Conditions for Ignorable Dropout in Shared-Parameter Models

Why Ignorability Needs a Careful Check

A joint model can represent repeated outcomes and dropout through a shared latent effect. That structure is useful when people with different latent outcome trajectories also have different dropout risks. But it has an important consequence: even if the outcome and dropout processes are conditionally independent given the latent effect, they can be associated after that effect is integrated out. Therefore, conditional independence alone does not establish that the dropout process can be ignored when estimating outcome parameters.

The question here is narrower and more precise than whether a joint model is scientifically plausible: under what conditions does the observed-data likelihood for the outcome parameters separate from the dropout mechanism? The answer depends on what the dropout process conditions on, whether it uses outcome-related latent effects, and whether outcome and dropout parameters are distinct.

Observed-Data Likelihood and Conditional Independence

For subject $i$, let $Y_i$ denote the full outcome vector, $Y_i^{\mathrm{obs}}$ its observed components, and $Y_i^{\mathrm{mis}}$ its missing components. Let $R_i$ denote the observation or dropout pattern, $X_i$ the observed covariates, and $b_i$ a latent effect. Write $\theta$ for outcome-model parameters, $\psi$ for dropout-model parameters, and $\eta$ for parameters of the latent-effect distribution.

The general chain rule, with the outcome process ordered before the dropout process, gives the conditional joint density $f_Y(y_i\mid b_i,X_i;\theta)f_D(r_i\mid y_i,b_i,X_i;\psi)$. This is a chain-rule factorization; it does not assert conditional independence. Conditional independence of the outcome and dropout processes given $b_i$ and $X_i$ is the additional assumption that the second factor does not depend on $y_i$.

Definition: Under conditional independence given the shared effect, the joint model has the form $$ f_Y(y_i\mid b_i,X_i;\theta)\, f_D(r_i\mid b_i,X_i;\psi)\, f_b(b_i\mid X_i;\eta). $$ The observed-data likelihood contribution is obtained by integrating over both the missing outcomes and the latent effect. Conditional independence given $b_i$ does not imply independence after integrating over $b_i$.

Under this conditional-independence model, integrating out missing outcomes first gives

L_i(\theta,\psi,\eta) = \int f_Y(y_i^{\mathrm{obs}}\mid b_i,X_i;\theta)\, f_D(r_i\mid b_i,X_i;\psi)\, f_b(b_i\mid X_i;\eta)\,db_i.

The outcome density in this expression is the marginal density of the observed outcomes conditional on $b_i$, after integrating over $Y_i^{\mathrm{mis}}$. In general, the integral is not the product of a marginal outcome likelihood and a marginal dropout likelihood: both factors are averaged over the same $b_i$. If $b_i$ affects both processes, the dropout pattern can change the distribution of latent effects among subjects with observed outcomes.

When the Dropout Mechanism Is Ignorable

For likelihood inference about $\theta$, a standard sufficient condition for ignorability is that the dropout mechanism is missing at random with respect to the full outcome data, and that its parameters are distinct from the outcome parameters. In this setting, missing at random means that the probability of the observed pattern can depend on $X_i$ and $Y_i^{\mathrm{obs}}$, but not on the unobserved outcomes $Y_i^{\mathrm{mis}}$ after conditioning on the observed information. Parameter distinctness means that the parameter spaces for the outcome and dropout models factor, with no shared unknown parameter linking their distributions.

To see the likelihood consequence, suppose the pattern probability can be written as $g(r_i\mid y_i^{\mathrm{obs}},X_i;\psi)$, with no dependence on $Y_i^{\mathrm{mis}}$. Then

\begin{aligned} L_i(\theta,\psi) &= \int f_Y(y_i^{\mathrm{obs}},y_i^{\mathrm{mis}}\mid X_i;\theta) \,g(r_i\mid y_i^{\mathrm{obs}},X_i;\psi)\,dy_i^{\mathrm{mis}}\\ &= g(r_i\mid y_i^{\mathrm{obs}},X_i;\psi)\, f_Y(y_i^{\mathrm{obs}}\mid X_i;\theta). \end{aligned}

The second equality follows because the dropout factor is constant with respect to the integration over missing outcomes. With distinct parameters, this factor does not change likelihood-based inference about $\theta$; it can be omitted for that purpose. The observed outcomes may still be associated with the dropout pattern because the mechanism depends on observed history. Ignorability is a claim about the likelihood factorization, not a claim that outcome and dropout indicators are marginally unrelated.

In a shared-parameter model, this argument applies if the dropout component does not depend on an outcome-related latent effect. For example, dropout that depends only on observed covariates or observed outcome history can satisfy the required factorization. A different special case is a dropout-only latent effect that is independent of the outcome-related latent effects, with separate distributional parameters: integrating over the independent effects can then preserve separate factors. The independence and parameter-separation conditions matter; merely giving the latent effects different names does not establish them.

Worked Example: Dropout Independent of the Outcome-Related Effect

Suppose a hypothetical study has a single outcome $Y_i\mid b_i\sim N(\mu+b_i,1)$, where $b_i\sim N(0,1)$. Let $R_i=1$ mean the outcome is observed. Assume $P(R_i=1\mid X_i)=0.75$, with this probability independent of $b_i$ and $Y_i$. The outcome parameter is $\mu$, and the observation probability is a separate dropout parameter.

Because the observation probability does not depend on $b_i$, the observed-data contribution for an observed value $y_i$ is

\begin{aligned} L_i &=\int f_Y(y_i\mid b_i;\mu)\,(0.75)\,f_b(b_i)\,db_i\\ &=0.75\int f_Y(y_i\mid b_i;\mu)f_b(b_i)\,db_i\\ &=0.75\,f_Y(y_i;\mu). \end{aligned}

For a hypothetical observed value $y_i=2$, the marginal outcome distribution is $Y_i\sim N(\mu,2)$, since the residual variance and random-intercept variance are both 1. Thus the contribution is $0.75$ times the $N(\mu,2)$ density evaluated at 2. The factor $0.75$ records the observation pattern but does not involve $\mu$. The assumptions justify omitting it for likelihood inference about $\mu$; they would not justify doing so if observation probability depended on $b_i$.

Why a Shared Effect Usually Prevents Ignorability

When dropout depends on an outcome-related latent effect, the observed-data likelihood generally retains the dropout component inside the integral. The same latent value that changes the outcome distribution also changes the probability of being observed. Consequently, the latent-effect distribution among subjects with a particular observed pattern can differ from its population distribution.

This is the distinction between conditional and marginal association. Conditional independence given $b_i$ says that knowing the outcome gives no additional information about dropout once $b_i$ is known. It does not say that dropout is independent of the outcome when $b_i$ is unobserved. The integration over $b_i$ can induce marginal association.

Worked Example: Conditional Independence but Informative Observation

Consider a hypothetical binary latent effect $B$, with $P(B=1)=P(B=-1)=0.5$. Given $B=1$, let $P(Y=1\mid B)=0.8$; given $B=-1$, let $P(Y=1\mid B)=0.2$. Let $R=1$ mean the outcome is observed, and suppose $P(R=1\mid B=1)=0.8$ and $P(R=1\mid B=-1)=0.2$. Assume $Y$ and $R$ are conditionally independent given $B$.

The marginal probability of a positive outcome is $P(Y=1)=0.5(0.8)+0.5(0.2)=0.50$. The marginal probability of observation is also $P(R=1)=0.5(0.8)+0.5(0.2)=0.50$. If the variables were marginally independent, their joint probability would be $0.50(0.50)=0.25$. Instead, conditional independence gives

\begin{aligned} P(Y=1,R=1) &=P(B=1)P(Y=1\mid B=1)P(R=1\mid B=1)\\ &\quad+P(B=-1)P(Y=1\mid B=-1)P(R=1\mid B=-1)\\ &=0.5(0.8)(0.8)+0.5(0.2)(0.2)\\ &=0.32+0.02=0.34. \end{aligned}

Thus $P(Y=1\mid R=1)=0.34/0.50=0.68$, not 0.50. The observed outcomes are enriched for positive values because the latent effect raises both the chance of a positive outcome and the chance of observation. Conditional independence given $B$ holds exactly, but omitting the dropout process would miss this selection.

A Useful Ignorable Case: Dropout Based on Observed History

Dropout need not be independent of all outcomes to be ignorable. It may depend on outcomes already observed, provided it does not depend on the still-unobserved outcomes after conditioning on the observed data, and provided the mechanism parameters are distinct. This is the likelihood form of the missing-at-random condition. In a shared-parameter specification, the key check is whether the dropout model also uses a latent effect that carries information about unobserved outcomes.

Worked Example: Dropout Depending on a Recorded Outcome

Suppose a hypothetical two-visit study records $Y_1$ for everyone who reaches the first visit. Let $Y_1\sim\mathrm{Bernoulli}(p)$. After observing $Y_1$, a subject may drop out before $Y_2$. Let the dropout probability be 0.30 when $Y_1=1$ and 0.10 when $Y_1=0$. Assume the dropout mechanism depends on $Y_1$ only, not on the missing $Y_2$, and has parameters distinct from those of the outcome model.

For a hypothetical subject with $Y_1=1$ who drops out before the second visit, take $p=0.60$. The joint contribution for the observed record is $P(Y_1=1)\,P(\text{dropout}\mid Y_1=1)=0.60(0.30)=0.18$. The factor involving dropout is 0.30, while the outcome factor for $p$ is 0.60. Since dropout depends only on the recorded $Y_1$, it does not require a model for the missing $Y_2$; the dropout factor can be omitted for likelihood inference about the outcome parameters under the stated distinctness condition. It must not be omitted if the goal is to estimate or describe the dropout mechanism itself.

Checking the Conditions and Comparing Approaches

Before calling dropout ignorable in a shared-parameter analysis, check the likelihood rather than relying on the label “conditional independence.” A practical sequence is:

Write the pattern contribution.
Identify which outcomes are observed and which are missing, and state what the dropout probability conditions on.

Inspect latent-effect dependence.
Ask whether dropout depends on any latent effect that also predicts the outcomes, or on a latent-effect distribution parameter that is part of the outcome model.

Check the likelihood factorization.
After integrating missing outcomes and latent effects, verify whether the likelihood is an outcome factor in $\theta$ times a dropout factor in $\psi$.

State the scope of omission.
If factorization and parameter distinctness hold, omission is justified for likelihood inference about outcome parameters, not for inference about dropout parameters or the joint distribution.

The main alternatives answer different questions. An outcome-only analysis that omits dropout is justified when the likelihood is ignorable under the conditions above. A joint shared-parameter likelihood retains the dropout factor when latent effects link the processes; this represents the selection mechanism but requires assumptions about the latent-effect distribution and dropout model. A sensitivity analysis can examine how conclusions change across plausible dependence assumptions. No approach makes an unidentified dependence structure empirically determined merely by fitting a more elaborate model.

Diagnostics can reveal problems without proving ignorability. Compare observed histories and dropout frequencies, check whether the assumed dropout model is compatible with recorded data, and examine whether fitted shared-effect associations are weakly identified or highly sensitive to model choices. A weak estimated association is not proof that dropout is ignorable: the relevant dependence may be difficult to estimate from observed data. Likewise, good fit to observed outcomes alone cannot validate assumptions about unobserved outcomes after dropout.

A common failure is to use a factorization that is valid only after conditioning on a latent effect, then treat its factors as marginally separate. The correction is to keep the integral explicit:

\int f_Y(y_i^{\mathrm{obs}}\mid b_i;\theta) f_D(r_i\mid b_i;\psi)f_b(b_i;\eta)\,db_i \ne \left\{\int f_Y(y_i^{\mathrm{obs}}\mid b_i;\theta)f_b(b_i;\eta)\,db_i\right\} \left\{\int f_D(r_i\mid b_i;\psi)f_b(b_i;\eta)\,db_i\right\}

in general. The equality requires additional conditions, such as the dropout factor being independent of the outcome-related latent effect, or an appropriate independence structure that separates the latent effects and their parameters. A second failure is to call a mechanism ignorable because it depends only on $b_i$, not directly on $Y_i$. If $b_i$ predicts the outcomes, dropout based on $b_i$ can still select subjects according to their outcome distribution.

Key takeaway: Likelihood ignorability requires more than conditional independence given shared effects. It requires a factorization of the observed-data likelihood for outcome and dropout parameters. Dropout that depends on outcome-related latent effects generally prevents that factorization; dropout based only on observed information can be ignorable when the parameter spaces are distinct.

Check Your Understanding

Answer each question, then compare with the solutions below.

Why does conditional independence of $Y_i$ and $R_i$ given $b_i$ not imply marginal independence?
A dropout probability depends on an observed covariate $X_i$, but not on $b_i$ or unobserved outcomes. What other condition is needed to omit its likelihood factor for inference about outcome parameters?
In the binary latent-effect example, calculate $P(Y=1\mid R=1)$ and explain its relevance to ignorability.
Suppose dropout depends on a latent effect that is independent of the outcome-related latent effect. Name two requirements for this to preserve likelihood factorization.

Solutions. (1) Conditioning on $b_i$ removes the association only when $b_i$ is known. Integrating over a latent effect that predicts both processes can induce marginal association, as shown by the joint probability 0.34 rather than the independence value 0.25. (2) The dropout and outcome parameter spaces must be distinct, with no shared unknown parameter linking the factors. (3) $P(Y=1\mid R=1)=P(Y=1,R=1)/P(R=1)=0.34/0.50=0.68$. The observed outcomes therefore have a different distribution from the full outcome population, so the dropout process is informative. (4) The latent effects must be independent in their joint distribution, and their distributional parameters must be separated so that integrating over them does not reintroduce a shared unknown parameter.