Measurement Error In Time-Varying Confounders

Why Measurement Error Matters for Later Treatment

A time-varying confounder can influence the next treatment decision and the outcome. If the covariate is measured imperfectly, adjusting for its recorded value does not generally adjust for the true covariate. The remaining association between true covariate and treatment can therefore leave residual confounding in an outcome regression.

The resulting bias is not governed by a universal rule that a treatment coefficient moves toward zero. Its direction depends on how the true confounder predicts treatment and outcome; its magnitude also depends on treatment variation and measurement precision. The derivations below isolate this mechanism in linear models. They do not imply that measurement error is the only possible source of bias in a longitudinal analysis.

Definition: Let $L$ be a true covariate, $W=L+U$ its measured value, and $U$ measurement error. The measurement is classically error-prone here when $U$ has mean zero and is independent of $L$, treatment-assignment errors, and outcome errors. A regression that adjusts for $W$ rather than $L$ may leave residual confounding if $L$ still predicts both treatment and outcome after adjustment for $W$.

A Single-Decision Derivation

First consider a centered, scalar covariate $L$, treatment $A$, and outcome $Y$, with the linear equations

A=\alpha L+E_A,\qquad Y=\tau A+\beta L+E_Y,\qquad W=L+U.

Assume $L,U,E_A,E_Y$ are mutually independent, have mean zero and finite variances, and let $\operatorname{Var}(L)=s>0$, $\operatorname{Var}(U)=r\geq 0$, and $\operatorname{Var}(E_A)=q>0$. The outcome error is independent of the regressors. These assumptions make the population least-squares coefficient well-defined. The coefficient $\tau$ is the treatment coefficient in the specified outcome equation; $\beta$ is the covariate's outcome coefficient.

Let $b_A$ denote the population coefficient of $A$ in the regression of $Y$ on an intercept, $A$, and $W$. The Frisch-Waugh-Lovell theorem says that this coefficient is the slope from regressing the part of $Y$ left after linear projection on $W$ on the part of $A$ left after linear projection on $W$. Since the part of $Y$ due to $\tau A$ contributes $\tau$ to that slope, the additional term is $\beta$ times the residual covariance of $A$ and $L$, divided by the residual variance of $A$.

Theorem: Under the assumptions above, the population coefficient from regressing $Y$ on $A$ and $W$ is $$ b_A=\tau+\frac{\beta\alpha s r}{q(s+r)+\alpha^2sr}. $$ The second term is the residual-confounding bias relative to $\tau$ in this specified regression.

To verify the formula, the linear projection coefficient of $L$ on $W$ is $\operatorname{Cov}(L,W)/\operatorname{Var}(W)=s/(s+r)$. Thus the residual from projecting $L$ on $W$ is $L-\frac{s}{s+r}W$, and its variance is

s-\frac{s^2}{s+r}=\frac{sr}{s+r}.

Because $E_A$ is independent of $L$ and $W$, the residual from projecting $A=\alpha L+E_A$ on $W$ has variance $q+\alpha^2sr/(s+r)$. Its covariance with the residual of $L$ is $\alpha sr/(s+r)$. The outcome equation therefore gives

b_A-\tau =\beta\frac{\alpha sr/(s+r)}{q+\alpha^2sr/(s+r)} =\frac{\beta\alpha sr}{q(s+r)+\alpha^2sr}.

The independent, mean-zero outcome error contributes no covariance with the treatment residual. The displayed expression follows, including the case $r=0$, in which the true covariate is measured exactly and the bias is zero. $\square$

Worked Example: Computing Residual Confounding

Suppose $s=4$, $r=1$, $q=1$, $\alpha=0.5$, $\beta=2$, and $\tau=1$. The treatment responds positively to $L$, and $L$ predicts a higher outcome after treatment adjustment. What coefficient does the regression on $A$ and $W$ produce?

The bias is

\frac{\beta\alpha sr}{q(s+r)+\alpha^2sr} =\frac{2(0.5)(4)(1)}{1(4+1)+(0.5)^2(4)(1)} =\frac{4}{5+1} =\frac{2}{3}.

Thus $b_A=1+\frac{2}{3}=\frac{5}{3}$. As a check, the residual variance of $L$ after projection on $W$ is $4(1)/5=0.8$. The residual covariance of $A$ and $L$ is $0.5(0.8)=0.4$, while the residual variance of $A$ is $1+(0.5)^2(0.8)=1.2$. The bias is $2(0.4)/1.2=2/3$, agreeing with the formula.

How Bias Changes With Measurement Precision

The error variance $r$ measures how noisy the recorded covariate is relative to the scale of $L$. Under the theorem's assumptions, the magnitude of the bias is

|b_A-\tau|=\frac{|\beta\alpha|sr}{qs+r(q+\alpha^2s)}.

For $s>0$ and $q>0$, this is zero at $r=0$, increases strictly with $r$ when $\alpha\beta\ne 0$, and approaches $\left|\beta\alpha s/(q+\alpha^2s)\right|$ as $r$ grows without bound. To see the monotonicity, write $C=|\beta\alpha|s$. The derivative with respect to $r$ is

\frac{d}{dr}\left\{\frac{Cr}{qs+r(q+\alpha^2s)}\right\} =\frac{Cqs}{[qs+r(q+\alpha^2s)]^2}.

This derivative is positive if $C>0$, since $q,s>0$, and is zero if $\alpha\beta=0$. If $s=0$, there is no variation in $L$ to leave unadjusted; if $\beta=0$ or $\alpha=0$, this particular confounding path also gives zero bias. The formula is signed: when $s,r,q>0$ and $\alpha\beta\ne0$, the sign of the bias is the sign of $\alpha\beta$, not necessarily positive. $\square$

Worked Example: The Direction Need Not Be Toward Zero

Let $s=2$, $r=3$, $q=2$, $\alpha=-1$, $\beta=3$, and $\tau=1$. The confounder predicts less treatment but a higher outcome. The bias is

b_A-\tau =\frac{(3)(-1)(2)(3)}{2(2+3)+(-1)^2(2)(3)} =\frac{-18}{10+6} =-\frac{9}{8}.

Consequently $b_A=1-\frac{9}{8}=-\frac{1}{8}$, which has the opposite sign from $\tau=1$. This is not a paradox: the fitted coefficient combines the treatment association with the residual association induced by the incompletely controlled confounder. The example also shows why “measurement error attenuates the treatment effect” is not a valid general description.

When the Confounder Follows Prior Treatment

The same calculation applies to a later decision after prior treatment has changed the confounder. Let $A_0$ be prior treatment, $L_1$ the later true covariate, $A_1$ later treatment, $W_1=L_1+U_1$ its measurement, and $Y$ the outcome. Consider

L_1=\gamma A_0+V,\qquad A_1=\delta A_0+\alpha L_1+E_1,\qquad Y=\theta A_0+\tau A_1+\beta L_1+E_Y.

Assume the variables are centered (or the equations include intercepts), $\operatorname{Var}(V)=s>0$, $\operatorname{Var}(E_1)=q>0$, and $\operatorname{Var}(U_1)=r\geq0$. Assume $V$ is independent of $A_0$ and of $E_1,U_1,E_Y$; assume $E_1,U_1,E_Y$ are mutually independent and independent of $A_0$. Also assume $A_0$ has positive variance and all variables have finite second moments. These conditions ensure that, after linear adjustment for $A_0$, the remaining variation in $L_1$ is $V$, with variance $s$. Independence of $V$ and $A_0$ is essential to that calculation.

Result: In the population regression of $Y$ on $A_0,A_1,W_1$, the coefficient $b_1$ of later treatment is $$ b_1=\tau+\frac{\beta\alpha s r}{q(s+r)+\alpha^2sr}. $$

For the derivation, first remove the linear projections on $A_0$. The residuals of $L_1,A_1,W_1$ are respectively $V,\alpha V+E_1,V+U_1$. The outcome residual is $\tau(\alpha V+E_1)+\beta V+E_Y$. Projecting the residuals of $A_1$ and $V$ on $V+U_1$ gives residual covariance $\alpha sr/(s+r)$ and residual treatment variance $q+\alpha^2sr/(s+r)$, as in the single-decision derivation. The outcome residual's $\tau$-part contributes $\tau$; its remaining $ \beta V$-part contributes the ratio of these two quantities multiplied by $\beta$. Simplifying yields the stated formula. Thus prior treatment may determine the level of the later confounder, but the residual-confounding calculation depends on the variation in $L_1$ left after accounting for prior treatment.

Worked Example: A Noisy Covariate in a Feedback Model

Suppose $A_0$ has variance 1, $L_1=A_0+V$ with $\operatorname{Var}(V)=4$, and the independent errors have variances $\operatorname{Var}(E_1)=1$ and $\operatorname{Var}(U_1)=1$. Let $A_1=0.3A_0+0.5L_1+E_1$, and $Y=2A_0+A_1+2L_1+E_Y$, where $E_Y$ is independent and mean zero. Here $\gamma=1$, $\alpha=0.5$, $\beta=2$, $\tau=1$, $s=4$, $q=r=1$.

After adjusting for $A_0$, the remaining covariate is $V$, and the remaining later treatment is $0.5V+E_1$. The bias in the coefficient of $A_1$ is

\frac{2(0.5)(4)(1)}{1(4+1)+(0.5)^2(4)(1)} =\frac{4}{6} =\frac{2}{3}.

The regression coefficient of $A_1$ is therefore $1+\frac{2}{3}=\frac{5}{3}$. If $V$ were not independent of $A_0$, its residual variance after adjustment for $A_0$ would not generally equal 4, so this numerical result would not follow from the stated variances alone.

Diagnostics, Alternatives, and Limitations

The derivation isolates one mechanism under a correctly specified linear model. In practice, the true covariate is often not fully observed, so the residual covariance in the proof cannot be calculated directly. Useful warning signs include a weak or inconsistent relationship between the recorded covariate and plausible indicators of the underlying state, substantial disagreement between repeated measurements, and sensitivity of the treatment coefficient to measurement procedures or validation data. These signs do not by themselves identify the bias or its direction.

A common mistake is to reason that adding the measured covariate to the regression has “controlled for” the confounder, or that measurement error must only shrink the treatment coefficient. The correction is to distinguish $L$ from $W$: adjustment removes the linear association represented by $W$, not necessarily the association with $L$. The formula shows that the signed bias depends on $\alpha\beta$, and can be positive, negative, or zero.

If replicate measurements or a validation sample are available, they can help estimate the measurement process and the relationship between $W$ and $L$. Regression calibration and likelihood-based measurement-error models are alternatives that explicitly model this relationship; their validity requires suitable measurement-model assumptions and adequate data. An ordinary regression or a weighting procedure that simply substitutes $W$ for $L$ does not automatically recover control for the true confounder. Conversely, this tutorial's formula is not a general correction formula: nonlinear relations, differential error, additional confounders, and model misspecification require separate analysis. The limitations of a linear coefficient also remain: even with error-free covariates, it need not summarize the effect of a treatment regime.

Key takeaway: A noisy measurement of a time-varying confounder can leave later treatment associated with the true covariate after adjustment. In the stated linear model, the remaining treatment-coefficient bias is $\beta\alpha sr/[q(s+r)+\alpha^2sr]$; its sign and size depend on the full set of parameters.

Check Your Understanding

Work through each question, then compare with the solutions below.

In the single-decision model, let $s=q=\alpha=\beta=1$, $r=1$, and $\tau=0.5$. Find the coefficient from the regression on $A$ and $W$.
Let $s=4$, $q=1$, $\alpha=0.5$, $\beta=2$, $r=0.25$, and $\tau=1$. Calculate the bias and fitted treatment coefficient.
Name conditions in the stated model under which the residual-confounding bias is zero, even if $r>0$.
Why is it not enough, in the feedback model, to specify that the measurement and treatment errors are mutually independent when calculating the variance of the later covariate after adjustment for $A_0$?

Solutions. (1) The bias is $1(1)(1)(1)/[1(1+1)+1(1)(1)]=1/3$, so the coefficient is $0.5+1/3=5/6$. (2) The bias is $2(0.5)(4)(0.25)/[1(4.25)+(0.25)(4)(0.25)]=1/4.5=2/9$; the coefficient is $1+2/9=11/9$. (3) Bias is zero if $r=0$, or if $\alpha=0$, or if $\beta=0$; it is also zero if $s=0$. (4) The residual variance of $L_1=\gamma A_0+V$ depends on the association between $V$ and $A_0$. Mutual independence among the errors does not establish that $V$ is independent of $A_0$. The stated derivation requires that independence, or a direct specification of the needed conditional moments.