Conditional Independence Given Shared Random Effects

Why Conditional Independence Matters

A shared-parameter model represents dependence between a person’s repeated outcomes and their dropout process through latent person-specific effects. The key claim is not that the outcomes and dropout are unrelated. Rather, it is that after conditioning on the shared effects and observed covariates, their remaining randomness is independent. This distinction determines how the joint model is built and what association it can represent.

Let $i$ index individuals and $j=1,\ldots,m_i$ index scheduled outcome occasions. Write $Y_i=(Y_{i1},\ldots,Y_{im_i})$ for the outcome vector, $D_i$ for the dropout process, $X_i$ for observed covariates, and $b_i$ for the individual’s shared random effects. The dropout process may be represented by a dropout time or by indicators across occasions; here $D_i$ denotes its full path.

Definition: Conditional independence of the outcome and dropout processes given shared effects and covariates is written $Y_i \perp D_i \mid (b_i,X_i)$. It means that, for values of $b_i$ and $X_i$ with positive probability or density, the conditional joint distribution factors as $f_{Y_i,D_i}(y,d\mid b_i,X_i)=f_{Y_i}(y\mid b_i,X_i)f_{D_i}(d\mid b_i,X_i)$.

The interpretation is specific: once the latent effects and covariates are fixed, knowing the dropout path supplies no further information about the outcome vector, and knowing the outcome vector supplies no further information about dropout. The shared effects can still link the processes. Individuals with different $b_i$ values may have different outcome distributions and different dropout risks.

What the Assumption Does—and Does Not—Say

The conditional-independence statement concerns the entire outcome vector $Y_i$ and the entire dropout process $D_i$. It does not, by itself, assert that the repeated outcomes are mutually independent given $b_i$ and $X_i$. An outcome model may include residual serial dependence while still satisfying $Y_i \perp D_i \mid (b_i,X_i)$.

A more restrictive model may additionally assume conditional local independence: the outcomes $Y_{i1},\ldots,Y_{im_i}$ are mutually independent given $b_i$ and $X_i$. Under both assumptions, the conditional joint distribution can be written as a product over occasions and a separate dropout factor. Without local independence, the outcome-vector distribution remains joint rather than factoring over $j$.

The covariates and shared effects must be stated clearly. If an observed covariate predicts both outcomes and dropout, omitting it can leave residual dependence that the shared effects are then asked to absorb. Likewise, conditioning on $b_i$ is a model statement about a latent quantity: it does not mean that $b_i$ is directly observed or that the assumption can be checked by simply stratifying the data on its true value.

The assumption also differs from a claim that dropout is unrelated to outcomes. Marginally, $Y_i$ and $D_i$ will generally be associated because both depend on $b_i$. The point of the shared effects is precisely to allow that association while positing independence of the remaining process-specific variation.

Worked Example: Conditional Independence but Marginal Association

Suppose a hypothetical population has a binary shared effect $B$, with $P(B=1)=P(B=0)=0.5$. Given $B$, two binary outcomes $Y_1,Y_2$ are mutually independent, each with probability 0.8 of being 1 when $B=1$, and 0.2 when $B=0$. Dropout $D$ is conditionally independent of both outcomes given $B$, with $P(D=1\mid B=1)=0.7$ and $P(D=1\mid B=0)=0.1$. All probabilities are hypothetical.

Within either $B$ stratum, the outcome and dropout probabilities multiply. For example, when $B=1$, $P(Y_1=1,D=1\mid B=1)=0.8(0.7)=0.56$. When $B=0$, the corresponding joint probability is $0.2(0.1)=0.02$. Thus the stated conditional independence holds in both strata.

After averaging over $B$, $P(Y_1=1)=0.5(0.8)+0.5(0.2)=0.5$, and $P(D=1)=0.5(0.7)+0.5(0.1)=0.4$. The joint probability is $P(Y_1=1,D=1)=0.5(0.56)+0.5(0.02)=0.29$. If $Y_1$ and $D$ were marginally independent, that joint probability would be $0.5(0.4)=0.20$, not 0.29. Their marginal covariance, using binary indicators, is $0.29-(0.5)(0.4)=0.09$.

The repeated outcomes are also marginally associated: $P(Y_1=1,Y_2=1)=0.5(0.8^2)+0.5(0.2^2)=0.34$, whereas $P(Y_1=1)P(Y_2=1)=0.25$. In this model, both forms of association arise from mixing over the shared effect; they do not contradict conditional independence.

Deriving the Joint Model

The conditional-independence assumption gives a direct factorization. Let $f_b(b\mid x)$ denote the distribution of the shared effects given covariates, $f_Y(y\mid b,x)$ the conditional outcome-vector distribution, and $f_D(d\mid b,x)$ the conditional dropout-process distribution. The joint distribution of the effects, outcomes, and dropout, conditional on covariates, is

f_{b,Y,D}(b,y,d\mid x) = f_b(b\mid x)f_Y(y\mid b,x)f_D(d\mid b,x).

This follows from the chain rule and the conditional-independence statement. First write the joint density as $f_b(b\mid x)f_{Y,D}(y,d\mid b,x)$; then replace the conditional joint factor by $f_Y(y\mid b,x)f_D(d\mid b,x)$. The observed-data joint distribution with the effects integrated out is therefore

f_{Y,D}(y,d\mid x) = \int f_b(b\mid x)f_Y(y\mid b,x)f_D(d\mid b,x)\,db.

For discrete effects, the integral is replaced by a sum over their possible values. This integration is the mathematical reason conditional independence does not usually imply marginal independence: the same $b$ appears in both conditional distributions before averaging.

If conditional local independence of repeated outcomes is also imposed, then $f_Y(y\mid b,x)=\prod_{j=1}^{m_i}f_{Y_{ij}}(y_{ij}\mid b,x)$. That product is an additional modeling choice, not a consequence of independence between $Y_i$ and $D_i$. A model that permits residual serial dependence instead uses a non-product $f_Y$, while retaining the factorization between the whole outcome vector and dropout.

Derivation: If $Y$ and $D$ are conditionally independent given $B,X$, then their conditional covariance is zero. The law of total covariance gives $$ \operatorname{Cov}(Y,D\mid X) = E\{\operatorname{Cov}(Y,D\mid B,X)\mid X\} + \operatorname{Cov}\{E(Y\mid B,X),E(D\mid B,X)\mid X\}. $$ The first term is zero, so marginal covariance is the covariance across shared-effect values of the two conditional means. It can be nonzero when those means vary together.

Worked Example: Covariance Induced by a Shared Effect

Consider a hypothetical model with $B\in\{-1,+1\}$, each value having probability 0.5. Let $Y_1=5+B+\epsilon_1$, $Y_2=7+2B+\epsilon_2$, where the residuals are mutually independent, have mean zero, and have variances 1 and 4. Let $D$ be a binary dropout indicator, independent of the residuals and outcomes conditional on $B$, with $P(D=1\mid B=+1)=0.8$ and $P(D=1\mid B=-1)=0.2$.

Given $B$, the outcome residuals are independent and the dropout draw is independent of them. Thus $\operatorname{Cov}(Y_1,Y_2\mid B)=0$, and each outcome is conditionally independent of $D$. Marginally, $E(B)=0$ and $\operatorname{Var}(B)=1$, so $E(Y_1)=5$ and $E(Y_2)=7$.

The covariance between the two outcomes is $\operatorname{Cov}(Y_1,Y_2)=\operatorname{Cov}(B,2B)=2\operatorname{Var}(B)=2$. Their variances are $1+1=2$ and $4+4=8$, respectively. Hence their correlation is $2/\sqrt{2(8)}=0.5$. The association is present even though their conditional covariance is zero.

For dropout, $E(D)=0.5(0.8)+0.5(0.2)=0.5$, and $E(BD)=0.5(1)(0.8)+0.5(-1)(0.2)=0.3$. Therefore $\operatorname{Cov}(Y_1,D)=\operatorname{Cov}(B+\epsilon_1,D)=0.3$, since the residual is independent of dropout and has mean zero. The outcome and dropout have positive marginal covariance, induced by their shared dependence on $B$.

Checking the Assumption and Comparing Alternatives

Conditional independence is a substantive restriction, not an automatic property of a model with random effects. A useful conceptual diagnostic is to ask whether the proposed $b_i$ and $X_i$ plausibly capture the common causes of outcomes and dropout. If dropout remains directly related to recent outcomes after accounting for those quantities, the factorization is not credible as stated. Because dropout can hide later outcomes, this concern cannot generally be settled using observed data alone; observed patterns can flag tension, but they do not prove the assumption.

It is also important not to confuse two different restrictions: conditional independence between the outcome process and dropout, and conditional independence among repeated outcomes. Remaining within-person outcome correlation after accounting for the proposed effects challenges a local-independence outcome model, but it does not by itself disprove independence between the outcome vector and dropout.

Worked Example: A Direct Outcome–Dropout Link Violates the Factorization

Suppose, hypothetically, that within a fixed covariate stratum and at a fixed shared-effect value, $Y$ is binary with $P(Y=1)=0.5$. Dropout is more likely after $Y=1$: $P(D=1\mid Y=1)=0.8$, compared with $P(D=1\mid Y=0)=0.2$. These probabilities describe a direct dependence remaining after conditioning on the proposed effects and covariates.

The marginal dropout probability within this stratum is $P(D=1)=0.5(0.8)+0.5(0.2)=0.5$. But $P(Y=1,D=1)=0.5(0.8)=0.4$, whereas conditional independence at this fixed stratum would require the product $P(Y=1)P(D=1)=0.5(0.5)=0.25$. Since $0.4\ne0.25$, the assumed conditional factorization fails.

One response would be to specify a dropout model that depends on the relevant outcome history, if that is scientifically justified and the necessary information is available. Another would be to reconsider the shared effects or covariates. Simply asserting that a random effect is present does not remove the observed conditional dependence.

An independent-process model is a stronger alternative: it omits the shared link and asserts no remaining association between outcome and dropout, even after marginalization under that model. A direct-dependence model instead lets dropout depend on outcome history or other observed outcome features beyond the shared effects. The shared-effects formulation lies between these choices: it permits marginal association through latent common causes, while ruling out residual dependence conditional on those causes.

The factorization is useful because it gives a transparent description of how the two processes are connected. Its limitation is equally important: the observed data may not determine whether an association is due to shared latent effects, direct outcome dependence, or a different latent-effect distribution. Thus the factorization should be interpreted as a model assumption, not as a conclusion established merely by fitting the model. The earlier material on shared random effects as links between outcome and dropout motivates the latent connection; this tutorial specifies the conditional-independence restriction that makes the joint distribution factor.

Common Pitfalls

Claiming marginal independence. Conditional independence given $b_i,X_i$ does not imply independence after integrating out $b_i$. Shared variation can induce marginal association.
Treating outcome local independence as automatic. Independence of the outcome vector and dropout does not require the repeated outcomes to factor over occasions.
Using a shared effect as a cure-all. If dropout depends on outcome history even after conditioning on the proposed effects and covariates, the stated factorization is false.
Calling the assumption empirically verified. Since dropout can conceal outcomes, observed-data diagnostics can reveal problems but cannot establish conditional independence for unobserved values.

Key takeaway: In a shared-parameter model, the outcome vector and dropout process are independent conditional on the shared effects and covariates, not generally independent marginally. The joint model follows by multiplying their conditional distributions with the shared-effect distribution and integrating over the latent effects.

Check Your Understanding

Answer each question, then compare with the brief solutions.

If $Y_i\perp D_i\mid(b_i,X_i)$, must $Y_i$ and $D_i$ be marginally independent? Explain.
Does the factorization between the outcome vector and dropout require outcomes to be mutually independent given $b_i,X_i$?
Suppose $E(Y\mid B)=2+B$ and $E(D\mid B)=0.4+0.1B$, with $\operatorname{Var}(B)=2$, and the conditional covariance is zero. What is the marginal covariance if these conditional means are the only $B$-dependent parts?
At fixed covariates and shared-effect value, $P(Y=1)=0.3$, $P(D=1)=0.5$, and $P(Y=1,D=1)=0.15$. Is conditional independence contradicted by these probabilities?

Solutions. (1) No. Integrating over shared effects can induce marginal association. (2) No. The outcome vector may have residual serial dependence; local independence is an extra assumption. (3) The covariance is $\operatorname{Cov}(2+B,0.4+0.1B)=0.1\operatorname{Var}(B)=0.2$. (4) No contradiction: $0.3(0.5)=0.15$, exactly the product required by conditional independence at that stratum.