Baseline Measurement Error and Regression to the Mean

Why a Stable Outcome Can Appear to Change

A participant whose recorded baseline outcome is unusually high may have a lower value at the next visit, even if the underlying state has not changed. The reverse can occur for an unusually low baseline. This pattern is called regression to the mean. It does not require a treatment effect, a placebo response, or a true biological change: it can arise because the baseline measurement combines a persistent individual state with measurement error.

In pharmacometrics, this distinction matters when baseline is used to describe an initial condition, when participants are selected because of an extreme baseline value, or when change from baseline is analyzed. A model that treats the recorded value as the exact underlying state can attribute measurement noise to apparent response. A latent-baseline model instead represents the underlying state and the error in its observation separately.

Definition: Let \(B_i\) be participant \(i\)'s latent baseline state: the underlying value relevant to the model but not observed without error. Let \(Y_{i0}\) be the recorded baseline measurement and \(Y_{i1}\) a later measurement. Regression to the mean is the tendency for the conditional expected later measurement, given an unusually high or low recorded baseline, to be closer to the population mean than that recorded baseline is.

A Simple Measurement-Error Model

To isolate the mechanism, suppose the underlying state is stable between baseline and follow-up. For one participant, write

B \sim N(\mu,\tau^2), \qquad Y_0 = B+\varepsilon_0,\qquad Y_1 = B+\varepsilon_1,

where \(B\) is the latent baseline, \(\mu\) is the population mean latent baseline, and \(\tau^2\) is the between-participant variance. The measurement errors \(\varepsilon_0\) and \(\varepsilon_1\) are normally distributed, \(\varepsilon_0,\varepsilon_1\sim N(0,\sigma^2)\), and are independent of \(B\) and of each other. Thus \(Y_0\) and \(Y_1\) have units of the outcome, while \(\tau^2\) and \(\sigma^2\) have squared outcome units. These assumptions describe a stable latent state plus independent measurement error; they do not describe every longitudinal outcome process.

Under this model, the recorded baseline has variance \(\tau^2+\sigma^2\). Its reliability, in this particular model, is the fraction of that variance attributable to differences in latent states:

\lambda = \frac{\tau^2}{\tau^2+\sigma^2}.

The conditional expected latent baseline given the recorded value \(Y_0=y_0\) is

E(B\mid Y_0=y_0) = \mu+\lambda(y_0-\mu).

Because the follow-up error has mean zero and is independent of the baseline information, \(E(Y_1\mid Y_0=y_0)=E(B\mid Y_0=y_0)\). Therefore,

E(Y_1-Y_0\mid Y_0=y_0) = -\frac{\sigma^2}{\tau^2+\sigma^2}(y_0-\mu).

When \(y_0\) is above \(\mu\), the expected change is negative; when \(y_0\) is below \(\mu\), it is positive. If measurement error is zero, then \(\lambda=1\) and this model predicts no regression to the mean. If measurement error is large relative to between-person variation, the recorded baseline is a less informative indicator of the latent state and the expected movement toward \(\mu\) is larger.

Deriving the Conditional Expectation

The conditional result follows from the joint normal model. Since \(Y_0=B+\varepsilon_0\), its mean is \(\mu\), and its variance is \(\tau^2+\sigma^2\). Its covariance with \(B\) is \(\operatorname{Cov}(B,Y_0)=\tau^2\). For jointly normal variables, the conditional mean is the mean of \(B\) plus the covariance divided by the variance of \(Y_0\), multiplied by the observed deviation of \(Y_0\) from its mean:

E(B\mid Y_0=y_0) = \mu+ \frac{\operatorname{Cov}(B,Y_0)}{\operatorname{Var}(Y_0)}(y_0-\mu) = \mu+\frac{\tau^2}{\tau^2+\sigma^2}(y_0-\mu).

Subtracting the observed baseline gives the expected change. Since \(E(Y_1\mid Y_0=y_0)=E(B\mid Y_0=y_0)\),

E(Y_1-Y_0\mid Y_0=y_0) = \mu+\lambda(y_0-\mu)-y_0 = -(1-\lambda)(y_0-\mu) = -\frac{\sigma^2}{\tau^2+\sigma^2}(y_0-\mu).

This derivation makes the source of the pattern explicit: the baseline deviation is partly latent signal and partly baseline-specific noise. The follow-up shares the latent signal but not the baseline measurement error.

Worked Example: One Unusually High Baseline

Suppose a hypothetical outcome is stable for each participant. Latent baselines have mean \(100\) points and standard deviation \(12\) points, while independent measurement errors have standard deviation \(8\) points. A participant has a recorded baseline of \(120\) points. Find the expected latent baseline, expected follow-up, and expected change conditional on this recorded baseline.

The variances are \(\tau^2=12^2=144\) points squared and \(\sigma^2=8^2=64\) points squared. Thus the reliability is

\lambda=\frac{144}{144+64}=\frac{144}{208}=0.6923.

The conditional expected latent baseline is

E(B\mid Y_0=120) =100+0.6923(120-100) =113.85\text{ points}.

The expected follow-up is also \(113.85\) points, because the follow-up measurement error has conditional mean zero. The expected change, using the unrounded conditional expectation, is

E(Y_1-Y_0\mid Y_0=120) =113.8462-120 \approx -6.154\text{ points}.

This is an expected change under the stated stable-state model, not evidence that the participant’s underlying state improved by \(6.154\) points. A single future measurement can differ substantially from its conditional expectation because it also contains follow-up measurement error.

Observed-Baseline and Latent-Baseline Formulations

An observed-baseline formulation uses the recorded value \(Y_{i0}\) directly as the participant’s baseline or initial condition. A latent-baseline formulation introduces an unobserved state \(B_i\), models how baseline measurements arise from that state, and carries uncertainty in \(B_i\) into predictions. They answer different modeling needs; neither is automatically correct in every application.

Formulation	What baseline represents	Useful when	Main concern
Observed baseline	The recorded baseline value is used directly	Measurement error is small, or the target is explicitly defined relative to the observed value	Baseline-specific error can induce apparent change and distort associations with change
Latent baseline	An underlying state generates one or more observed baseline measurements	Measurement error matters and repeated measurements or other information can identify its contribution	Requires defensible assumptions and enough data to distinguish latent variation from measurement error

The observed-baseline formulation may be reasonable when the measurement is sufficiently precise for the purpose, or when the scientific estimand is specifically change from the recorded value. It should not be interpreted as proof that the recorded baseline equals the underlying biological state. The latent formulation is more explicit about that distinction, but it does not make the latent state directly observable; its estimate depends on the model and available data.

Worked Example: Selection Above a Baseline Threshold

Consider a hypothetical stable outcome with latent baseline mean \(100\) points and between-person standard deviation \(6\) points. Measurement error has standard deviation \(8\) points. Suppose participants are selected because their recorded baseline exceeds \(110\) points. Estimate the mean recorded baseline among selected participants and their expected mean follow-up.

The recorded baseline has mean \(100\) and standard deviation \(\sqrt{36+64}=10\) points. The selection threshold is one standard deviation above the mean, so \(z=(110-100)/10=1\). For a normally distributed variable truncated below at \(z\), the conditional mean is the population mean plus the standard deviation times \(\phi(z)/[1-\Phi(z)]\), where \(\phi\) and \(\Phi\) are the standard normal density and cumulative distribution functions. At \(z=1\), \(\phi(1)\approx0.2420\), \(1-\Phi(1)\approx0.1587\), and their ratio is about \(1.525\). Therefore,

E(Y_0\mid Y_0>110) =100+10(1.525) \approx115.3\text{ points}.

The reliability is \(36/(36+64)=0.36\). Conditional on a particular observed baseline, the expected follow-up is \(100+0.36(Y_0-100)\). Averaging this expression over the selected group gives

E(Y_1\mid Y_0>110) =100+0.36(115.25-100) \approx105.5\text{ points}.

The expected mean change in this selected group is approximately \(105.49-115.25=-9.76\) points. The selected participants have not been shown to improve biologically: the selection rule preferentially includes positive baseline measurement errors, which are not repeated at follow-up.

Worked Example: Two Replicate Baseline Measurements

In another hypothetical setting, let latent baselines have mean \(50\) outcome units and variance \(36\), with independent measurement errors of variance \(64\). A participant has two baseline replicates, \(60\) and \(58\) units. Under the same normal model, find the posterior expected latent baseline and compare it with using only the first replicate.

The replicate average is \((60+58)/2=59\) units. Averaging two independent errors reduces the variance of their mean to \(64/2=32\). The conditional mean of the latent state is therefore

E(B\mid \bar{Y}_0=59) =50+\frac{36}{36+32}(59-50) =50+\frac{36}{68}(9) \approx54.76\text{ units}.

With only the first replicate, the measurement-error variance remains \(64\), so the conditional mean would be \(50+[36/(36+64)](60-50)=53.60\) units. The two replicates provide more information about the latent state; their expected follow-up, under the stable-state assumptions, is \(54.76\) units. Relative to the replicate average of \(59\) units, the expected change is approximately \(54.76-59=-4.24\) units. Replication reduces uncertainty about the state but does not make the measurements error-free.

Diagnostics, Assumptions, and Failure Modes

The formulas above rely on a stable latent state, additive unbiased measurement error, independent baseline and follow-up errors, a normal distribution for latent baselines, and a common measurement-error variance. Before interpreting an apparent change as regression to the mean, consider whether these conditions are plausible. A true time trend, treatment effect, placebo effect, or disease progression can coexist with regression to the mean; the simple model attributes all expected change to measurement error only because it assumes no true change.

Useful checks include comparing replicate baseline measurements, examining whether residual variability changes with the outcome level, and checking whether baseline and follow-up errors could be correlated because of shared assay batches or systematic calibration. A systematic shift between visits is not mean-zero independent noise. Likewise, values near a lower or upper measurement limit may have non-normal or censored errors. If error variance differs by outcome level, the constant-\(\sigma^2\) model may misrepresent the degree of shrinkage.

Selection also matters. Regression to the mean is expected when selection uses an extreme noisy measurement, but not every selected cohort will show the same numerical change. The expected amount depends on the latent variation, measurement error, selection rule, and any genuine time course. A simple before-and-after comparison cannot separate these contributions by itself.

Common Mistakes and Practical Limits

Calling regression to the mean a treatment effect. An extreme recorded baseline followed by a less extreme value is compatible with a stable underlying state. Attribution to treatment requires evidence that distinguishes treatment-associated change from measurement error and other time effects.
Treating baseline error as if it were follow-up error too. In the model, the baseline error \(\varepsilon_0\) is not shared with the independent follow-up error \(\varepsilon_1\). The baseline-specific error therefore contributes to the apparent change.
Assuming every observed decrease is regression to the mean. The conditional expectation describes an average over repeated observations under assumptions. It does not determine why a particular participant changed, nor does it rule out true progression or response.
Assuming a latent-baseline model solves identifiability automatically. One observed measurement per person may not provide enough information to estimate both between-person variation and measurement-error variance without additional structure or information.
Ignoring the estimand. Change relative to a recorded baseline and change relative to an estimated latent state are different quantities. State which one is modeled and interpreted.

An observed-baseline approach can be transparent and useful when measurement error is negligible or when the recorded baseline defines the scientific target. A latent-baseline approach is preferable when measurement error is substantial and can be supported by replicate measurements, external reliability information, or a sufficiently structured model. Both approaches remain limited by the quality and design of the data: no formulation can recover an underlying state from noise without assumptions.

Key takeaway: Under a stable-state normal model with independent additive measurement error, the expected follow-up conditional on an observed baseline is pulled toward the population mean. Quantify that pull using the ratio of between-person variance to total observed variance, and distinguish this statistical pattern from true treatment or disease change.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

Latent baselines have mean \(120\) points and variance \(100\), and measurement errors have variance \(25\). For a recorded baseline of \(130\) points, what are the expected latent baseline and expected change under the stable-state model?
In the model above, what happens to regression to the mean as measurement-error variance approaches zero?
Why can selecting participants with unusually high recorded baselines lead to a lower average follow-up even in the absence of treatment?
What additional information do two independent baseline replicates provide compared with one measurement?

Solutions. 1. The reliability is \(100/(100+25)=0.8\). The expected latent baseline is \(120+0.8(130-120)=128\) points; expected follow-up is also \(128\) points, so expected change is \(128-130=-2\) points. 2. As measurement-error variance approaches zero, reliability approaches one and the expected change approaches zero for any fixed observed deviation, because the observation increasingly reflects the latent state. 3. Selection on a high recorded value tends to select positive baseline errors as well as high latent states. Independent follow-up error does not repeat those positive baseline errors, so the selected group’s expected follow-up is less extreme. 4. Independent replicates reduce the error variance of their average and provide more information about the latent baseline; they do not eliminate uncertainty or guarantee that the model assumptions hold.