Measurement Error In Longitudinal Omics Predictors

Why Repeated Assays Can Distort Survival Associations

A time-updated survival analysis relates the hazard at time $t$ to predictor values available up to that time. In omics studies, however, the assay value is not the biological quantity itself: it also contains technical error and short-term variation. If the observed assay is entered directly into a Cox model, the resulting association may be biased, even when the assay error has mean zero. The problem is especially important when the biological signal changes over time, replicate assays are sparse, or the measurement error is large relative to between-person variation.

The earlier material on time-dependent omics covariates in Cox models defines how a time-varying predictor enters the hazard, and the material on choosing measurement windows addresses which assays are available to represent exposure at a given time. Here the concern is different: even after selecting an appropriate time window, how does error in the assay affect the estimated association?

Definition: Let $X_i(t)$ be the latent, or true-value, omics predictor for individual $i$ at time $t$, and let $W_{ij}$ be the $j$th observed assay at time $t_{ij}$. A basic additive measurement model is $$ W_{ij}=X_i(t_{ij})+U_{ij}, $$ where $U_{ij}$ is measurement error. A time-updated survival model may specify $$ \lambda_i(t\mid X_i(t))=\lambda_0(t)\exp\{\beta X_i(t)+\boldsymbol{\gamma}^{\mathsf T}\mathbf Z_i(t)\}, $$ where $\lambda_0(t)$ is the baseline hazard, $\mathbf Z_i(t)$ denotes other covariates, and $\beta$ is the log hazard ratio for a one-unit increase in the true predictor, conditional on the other covariates.

The additive model is a useful starting point, not a universal assay model. It commonly assumes that, conditional on the latent value, error has mean zero, is independent across replicate assays, and is independent of survival and covariates. In a simple homoscedastic version, $E(U_{ij}\mid X_i(t_{ij}))=0$ and $\operatorname{Var}(U_{ij}\mid X_i(t_{ij}))=\sigma_U^2$. Batch effects, detection limits, intensity-dependent error, or errors correlated over time require a richer measurement model. The survival model also requires its usual assumptions, including an appropriately specified hazard relation and censoring that is non-informative conditional on included information.

What Goes Wrong With the Naive Cox Model?

A naive analysis substitutes $W_{ij}$, or a time-updated value constructed from the observed assays, for $X_i(t)$ and fits the survival model as if the observed value were error-free. Under simple classical error, an association is often pulled toward zero: people with high true values are mixed with people whose measurements happen to be low, and vice versa. But the familiar reliability-ratio attenuation formula belongs to particular linear-model settings. A time-dependent Cox coefficient does not generally equal the true coefficient multiplied by a reliability ratio. Risk-set changes, time variation, covariate adjustment, and the error structure can produce more complicated bias; it is not safe to treat a noisy Cox estimate as predictably attenuated.

A useful summary of assay quality in a simple setting is the reliability ratio, $$ \lambda=\frac{\sigma_X^2}{\sigma_X^2+\sigma_U^2}, $$ where $\sigma_X^2$ is the between-person variance in true values at the relevant time and $\sigma_U^2$ is the assay-error variance. This ratio is between zero and one when both variances are nonnegative and their sum is positive. It is not a universal correction factor for a survival coefficient. Reliability can also vary over time or across the range of the predictor.

Worked Example: Replicates Improve Reliability

Suppose a hypothetical validation study estimates a between-person variance of $0.64$ assay units squared and a within-person technical error variance of $0.36$ assay units squared. Assume the two replicate errors are independent, have mean zero, and have the same variance. Calculate the reliability of one assay and of the mean of two replicates.

For one assay, the reliability ratio is $$ \lambda_1=\frac{0.64}{0.64+0.36}=0.6400. $$ Thus, in this simplified measurement model, 64% of the observed variance is attributable to between-person differences in true values.

The mean of two independent replicates has error variance $0.36/2=0.18$. Its reliability is therefore $$ \lambda_2=\frac{0.64}{0.64+0.18}=\frac{0.64}{0.82}=0.7805. $$ Averaging replicates improves reliability from $0.6400$ to $0.7805$. This describes measurement precision, not a formula for multiplying or dividing a time-dependent Cox coefficient.

The calculation relies on independent replicate errors and a common error variance. If the replicates share a batch effect, averaging them does not remove that shared component, so the $0.36/2$ calculation would be inappropriate.

Regression Calibration

Regression calibration replaces the unobserved predictor with its conditional expectation given the available, error-prone measurements and other permitted information. Write this calibrated value as $$ m_i(t)=E\{X_i(t)\mid \text{assays observed for individual }i\text{ by time }t,\ \mathbf Z_i(t)\}. $$ The phrase “by time $t$” matters: using a future assay to calibrate an earlier risk time can introduce information unavailable then and distort the intended time-updated association. The calibration model may use replicate assays, measured covariates, and a model for the trajectory; its adequacy depends on the data and assumptions.

The method is an approximation for a Cox model, rather than a general identity. Conditional on observed assays, the hazard involves an expectation of an exponential: $$ E\{\exp(\beta X_i(t))\mid W\}, $$ which is not generally equal to $\exp\{\beta E(X_i(t)\mid W)\}$. Regression calibration substitutes the latter expression. It can work well when uncertainty about $X_i(t)$ is modest and the association is not too strong, but those conditions should not be assumed automatically.

Justification: For a small value of $\beta X_i(t)$, the first-order expansion $\exp(\beta X_i(t))\approx 1+\beta X_i(t)$ gives $$ E\{\exp(\beta X_i(t))\mid W\}\approx 1+\beta E\{X_i(t)\mid W\}. $$ Applying the same first-order approximation to $\exp\{\beta m_i(t)\}$ gives $1+\beta m_i(t)$, where $m_i(t)=E\{X_i(t)\mid W\}$. The two expressions therefore agree to first order. Higher-order terms, including predictor uncertainty, explain why the replacement is approximate rather than exact.

Worked Example: Calibrating a Time-Available Predictor

Consider a hypothetical landmark time at which the true omics value $X$ has a normal distribution with mean $2.0$ units and variance $0.64$ units squared. Two assays available by that time have independent additive errors, each with variance $0.36$. For observed values $W_1=2.4$ and $W_2=2.0$, calculate the conditional mean of $X$.

The replicate mean is $\bar W=(2.4+2.0)/2=2.2$ units. Its error variance is $0.36/2=0.18$. Under the stated normal model, the conditional mean is $$ E(X\mid W_1,W_2)=2.0+\frac{0.64}{0.64+0.18}(2.2-2.0). $$ The weight is $0.64/0.82=0.7805$, so the calibrated predictor is $$ 2.0+0.7805(0.2)=2.156\text{ units}. $$ The calibration shrinks the noisy replicate mean toward the population mean because the assays are imperfect. This calibrated value could be used as a time-updated covariate at this landmark, provided the model and timing assumptions are appropriate.

For illustration, if a fitted regression-calibrated Cox model estimated $\beta=0.40$ per assay unit, the hazard ratio comparing calibrated values $2.156$ and $1.156$ units, a one-unit difference, would be $\exp(0.40)=1.492$. This is an estimated conditional association, not evidence that changing the omics value causes a change in hazard. The example does not imply that the same coefficient would result from fitting the naive model.

In practice, the calibration model should reflect the temporal structure of the assay data. A mixed model or other trajectory model can provide conditional predictions and uncertainty, but its assumptions must match the sampling and error process. A single reliability ratio may be inadequate when error varies by time, assay batch, or true-value level.

Latent True-Value Models

A latent true-value approach specifies a model for the unobserved trajectory and a measurement model for the assays, then accounts for uncertainty in the trajectory when estimating the survival association. Unlike simple regression calibration, it does not merely substitute a conditional mean into the hazard. Conceptually, it averages the survival contribution across plausible true trajectories consistent with the observed assays. This can represent the nonlinear effect of uncertainty on the hazard, but it requires stronger modeling and computational work.

For example, let $X_i(\cdot)$ denote the predictor trajectory, let $W_i$ denote the available assay data, and let $L_i^{S}\{X_i(\cdot)\}$ be the survival contribution conditional on a trajectory. If the measurement model has likelihood $L_i^{M}\{W_i\mid X_i(\cdot)\}$ and the trajectory model has density $p\{X_i(\cdot)\}$, an observed-data contribution has the form $$ L_i=\int L_i^{S}\{X_i(\cdot)\}L_i^{M}\{W_i\mid X_i(\cdot)\}p\{X_i(\cdot)\}\,dX_i(\cdot). $$ The integral averages over possible true trajectories rather than treating one noisy assay sequence as truth. The model must state how the trajectory is distributed, how assay errors behave, and how the survival process depends on the trajectory. If visit timing, dropout, or post-event measurement availability is informative, additional modeling is needed; treating the measurement process as ignorable is not justified merely by writing this integral.

Worked Example: Integrating Uncertainty in a Latent Value

At a hypothetical risk time, suppose $X\mid W$ is normally distributed with mean $0.60$ units and variance $0.50$ units squared. Assume the instantaneous hazard multiplier for a one-unit increase in $X$ has coefficient $\beta=0.70$ per unit, with other covariates held fixed. Calculate the conditional expected hazard multiplier under this distribution and compare it with plug-in regression calibration.

For a normal random variable $X\sim N(m,v)$, its exponential moment is $E\{\exp(\beta X)\}=\exp(\beta m+\beta^2v/2)$. Thus, $$ E\{\exp(0.70X)\mid W\} =\exp\{0.70(0.60)+(0.70^2)(0.50)/2\} =\exp(0.42+0.1225) =\exp(0.5425) =1.720. $$ A plug-in calculation using only the conditional mean gives $$ \exp\{0.70(0.60)\}=\exp(0.42)=1.522. $$ The difference arises because the latent-value calculation includes uncertainty in $X$. It is a conditional expected instantaneous multiplier under the specified normal model. It is not, by itself, a complete survival-model estimate: over time, survival and changing risk sets can alter the distribution of latent values among those still at risk.

A latent true-value model is not automatically a shared-random-effects joint model. The focus here is correction for predictor measurement error. Modeling a shared latent effect that jointly drives the longitudinal trajectory and survival, and deriving the associated joint likelihood and interpretation, requires additional assumptions and is a distinct modeling task.

Diagnostics, Alternatives, and Limitations

No correction rescues a badly specified measurement model. Assess whether replicates support the assumed error variance, whether errors are plausibly independent, and whether calibration residuals show systematic patterns by assay level, batch, or time. Compare predicted and observed replicate distributions and, where possible, evaluate calibration in data not used to fit the calibration model. A low reliability ratio warns that the observed assays contain little information about between-person true values; it does not by itself determine the size or direction of Cox-model bias.

Check the timing carefully. A predictor used at time $t$ should be constructed from information available by $t$, consistent with the chosen measurement window. Check whether the measurement error is differential: if error depends on disease status, future survival, or a covariate related to survival after conditioning on the true value, the classical-error calibration may fail. Also examine the proportional-hazards and functional-form assumptions for the survival model; correcting assay error does not correct those separate model failures.

Approach	What it does	Main caution
Naive time-dependent Cox model	Uses observed assays as if they were true values.	Can be biased by measurement error; the direction and size are not universally determined by reliability.
Regression calibration	Uses the conditional expected true value based on available assay information.	Approximate for a Cox hazard; depends on a well-specified calibration model and adequate information.
Latent true-value model	Models assays and latent trajectories, integrating survival contributions over trajectory uncertainty.	More computationally demanding and sensitive to assumptions about trajectories, errors, and observation processes.

These methods also differ in what uncertainty they propagate. A plug-in calibration analysis may understate uncertainty if the calibration model is treated as known; joint estimation or resampling that refits calibration can better reflect that source of variation. A latent model propagates trajectory uncertainty through its likelihood, but this advantage depends on the model being credible. When replicate data or external validation data are absent, measurement-error parameters may be weakly identified, and sensitivity analyses over plausible error variances can be more informative than a single apparently precise correction.

A common mistake is to say, “The reliability is $0.64$, so divide the Cox coefficient by $0.64$ to recover the true association.” That correction is not generally valid for a time-dependent Cox model. Reliability summarizes a particular variance decomposition; it does not encode the risk-set process, time-varying error, or covariate adjustment. Use a justified calibration or latent model, state its assumptions, and interpret its result as model-dependent.

Key takeaway: Repeated assays do not make the true trajectory directly observable. A time-updated survival analysis should distinguish latent values from observed noisy assays, use only information available at each risk time, and match its correction method to a defensible measurement-error model.

Check Your Understanding

Answer each question before reviewing the solutions.

In $W_{ij}=X_i(t_{ij})+U_{ij}$, what does each term represent, and what is meant by classical additive error?
A calibration model estimates $E\{X_i(t)\mid W_i(t+1)\}$ using an assay collected after time $t$. What is the concern?
Why is the reliability ratio not generally a correction factor for a time-dependent Cox coefficient?
Under $X\mid W\sim N(0.5,0.25)$ and $\beta=0.4$, calculate $E\{\exp(\beta X)\mid W\}$ and the plug-in value $\exp\{\beta E(X\mid W)\}$.
Name one diagnostic or sensitivity analysis that can reveal a fragile measurement-error correction.

Solutions. (1) $W_{ij}$ is the observed assay, $X_i(t_{ij})$ is the latent true value at its collection time, and $U_{ij}$ is measurement error. Classical additive error usually assumes mean-zero error independent of the true value, with a specified variance structure. (2) The future assay may not be available at the earlier risk time; using it can create look-ahead information and change the intended time-updated predictor. (3) The reliability ratio describes a variance decomposition, whereas a Cox coefficient also depends on time variation, risk sets, adjustment, and the error structure; no universal multiplication or division rule follows. (4) The integrated value is $\exp\{0.4(0.5)+0.4^2(0.25)/2\}=\exp(0.22)=1.246$. The plug-in value is $\exp\{0.4(0.5)\}=\exp(0.20)=1.221$. (5) Suitable answers include checking replicate residuals by batch and time, validating calibration predictions in held-out data, or repeating the analysis over plausible measurement-error variances.