Full-Information Maximum Likelihood Path¶
Method — instantiates Missingness-Aware Estimator Selection
Uses likelihood-based estimation with incomplete observed data when model and missingness assumptions are appropriate.
Full-Information Maximum Likelihood (FIML) estimates a model's parameters directly from the incomplete data by maximizing the likelihood of whatever each case actually contributes — a case with three of five variables observed contributes its three-dimensional marginal density, not a discarded row and not a filled-in guess. Its defining move is that it never completes the data at all: there is one model, one likelihood, and every observed value is used exactly once, so estimation and missing-data handling are the same step rather than two. This is what makes it FIML and not an imputation-then-analyze pipeline — there are no completed datasets to pool and no separate fill-in model whose errors must be tracked.
Example¶
A developmental-psychology team fits a structural equation model of how reading achievement and self-efficacy co-develop across grades 3, 5, and 7. Attrition is heavy: families move, and some children are tested at only one or two waves. Rather than drop anyone or impute the missing waves, they estimate the SEM by FIML — each child contributes the joint density over exactly the waves they completed, and the software maximizes the summed log-likelihood over all children at once. Because attrition here is plausibly related to observed early scores and to family variables the study measured, the team adds those variables into the model as auxiliary correlates — predictors that carry no substantive hypothesis but sharpen the ignorability assumption and recover efficiency. The estimate that emerges uses the partial records fully, under an explicitly stated assumption that, given the observed scores and correlates, the pattern of who was tested when carries no further information about the growth parameters.
How it works¶
- Write one likelihood. Specify the full joint model for all variables of interest.
- Score each case by its observed part. For every unit, evaluate the likelihood contribution of the variables it actually has, integrating over what it lacks.
- Maximize the sum. Find the parameters that best explain all partial records jointly — no deletion, no imputation, no pooling.
- Add auxiliary correlates. Fold in variables that predict missingness or the incomplete outcomes to make the ignorability assumption more credible and tighten standard errors.
Tuning parameters¶
- Auxiliary-variable set — how many missingness-relevant correlates to include; more can strengthen the ignorability assumption and efficiency, but a bloated set adds parameters and convergence trouble.
- Distributional model — the assumed joint form (e.g. multivariate normal versus a robust alternative); a wrong shape biases estimates because FIML leans entirely on it.
- Standard-error method — model-based versus robust/sandwich errors, trading assumptions for coverage under mild misspecification.
- Estimated-parameter parsimony — a saturated versus constrained model, balancing fidelity against convergence with sparse patterns.
When it helps, and when it misleads¶
FIML shines when the analysis is already a single coherent likelihood model — SEM, mixed models, latent-variable growth curves — and the mechanism is plausibly ignorable given observed data. It uses every scrap of partial information without the bookkeeping of multiple completed datasets, and it is exact rather than simulation-based, so two analysts get the same number.[n1]
Its failure mode is that all its robustness comes from one model being right: misspecify the joint distribution and there is no second line of defense, unlike an estimator that hedges across two models. It quietly assumes ignorability, so if missingness depends on the unobserved values themselves, FIML is confidently biased — the classic misuse is to reach for it because "it uses all the data" while never checking whether the mechanism is actually MAR. The guarding discipline is to state the mechanism assumption out loud, enrich it with auxiliary correlates, and hand any residual MNAR worry to a sensitivity method rather than trusting the likelihood to absorb it.
How it implements the components¶
missingness_mechanism_assumption_frame— FIML's validity rests on an explicitly stated ignorability (MAR) assumption; the path makes that assumption the visible precondition for trusting the likelihood, rather than burying it in software defaults.auxiliary_variable_strategy— it folds missingness-relevant correlates into the estimated model (a "saturated correlates" tactic) so that conditioning on observed information makes the ignorability claim more defensible and the estimates more efficient.
It does not choose among competing estimators or reconcile a separate imputation model with the analysis model — estimator_identification_match and imputation_model_congruence_check are Doubly Robust Missingness Adjustment's; FIML commits to a single joint likelihood and gains robustness only if that one model is right, where the doubly robust method hedges across two.
Related¶
- Instantiates: Missingness-Aware Estimator Selection — the likelihood-based path for an ignorable, model-based analysis.
- Consumes: Process-Based Missingness Audit surfaces the process correlates that become FIML's auxiliary variables.
- Sibling mechanisms: Doubly Robust Missingness Adjustment · Inverse-Probability Weighting Model · MCAR Diagnostic Test and Balance Review · Missingness Indicator Matrix · Multiple Imputation Workflow · Pattern-Mixture Sensitivity Model · Process-Based Missingness Audit · Selection-Model Sensitivity Analysis · Tipping-Point Analysis
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Full-Information Maximum Likelihood Path operates as a computation, comparison, model, or analytic representation used to infer, estimate, or choose because it uses likelihood-based estimation with incomplete observed data when model and missingness assumptions are appropriate.
Independent corroboration: The frozen evidence defines Full-Information Maximum Likelihood Path as 'Uses likelihood-based estimation with incomplete observed data when model and missingness assumptions are appropriate', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: FIML is a likelihood-based missing-data estimator developed in statistical inference.
Review resolution: Both reviewers agree that statistics_experimental_design is primary: FIML is a likelihood-based missing-data estimator developed in statistical inference. I retain no alternate origin domains only where the reviewers identify formative lineage, not every later application. I resolve origin_mode as single_lineage because the mechanism is best explained as one continuous professional lineage. I resolve domain_reach as multi_domain because it transfers across several fields but is not a domain-free primitive. Encyclopedia synthesis is false because the exact generalized packaging is already established enough that encyclopedia-specific synthesis is not required.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
FIML and Multiple Imputation Workflow target the same MAR problem and often agree numerically; the practical difference is bookkeeping. FIML solves estimation and missingness in one likelihood, so it is clean when the analysis is a single model, but it cannot mix a Gaussian estimation model with, say, a logistic auxiliary the way imputation can — that flexibility is imputation's edge.
[n1] Ignorability (in Rubin's sense) — the condition under which the missing-data mechanism can be ignored for likelihood-based estimation because, given the observed data, the probability of missingness carries no further information about the parameters. FIML's consistency depends on it. ↩