Conditional Independence Assumptions

Why Conditional Independence Matters

A biomarker measured over time and a time-to-event outcome can be associated for several reasons. A subject’s underlying health may shape both the biomarker trajectory and the event risk. A joint model represents this shared subject-specific information through covariates, random effects, and a latent trajectory. Its conditional independence assumptions specify what remains unrelated once that information is accounted for.

This is a statement about the model, not a claim that marker values and event times are unrelated in the observed population. In fact, their marginal association is often one reason to use a joint model. The assumption is that, after conditioning on the specified latent subject information, the marker measurements do not provide additional information about the event outcome, and vice versa.

Definition: For subject $i$, let $Y_i=(Y_{i1},\ldots,Y_{in_i})$ denote the repeated marker measurements, $X_i$ the covariates, $b_i$ the shared random effects, and $m_i(\cdot)$ the subject’s latent trajectory. Let $T_i$ denote the event time. The marker–event conditional independence assumption is $$ Y_i \perp\!\!\!\perp T_i \mid X_i,b_i,m_i(\cdot). $$ It says that, for fixed covariates and latent subject information, the conditional distribution of the marker data does not depend further on the event time, and the conditional distribution of the event time does not depend further on the marker data.

The latent trajectory $m_i(\cdot)$ describes the underlying marker course, rather than the noisy values actually recorded at visits. In many shared-random-effects models it is determined by the covariates and random effects, so including it in the conditioning information can be mathematically redundant. It remains useful to name it because it clarifies the distinction between the underlying course and the observed measurements. This assumption does not specify a particular way for the event model to depend on that course.

Two Assumptions That Should Not Be Confused

The marker–event assumption concerns the relationship between the marker data as a whole and the event outcome, conditional on the shared latent information. A separate assumption is often used for the repeated measurements within the marker data: measurement errors at different visits are conditionally independent given the covariates and random effects. If that additional assumption is made, the marker contribution can be written as a product across visits.

f(Y_i,T_i \mid X_i,b_i,m_i) = f(Y_i \mid X_i,b_i,m_i)\, f(T_i \mid X_i,b_i,m_i), \qquad f(Y_i \mid X_i,b_i,m_i) = \prod_{j=1}^{n_i} f(Y_{ij} \mid X_i,b_i,m_i).

The first equality expresses marker–event conditional independence. The second requires the additional within-marker conditional independence assumption. The first does not imply the second: the marker observations may still have correlated errors across visits while their joint distribution remains conditionally independent of the event time. In that case, use a joint conditional distribution for $Y_i$ rather than a product over visits.

Conversely, conditional independence of marker measurements across visits does not by itself establish their conditional independence from the event outcome. Each assumption concerns a different relationship and needs its own scientific and modeling justification.

Important distinction: Conditional independence given $b_i$ and $m_i(\cdot)$ does not say that $T_i$ is independent of the latent trajectory or shared random effects. The event submodel may depend on those quantities. It says that the observed marker measurements have no additional association with the event outcome once the shared latent information is fixed.

Worked Examples

Worked Example: Multiplying Conditional Contributions

Suppose that, for a hypothetical subject with fixed covariates and latent subject information, the probability of a positive marker result at visit 1 is $0.80$, the probability of a positive result at visit 2 is $0.60$, and the probability of an event by a specified time is $0.30$. Assume both conditional independence assumptions: marker results are independent across these visits, and the marker data are independent of the event outcome, conditional on the given latent information. Find the probability that both results are positive and the event occurs by that time.

Under the stated assumptions, the conditional probability is the product:

P(Y_{i1}=1,Y_{i2}=1,\text{ event by the specified time}\mid X_i,b_i,m_i) =0.80 \times 0.60 \times 0.30 =0.144.

Thus the conditional probability is $0.144$, or $14.4\%$. This calculation is justified only because the example explicitly assumes both independence across marker visits and independence between the marker data and event outcome, given the latent information. If only marker–event independence were assumed, the two marker probabilities could not necessarily be multiplied.

Worked Example: Conditional Independence Can Coexist With Marginal Association

Consider a hypothetical binary shared random effect $B$, with $P(B=1)=P(B=-1)=0.5$. Let a marker measurement be $Y=B+\varepsilon$, where $E(\varepsilon\mid B)=0$. Let $D=1$ indicate an event by a fixed horizon, with $P(D=1\mid B=1)=0.80$ and $P(D=1\mid B=-1)=0.20$. Suppose the marker error and event outcome are independent given $B$. Calculate the covariance between $Y$ and $D$ without conditioning on $B$.

First, $E(Y)=E(B)+E(\varepsilon)=0$, and

$$ E(D)=0.5(0.80)+0.5(0.20)=0.50. $$

Since the marker error has conditional mean zero and is conditionally independent of $D$, it contributes zero to $E(YD)$. Therefore,

$$ E(YD)=E(BD) =0.5(1)(0.80)+0.5(-1)(0.20) =0.30. $$

It follows that

\operatorname{Cov}(Y,D) =E(YD)-E(Y)E(D) =0.30-(0)(0.50) =0.30.

The marker and event indicator are conditionally independent given $B$, yet their marginal covariance is $0.30$. Subjects with different values of the shared effect have different expected marker values and event probabilities. The association arises through that shared subject-level information; it does not contradict conditional independence.

Worked Example: An Unmodeled Shared Cause Can Break the Assumption

Suppose the shared random effects and covariates are fixed, but an unmeasured acute condition $U$ remains. Let $P(U=1)=P(U=0)=0.5$. Given $U$, assume the marker indicator $Y$ and event indicator $D$ are independent, with $P(Y=1\mid U=1)=0.80$, $P(Y=1\mid U=0)=0.20$, $P(D=1\mid U=1)=0.70$, and $P(D=1\mid U=0)=0.10$. Check whether $Y$ and $D$ are independent conditional only on the stated shared effects and covariates.

Because $U$ is not included among the conditioning variables, average over its two values. The marginal probabilities are

P(Y=1)=0.5(0.80)+0.5(0.20)=0.50, \qquad P(D=1)=0.5(0.70)+0.5(0.10)=0.40.

The joint probability is

$$ P(Y=1,D=1) =0.5(0.80)(0.70)+0.5(0.20)(0.10) =0.28+0.01 =0.29. $$

If $Y$ and $D$ were independent given the modeled information, their joint probability would be $0.50 \times 0.40=0.20$. Since $0.29 \ne 0.20$, the assumed conditional independence fails. Conditional independence given $U$ does not imply conditional independence after $U$ is omitted. A measured and scientifically justified $U$ could be incorporated into the conditioning information; otherwise, the model may need a different representation of residual dependence.

Why Marginal Dependence Remains

A shared random effect is unobserved in the population-level distribution. Conditional independence holds at a fixed value of that effect, but population-level probabilities average over its distribution. As a result, the marker and event outcomes can be associated after the random effects are integrated out.

The covariance identity makes this precise. Write $W_i=(X_i,b_i,m_i(\cdot))$. For variables with finite second moments,

\operatorname{Cov}(Y_i,D_i\mid X_i) = E\!\left[\operatorname{Cov}(Y_i,D_i\mid W_i)\mid X_i\right] + \operatorname{Cov}\!\left(E[Y_i\mid W_i],E[D_i\mid W_i]\mid X_i\right).

If $Y_i$ and $D_i$ are conditionally independent given $W_i$, the first term is zero. The second term need not be zero: the conditional mean marker level and conditional event probability can both vary with $b_i$ or the latent trajectory. This is how a model can assert conditional independence while representing marginal association.

The identity follows by writing each variable as its conditional mean plus its deviation from that mean. The product of the two deviations has conditional expectation equal to the conditional covariance; the product of the conditional means contributes their covariance across values of $W_i$. Taking the conditional expectation given $X_i$ gives the two terms above.

Checking the Assumptions and Comparing Alternatives

Conditional independence is a modeling assumption, not a property that can usually be verified directly from observed data. The random effects and latent trajectory are not observed, so a simple observed correlation between marker and event outcomes is not evidence against the assumption. That association may be exactly what shared latent subject information is intended to represent.

Useful checks focus on whether the model leaves systematic dependence unexplained. For example, one can examine whether marker residuals show time-patterned association after accounting for the fitted longitudinal structure, whether event predictions are poorly calibrated across marker-history patterns, and whether conclusions change under plausible alternative dependence structures. Such checks are informative but not decisive: latent effects are estimated, and residual or prediction checks cannot prove conditional independence.

A common alternative is to allow correlated marker measurement errors across visits while retaining marker–event conditional independence. This changes the conditional marker distribution, not the central marker–event assumption. Another alternative is to introduce a further shared latent factor or a joint residual structure when science suggests a source of dependence not captured by the existing random effects. A model with separate random effects for marker and event processes may be suitable when no shared latent subject factor is justified, but it does not represent the same shared-source assumption. These approaches answer different scientific questions and should not be treated as interchangeable computational options.

When event times are censored, assumptions about the censoring process also matter. Conditional independence between marker data and the event time does not itself guarantee that censoring is independent or noninformative. The event and censoring assumptions must be considered separately, using the relevant covariates and latent information where justified.

Name the conditioning information.
State which covariates, shared random effects, and features of the latent trajectory are being held fixed. Do not describe the assumption merely as “the two processes are independent.”

Separate the independence claims.
Specify marker–event conditional independence separately from any assumption of independence among repeated marker measurements.

Assess what could remain unmodeled.
Consider whether a time-varying condition, shared measurement error, or other subject-level feature could still connect marker observations and event outcomes after conditioning.

Check sensitivity and model fit.
Look for systematic residual patterns and compare conclusions under scientifically plausible alternatives. Treat these checks as evidence about model adequacy, not proof of the assumption.

Common Pitfall and Limitations

Wrong reasoning: “The biomarker and event time are associated, so they cannot be conditionally independent.” This confuses marginal association with conditional independence. If subjects with different shared random effects have different biomarker levels and event risks, averaging over those effects can produce a strong marginal association.

Correction: Ask whether the association remains after conditioning on the specified shared random effects, covariates, and latent trajectory. The assumption concerns that conditional relationship, not the raw relationship. Also avoid the opposite error: a shared random effect does not automatically capture every source of dependence. An unmodeled common cause or correlated measurement errors may leave conditional dependence in the model.

The assumption simplifies the representation of the joint data and gives the shared latent information a clear role, but it can be restrictive. It may be implausible when marker errors and event occurrence share an additional cause, or when the marker process has residual serial dependence that is ignored. Conclusions can then depend on the chosen latent structure. Scientific justification, model checking, and sensitivity analysis are therefore important companions to the factorization.

Check Your Understanding

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

If marker measurements at separate visits are conditionally independent given the random effects, does this alone imply conditional independence between the marker data and event time?
Suppose marker data and event status are conditionally independent given a shared random effect. Can they have nonzero marginal covariance? Explain.
For one subject, the conditional probability of a marker pattern is $0.25$, and the conditional probability of an event pattern is $0.40$. Under marker–event conditional independence, what is the joint conditional probability of both patterns?
Give one reason the conditional independence assumption could fail even after conditioning on the modeled random effects.
Does a marginal marker–event association by itself show that conditional independence is false?

Solutions. (1) No. Independence across marker visits and independence between marker data and the event are separate assumptions. (2) Yes. Different values of the shared effect can change both conditional means or event probabilities, producing marginal association after averaging over that effect. (3) $0.25 \times 0.40=0.10$, or $10\%$. (4) For example, an unmeasured acute condition could affect both the marker and event outcome; correlated measurement errors could also matter if an independence-across-visits assumption is being used. (5) No. The association may arise through the shared latent information, even when the conditional independence assumption holds.