Statistics Glossary → Multiple Imputation
Multiple Imputationanalysis
A method that fills in missing values many times, each time with slightly different predictions based on the observed data, repeats the analysis in every completed data set and combines the results so that they also reflect the uncertainty created by the missingness.
Concept
Missing values are unavoidable in real data: patients skip some tests and participants leave some questions blank. Deleting incomplete rows, the most common fix, both shrinks the sample and biases the results if the missingness is not random. Filling each missing value with a single guess (e.g. the mean) treats invented values as if they were real and makes standard errors too small. Multiple imputation tackles both problems together.
The method has three steps. (1) Imputation: each missing value is filled in m times from models built on the other variables, with random variability added, producing m completed data sets. (2) Analysis: the desired model (e.g. a regression) is fitted in each data set separately. (3) Pooling: by Rubin's rules, the combined estimate is the average of the m estimates, and the total variance combines the average within-set variance (Ū) with the between-set variance (B): T = Ū + (1 + 1/m)·B. The difference between the sets represents the extra uncertainty that comes from the missing data.
The key assumption is that data are missing at random (MAR): the probability that a value is missing can be explained by the observed variables in the model and does not depend on the missing value itself. This cannot be proven from the data; if missingness is thought to depend on the missing value itself (MNAR), sensitivity analyses should be done. The imputation model should include all variables in the analysis model, including the outcome; otherwise relationships are weakened towards zero.
When to use it
- When the variables used in the analysis have more than a few per cent missing values and deleting incomplete rows noticeably shrinks the sample or may create bias.
- Multiple imputation suits missingness related to observed variables (MAR); if missingness is very small and completely random (MCAR), deleting incomplete rows can also give unbiased results.
- If the model is a structural equation or path model, full information maximum likelihood (FIML) is an alternative; if missingness is thought to depend on the missing value itself (MNAR), use a δ-adjustment sensitivity analysis or selection / pattern-mixture models.
An everyday example
A family health centre studies how blood pressure relates to age, body mass index and smoking in 300 patients. But body mass index is missing for 25% of patients and smoking status for 10%. Deleting incomplete rows would leave 200 patients. Multiple imputation fills in the missing values 20 times, each slightly differently, based on the patients' age, blood pressure and other information; the results of the 20 regressions are combined, and the standard errors also include the uncertainty created by the missing information.
Parameters in MerQur
Dependent variable (DV), Independent variables
Concept: These are the dependent and independent variables of the analysis model. All of them, including the outcome, also enter the imputation model, so that their relationships are preserved in the filled-in values.
Everyday example: Blood pressure is the dependent variable; age, body mass index and smoking are the independent variables. Blood pressure is also used when filling in missing body mass index; otherwise the relationship between the two would look weakened.
What it does: The variables of the analysis model. A binary (0/1) dependent variable and, in FCS, text (categorical) independent variables can be selected.
Effect on the result: All imputation and pooling change.
On the validation data: nhanes chl ~ age + bmi + hyp; in the colon logistic model the nodes coefficient is 0.1984 (SE 0.0270); for boys the categorical independent variable (gen) is used in FCS.
Auxiliary variables
Concept: Auxiliary variables do not enter the analysis model but help predict the missing values. Adding auxiliary variables related to the missingness or to the incomplete variable makes the MAR assumption more plausible and makes the imputations more precise.
Everyday example: Waist circumference is not part of the analysis but is strongly related to body mass index. Added as an auxiliary variable, it lets the missing body mass index values be filled in far more accurately.
What it does: Variables that do not enter the analysis but are predictors in the imputation model (they strengthen the MAR assumption).
Effect on the result: The imputed values and therefore the pooled estimates change.
On the validation data: nhanes chl ~ age + bmi, hyp auxiliary: bmi 2.7755 (SE 2.537, λ 0.292).
m (number of imputations) (2 to 100; default 5)
Concept: The number of completed data sets to create. As m grows, the between-set variance is estimated more stably and the results are less affected by randomness; with a high missing rate, a larger m is needed.
Everyday example: With 25% of body mass index missing and only 5 imputations, the coefficient and its standard error can shift noticeably each time the analysis is rerun. 20 or more imputations reduce this wobble.
What it does: The number of imputations. The between-imputation variance B (with m-1 df) and Rubin's total variance T = Ū + (1 + 1/m)B depend on this number.
Effect on the result: The pooled estimate, SE and degrees of freedom change.
On the validation data: BayesianRidge, bmi: m = 2: 3.345 (SE 2.465, df 12.39); 5: 2.651 (2.512, df 14.08); 20: 4.252 (3.573, df 7.62); 50: 4.846 (3.402, df 9.03).
Max iterations (default 10)
Concept: In chained imputation, each incomplete variable is predicted in turn from the others, and this cycle is repeated many times. The number of iterations should be enough for the chain to reach a stable state (convergence) independent of its starting values.
Everyday example: In the first cycle, missing body mass index is predicted from still roughly filled smoking data; in later cycles both variables are corrected against each other. Ten cycles are often enough; if the convergence trace forms a flat band, the chain has settled.
What it does: max_iter of the IterativeImputer in BayesianRidge; the chain length in FCS.
Effect on the result: The imputed values change.
On the validation data: BayesianRidge, bmi: 5.201 at 1 iteration (λ 0.643); 2.651 at 10 (λ 0.178); 4.726 at 30 (λ 0.567). In the BayesianRidge engine, single-seed results are very sensitive to iteration and seed.
Seed
Concept: The seed is the starting point of the random imputations. The same seed reproduces the same imputations and results; a different seed changes the results slightly, and the size of that change also shows whether m is large enough.
Everyday example: The seed is reported so that the coefficients can be reproduced. If the body mass index coefficient changes noticeably with a different seed, more imputations are needed.
What it does: seed + i is used for the i-th imputation.
Effect on the result: The imputed values change; the pooling rule does not change.
On the validation data: BayesianRidge: seed 42: bmi SE 2.512 (λ 0.18); seed 7: SE 4.811 (λ 0.81). The comparison with the references is made over a 30-seed distribution.
Imputation method (IterativeImputer + BayesianRidge (numeric only) / FCS/MICE - numeric: PMM; categorical: logreg/polyreg (default) / FCS/MICE - numeric: norm (Bayesian linear); categorical: logreg/polyreg)
Concept: Sets the model used to fill in missing values. Predictive mean matching (PMM) borrows the real value of one of a few observed people closest to the prediction and so produces only realistic values; the Bayesian linear (norm) method draws from a normal distribution; categorical variables are filled with logistic models.
Everyday example: With PMM, a missing body mass index is filled with the real value of one of several patients of similar age and blood pressure, so an unrealistic value such as 12 cannot appear. Smoking (yes/no) is filled with a logistic model.
What it does: The default FCS is PMM (5 candidates, type 1 matching). FCS norm is Bayesian linear regression (Rubin 1987). For categorical variables, logreg/polyreg are used. The BayesianRidge option does posterior sampling with chained Bayesian ridge regression (numeric only).
Effect on the result: The distribution of the imputed values and the coefficients change.
On the validation data: nhanes, 30-seed mean, bmi: BayesianRidge 5.30 ± 1.37 (SE 3.66); FCS norm 6.98 ± 0.67. With the default PMM, bmi 6.454 (SE 1.900). BayesianRidge attenuates the coefficients in a small sample and inflates the SE.
→ Multiple imputation by chained equations (FCS/MICE) and PMM
Analysis model (Linear regression (OLS, default) / Logistic regression (binary DV) / Poisson regression (count DV))
Concept: The analysis model fitted in each completed data set: linear regression for a continuous outcome, logistic for a binary outcome and Poisson for a count outcome.
Everyday example: If the outcome is the blood pressure value, linear regression is chosen; if it is 'hypertension yes / no', logistic regression; if it is the number of emergency visits in the past year, Poisson regression.
What it does: The model estimated in each completed dataset.
Effect on the result: The pooled coefficients change.
On the validation data: colon logistic: nodes 0.1984 (SE 0.0270). In the combination of airquality Poisson with BayesianRidge, the imputed Ozone comes out negative (smallest -49.547) and an explicit error is given; in the default PMM, the imputed values come from observed values.
Restrict imputed values to the observed range (min/max) and Round imputed values of integer-valued variables
Concept: Keeps the filled-in values between the observed minimum and maximum and rounds them for integer-valued variables. This prevents unrealistic values; but since squeezing and rounding can distort the distribution somewhat, it should be used when needed.
Everyday example: A draw from a normal distribution may produce a body mass index of 8 when observed values range from 18 to 42. Restricting to the range pulls it up to 18; for an integer variable such as number of children, 2 is written instead of 2.4.
What it does: Squeezes the imputed values into the observed [min, max] range (mice post squeeze); rounds in variables whose observed values are integers.
Effect on the result: The imputed values change. In BayesianRidge the constraint is applied after imputation, in FCS at every update.
On the validation data: With clipping, chl [113, 284], bmi [20.4, 35.3], hyp [1, 2]; all imputed cells are within the range. With rounding, chl and hyp are integers.
Chart (Pooled coefficients (default) / Convergence trace (FCS: mean/SD × iteration) / Observed vs imputed distributions) and Save the m completed datasets (long-format CSV, .imp/.id)
Concept: The pooled coefficients plot summarises the results; the convergence trace shows whether the chain has stabilised, and comparing observed with imputed distributions shows whether the filled-in values are plausible. The completed data sets can be exported for use in other programs.
Everyday example: If the imputed body mass index values have a distribution similar to the observed ones, the imputation is plausible. If the imputed values are clearly higher this need not be an error: if missingness is more common among overweight patients, higher imputed values are expected.
What it does: Chart (pooled coefficients, convergence trace, observed and imputed distributions); exporting the imputed dataset as CSV.
Effect on the result: It does not change the numerical results. The exported CSV (m·n rows) matches the long-format output of the engine exactly.
On the validation data: None.
→ Multiple imputation by chained equations (FCS/MICE) and PMM
Confidence Level (%90 / %95 / %99 / Other (custom))
Concept: Sets the level of the confidence interval for the pooled coefficients. The interval includes both the within-set uncertainty and the between-set uncertainty caused by the missing data, and is built from a t distribution with correspondingly adjusted degrees of freedom.
Everyday example: The 95% interval for the body mass index coefficient may differ from one based only on patients with complete records: it can narrow because more patients are used, but widen if the imputations differ a lot.
What it does: The pooled coefficient interval (t, Barnard-Rubin df). The pooled R² interval is always 95% (stated in the text).
Effect on the result: Only the intervals change.
On the validation data: bmi CI: 95% [-2.734, 8.037], 90% [-1.772, 7.075], 85.5% [-1.225, 6.528].
Monotone imputation (sequential regression), MNAR sensitivity variable and MNAR δ shifts
Concept: With monotone missingness, the variables can be put in an order such that anyone missing one variable is also missing all later ones; imputation can then be done in a single pass. In the MNAR sensitivity analysis the imputed values are shifted by δ, and the shift at which the conclusion changes (the tipping point) is examined.
Everyday example: If overweight patients are thought to avoid being weighed, the missing body mass index values may be higher than the observed ones. δ = +2 and +4 are added to the imputed values and the analysis repeated; if the conclusion changes only at a large shift such as +6, the finding is robust to departures from MAR.
What it does: In a monotone missing pattern, the variables are ordered by the number of missing values and imputed in a single pass using only the preceding variables; if the pattern is not monotone, an explicit error is given. In the MNAR option, δ is added to the imputed values of the selected variable and a sensitivity (tipping-point) table with Rubin pooling is given for each δ; the main result remains δ = 0 (MAR). It works only with the FCS engine.
Effect on the result: The imputed values and the pooled coefficients as a function of δ change.
On the validation data: Seed 7, m = 5, norm, monotone imputation: Intercept 0.736790966554 (SE 1.7424873166054, df 132.58), x1 0.345351605735 (df 128.65), x2 0.545344277151 (df 49.55); pooled by Rubin's rules. δ: the imputed values of the last variable shift by exactly δ.
Missing value analysis: t tests and categorical tables (SPSS MVA /TTEST /TABULATE)
Concept: Examines the pattern of missingness: people with and without a value missing on one variable are compared on the other variables. Clear differences show that the data are not missing completely at random (MCAR); no difference does not prove MAR.
Everyday example: If patients missing body mass index have a mean age of 62 against 48 for the others, missingness is related to age and is not MCAR. Including age in the imputation model is then especially important.
What it does: For every numeric variable with 5% or more missingness, the Welch t, df, p, group n's and means on the other numeric variables by the missingness indicator; for categorical variables, a cross-tabulation per indicator. The output is appended under the Little MCAR row.
Effect on the result: Only adds diagnostic output.
On the validation data: nhanes2: for chl by bmi missingness t = 0.5002708691, df = 12.2747293162, p = 0.6257327768; for bmi by chl missingness t = -0.0657210688, p = 0.9498934334. All cross-table n's and percentages are the same in the four programs.
Example on real data
Data: Schafer 1997, R mice nhanes
Variables: dependent chl (cholesterol), independent age (age group), bmi (body mass index), hyp (hypertension)
- Question: In the nhanes example with missing data (n = 25; only 13 complete cases), is cholesterol (chl) related to age group (age), body mass index (bmi) and hypertension (hyp)?
- Missing data: chl 40%, bmi 36%, hyp 32% missing; Little MCAR test χ² = 8.000, df = 9, p = 0.5342. MCAR is not rejected, but this does not prove MAR; multiple imputation rests on the MAR assumption.
- Method: m = 5 completed datasets are generated with FCS (MICE) PMM, OLS is estimated in each, and the results are pooled with Rubin's rules (Barnard-Rubin df).
- Result: bmi (β = 6.4539, SE = 1.9002, p = 0.0044) and age (β = 33.70, SE = 11.79, p = 0.0169) are significantly and positively related to cholesterol; hyp is not significant (p = 0.7245).
- Model fit: Pooled R² = 0.5346, 95% CI [0.1506, 0.8033]; D1 test for all slopes F(3, 14.66) = 5.7418, p = 0.0083.
- Loss of information and comparison: The largest λ is for age, 0.309 (moderate loss of information, m ≥ 10 may be considered). Listwise deletion overstates the age effect at 55.21; imputation uses all cases and gives a more cautious estimate.
- APA: In the regression pooled with multiple imputation (m = 5), bmi, b = 6.45, SE = 1.90, t(13.77) = 3.40, p = .004, and age, b = 33.70, SE = 11.79, t(10.09) = 2.86, p = .017, significantly predicted cholesterol.
Related concepts
Multiple imputation by chained equations (FCS/MICE) and PMMMissing-data mechanisms (MCAR, MAR, MNAR)Rubin's rules (pooling imputation results)Auxiliary variables (imputation model)Unconditional exact tests (Barnard, Boschloo)Convergence and maximum iterationsSensitivity analysisDistribution family (GLM)
The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.