Jacobson-Truax — Reliable Change & Clinical Significance

A significant mean change in a therapy study says nothing about whether an individual patient actually feels better after treatment. Jacobson & Truax (1991) resolve this with two questions per person: (1) Is the change reliable? (2) Is it clinically meaningful?

The measurement error (standard error of measurement, SE_M) results from spread and reliability:

If |RCI| > 1.96, the change exceeds the measurement error with 95% confidence. Equivalently: the raw difference must be larger than the critical value = 1.96 · S_diff. The SD of the clinical population (SD_d) is used here as the pretest SD.

a uses only the clinical sample (when no healthy norm is available) and is conservative. b uses only the healthy norm. c is the likelihood-weighted intersection of both distributions. d is a pragmatic addition (see below).

Criterion a is conservative. It uses only the clinical sample (a = M_d − 2·SD_d) and requires the post value to leave the clinical distribution downward — essentially returning almost to the healthy mean. Sensible when no healthy norm sample is available; tends to be strict, though, so genuine improvements are less often classified as "clinically meaningful".

Criterion b is often too liberal. "Within 2 SD of the healthy" corresponds to the 97.5th percentile of the healthy distribution — a very wide net. When the clinical and healthy distributions overlap (the usual case for symptom scales), many patients already count as "functional" before therapy, and the criterion contributes little information.

Criterion c is usually the best choice. It automatically lands where both populations are equally likely — typically close to clinically meaningful thresholds. Important: in a JT analysis you always have pretest data for the clinical group — so you can always estimate M_d and SD_d and compute c from them. The excuse "only a healthy norm is available → I must use b" is rarely truly necessary.

The "2-SD" factor is convention, not law. With 1 or 1.28 SD, b would come closer to validated clinical thresholds.

d is not part of the original procedure by Jacobson & Truax — but in practice it's often the most transparent and best-justified solution: instead of a statistical 2-SD value, you plug in a clinically validated threshold directly from a test manual or the literature.

Example SCL-90-S: the authors define a case at GSI-T > 60. Setting this value as d=60 is traceable and manual-conformant — reviewers generally accept this, precisely because it rests on the original source.

Note — d doesn't exist directly in JTRCI: the R package JTRCI only knows the criteria a, b, and c. If you still want to reach a fixed cutoff there, you can "trick" it by choosing the population parameters accordingly. Example SCL-90-S with target cutoff 60: take criterion b, leave the healthy norm at M_f = 50 as in the original, but set SD_f = 5 — that gives b = M_f + 2·SD_f = 50 + 2·5 = 60. This works, but is harder for outsiders to follow. Option d in this tool makes the same step explicit and transparent.

Healthy M_f=50 / SD_f=10, clinical M_d=65 / SD_d=8. From this:

In an outpatient clinic with a pretest mean of T≈65, this nicely shows the weaknesses: b=70 would sit above almost all patients (barely anyone counts as a case), a=49 requires an almost complete return to the healthy mean, c≈58 comes close but misses the official threshold. d=60 directly reflects the manual-conformant definition.

This tool follows exactly the classification logic of the R package JTRCI (A.-W. Kruijt) — the same grouping reported in many publications. The decision is made on two quantities: the position of the post value relative to the cutoff, and the RCI. The five groups:

Recovered: post in the functional range AND reliable improvement (|RCI| ≥ 1.96).
Not reliably recovered: post in the functional range, but change within the measurement-error band (|RCI| < 1.96).
Improved: reliable improvement, but post still in the clinical range.
Unchanged: post in the clinical range, no reliable change.
Deteriorated: reliable change in the unfavorable direction — overrides all other categories.

Why "not reliably recovered"? Someone can end up below the cutoff without the change exceeding the measurement error — the apparent recovery could therefore be measurement noise. JTRCI makes this uncertainty visible as its own group, instead of hiding it under "unchanged" (as the classic four-cell scheme does).

Caution regarding pre-status: the classification — like JTRCI — uses only the post value and RCI, not whether someone was even in the clinical range to begin with. Someone who already starts below the cutoff before treatment cannot, strictly speaking, "recover" at all; such cases should be interpreted with caution (JTRCI flags them separately in the plot). For strict comparability, though, this tool deliberately follows the package logic.

Source: JTRCI — github.com/AWKruijt/JT-RCI. RCI convention there: RCI = (post − pre) / S_diff; under "lower = better", a negative RCI means improvement.

The x-axis shows the pre value, the y-axis the post value. The diagonal (y = x) is the line of no change. The gray reliable-change band marks the range in which changes cannot be distinguished from measurement error (width = ± critical value). The cutoff line separates functional from dysfunctional post values. Each point's position relative to the band and cutoff determines its category.

Using the two sliders Pre Value and Post Value at the very bottom of the sidebar, you set a single example patient and see live how the Jacobson-Truax logic classifies them. This is independent of the simulated sample and is meant for deliberately trying out individual cases.

Directly below the sliders, two displays appear:

① The colored category field shows the classification (Recovered / Not Reliably Recovered / Improved / Unchanged / Deteriorated) in the corresponding category color — the same color the patient appears in on the JT plot.

② The RCI line gives the concrete Reliable Change Index, whether the change is reliable (reliable = outside the measurement-error band, n.s. = inside), the raw difference Δ in points, and the critical value (± crit.) the difference must exceed.

In the plot, the focus patient appears as a large, outlined point with a dashed connecting line to the diagonal — this line visually shows the magnitude of their change (the longer, the larger the pre-post difference). This lets you try out, for example: at what post value does a patient tip from "Improved" to "Recovered"? When does the change just barely fall short of reliable change? Move the cutoff and RCI settings and watch the point change category.

BDI-II: scale 0–63 · r_tt = .92 · healthy M=8 / SD=7 · clinical M=30 / SD=10. From this, SE_M ≈ 2.83, S_diff ≈ 4.0, critical value ≈ 7.8 points, cutoff c ≈ 17. A BDI-II change must therefore be about 8 points to be reliable. Criterion d is preset to 20 — the threshold often used as clinically relevant (transition moderate/severe depression).

SCL-90-S · GSI: T scores 30–80 · r_tt = .90 · healthy M=50 / SD=10 · clinical M=65 / SD=8 · cutoff d = 60 (manual caseness). From this, SE_M ≈ 2.53, critical value ≈ 7.0 T points. This shows why the a/b/c automation alone misses the manual threshold.

Jacobson, N.S. & Truax, P. (1991). Clinical significance: A statistical approach to defining meaningful change in psychotherapy research. Journal of Consulting and Clinical Psychology, 59(1), 12–19.
Beck, A.T., Steer, R.A. & Brown, G.K. (1996). Manual for the Beck Depression Inventory-II. Psychological Corporation.
Franke, G.H. (2002). SCL-90-R. Die Symptom-Checkliste von L.R. Derogatis (2nd ed.). Beltz Test.