only baseline and final SDs.

Influence of the correlation coefficient on the estimate

The correlation coefficient has a large effect on the final estimate because it enters the formula as a subtracted term multiplied by both SDs; even a moderate change in its assumed value can therefore shift the resulting change-score SD substantially, particularly when the baseline and final SDs are of similar magnitude. Clifton and Clifton's analysis notes that a change score tends to be statistically reasonable, meaning that its SD remains small relative to the baseline and final SDs, specifically when the correlation is high, for instance above about 0.8, which tends to hold for stable, chronic conditions in which an individual's relative position in the population changes little between the two measurement points1. When the correlation is low, the variance of the change score can exceed that of the final measurement itself, making the change score a noisier and less statistically powerful outcome than the final measurement alone would have been.

Weir and colleagues' 2018 systematic review situates this sensitivity within a broader set of documented methods for handling a missing change-score SD, describing two specific Bayesian approaches, developed by Abrams and colleagues and by Nixon and colleagues respectively, that explicitly model the correlation as a quantity to be estimated from the available data rather than assumed as a fixed input2. The Abrams approach places a vague Uniform(0,1) prior on the correlation within a Bayesian meta-analysis framework, which allows the data to inform the estimate of a plausible value, while noting that results remain sensitive to the choice of prior, especially when few studies are available to inform it2. Both approaches exist because treating the correlation as a single, precisely known number when it is uncertain risks producing a misleadingly precise change-score SD.

Alternative methods when no usable correlation coefficient is available

When no study in the relevant literature reports a usable correlation coefficient, several alternative approaches exist that do not depend on it. Hozo, Djulbegovic, and Hozo's widely used method estimates a sample's mean and variance directly from its median, range, and sample size, which can substitute for a full SD calculation when a study reports only these simpler summary statistics rather than a mean and SD3. Wan and colleagues' later extension of this approach adds formulas that also incorporate the interquartile range where it is reported alongside the median and range, generally producing a more accurate estimate than the range-only method when that additional information is available4.

Weir and colleagues' review also describes single-imputation approaches that borrow an observed correlation, or an observed change-score SD, from a different but sufficiently similar trial already included in the same meta-analysis; this avoids the correlation question by substituting a directly observed figure from comparable evidence for an assumed or statistically estimated one2. Whichever method is used, the Cochrane Handbook's guidance on missing standard deviations recommends that any imputed or estimated value be clearly identified as such in the methods of a meta-analysis and, ideally, tested in a sensitivity analysis that examines whether the overall pooled result changes meaningfully under a different, equally plausible correlation or imputation method; this is the most direct means of confirming that a single uncertain input has not determined the main conclusion of the review5.

Worked example of the effect of the correlation coefficient

Applying Clifton and Clifton's formula to a specific pair of values demonstrates the disproportionate influence of the correlation. Consider a trial that reports a baseline SD of 10 and a final SD of 10 for a continuous outcome. With a correlation of 0.8, a plausible figure for a stable, chronic measurement, the change-score variance is 10 squared plus 10 squared minus two times 0.8 times 10 times 10, which equals 100 plus 100 minus 160, or 40, giving a change-score SD of about 6.3, noticeably smaller than either the baseline or final SD1. With a correlation of only 0.3, representing a less stable measurement in which an individual's relative position shifts more between the two time points, the same calculation gives 100 plus 100 minus 60, or 140, a change-score SD of about 11.8, which is larger than either original SD.

In this example, reducing the assumed correlation from 0.8 to 0.3 while holding both original SDs constant nearly doubles the resulting change-score SD, from about 6.3 to about 11.8, which is sufficient to widen a confidence interval noticeably or to alter a statistical significance result in a meta-analysis based on that estimate. It is precisely this sensitivity that Weir and colleagues' review and the Cochrane Handbook's guidance address when they recommend sourcing a correlation from comparable evidence wherever possible and performing the calculation under more than one plausible value whenever a precisely known figure is not available, rather than selecting a single convenient number and treating the resulting change-score SD as settled25. The same arithmetic also explains why two meta-analysts working from the same pair of baseline and final SDs can legitimately arrive at different change-score estimates without either having made an error: the discrepancy arises from the correlation value each assumed. For this reason, that assumption should be stated explicitly in the methods section rather than embedded in a spreadsheet formula that later readers of the published review will not see.

Key points

Key points

The SD of a change score can be estimated from the baseline SD, the final SD and a correlation coefficient using the variance sum law. Because the correlation has a disproportionate effect on the result, any imputed or borrowed correlation value should be reported explicitly and tested in a sensitivity analysis rather than treated as a precisely known input.

Practical application

Practical application

  1. The relevant literature should be searched for any study that reports baseline, final and change-score SDs together. The rearranged formula of Clifton and Clifton can then be used to derive an implied correlation from that study and to apply it to similar trials that lack the change-score SD.
  2. A sensitivity analysis should be conducted using at least two plausible correlation values. The analyst should determine whether the pooled result of the meta-analysis changes meaningfully between a conservative and a more optimistic correlation assumption before relying on a single estimate.
  3. A range- or quartile-based method should be used only when no SD, mean or usable correlation is reported. The formulas of Hozo, Djulbegovic and Hozo and of Wan and colleagues are a fallback for gaps in summary statistics, not a substitute for the variance sum law when the required SDs are available.

Frequently asked questions

The baseline SD, the final (post-intervention) SD, and a correlation coefficient between the two measurements are required; the correlation is most often missing and must be sourced from a similar study, estimated statistically, or assumed, with a sensitivity analysis to assess the impact of that assumption.

Because the correlation enters the formula as a subtracted term multiplied by both SDs, an incorrect assumption can substantially over- or underestimate the change score's true variability, which is why Clifton and Clifton's guidance and the Cochrane Handbook both recommend sensitivity testing rather than relying on a single assumed value.

No. Clifton and Clifton's analysis notes that high correlations, above roughly 0.8, are plausible for stable, chronic conditions but should not be assumed automatically for conditions or outcomes where an individual's relative standing could plausibly shift substantially between measurements.

Only indirectly, by borrowing an observed change-score SD or correlation from a sufficiently similar trial, or by using a different method entirely, such as Hozo, Djulbegovic, and Hozo's or Wan and colleagues' range- or quartile-based formulas, if the required SDs were never reported.

References

  1. Clifton L, Clifton DA. The correlation between baseline score and post-intervention score, and its implications for statistical analysis. Trials. 2019;20:43. read 10 October 2026. View on Trials
  2. Weir CJ, Butcher I, Assi V, Lewis SC, Murray GD, Langhorne P, Brady MC. Dealing with missing standard deviation and mean values in meta-analysis of continuous outcomes: a systematic review. BMC Medical Research Methodology. 2018;18(1):25. read 10 October 2026. View on BMC Medical Research Methodology
  3. Hozo SP, Djulbegovic B, Hozo I. Estimating the mean and variance from the median, range, and the size of a sample. BMC Medical Research Methodology. 2005;5:13. read 10 October 2026. View on BMC Medical Research Methodology
  4. Wan X, Wang W, Liu J, Tong T. Estimating the sample mean and standard deviation from the sample size, median, range and/or interquartile range. BMC Medical Research Methodology. 2014;14:135. read 10 October 2026. View on BMC Medical Research Methodology
  5. Higgins JPT, Li T, Deeks JJ (editors). Chapter 6: Choosing effect measures and computing estimates of effect. In: Cochrane Handbook for Systematic Reviews of Interventions, Version 5.1.0. The Cochrane Collaboration; 2011. read 10 October 2026. View on Cochrane Training
  6. Sterne JAC, Sutton AJ, Ioannidis JPA, et al. Recommendations for examining and interpreting funnel plot asymmetry in meta-analyses of randomised controlled trials. BMJ. 2011;342:d4002. read 10 October 2026. View on The BMJ
  7. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. read 10 October 2026. View on The BMJ
  8. Schulz KF, Altman DG, Moher D; CONSORT Group. CONSORT 2010 Statement: updated guidelines for reporting parallel group randomised trials. BMJ. 2010;340:c332. read 10 October 2026. View on The BMJ
  9. Hazra A, Gogtay N. Biostatistics Series Module 2: Overview of Hypothesis Testing. Indian Journal of Dermatology. 2016;61(2):137-145. read 10 October 2026. View on Indian Journal of Dermatology
  10. Egger M, Sterne JAC. Funnel plots for detecting bias in meta-analysis: guidelines on choice of axis. Journal of Clinical Epidemiology. 2001;54(10):1046-1055. read 10 October 2026. View on Journal of Clinical Epidemiology

Every source above is a peer-reviewed article, a publisher’s record of one, or a standard issued by the body that owns it. Each carries the date it was last read. Where a standard moves, this article is revised rather than left up with a note.

How to cite this article

Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.

Pubrica Academy. (2026, October 10). Estimating the Standard Deviation of a Change Score from Baseline and Post-Intervention Values. Pubrica. https://pubrica.com/academy/calculation-of-standard-deviation-for-change-group/

This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.

Change-Score SD Estimation Worksheet

This checklist applies when a study reports baseline and final SDs but not the change-score SD.

[ ] Recorded the baseline SD and final (post-intervention) SD from the study
[ ] Searched for a correlation coefficient reported in the same or a similar study
[ ] If unavailable, derived an implied correlation from a study reporting all three SDs
[ ] Applied SD(change) = sqrt(SD(baseline)^2 + SD(final)^2 - 2 x r x SD(baseline) x SD(final))
[ ] Ran a sensitivity analysis with at least two plausible correlation values
[ ] Flagged the imputed/estimated value explicitly in the methods section

Pubrica AcademyWritten and maintained by Pubrica’s subject matter experts. Articles are revised when the underlying standards change, and each article states the date on which it was last reviewed.

How Pubrica can help

Pubrica’s biostatistics team estimates missing change-score SDs using the variance sum law, identifies or derives a defensible correlation value, and conducts the sensitivity analyses needed to show whether that choice determines the conclusion of the review.

Find something else

Search the Pubrica Academy

Access is free and requires no account. Enter a topic, or open the full library to filter by category, sub-topic or resource type.

Assessment of the correlation assumptions underlying a change-score SD

A scoping assessment is free and takes two business days. It provides a written assessment of what the work requires and which specialist would undertake it.

Get a free quote

Regulations

Check your project against this rule

  1. 1Your work
  2. 2Your details
  3. 3Send
What do you need?

Material you send is held under signed confidentiality, with access limited to the assigned team and encrypted transfer. It is not used to train anything and is not passed to third-party tools. Ask for the agreement before you send anything you would rather cover first.