FIML vs. Multiple Imputation: Which Missing-Data Method Should You Use?

2 min read ▶ Listen to this post

Full-information maximum likelihood (FIML) and multiple imputation (MI) are the two standard, defensible ways to handle missing data under the missing-at-random (MAR) assumption — meaningfully better than the older approaches (listwise deletion, mean substitution) that are still common in practice but generally indefensible in a reviewed methods section. Choosing between FIML and MI isn't about which is more rigorous; both are, when used correctly. It's about which fits your model and your data.

How they differ

FIML uses all available data directly within a single model — common in structural equation modeling and multilevel software — estimating parameters using every case's observed values without ever creating a "completed" dataset. There's no imputation step to run separately; it's built into the estimation itself.

Multiple imputation creates several complete versions of the dataset, each with missing values filled in using a model that reflects the uncertainty about what the true value might be, runs the intended analysis on each completed dataset, and pools the results using Rubin's rules. The multiple datasets exist specifically so the pooled result reflects the genuine uncertainty from having imputed rather than observed the missing values.

When FIML is the natural choice

• You're already running an SEM or multilevel model in software that supports FIML natively (Mplus, lavaan, and most multilevel packages do).
• You want one model, one estimation step, and don't need the missing-data-handling to also serve variables outside your final analysis model.

When multiple imputation is the better fit

• Your final model isn't SEM-based, or your software doesn't support FIML for the model you're running.
• You want to use auxiliary variables — predictors of missingness that aren't part of your substantive model — to improve the imputation. MI handles this more flexibly than FIML.
• You have a mix of variable types (continuous, categorical, count) where a single likelihood-based model would be awkward to specify, but a variable-by-variable imputation model handles each appropriately.
• You need to run several different analyses on the same dataset — impute once, analyze many ways, rather than rebuilding the missing-data handling for each model.

What matters more than the choice itself

Both methods assume the data are missing at random conditional on the variables in the model (or imputation model). That assumption is usually more consequential than which of the two methods you pick — and it's worth stating and defending explicitly in a methods section, since a reviewer who understands missing-data methods will ask about it either way.

If you're deciding how to handle missing data in an active analysis, this is a conversation worth having before you commit to a method.

One email whenever we publish something new. No spam, unsubscribe anytime.