Extreme baselines: separate regression to the mean from recovery
A group selected because its latest measurement was unusually poor can improve on the next measurement even when nothing has improved in the underlying process. Selection partly captures temporary noise. On remeasurement, that temporary component need not remain as extreme.
Regression to the mean is therefore a problem for interpreting changes after an extreme baseline. It is not a claim that every individual reverses, that businesses cannot recover, or that market prices follow a profitable reversal rule. The research task is to separate the selection mechanism, the repeated-measurement model and evidence for a real change.
Preserve the rule that created the extreme group
Suppose a hypothetical research team flags document-processing cases with unusually high standardized delay scores. Higher scores mean slower processing. The team subsequently introduces a review change and observes lower scores. Before attributing the improvement, save the original eligibility threshold, the measurement used for selection, its date and the scheduled follow-up.
Do not reconstruct the threshold from whichever cases were eventually discussed. A rule based on one extreme measurement differs from a rule based on a persistent multi-period pattern. Likewise, selecting a minimum or maximum over many attempts creates a different conditional distribution from taking a single prespecified baseline.
Martin Bland’s original explanation, especially the sections on repeated measurements, concurrent controls and duplicate baselines, shows why an extreme selected group can change without a treatment effect. The model and numbers below are independently constructed, not estimates of a real workflow intervention.
Build a stable process with no intervention effect
Let Q be a normally distributed, stable underlying standardized delay component with mean zero and variance 0.6. Let e1 and e2 be independent zero-mean normal measurement components, each with variance 0.4 and independent of Q. Set Z1=Q+e1 and Z2=Q+e2. No intervention changes Q or either measurement mechanism.
Both observations have mean zero and variance one. Their covariance is Var(Q)=0.6, so their correlation is 0.6. They form a joint normal repeated-measurement example with the same marginal distribution at both times.
For jointly standard normal observations with correlation rho, the conditional mean is E[Z2|Z1=z]=rho z. At an observed baseline z=2, the expected next value is 0.6×2=1.2. The expected change is 1.2−2=−0.8, an apparent improvement despite the complete absence of an intervention effect.
The selected case was not necessarily a stable two-standard-deviation problem. Its initial value contains information about both Q and an unusually high e1. In this model, the expected stable component conditional on Z1=2 is 1.2; the next independent measurement does not preserve the initial measurement’s unusually high residual.
The conditional SD of Z2 given Z1 is sqrt(1−0.6²)=0.8. A conditional mean of 1.2 is not an individual prediction of exactly 1.2, and some cases can become more extreme. The statement concerns conditional expectation, not a deterministic recovery path.
Do not confuse a point baseline with threshold selection
The calculation at Z1=2 describes the conditional mean at that value. A group selected by Z1>2 has an average baseline above 2, not exactly 2. Under the same standard normal model, its mean baseline is about 2.3732, using the normal upper-tail mean.
Its expected follow-up mean is 0.6×2.3732≈1.4239, with expected change about −0.9493. Using the point calculation 1.2 as the selected group’s mean would mix two different conditioning events. Save the threshold and the selected group’s actual baseline distribution before choosing a calculation.
Selection at the low extreme reverses the direction. A case with Z1=−2 has expected next value −1.2, an expected increase 0.8 under the stable model. The common feature is movement of the conditional expectation toward the population mean, not a universal downward trend.
The needed correlation concerns the relevant repeat interval and measurement process. A same-afternoon correlation may not describe a six-month remeasurement. Estimating the correlation only within a truncated extreme subgroup can also differ from the population relationship needed for the model. Retain evidence from an appropriate unselected repeat sample when available.
A real shift can coexist with regression to the mean
For a separate illustrative boundary, suppose a genuine common change reduces the second measurement by 0.5 while the original stable model otherwise remains intact. At Z1=2, the new conditional mean is 1.2−0.5=0.7. The expected observed change is −1.3, combining −0.8 from extreme-baseline conditioning with −0.5 from the stipulated shift.
This calculation demonstrates coexistence, not a method for recovering an unknown effect from a single observed change. A follow-up value 0.7 could also arise from sampling variation or a different measurement process. Evidence is needed to identify the shift rather than merely assume it.
An entire population can also genuinely improve, or the metric’s definition can change. Those are different mechanisms from stable-distribution regression to the mean. Check unselected comparison cases, measurement versions and calendar conditions before importing the simple conditional formula.
Design the comparison before the follow-up arrives
A concurrent randomized comparison among cases satisfying the same extreme-baseline rule can distinguish a treatment contrast from improvement shared by both groups, provided the trial’s assumptions and observation quality are maintained. Comparing extreme treated cases with ordinary untreated cases does not provide that safeguard; their selection mechanisms differ.
Repeated baseline measurements can reveal whether the selected extremity persists. Preserve one selection measurement and an independently collected baseline where the design calls for it. A cached duplicate supplies no independent check, and another reading from the same short-lived common shock may have less independence than intended.
Do not claim that averaging baselines makes all bias disappear. It changes the selection measure and its reliability, so the resulting group and conditional calculation change as well. Specify whether eligibility uses the first reading, an average or a sustained threshold, and retain all scheduled measurements rather than only those that justify inclusion.
Execute an extreme-baseline audit
| Step | Evidence to preserve | Response to a gap |
|---|---|---|
| Recover selection | threshold, metric, dates and eligible population | do not infer a causal before/after effect |
| Check repetition | independent readings, versions and interval | separate caching and common shocks |
| Specify the model | distribution, means, variances and correlation | label the numeric result conditional |
| Match conditioning | exact value versus selected tail | use the correct baseline distribution |
| Assess real change | concurrent comparison and process evidence | retain alternative explanations |
| Report | observed change and identified treatment contrast separately | avoid calling all change an effect |
If a provider revised historical scores, reconstruct selection using the information actually available at the original decision date. A corrected score may be useful for a new analysis, but it should not silently rewrite why a case was selected. Likewise, missing follow-up for the worst cases can create additional selection bias that regression-to-the-mean arithmetic does not repair.
When a repeated-measurement model is implausible, do not force a numerical subtraction from every observed change. Use it as a diagnostic explanation, collect better baseline information or redesign the comparison. Stable means, normality and the relevant correlation are substantive assumptions, not features guaranteed by standardizing a column.
Update the judgment when the selection record changes, new independent baselines alter the estimated repeat relationship, or a well-designed comparison supplies evidence of a genuine shift. Preserve the original observed change and distinguish it from the causal quantity the study can identify. That leaves room for real improvement while preventing an extreme starting point from receiving all the explanatory credit.
Check the algebra of baseline and change
In the stable standardized model, the unconditional mean change is zero. Yet Cov(Z1,Z2−Z1)=0.6−1=−0.4. The variance of change is 1+1−2×0.6=0.8, so baseline and change have correlation about −0.4472. A negative baseline-change association therefore appears without any true change mechanism.
This is not a prohibition on analyzing change in a suitable randomized design. It is a warning against interpreting an extreme-baseline versus subsequent-change scatterplot as proof that the worst cases responded most strongly to an intervention. The baseline appears with opposite sign in the change score, and the selection rule further changes the conditional distribution. Save the repeated raw measurements as well as the derived difference so another reviewer can reproduce that coupling. Compare equivalent eligibility groups and the planned causal contrast before describing differential response.
FAQ
Does regression to the mean guarantee that every extreme case improves?
No. It is a statement about conditional expectations in a specified process. Individual follow-up measurements remain variable.
Can it explain every business recovery?
No. Real changes, changing populations and measurement revisions can coexist. Their contributions require separate evidence and an appropriate comparison design.