1 Concept and motivation
Full-information maximum likelihood (FIML) is a likelihood-based estimation method for statistical models with incomplete observations. In settings such as structural equation modeling (SEM) and other latent variable frameworks, some individuals may have missing values for one or more variables. FIML addresses this by evaluating the contribution of each case to the likelihood using only the variables that are observed for that case, while mathematically integrating over the missing components implied by the model.
1.1 Relation to maximum likelihood estimation
Standard maximum likelihood estimation (MLE) assumes that each observation provides a full vector of data, allowing the likelihood to be computed directly from the model’s density. When observations are incomplete, the likelihood for a case cannot be computed in the same way because the missing entries are unknown. FIML extends the MLE idea by constructing, for each pattern of missingness, a likelihood contribution that is consistent with maximum likelihood under partial information. The overall objective sums (or averages) these case-wise contributions across the dataset and estimates parameters by maximizing the resulting likelihood.
1.2 Why “full information” matters with missing data
“Full information” refers to the idea that the estimation procedure uses all available observations and all observed parts of each observation, rather than discarding incomplete cases or requiring a preliminary imputation step. For example, if a person has missing values on one variable but observed values on others, FIML still contributes to estimating parameters through the likelihood evaluated for the observed subset. This typically improves efficiency and reduces bias compared with methods that ignore incomplete cases.
1.3 Comparison to complete-case analysis and single imputation
Complete-case analysis restricts estimation to observations without any missing values. This can substantially reduce sample size and may introduce bias if the missingness mechanism is related to unobserved data. Single imputation fills in missing values once (often via regression or similar techniques) and then treats the filled dataset as if it were complete; while convenient, it can underestimate uncertainty because it ignores imputation variability. FIML is designed to avoid both the information loss of complete-case analysis and the additional uncertainty handling shortcomings of single imputation by embedding missing-data handling into the likelihood itself.
2 Mathematical foundation
The core mathematical feature of FIML is that it defines the likelihood for partially observed data by integrating the model’s joint density over the unobserved portions. This yields an objective function that is compatible with maximum likelihood.
2.1 Likelihood with partially observed data
Suppose a model defines a joint distribution for a vector of variables \( (y_{\text{obs}}, y_{\text{mis}}) \) with parameters \( \theta \). If, for a particular individual, only \( y_{\text{obs}} \) is observed and \( y_{\text{mis}} \) is missing, the likelihood contribution is based on the marginal density of the observed part: \[ L_i(\theta) = p(y_{\text{obs},i}\mid \theta). \] This marginal density is obtained from the joint model by integrating out the missing components.
2.2 Partitioning observed and missing variables
Let the complete set of variables in the model be partitioned into observed and missing components for each case. In practice, missingness patterns can differ across individuals: one person may have missing values in a subset of outcomes, another may miss different variables, and so on. FIML handles these variations by computing a likelihood contribution appropriate to each case’s observed subset, with the implied missing part integrated out.
2.3 Integrating over missing components
Given a joint density \( p(y_{\text{obs}}, y_{\text{mis}} \mid \theta) \), the observed-data likelihood is \[ p(y_{\text{obs}}\mid \theta) = \int p(y_{\text{obs}}, y_{\text{mis}} \mid \theta)\, d y_{\text{mis}}. \] In many SEM-like settings, the integral can be expressed analytically under distributional assumptions (for example, multivariate normality). When analytic integration is not available, numerical methods or approximations may be used.
2.4 Objective function and parameter maximization
The full objective function sums the log-likelihood contributions over all individuals: \[ \ell(\theta)=\sum_{i=1}^N \log p(y_{\text{obs},i}\mid \theta). \] Parameter estimates are obtained by maximizing \( \ell(\theta) \). Because each case contributes based on its observed subset, the fitted parameters reflect information pooled across all available parts of the dataset.
3 Missing-data assumptions and implications
FIML’s validity depends on assumptions about the missingness mechanism. The method can be interpreted as producing consistent estimates when the missing-data process satisfies conditions that relate the probability of missingness to observed (rather than unobserved) information.
3.1 Missing completely at random (MCAR)
Under missing completely at random (MCAR), the probability that a value is missing is independent of both observed and unobserved data. When MCAR holds, using likelihood with partial information typically yields unbiased and consistent estimates, and complete-case analysis becomes less problematic (though still often less efficient).
3.2 Missing at random (MAR)
Under missing at random (MAR), the probability of missingness may depend on observed variables but not on the missing values themselves after conditioning on the observed information. In many applications, MAR is the key assumption under which FIML is favored over complete-case deletion. Practically, MAR implies that including relevant observed covariates and outcomes in the model can help satisfy the condition by accounting for systematic differences between cases with different missingness patterns.
3.3 Missing not at random (MNAR) overview
Missing not at random (MNAR) means the missingness depends on the unobserved values even after conditioning on observed data. In this case, the likelihood constructed from observed parts may not represent the true missingness mechanism, and FIML may yield biased estimates unless the model explicitly accounts for the missing-data process. MNAR scenarios often require specialized modeling strategies, sensitivity analyses, or auxiliary variables that inform the missingness mechanism.
3.4 Practical diagnostics for missingness patterns
Missingness is rarely “diagnosed” definitively, but patterns can be investigated. Common diagnostics include comparing observed distributions across missingness groups (e.g., whether observed outcomes differ by whether a related variable is missing), examining missingness rates by key covariates, and testing whether missingness can be explained using observed predictors. While these checks cannot prove MAR, they can guide model specification and motivate sensitivity analyses.
4 FIML in latent variable and structural models
FIML is widely used in latent variable modeling because such models naturally express dependencies among observed indicators, latent factors, and covariates. The missing-data handling can then be integrated directly into estimation.
4.1 Application in structural equation modeling (SEM)
In SEM, observed indicators are linked to latent variables via measurement models, while latent variables may be connected through structural relations. With missing data on indicators, outcomes, or covariates, FIML computes the likelihood for each person by evaluating the model-implied distribution of the observed indicators and integrating out missing components. This approach avoids throwing away cases due to partial missingness and supports consistent estimation under appropriate missingness assumptions.
4.2 Handling multivariate outcomes and covariates
Many SEM applications involve multivariate outcomes and multiple covariates. Missingness can occur in any subset of these variables. FIML accommodates this by using the multivariate joint distribution implied by the model: if some outcomes are observed and others are absent for a given individual, their likelihood contribution is based on the marginal distribution of the observed components conditional on parameters.
4.3 Estimating measurement and structural components
Because SEM parameters jointly determine the implied joint distribution, the estimation of measurement parameters (e.g., factor loadings, measurement error variances) and structural parameters (e.g., regressions among latent variables) is influenced by all available data. Cases contribute information wherever their observed variables intersect the model structure, enabling simultaneous estimation without separately imputing missing measurements or outcomes.
4.4 Cross-equation information sharing
One benefit of latent variable frameworks is that indicators and outcomes can inform each other through shared latent constructs. When missingness affects one part of the system, observed variables tied to the same latent structure can still provide information about parameters elsewhere. This “information sharing” is a defining feature of full-information likelihood: it leverages the model’s network of relationships rather than treating each equation or variable in isolation.
5 Estimation details and implementation
Implementation details determine how reliably FIML estimates converge and how accurately uncertainty is quantified.
5.1 Starting values and optimization strategies
FIML estimation is typically carried out using nonlinear optimization. Choosing suitable starting values can be important, especially for complex SEMs with many parameters or constrained forms. Optimization strategies may include gradient-based algorithms and step-size controls, and software often provides multiple methods. Poor initialization can lead to local maxima, slow convergence, or unstable estimates.
5.2 Convergence checks and numerical stability
Convergence is usually assessed through criteria such as changes in parameter values, changes in the objective function, and gradient norms. Numerical stability can be affected by scaling of variables, extreme missingness rates, near-nonidentification, or ill-conditioned covariance matrices. Robust pre-processing (e.g., scaling and appropriate model constraints) and careful monitoring of optimizer output help ensure that reported estimates correspond to a valid maximum.
5.3 Standard errors and variance estimation
Standard errors quantify estimation uncertainty and often rely on asymptotic theory. In likelihood-based settings, one may use information-matrix-based variance estimates when regularity conditions are plausible. In SEM implementations, variance estimation can also involve numerical derivatives and finite-sample corrections depending on the estimator and software configuration.
5.4 Robust (sandwich/heteroskedasticity-consistent) options
When distributional assumptions are questionable or heteroskedasticity is present, robust variance estimators can be used. “Sandwich” (heteroskedasticity-consistent) standard errors are a common option in likelihood-based frameworks. These typically preserve consistency under weaker conditions about the error distribution, though they may increase standard error magnitude and can be sensitive to model misspecification or small samples.
6 Model evaluation and fit
FIML produces parameter estimates and likelihood-based diagnostics. Model evaluation must account for both model structure and missingness handling.
6.1 Likelihood-based fit comparisons
With FIML, models can be compared using differences in (log-)likelihood when nested models are considered and the relevant regularity conditions hold. Such comparisons assess whether adding parameters significantly improves explanatory power in the observed-data likelihood sense. This is often done using likelihood ratio tests or related statistics in SEM software.
6.2 Information criteria (e.g., AIC/BIC) and interpretation
Information criteria such as AIC and BIC trade off model fit and complexity. They are commonly based on the maximized log-likelihood and a penalty term for the number of parameters. Lower values generally indicate a preferred balance, but interpretation should consider sample size effects, model identifiability, and whether the likelihood calculations correspond to the same missing-data handling framework across compared models.
6.3 Residuals, diagnostics, and influence concepts
Residual-based diagnostics may be adapted to the model-implied covariance structure and to the observed-data likelihood context. Influence diagnostics assess whether particular cases with distinctive observed patterns (including extreme missingness patterns) disproportionately affect the fitted model. While exact influence measures can vary by software and model class, the general goal is to identify sensitivity to outliers or influential observations.
6.4 Sensitivity checks under different missing-data handling
Even when FIML is the primary approach, sensitivity analyses can strengthen conclusions. Common checks include comparing results from alternative missingness treatments (e.g., complete-case estimates as a benchmark), altering model specifications for variables related to missingness, or exploring plausible MNAR scenarios via targeted perturbations. Agreement across reasonable alternatives supports robustness, while divergence highlights potential missing-data dependence.
7 Special cases and alternatives
FIML sits among several strategies for handling missing observations. Its relatives differ in whether missing-data integration is performed directly in the likelihood, by iterating between imputation-like and parameter steps, or by combining imputed datasets.
7.1 Expectation-maximization (EM) vs FIML
The expectation-maximization (EM) algorithm is often used to fit models with missing data by alternating between computing expected sufficient statistics (E-step) and maximizing parameters (M-step). EM and FIML can be related in spirit, but in many implementations FIML aims to maximize the observed-data likelihood directly, whereas EM targets the same optimum indirectly through iterative steps. EM may be convenient when integrals are difficult, but it can converge slowly near the optimum.
7.2 Multiple imputation vs FIML
Multiple imputation (MI) fills in missing values multiple times to create several completed datasets, estimates model parameters in each, and combines results while accounting for imputation uncertainty. Compared with FIML, MI can be more flexible for complex missing-data mechanisms and can be easier to adapt when models do not naturally provide a tractable observed-data likelihood. However, MI requires careful imputation model specification and congeniality with the analysis model.
7.3 Weighted likelihood and related approaches
Weighted likelihood methods modify the likelihood contribution of each case using weights that reflect missingness patterns or assumptions about selection. These approaches can accommodate some forms of missingness by reweighting contributions, but they depend on modeling or specifying parts of the missingness mechanism. Compared with FIML, weighted methods may be conceptually and practically more complex depending on the setting.
7.4 Bayesian alternatives (conceptual overview)
Bayesian approaches handle missing data by treating missing values and parameters as random variables, using priors and posterior computation to integrate uncertainty. Missing-data integration is performed through the posterior distribution rather than classical likelihood maximization. These methods can directly incorporate MNAR mechanisms via appropriate modeling, at the cost of increased computational and modeling complexity.
8 Assumptions and limitations
FIML is powerful, but it rests on modeling and missingness assumptions that can fail in practice. Understanding the boundaries helps avoid overconfident interpretation.
8.1 Distributional requirements (e.g., normality approximations)
Many SEM/FIML implementations assume that continuous variables are approximately multivariate normal or rely on asymptotic normality approximations. Violations can affect standard errors, likelihood-based tests, and fit indices. Robust variance estimators and alternative distributional modeling choices can mitigate some issues, though they do not fully replace correct model specification.
8.2 Treatment of categorical variables (conceptual notes)
Categorical variables introduce additional complications because their likelihood contributions depend on discrete outcome probabilities rather than continuous densities. Many software packages extend FIML-like ideas to categorical outcomes using appropriate generalized modeling (e.g., latent response or categorical likelihood formulations). Conceptually, the same principle remains: compute likelihood contributions using observed categories and integrate over missing categories as implied by the model.
8.3 Identification considerations in structural/latent models
Even with correct missingness handling, latent variable models must be identified. If the model is underidentified or near nonidentification, missing-data likelihood contributions may not stabilize estimation, leading to large standard errors or unstable parameter estimates. Checking identification, ensuring adequate sample size, and using sensible parameter constraints remain essential.
8.4 When FIML may be unreliable
FIML can be unreliable when missingness is extensive in key variables, when the missingness mechanism strongly deviates from MAR and MNAR is not modeled, or when the joint model is misspecified (e.g., incorrect functional forms or omitted relationships). Small samples with complex missingness patterns can also undermine asymptotic approximations used for uncertainty quantification.
9 Reporting and reproducibility
Transparent reporting allows readers to evaluate whether FIML assumptions are plausible and whether results can be reproduced.
9.1 What to report about missing-data conditions
Reports should describe which variables have missing values, the overall missingness rates, and how missingness patterns differ across variables. If missingness is studied in a MAR-consistent way, authors typically list which covariates or auxiliary variables were included to support the missingness assumption.
9.2 Software and settings (e.g., estimators, robust options)
Results should state the software package and estimation settings used, including the estimator type, any robust standard error options, and choices related to distributional assumptions. Because different packages may compute observed-data likelihoods and standard errors differently, these details help interpret results across studies.
9.3 Documentation of assumptions and diagnostics
Authors should summarize diagnostics or at least qualitative findings about missingness patterns, such as whether missingness appears related to observed predictors. When sensitivity analyses are performed, the report should specify the alternative missing-data handling approaches and the nature of MNAR explorations.
9.4 Interpreting results responsibly
Interpretation should distinguish between evidence from the structural/measurement model and uncertainty introduced by missingness. Even when the method uses all available information, the conclusions depend on model correctness and missingness assumptions. Presenting fit statistics, uncertainty, and sensitivity checks can prevent overinterpretation of fragile findings.
10 Practical workflow examples
The following workflow outlines how FIML is typically used in applied modeling, with emphasis on model specification, estimation, interpretation, and communication.
10.1 Defining the model and identifying missingness structure
The process begins by specifying the substantive model (measurement and structural components in SEM) and then documenting the missingness structure: which variables are missing, how frequently they are missing, and whether missingness patterns suggest dependence on observed covariates. This informs which variables should be included in the analysis model to make MAR more plausible.
10.2 Running FIML estimation in typical SEM workflows
Next, the researcher runs SEM with an FIML-capable estimator or a missing-data option that computes observed-data likelihood contributions. During estimation, the optimizer output is monitored for convergence, and the analyst checks whether the solution appears stable. If convergence issues arise, adjustments to starting values, scaling, or model constraints may be needed.
10.3 Interpreting parameter estimates with incomplete data
Parameter estimates are interpreted as model-based quantities estimated using all observed portions of each case. It is appropriate to interpret coefficients and loadings with reference to the substantive theory, while also acknowledging that missingness affects which information enters each likelihood contribution. Large differences between estimates from FIML and more limited approaches can signal sensitivity to missingness assumptions or model specification.
10.4 Communicating uncertainty and model fit with missingness
Finally, the report communicates fit indices and uncertainty measures while noting the missing-data handling method. If robust standard errors or sensitivity analyses are used, their results should be clearly summarized. Where possible, authors explain whether conclusions remain consistent under alternative reasonable assumptions, supporting the credibility of the findings.