How to Combine Subscale Means & SDs in One Group (With Sensitivity Analysis)

Pool subscale scores from the same participants into one composite total mean and SD — with built-in r = 0.25 / 0.50 / 0.75 sensitivity analysis. For sums of subscales only; not for averaging paired body-side measurements.

You're extracting outcomes for a meta-analysis and the trial reports several subscales of the same instrument separately. SF-36 Physical and Mental Component Summaries. HADS Anxiety and Depression. Fugl-Meyer upper-limb and lower-limb motor subscales that sum into the total motor score. The synthesis you're running needs one composite per participant — but the paper only gives you the parts. What now?

The Combine Mean & SD for the Same Group (Subscales) calculator does this in one click, with a built-in sensitivity grid so reviewers can see exactly how robust your pooled SD is to the correlation you assume.

Open the calculator →

When to use this — and when not to.
✅ Use this calculator when the composite outcome is defined as a sum of subscales on the same participants (e.g., SF-36 PCS + MCS into a global score, HADS anxiety + depression into HADS total, Fugl-Meyer motor subscales summed into the total motor score).
❌ Do not use this calculator for paired body-side measurements (e.g., left vs right hemisphere motor scores, left vs right limb strength) when the per-participant composite is an average of the two values rather than a sum. For an average, compute each participant's mean of the two values first, then summarize across participants with a standard mean/SD calculation. Applying this sum-based formula to values you actually want to average will inflate the composite mean and SD.

Why you can't just add the SDs

Adding means is fine — that's just algebra. Adding SDs is where most reviewers slip. When two subscale scores come from the same people, their errors are correlated. A participant who scores high on SF-36 Physical also tends to score high on Mental. That correlation shows up in the variance of the sum, and ignoring it gives a pooled SD that is systematically wrong.

This is a different problem from combining independent groups (e.g., two arms of a trial, or two cohorts). There, participants are different people, so the covariance term drops to zero and a simple weighted formula applies. For same-group subscale pooling, you need the covariance — which means you need the correlation between subscales.

The formula in plain English

For k subscales measured on the same participants and summed into a total:

  • Combined mean: M = m₁ + m₂ + … + mₖ
  • Combined variance: Var = Σ sdᵢ² + 2 · Σᵢ<ⱼ rᵢⱼ · sdᵢ · sdⱼ
  • Combined SD: √Var

The first term is the sum of individual variances. The second is the covariance correction — twice the sum of every pair's correlation times the product of their SDs.

Cochrane connection. This is the same algebra the Cochrane Handbook uses for the SD of a change from baseline — but with a plus sign on the cross-term instead of a minus. Subtracting paired measurements removes covariance; summing them adds it. Same formula family, opposite sign.

Why sensitivity analysis with r = 0.25 / 0.50 / 0.75

Here's the practical problem: the source paper almost never reports the correlation between its subscales. You're forced to assume one. Pick a single value and a reviewer will (correctly) ask how the result holds up if you guessed wrong.

The calculator solves this by running the math four times in parallel:

  • r = 0.25 — weakly correlated subscales (a defensible lower bound)
  • r = 0.50 — moderate (the Cochrane default for unknown correlations)
  • r = 0.75 — strongly correlated (a defensible upper bound)
  • User r — your own value when a validation study or a methods paper gives you one

You report the pooled SD across the range. If the meta-analytic effect doesn't change meaningfully between r = 0.25 and r = 0.75, your conclusion is robust to the assumption. If it does, you have a real source of uncertainty to flag in the discussion.

Worked example 1 — SF-36 Physical + Mental

A trial reports SF-36 component scores for the treatment arm (n = 80):

SubscaleMeanSD
Physical Component Summary4810
Mental Component Summary529

Combined mean is straightforward: 48 + 52 = 100.

For the combined SD, the variance terms give 10² + 9² = 181. The cross-term contributes 2 · r · 10 · 9 = 180 · r. So:

rVarianceCombined SD
0.25181 + 45 = 22615.0
0.50181 + 90 = 27116.5
0.75181 + 135 = 31617.8

The pooled SD spans 15.0 to 17.8 — about a 19% range. If your effect size's confidence interval doesn't cross the null at either bound, you're fine. If it flips signs, you need to discuss the correlation assumption explicitly.

Worked example 2 — HADS Anxiety + Depression

A trial reports HADS sub-scores for the intervention arm (n = 60). HADS total is defined as the sum of the two subscales:

SubscaleMeanSD
HADS Anxiety83
HADS Depression62.5

Combined mean: 8 + 6 = 14 (HADS total). Variance sum: 3² + 2.5² = 15.25. Cross-term: 2 · r · 3 · 2.5 = 15 · r.

rCombined SD
0.254.36
0.504.77
0.755.15

The same machinery works for any multi-subscale instrument where the total is defined as a sum: SF-36 component summaries, HADS anxiety + depression, Fugl-Meyer motor subscales, PHQ subscores, or any multi-domain PRO whose total score is the sum of its parts.

When in doubt: sum vs. average for body-side measurements

Some studies report a measurement separately for the left and right side of the body — left/right hemisphere motor scores, left/right grip strength, left/right limb function. If your synthesis treats each participant's composite as the average of the two sides (the usual convention for bilateral outcomes), this calculator is not appropriate — unless you are sure the paired measurement should be calculated by sum, not average. When you are sure the outcome is a sum, you can use this tool; otherwise it does not apply.

Common mistakes

  • Using it for averaged body-side measurements. Left/right hemisphere or left/right limb outcomes whose composite is an average per participant should not be summed with this tool unless you are sure the outcome is genuinely a sum, not an average.
  • Using it for independent groups. If your two rows are different cohorts or trial arms, you want Combine Independent Groups instead. The same-group formula will over-estimate the pooled SD because the covariance term is spurious.
  • Reporting only r = 0.5 with no sensitivity range. Cochrane recommends r = 0.5 as a default when nothing is known, not as a final answer. Always show the three-point grid in supplementary material.
  • Ignoring SDs that look "too small". If the source paper reports an implausibly small composite SD, check whether they used a same-group formula with a high assumed r. Re-pool with r = 0.5 and compare.

User r vs. the sensitivity grid

Use the user r column when you have a real number — for example, a psychometric validation study reports r = 0.62 between SF-36 PCS and MCS. Use the 0.25 / 0.50 / 0.75 grid as the sensitivity envelope you report in the paper. They're complementary, not alternatives: the user r is your best estimate, the grid is your robustness check.

Try it now

Enter one row per subscale, fill mean, SD, and (optionally) your user r. The calculator returns the combined mean and four combined SDs in one pass — ready to paste into your forest plot or supplementary table.

Open the Combine Mean & SD (Same Group) calculator →

Related tools: Change in Mean & SD from Baseline · Combine Independent Groups · More articles

Step-by-step worked example (with full arithmetic)

Let's walk through every number for a single trial so you can reproduce the calculator's output by hand. Suppose a stroke-rehabilitation RCT reports the treatment arm (n = 60) on three Fugl-Meyer motor subscales that the protocol sums into a single total motor score:

SubscaleMean (mᵢ)SD (sdᵢ)
Upper Limb Motor3410
Lower Limb Motor227
Coordination & Speed145

Step 1 — Combined mean

Sum the subscale means: M = 34 + 22 + 14 = 70.

Step 2 — Sum of individual variances

Square each SD and add: 10² + 7² + 5² = 100 + 49 + 25 = 174. This is the "no-correlation" baseline variance.

Step 3 — List every pair

With 3 subscales there are C(3,2) = 3 pairs. For each pair compute 2 · sdᵢ · sdⱼ (the coefficient that gets multiplied by r):

  • Upper × Lower: 2 · 10 · 7 = 140
  • Upper × Coordination: 2 · 10 · 5 = 100
  • Lower × Coordination: 2 · 7 · 5 = 70
  • Sum of pair coefficients: 140 + 100 + 70 = 310

Step 4 — Apply each correlation

Cross-term contribution at correlation r is 310 · r. Total variance is 174 + 310 · r; combined SD is its square root.

rCross-term (310 · r)Total varianceCombined SD = √Var
0.2577.5251.515.86
0.50155.0329.018.14
0.75232.5406.520.16
User r = 0.60186.0360.018.97

Step 5 — Read the spread

The combined SD moves from 15.86 (r = 0.25) to 20.16 (r = 0.75) — roughly a 27% range. Report all four values in supplementary material and use the r = 0.50 result (or your user r) as the primary input to the forest plot.

Reproduce this in the calculator →

Frequently asked questions

What inputs does the calculator need?

One row per subscale, each with a mean, a standard deviation (SD > 0), and an optional user correlation r between −1 and 1. The number of participants (n) doesn't enter the formula here — it stays the same for every subscale because they were measured on the same people.

Can I use this for left vs right hemisphere (or left/right limb) measurements?

Only if the published outcome is genuinely defined as the sum of the two sides. Don't use this calculator for left/right body sides or paired measurements where the per-person composite is an average — unless you are sure the outcome should be calculated by sum, not average. If you're sure it's a sum, this calculator applies; otherwise it does not.

What if my source paper gives SE or 95% CI instead of SD?

Convert first. Use our SE→SD or CI→SD converters on each row, then paste the SDs into this calculator. Don't mix SE and SD in the same column.

What should I use for r when the paper doesn't report it?

Cochrane's default is r = 0.5 for unknown within-subject correlations. The calculator's three-point grid (0.25 / 0.50 / 0.75) covers the realistic envelope. If a published validation study for the instrument reports a specific correlation, enter it as the user r and treat the grid as your sensitivity check.

Can I use this to combine arms of a trial (e.g., placebo + treatment)?

No — those are independent groups (different participants). Use the Combine Independent Groups calculator, which uses a weighted formula with no correlation term.

Does this work for more than two subscales?

Yes. The formula generalises to any number of subscales: sum all the variances, then add 2 · r · sdᵢ · sdⱼ for every pair. The calculator handles any number of rows you add.

What if my subscales are on different scales (e.g., 0–100 vs. 0–10)?

The math still runs, but the composite may not be meaningful. Standardise first (e.g., z-scores or percent-of-max) before pooling subscales measured on incompatible scales.

Why is my combined SD smaller than I expected?

Two common causes: (1) you accidentally used a negative correlation, which subtracts the cross-term; (2) one of your SDs is actually an SE. Re-check both inputs.

Why is my combined SD larger than expected?

Usually because you assumed a high correlation (r ≥ 0.8) when the subscales are only weakly related. Drop to r = 0.5 and compare; if the result still looks off, sanity-check by computing the "no-correlation" SD: √(Σ sdᵢ²).

How do I cite the methodology?

Reference the Cochrane Handbook (Chapter 6, formula for SD of a sum/difference of paired measurements). The calculator uses the same algebra with a plus sign on the cross-term because we are summing rather than subtracting.

Common mistakes to avoid

  • Using it on averaged body-side measurements — left/right hemisphere or left/right limb outcomes whose composite is an average per participant should not be summed with this tool unless you are sure the outcome is genuinely a sum, not an average.
  • Using it on independent groups — wrong tool; use Combine Independent Groups.
  • Reporting only one r — always include the sensitivity range in supplementary material.
  • Mixing SE and SD across rows — convert everything to SD first.
  • Forgetting to check sign on the cross-term — this calculator adds the cross-term (sum); the change-from-baseline calculator subtracts it (difference).
  • Pooling incommensurable scales — standardise subscales measured on very different ranges before combining.

More articles