Exposure Measurement Error in Interaction Studies

Why Exposure Measurement Matters for Interaction

A gene–environment interaction estimate compares exposure effects across genotype groups, or equivalently measures how the exposure effect changes with genotype under a specified model and scale. If the exposure is measured imperfectly, the comparison can be distorted. A single overall statement such as “measurement error attenuates effects” is not enough: the error can differ across genotype groups, and even identical error processes can have different consequences when exposure distributions differ.

The discussion below uses a binary genotype indicator \(G\), with \(G=0\) and \(G=1\), and distinguishes the true exposure \(X\) from its measured version \(W\). For a continuous outcome, a simple linear mean model makes the interaction especially transparent. The principles also apply to other outcomes, but corrections must match the outcome model and the interaction scale; the linear-model formulas below are not automatic corrections for logistic or risk-scale interactions.

Definition: Exposure measurement error occurs when the recorded exposure \(W\) differs from the exposure \(X\) that defines the scientific question. Random error refers to variation in that discrepancy; misclassification refers to recording an exposure category incorrectly. Measurement is differential when its error distribution depends on the outcome, genotype, or other relevant variables, rather than being conditionally independent of them under the specified measurement model.

Random Error in a Continuous Exposure

Suppose the true conditional mean is linear within genotype groups:

E(Y\mid X,G=g)=a_g+\beta_g X, \qquad \beta_0=\beta_X,\quad \beta_1=\beta_X+\beta_{GX}.

Here \(\beta_{GX}=\beta_1-\beta_0\) is the difference between the two true exposure slopes, in outcome units per exposure unit. This interpretation is tied to the linear outcome scale. If \(W\) replaces \(X\) in the analysis, the estimated slopes need not equal these true slopes.

A commonly studied reference case is classical additive error, \(W=X+U\). The assumptions required for the usual reliability-ratio result must hold within each genotype group. In particular, \(U\) has conditional mean zero, is uncorrelated with \(X\), and is uncorrelated with the outcome residual within each group. The outcome residual must also be uncorrelated with \(X\) within each group. These conditions are stronger than corresponding marginal conditions across the whole sample.

Classical-error result: Let \(\varepsilon=Y-(a_g+\beta_gX)\) within genotype group \(g\). If \(W=X+U\), and within each group \(E(U\mid G=g)=0\), \(\operatorname{Cov}(X,U\mid G=g)=0\), \(\operatorname{Cov}(U,\varepsilon\mid G=g)=0\), and \(\operatorname{Cov}(X,\varepsilon\mid G=g)=0\), then the population slope of \(Y\) regressed on \(W\) within group \(g\) is \(\beta_g^{(W)}=\lambda_g\beta_g\), where \(\lambda_g=\operatorname{Var}(X\mid G=g)/\{\operatorname{Var}(X\mid G=g)+\operatorname{Var}(U\mid G=g)\}\). This \(\lambda_g\), greater than zero and at most one when the variances are finite and the true exposure variance is positive, is the group-specific reliability ratio; it is strictly less than one if the error variance is positive.

Why the Reliability Ratio Changes the Interaction

The result follows directly from the population regression slope, \(\operatorname{Cov}(Y,W\mid G=g)/\operatorname{Var}(W\mid G=g)\). Under the stated assumptions, the numerator is \(\beta_g\operatorname{Var}(X\mid G=g)\), while the denominator is \(\operatorname{Var}(X\mid G=g)+\operatorname{Var}(U\mid G=g)\). Thus their ratio is \(\lambda_g\beta_g\). This derivation uses conditional covariances in each genotype group; marginal lack of correlation between \(X\) and \(U\) does not suffice.

\begin{aligned} \operatorname{Cov}(Y,W\mid G=g) &=\operatorname{Cov}(a_g+\beta_gX+\varepsilon,X+U\mid G=g)\\ &=\beta_g\operatorname{Var}(X\mid G=g),\\ \operatorname{Var}(W\mid G=g) &=\operatorname{Var}(X\mid G=g)+\operatorname{Var}(U\mid G=g),\\ \beta_g^{(W)} &=\frac{\operatorname{Cov}(Y,W\mid G=g)} {\operatorname{Var}(W\mid G=g)} =\lambda_g\beta_g. \end{aligned}

Consequently, the observed difference in slopes is \(\beta_1^{(W)}-\beta_0^{(W)}=\lambda_1(\beta_X+\beta_{GX})-\lambda_0\beta_X\). If the reliability ratios are equal, this difference is \(\lambda\beta_{GX}\): the interaction is attenuated toward zero under this model. If they differ, the observed interaction is not generally a simple attenuated version of the true interaction. It can be biased in either direction, or appear when the true interaction is zero.

Worked Example: Different Reliability Can Reverse an Interaction

Consider a hypothetical study with a continuous outcome and exposure. The true exposure slope is \(2\) outcome units per exposure unit for \(G=0\), and \(3\) for \(G=1\), so the true interaction contrast is \(3-2=1\). Assume the conditional classical-error conditions above hold in each group. In both groups let \(\operatorname{Var}(X\mid G=g)=1\), but let \(\operatorname{Var}(U\mid G=0)=0.25\) and \(\operatorname{Var}(U\mid G=1)=1\).

For \(G=0\), \(\lambda_0=1/(1+0.25)=0.8\), so the expected measured-exposure slope is \(0.8(2)=1.6\). For \(G=1\), \(\lambda_1=1/(1+1)=0.5\), so that slope is \(0.5(3)=1.5\). The observed interaction contrast is \(1.5-1.6=-0.1\) outcome units per exposure unit. Thus an underlying positive contrast of \(1\) has become a slightly negative contrast. The calculation does not imply that the true interaction is negative; it shows how the differing error variances distort the comparison.

Even equal error variance does not guarantee equal reliability: the true exposure variance can differ by genotype. Conversely, genotype-specific reliability can arise from different instruments, reporting practices, or exposure distributions. The assumptions should therefore be assessed within the genotype groups and, where needed, within important covariate strata. If the error is related to the outcome residual or the true exposure after conditioning as specified, the reliability formula no longer follows.

Binary Exposure Misclassification

For a binary true exposure \(D\), let \(D^*\) be the recorded classification. Within group \(g\), sensitivity is \(Se_g=P(D^*=1\mid D=1,G=g)\), and specificity is \(Sp_g=P(D^*=0\mid D=0,G=g)\). Under nondifferential misclassification for this calculation, the recorded classification is independent of \(Y\) conditional on \(D\) and \(G\). This assumption still permits sensitivity and specificity to differ by genotype.

Suppose the true mean outcome is linear in \(D\) within group \(g\), with exposure contrast \(\delta_g=E(Y\mid D=1,G=g)-E(Y\mid D=0,G=g)\), and let \(q_g=P(D=1\mid G=g)\). Then \(q_g^*=P(D^*=1\mid G=g)=Se_gq_g+(1-Sp_g)(1-q_g)\). The regression slope using \(D^*\) is the difference in mean outcomes across recorded categories. When \(0<q_g<1\) and \(0<q_g^*<1\), so both the true exposure contrast and the recorded-exposure slope are defined, the slope can be expressed as:

\frac{\operatorname{Cov}(Y,D^*\mid G=g)} {\operatorname{Var}(D^*\mid G=g)} = \frac{\delta_g(Se_g+Sp_g-1)q_g(1-q_g)} {q_g^*(1-q_g^*)}.

This result follows because \(E(D^*\mid D,G=g)=(1-Sp_g)+(Se_g+Sp_g-1)D\). Conditional independence of \(D^*\) and \(Y\) given \(D,G\) gives \(\operatorname{Cov}(Y,D^*\mid G=g)=\delta_g(Se_g+Sp_g-1)q_g(1-q_g)\); the variance of the binary \(D^*\) is \(q_g^*(1-q_g^*)\). When exposure prevalence, sensitivity, or specificity differs by genotype, the distortion can differ across groups and alter the interaction contrast.

Worked Example: Genotype-Specific Misclassification Creates a Contrast

In a hypothetical setting, suppose the true exposure has prevalence \(q_0=q_1=0.5\) and the true outcome contrast is \(\delta_0=\delta_1=2\) outcome units. There is no true exposure-by-genotype interaction. Assume exposure classification is independent of the outcome conditional on true exposure and genotype. For \(G=0\), take \(Se_0=0.9\) and \(Sp_0=0.9\); for \(G=1\), take \(Se_1=0.6\) and \(Sp_1=0.9\).

For \(G=0\), \(q_0^*=0.9(0.5)+0.1(0.5)=0.5\). The recorded-exposure slope is \(2(0.9+0.9-1)(0.5)(0.5)/\{(0.5)(0.5)\}=1.6\) outcome units. For \(G=1\), \(q_1^*=0.6(0.5)+0.1(0.5)=0.35\), so the slope is \(2(0.6+0.9-1)(0.5)(0.5)/\{(0.35)(0.65)\}=0.25/0.2275\approx1.099\) outcome units. The measured interaction contrast is approximately \(1.099-1.6=-0.501\) outcome units, despite equal true exposure effects. This is distortion from genotype-specific classification performance, not evidence of biological effect modification.

Differential Error and Its Consequences

Differential measurement error is broader than genotype-specific sensitivity or reliability. For example, exposure reporting may differ between outcome cases and non-cases, or the measurement process may depend on genotype after conditioning on true exposure. In these circumstances, the assumptions used in the preceding formulas fail. Error can create an exposure association, change its direction, or alter the estimated interaction; there is no universal direction of bias.

The word “differential” must be defined relative to a model. An error process can be nondifferential conditional on \(D\) and \(G\), yet have different sensitivity across genotype groups, as in the worked example. Conversely, an instrument can have the same marginal accuracy in both groups but still be related to the outcome within exposure and genotype strata. A useful analysis specifies the distribution of the recorded measure conditional on the true exposure and the variables that could affect measurement.

Worked Example: Validation Data and a Sensitivity Range

Suppose a hypothetical main study estimates slopes using a noisy continuous exposure: \(b_0^{(W)}=1.2\) and \(b_1^{(W)}=1.4\), in outcome units per recorded exposure unit. A validation substudy obtains a higher-quality exposure measure \(X\) on a subset of participants. Assume that it is representative within genotype groups and supports the classical-error model. Estimated variances in the validation data are \(\operatorname{Var}(X\mid G=0)=3\), \(\operatorname{Var}(U\mid G=0)=1\), \(\operatorname{Var}(X\mid G=1)=2\), and \(\operatorname{Var}(U\mid G=1)=2\).

The corresponding estimated reliability ratios are \(\hat{\lambda}_0=3/(3+1)=0.75\) and \(\hat{\lambda}_1=2/(2+2)=0.50\). Under the model, divide each observed slope by its group-specific reliability: the corrected slopes are \(1.2/0.75=1.60\) and \(1.4/0.50=2.80\). The corrected interaction contrast is \(2.80-1.60=1.20\) outcome units per exposure unit. This correction is conditional on the measurement assumptions; it is not justified merely because a validation subset exists.

To show dependence on the error assumptions, suppose plausible reliability ranges are \(0.70\) to \(0.80\) for \(G=0\) and \(0.45\) to \(0.60\) for \(G=1\). The corrected contrast over this grid ranges from \(1.4/0.60-1.2/0.70\approx0.619\) to \(1.4/0.45-1.2/0.80\approx1.611\) outcome units per exposure unit. This is a sensitivity range generated by the selected reliability values, not a confidence interval. It does not include sampling uncertainty unless that uncertainty is also propagated.

Using Validation Substudies and Sensitivity Analysis

A validation substudy measures the exposure more accurately in at least a subset of the main study. Its usefulness depends on the reference measure being sufficiently accurate for the scientific purpose, the subset representing the main study with respect to error processes, and the availability of enough information across genotype and relevant covariate groups. If validation participation depends on exposure, outcome, or genotype, the design and analysis must account for that selection. Validation estimates also have uncertainty, which should not be treated as zero.

Possible analysis strategies include estimating an explicit measurement model, using likelihood methods that combine main-study and validation data, or using a correction appropriate to the specified outcome model. For continuous exposures under classical error, group-specific reliability corrections have a clear interpretation. For binary exposures, sensitivity and specificity can inform a misclassification correction. More complex differential error generally requires a model for the error process, rather than one pooled reliability number. When the reference measure is itself imperfect, its error must also be considered.

Without validation data, sensitivity analysis asks how the interaction estimate would change across plausible values for error parameters. The analyst should state which parameters vary, why their values are plausible, which assumptions are held fixed, and whether the reported range is a point-estimate range or includes uncertainty. Such an analysis does not identify the true interaction by itself. Its value is to reveal whether the substantive conclusion depends strongly on unverified measurement assumptions.

Specify the target.
Define the true exposure, genotype contrast, outcome scale, and interaction estimand before describing error.

Describe measurement conditionally.
State how the distribution of \(W\) given \(X\) may vary by genotype, outcome, or covariates, and identify the assumptions needed for the chosen correction.

Use observed or plausible error information.
Estimate error parameters from suitable validation data when possible; otherwise choose and justify sensitivity-analysis values.

Re-estimate and report the interaction.
Apply the correction within relevant groups, propagate uncertainty where feasible, and distinguish corrected estimates from assumptions-driven sensitivity ranges.

Common Mistake and Key Takeaway

Wrong reasoning: “The exposure error is random overall and unrelated to the outcome, so the interaction must just be attenuated toward zero.” This does not follow. For the reliability-ratio derivation, the required uncorrelatedness conditions must hold within each genotype group. Even when they do, group-specific reliability ratios can distort the slope difference in either direction. Binary misclassification can also distort effects differently as a function of prevalence and classification accuracy.

Correction: describe and assess measurement error within the genotype groups used to define the interaction. State the assumptions behind any correction, and do not transfer a formula from one outcome scale or exposure type to another without justification. An estimate that changes after correction is not automatically the truth; it is an estimate under a more explicit measurement model.

Key takeaway: Exposure error affects the comparison that defines gene–environment interaction. A validation substudy can help estimate that error, while sensitivity analysis can show how conclusions depend on it; neither removes the need to state and defend the measurement assumptions.

Check Your Understanding

Answer each question using the assumptions and scales stated in the tutorial.

In the continuous classical-error model, why must \(\operatorname{Cov}(X,U\mid G=g)=0\) hold within each genotype group rather than only in the pooled sample?
If both genotype groups have the same reliability ratio, what happens to the linear exposure-slope interaction under the stated model? What additional issue arises if the reliability ratios differ?
For a binary exposure, name the quantities needed to calculate the recorded-exposure slope within a genotype group under nondifferential misclassification.
In the binary-exposure worked example, which feature of the measurement process produced an apparent interaction when the true exposure effects were equal?
Why is a sensitivity range based on plausible reliability ratios not automatically a confidence interval?