Tutorials › Pattern-Mixture and Selection Models › Ignorability Conditions For Likelihood-Based Outcome Analysis

Home  /  Tutorials  /  Mechanisms and Ignorability in Longitudinal Studies

Mechanisms and Ignorability in Longitudinal Studies · Tutorial 16 of 499

Ignorability Conditions For Likelihood-Based Outcome Analysis

Understand when an outcome likelihood can be used without modeling dropout, and why MAR alone is not enough.

Intermediate 9 min read

What You'll Learn

  • Define MAR for a response pattern using the outcomes observed and missing under that pattern.
  • Explain parameter distinctness and why it is separate from MAR.
  • Factor the observed-data likelihood under the ignorability conditions.
  • Work through examples where ignorability holds, MAR fails, or distinctness fails.
  • Recognize limitations: ignorability does not guarantee identification or a correct outcome model.

Why Ignorability Is a Likelihood Question

When some longitudinal outcomes are missing, an analysis may model both the outcomes and the process that determines which outcomes are observed. Under specific conditions, likelihood-based inference on outcome parameters can instead use the observed-data outcome likelihood without including a model for dropout. This is called ignoring the missingness mechanism for that inference. It is a statement about a likelihood and its parameters, not a claim that dropout is unimportant or unrelated to outcomes.

The conditions matter. Missing at random (MAR) is one requirement, but MAR alone does not suffice: the parameters governing the outcomes and the missingness process must also be distinct. Even when both conditions hold, the outcome model must be suitable for the scientific question, and the observed data must contain enough information to identify its parameters.

Definition: Let \(Y\) denote the full outcome vector for a subject, \(R\) the response pattern, and \(X\) observed covariates included in the analysis. Write \(Y_{\mathrm{obs}}\) and \(Y_{\mathrm{mis}}\) for the components of \(Y\) observed and missing under the realized pattern \(R\). The dropout or response mechanism is MAR if, for every supported pattern, \(g_{\psi}(R\mid Y_{\mathrm{obs}},Y_{\mathrm{mis}},X)=g_{\psi}(R\mid Y_{\mathrm{obs}},X)\). Thus, conditional on the observed outcomes and covariates, the probability of the realized pattern does not depend on the unobserved outcomes.

The conditioning set is pattern-specific: outcomes observed for one response pattern may be missing for another. MAR is therefore a claim about how the response process depends on the full data relative to the observed data, not a claim that dropout has no predictors. For example, dropout may depend strongly on an observed baseline severity score and still be MAR if it does not additionally depend on unobserved outcomes after conditioning on the score and the observed outcome history.

Let \(f_{\theta}(Y\mid X)\) be the outcome model, indexed by outcome parameter \(\theta\), and let \(g_{\psi}(R\mid Y,X)\) be the response model, indexed by mechanism parameter \(\psi\). Parameter distinctness means that \(\theta\) and \(\psi\) can vary independently: the joint parameter space is a Cartesian product \(\Theta\times\Psi\), with no shared parameter or constraint linking them. Distinctness is a property of the model specification, not a statement that the estimates will be statistically independent.

Observed-Data Likelihood and Ignorability

For a subject with observed pattern \(R\), the contribution to the observed-data likelihood is obtained by summing or integrating over the missing outcomes. For continuous missing outcomes, it has the form below; for discrete outcomes, the integral is replaced by a sum.

$$ L(\theta,\psi;Y_{\mathrm{obs}},R\mid X) = \int f_{\theta}(Y_{\mathrm{obs}},Y_{\mathrm{mis}}\mid X) g_{\psi}(R\mid Y_{\mathrm{obs}},Y_{\mathrm{mis}},X) \,dY_{\mathrm{mis}}. $$

Under MAR, the mechanism factor in the integrand does not vary over the missing outcomes being integrated out. It can therefore be taken outside the integral. The result is

$$ L(\theta,\psi;Y_{\mathrm{obs}},R\mid X) = g_{\psi}(R\mid Y_{\mathrm{obs}},X) \int f_{\theta}(Y_{\mathrm{obs}},Y_{\mathrm{mis}}\mid X)\,dY_{\mathrm{mis}} = g_{\psi}(R\mid Y_{\mathrm{obs}},X) f_{\theta,\mathrm{obs}}(Y_{\mathrm{obs}}\mid X). $$

Here \(f_{\theta,\mathrm{obs}}\) is the outcome model marginalized over the missing components. For independent subjects, the full likelihood is the product of these contributions. If the parameters are distinct, the mechanism factor can be treated as a nuisance factor when making likelihood-based inferences about \(\theta\): it has no \(\theta\) dependence, and its parameter can vary freely while \(\theta\) is varied. The outcome likelihood to use is the likelihood of the observed outcomes, with the missing outcomes integrated out—not a likelihood that pretends the missing outcomes were observed.

This factorization gives the practical justification for likelihood ignorability. It does not establish that an outcome model is correctly specified or that \(\theta\) is identifiable. Nor does it say that every estimator using only observed records is valid. An outcome likelihood must represent the relevant outcome distribution, including covariates needed to describe it.

Worked Example: MAR Dropout Depending on Baseline Group

Suppose a hypothetical study has 20 independent participants in two baseline groups, \(X=0\) and \(X=1\), with 10 participants per group. At the outcome visit, 6 outcomes are observed in group \(X=0\), of which 4 are successes; in group \(X=1\), 5 outcomes are observed, of which 3 are successes. The remaining outcomes are missing. Assume a binary outcome with \(P(Y=1\mid X=x)=\theta_x\), and a response probability \(\rho_x=P(R=1\mid X=x)\) that depends on baseline group but not on the outcome.

The response model is MAR because \(X\) is observed and the response probability does not depend on \(Y_{\mathrm{mis}}\). It is distinct from the outcome model if \(\rho_0,\rho_1\) vary freely, separately from \(\theta_0,\theta_1\). The observed-data likelihood is

$$ L= \theta_0^4(1-\theta_0)^2 \theta_1^3(1-\theta_1)^2 \rho_0^6(1-\rho_0)^4 \rho_1^5(1-\rho_1)^5. $$

The first two factors give the outcome likelihood; the last two give the response likelihood. Maximizing the outcome part gives \( \hat{\theta}_0=4/(4+2)=4/6=0.6667 \) and \( \hat{\theta}_1=3/(3+2)=3/5=0.6000 \). The response estimates are \( \hat{\rho}_0=6/10=0.6000 \) and \( \hat{\rho}_1=5/10=0.5000 \), but they are not needed to maximize the outcome likelihood under these assumptions. These estimates describe the success proportions among observed outcomes within each group; the example does not imply that a complete-case analysis is generally valid under every outcome model.

Conditions That Can Break Ignorability

A failure of either requirement changes the likelihood argument. If MAR fails, the response probability can vary with the unobserved outcomes inside the integral, so the observed-data likelihood may depend jointly on outcome and mechanism parameters. If distinctness fails, even a MAR response factor can involve the outcome parameter through a shared parameter or constraint. In either case, omitting the mechanism may change likelihood-based inference on \(\theta\).

Worked Example: Outcome-Dependent Response Is Not MAR

Consider a hypothetical binary outcome \(Y\sim\mathrm{Bernoulli}(\theta)\), with response indicator \(R=1\) when \(Y\) is observed. Suppose the known response probabilities are \(P(R=1\mid Y=1)=0.8\) and \(P(R=1\mid Y=0)=0.4\). The response probability depends on the outcome that is missing when \(R=0\), so the mechanism is not MAR for this pattern.

For an observed success, the joint probability of \(R=1,Y=1\) is \(0.8\theta\); for an observed failure, it is \(0.4(1-\theta)\). When the outcome is missing, the probability of \(R=0\) requires summing over both possible outcomes:

$$ P(R=0)=0.2\theta+0.6(1-\theta). $$

Suppose a hypothetical sample of 100 participants contains 40 observed successes, 20 observed failures, and 40 missing outcomes. The observed-data likelihood is

$$ L(\theta) = (0.8\theta)^{40} \{0.4(1-\theta)\}^{20} \{0.2\theta+0.6(1-\theta)\}^{40}. $$

At \(\theta=0.5\), the three observed-category probabilities are \(0.8(0.5)=0.4\), \(0.4(0.5)=0.2\), and \(0.2(0.5)+0.6(0.5)=0.4\). The likelihood is therefore \(0.4^{40}0.2^{20}0.4^{40}\). The missing-outcome term depends on \(\theta\), so it cannot be discarded as a factor involving only the mechanism. The known response probabilities do not make the mechanism ignorable: the issue is the dependence on the unobserved outcome.

Worked Example: MAR but Parameters Are Not Distinct

Suppose, in a hypothetical study, \(Y\sim\mathrm{Bernoulli}(\theta)\) and \(R\sim\mathrm{Bernoulli}(\theta)\), independently given the model, so the response probability is also \(\theta\). Response is independent of the outcome, hence MAR, but the two parts of the model share the same parameter. The parameter space is not a product of freely varying outcome and mechanism parameters.

In a sample of 20, suppose 12 outcomes are observed, with 7 successes and 5 failures, and 8 outcomes are missing. The joint observed-data likelihood is

$$ L(\theta)= \theta^7(1-\theta)^5 \theta^{12}(1-\theta)^8 = \theta^{19}(1-\theta)^{13}. $$

Its maximum occurs at \( \hat{\theta}=19/(19+13)=19/32=0.5938 \). If the response part were omitted, the outcome-only likelihood would instead yield \(7/12=0.5833\). The difference arises because the observed response count also informs the shared parameter. MAR holds, but distinctness does not, so the mechanism cannot be ignored for likelihood inference on this \(\theta\).

Diagnostics, Alternatives, and Limits

MAR cannot generally be verified from observed data alone: the missing outcomes needed to check dependence on \(Y_{\mathrm{mis}}\) are not observed. Diagnostics can still identify problems with the working model or with its observable implications. Useful checks include comparing response rates across observed baseline groups and outcome histories, checking whether the proposed response model leaves systematic patterns in the observed records, and confirming that the outcome and response parameters have not been linked inadvertently. Such checks may expose poor fit or coding errors; they do not prove MAR.

Likelihood ignorability is not the same as several common alternatives. A complete-case analysis discards subjects with missing outcomes and can lose information or be biased unless its own assumptions justify it. A joint selection-model analysis specifies both the outcome distribution and the response mechanism; it is needed when the mechanism is not ignorable under the stated model. A pattern-mixture analysis models outcome distributions by response pattern. These approaches organize the joint model differently, but none removes the need to state assumptions. The relevant choice depends on the scientific question and the plausibility of the assumptions, not simply on which model is easiest to fit.

Ignorability also does not guarantee identification. For example, a covariate stratum with no observed outcomes may provide no direct information about its outcome distribution without additional structure. Model misspecification, dependent subjects not represented in the likelihood, or a response pattern outside the model's support can also undermine inference. Ignorability is a condition for omitting a mechanism from a particular likelihood analysis; it is not a general certificate of validity.

Common pitfall: “The dropout process is MAR, so I can always omit it.” This reasoning leaves out parameter distinctness. The correction is to check both whether response depends only on observed data and whether the outcome and response parameters vary independently. If either condition fails, the observed-data likelihood may couple the two parts.
1
Specify the conditioning information.
State which covariates and outcomes are observed for each response pattern, and define the outcome and response models.
2
Assess the two ignorability conditions.
Ask whether the response probability depends on missing outcomes after conditioning on observed data, and whether the parameter spaces are variation-independent.
3
Use the appropriate likelihood.
If both conditions hold, maximize the observed-data outcome likelihood for outcome parameters. Otherwise, retain or revise the joint model rather than silently dropping the response process.

Check Your Understanding

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

  1. A response probability depends on a recorded baseline covariate and the last observed outcome, but not on any missing outcomes. Is that dependence alone inconsistent with MAR?
  2. What does parameter distinctness require, and why is it relevant even if the mechanism is MAR?
  3. For an observed-data likelihood, what is done with outcomes that are missing under the realized response pattern?
  4. A model has MAR response, but an outcome parameter is constrained to equal a response parameter. Can the response part automatically be ignored for inference on that shared parameter?

Solutions. (1) No. The baseline covariate and last observed outcome are part of the observed information; dependence on them is compatible with MAR if the response probability does not additionally depend on missing outcomes. (2) The outcome and response parameters must be freely variable in separate parameter spaces, without shared parameters or linking constraints. Otherwise, the response likelihood can contain information about the outcome parameter. (3) Sum or integrate the outcome model over the missing components to obtain the observed-data outcome likelihood. (4) No. MAR addresses dependence on missing outcomes, but the shared parameter violates distinctness, so the response likelihood may affect inference on that parameter.

Previous

Next

Back to Tutorials
Clinical Biostats