mimar
Compact Multiple Imputation, Assessment, and Reporting
Provides compact tools for missing-data analysis, including artificial amputation, chained single and multiple imputation, statistical and machine-learning-based imputation methods, diagnostic evaluation, and post-imputation pooling.
mimar
mimar implements a compact chained-imputation workflow in R for
missing-data analysis, artificial amputation, native and learner-backed
single and multiple imputation, diagnostic evaluation, and
post-imputation pooling.
The package is built around a complete missing-data workflow: describe the missingness, create benchmark amputations when needed, impute with native or learner-backed update rules, inspect diagnostics, evaluate recovered cells when truth is available, and pool post-fit quantities. The goal is a concise grammar for the whole workflow, not a replacement for every specialist feature in larger imputation systems.
The package owns the imputation loop. Every imputer, whether implemented natively or backed by a learner package, is called the same way:
impute(data, imputer = "pmm", m = 5, maxit = 5, seed = 1)
impute(data, imputer = "rf", m = 5, seed = 1)
impute(data, imputer = "xgboost", m = 5, seed = 1)
There is no dependency on funcml. Learner-backed imputers call their
original packages directly, and those backend packages are hard
dependencies so users can run any registered imputer without manually
resolving learner installations.
Installation
Install the released version from CRAN:
install.packages("mimar")
The CRAN package page is https://CRAN.R-project.org/package=mimar.
You can install the development version from GitHub:
install.packages("remotes")
remotes::install_github("ielbadisy/mimar")
Then load the package:
library(mimar)
Quick use
For normal use, impute() is the only function you need. The input data
can contain NA, and the completed outputs returned by complete() do
not. Set verbose = TRUE when you want a concise progress log for the
chained imputation workflow.
Tabular results (describe(), complete(), evaluate(), pool(),
imputer_registry()) are returned as basetable objects, which behave
like data frames.
i <- impute(a, imputer = "knn", m = 5, maxit = 5, seed = 1)
complete(i, 1)
complete(i, "all")
Grammar
describe()
ampute()
imputer_registry()
imputer()
impute()
complete()
evaluate()
pool()
pool_lm(), pool_glm(), pool_coxph(), pool_survreg(), pool_clogit(), pool_panglm()
pool_test()
nelsonaalen(), ipcw()
plot()
Short example
library(mimar)
set.seed(1)
dat <- data.frame(
age = rnorm(120, 50, 10),
bmi = rnorm(120, 25, 4),
sex = factor(sample(c("F", "M"), 120, TRUE)),
group = factor(sample(c("A", "B", "C"), 120, TRUE)),
smoker = sample(c(TRUE, FALSE), 120, TRUE)
)
a <- ampute(
dat,
prop = 0.25,
mechanism = "MAR",
target = c("bmi", "group"),
by = c("age", "sex"),
seed = 1
)
i <- impute(a, imputer = "knn", m = 5, maxit = 5, seed = 1, ncore = 2)
complete(i, 1)
#> # basetable: 120 x 5
#> age bmi sex group smoker
#> 1 43.73546 22.97617 M C FALSE
#> 2 51.83643 30.37216 F A FALSE
#> 3 41.64371 24.14168 F B TRUE
#> 4 65.95281 24.28177 F C TRUE
#> 5 53.29508 24.59924 F A FALSE
#> 6 41.79532 27.85067 M C TRUE
#> 7 54.87429 24.70574 M B TRUE
#> 8 57.38325 24.84946 F B FALSE
#> 9 55.75781 22.27336 F B FALSE
#> 10 46.94612 30.76463 M A FALSE
#> # 110 more rows
summary(i)
#> mimar imputation summary
#> # basetable: 1 x 11
#> rows columns n_imputations imputer maxit ncore stochastic
#> 1 120 5 5 knn 5 2 TRUE
#> total_missing_before total_imputed remaining_missing variables_imputed
#> 1 53 53 0 2
#>
#> Variables:
#> # basetable: 5 x 9
#> variable type method n_missing_before prop_missing_before n_imputed
#> 1 age numeric none 0 0.0000000 0
#> 2 bmi numeric knn 26 0.2166667 26
#> 3 sex factor none 0 0.0000000 0
#> 4 group factor knn 27 0.2250000 27
#> 5 smoker logical none 0 0.0000000 0
#> prop_imputed remaining_missing between_imputation_sd
#> 1 0 0 0.000000
#> 2 1 0 0.341076
#> 3 0 0 0.000000
#> 4 1 0 0.000000
#> 5 0 0 0.000000
evaluate(i)
#> mimar imputation evaluation
#> # basetable: 1 x 4
#> n_imputations imputer total_missing evaluated_cells
#> 1 5 knn 53 53
plot(i, type = "density")
Imputers
Inspect available imputers with:
imputer_registry()
#> # basetable: 25 x 10
#> imputer implementation package supports_numeric supports_binary
#> 1 mean mimar internal TRUE TRUE
#> 2 median mimar internal TRUE TRUE
#> 3 mode mimar internal TRUE TRUE
#> 4 naive mimar internal TRUE TRUE
#> 5 norm mimar internal TRUE TRUE
#> 6 pmm mimar internal TRUE TRUE
#> 7 spmm mimar internal TRUE TRUE
#> 8 logreg mimar internal TRUE TRUE
#> 9 polyreg mimar internal TRUE TRUE
#> 10 rf wrapped ranger TRUE TRUE
#> supports_multiclass stochastic
#> 1 TRUE FALSE
#> 2 TRUE FALSE
#> 3 TRUE FALSE
#> 4 TRUE FALSE
#> 5 TRUE TRUE
#> 6 TRUE TRUE
#> 7 TRUE TRUE
#> 8 TRUE TRUE
#> 9 TRUE TRUE
#> 10 TRUE TRUE
#> description available status
#> 1 Mean imputation for numeric targets TRUE available
#> 2 Median imputation for numeric targets TRUE available
#> 3 Mode imputation for categorical targets TRUE available
#> 4 Median/mode chained baseline TRUE available
#> 5 Bayesian normal-style linear regression draw TRUE available
#> 6 Predictive mean matching TRUE available
#> 7 Single-step predictive mean matching TRUE available
#> 8 Binary logistic regression draw TRUE available
#> 9 One-vs-rest multinomial logistic regression draw TRUE available
#> 10 MissForest-style random forest imputer TRUE available
#> # 15 more rows
describe("imputers")
#> mimar available imputers
#> # basetable: 25 x 10
#> imputer implementation package supports_numeric supports_binary
#> 1 mean mimar internal TRUE TRUE
#> 2 median mimar internal TRUE TRUE
#> 3 mode mimar internal TRUE TRUE
#> 4 naive mimar internal TRUE TRUE
#> 5 norm mimar internal TRUE TRUE
#> 6 pmm mimar internal TRUE TRUE
#> 7 spmm mimar internal TRUE TRUE
#> 8 logreg mimar internal TRUE TRUE
#> 9 polyreg mimar internal TRUE TRUE
#> 10 rf wrapped ranger TRUE TRUE
#> supports_multiclass stochastic
#> 1 TRUE FALSE
#> 2 TRUE FALSE
#> 3 TRUE FALSE
#> 4 TRUE FALSE
#> 5 TRUE TRUE
#> 6 TRUE TRUE
#> 7 TRUE TRUE
#> 8 TRUE TRUE
#> 9 TRUE TRUE
#> 10 TRUE TRUE
#> description available status
#> 1 Mean imputation for numeric targets TRUE available
#> 2 Median imputation for numeric targets TRUE available
#> 3 Mode imputation for categorical targets TRUE available
#> 4 Median/mode chained baseline TRUE available
#> 5 Bayesian normal-style linear regression draw TRUE available
#> 6 Predictive mean matching TRUE available
#> 7 Single-step predictive mean matching TRUE available
#> 8 Binary logistic regression draw TRUE available
#> 9 One-vs-rest multinomial logistic regression draw TRUE available
#> 10 MissForest-style random forest imputer TRUE available
#> # 15 more rows
Core native imputers:
mean,median,modenaive: median/mode chained baselinenorm: linear normal drawpmm,spmm: predictive mean matchinglogreg: binary logistic regression drawpolyreg: one-vs-rest multinomial drawknn: nearest-neighbor donor imputationhotdeck: stochastic donor imputation
Learner-backed imputers:
rf: MissForest-style chained random forest imputer throughrangerranger: random forest throughrangerrpart: tree imputer throughrpartnbayes: naive Bayes throughnaivebayessvm: support vector machine throughe1071bart: Bayesian additive regression trees throughBARTglmnet: penalized regression throughglmnetfastgbm: gradient boosting throughfastgbmxgboost: gradient boosted trees throughxgboostdensemlp: dense multilayer perceptron throughdensemlpfamd: FAMD-assisted donor imputation throughmissMDAsuperlearner,sl: cross-validated Super Learner-style ensemble imputermissknn: whole-table masked k-nearest-neighbor imputation throughmissknn, run in a single pass instead of the chained loop
Imputer names are strict: use the names shown by imputer_registry().
Learner-backed imputers are applied as requested to numeric, binary, and
multiclass targets; mimar does not silently swap them for another
imputer inside benchmark runs.
Parallel imputation
The ncore argument runs independent completed datasets in parallel.
The parallel boundary is the outer imputation index: each completed
dataset gets a deterministic seed offset, so a fixed seed, m,
maxit, and imputer remain reproducible.
i <- impute(a, imputer = "knn", m = 5, maxit = 5, seed = 1, ncore = 2)
Use ncore = 1 for sequential execution, small examples, and the most
conservative behavior in constrained environments.
Tuning imputers
Learner-backed imputers expose their hyperparameters through imputer()
or directly through ... in impute(). Donor-based imputers use the
explicit donors argument.
rf_spec <- imputer("rf", num.trees = 500)
xgb_spec <- imputer("xgboost", nrounds = 100, max_depth = 3)
i1 <- impute(a, imputer = rf_spec, m = 5, maxit = 5, seed = 1)
i2 <- impute(a, imputer = "xgboost", m = 5, maxit = 5, seed = 1,
nrounds = 100, max_depth = 3)
i3 <- impute(a, imputer = "knn", m = 5, maxit = 5, seed = 1, donors = 10)
The same hyperparameter set is reused across all incomplete variables that a given imputer supports, which keeps the full chained-imputation pipeline reproducible and easy to tune.
Super Learner imputation
superlearner combines candidate imputers by cross-validating them on
observed cells, assigning non-negative loss-based weights, and using the
weighted ensemble inside the chained-imputation loop.
sl <- imputer(
"superlearner",
library = c("pmm", "knn", "rpart"),
folds = 5,
metalearner = "inverse_loss"
)
i_sl <- impute(a, imputer = sl, m = 5, maxit = 5, seed = 1)
The short alias imputer = "sl" is equivalent to
imputer = "superlearner".
Diagnostic Plots
plot() methods return ggplot objects. For mimar_imputation
objects, the main diagnostic types are:
plot(i) # imputed cell counts
plot(i, type = "missing") # observed/imputed cell map
plot(i, type = "trace", statistic = "mean") # convergence-screening trace
plot(i, type = "density", variable = "bmi") # line-only density overlays
plot(i, type = "boxplot", variable = "bmi") # observed vs imputation 1:m
plot(i, type = "strip", variable = "bmi") # individual values by imputation
Formula diagnostics are available for bivariate and categorical checks:
plot(i, type = "xy", formula = bmi ~ age | sex)
plot(i, type = "proportion", formula = group ~ sex)
For type = "xy", formulas use y ~ x or y ~ x | group. For
type = "proportion", formulas use categorical_variable ~ strata.
Density diagnostics use line-only overlays so several imputations remain
visible rather than obscuring each other with filled areas.
Chained Imputation Model
Let X be an n x p data frame and let R_ij = 1 when cell (i, j)
is missing. For each incomplete variable X_j:
O_j = {i : R_ij = 0}are the observed rowsM_j = {i : R_ij = 1}are the missing rows
At each chained update, mimar fits an imputer-specific model from the
observed rows and then predicts the missing rows from the current
completed data. In compact form:
fit model on X_-j, O_j -> X_j, O_j
update X_j, M_j using the fitted model
Multiple imputation repeats the same chained procedure m times with
controlled seeds, bootstrap samples of observed rows, and stochastic
prediction where supported.
Learner-backed imputers are practical stochastic update rules inside this chained workflow. They can improve predictive recovery, but users should still inspect trace, distribution, categorical-proportion, and downstream sensitivity diagnostics rather than assuming every learner automatically supplies proper multiple-imputation uncertainty for every analysis.
Algorithm
Input: X, R, h, m, T
Initialize: X~(0) <- init(X)
For k = 1,...,m:
X~_k(0) <- X~(0)
For t = 1,...,T:
For each incomplete variable j:
B_j <- bootstrap sample of O_j
fit h on X~_k, B_j, -j and X_Bj,j
update missing rows M_j using the fitted model
restore observed rows O_j to their original values
Return: {X~_1(T), ..., X~_m(T)}
Evaluation
When imputation is run on an ampute() object, evaluate() uses the
retained truth and scores only artificially removed cells. Numeric
recovery reports RMSE, MAE, bias, and correlation. Categorical recovery
reports accuracy and balanced accuracy.
Imputation with time-to-event data
When a data set carries a follow-up time and an event indicator, the raw survival time must not be used as a predictor for imputing other, partially observed variables: it biases the imputation model. Following White and Royston (2009), represent the outcome with the Nelson-Aalen cumulative hazard plus the event indicator instead.
nelsonaalen() builds that predictor. With no auxiliary variables it
matches mice::nelsonaalen(); aux makes the estimate conditional,
using strata for categorical variables or a Cox model otherwise.
set.seed(1)
n <- 200
surv <- data.frame(
x = rnorm(n),
sex = factor(sample(c("F", "M"), n, TRUE))
)
lp <- 0.4 * surv$x + 0.3 * (surv$sex == "M")
event_time <- rexp(n, rate = exp(lp) / 8)
cens_time <- rexp(n, rate = 1 / 12)
surv$time <- pmin(event_time, cens_time)
surv$status <- as.integer(event_time <= cens_time)
surv$H0 <- nelsonaalen(surv, "time", "status")
surv$H0_bysex <- nelsonaalen(surv, time, status, aux = "sex")
surv$x[sample(n, 40)] <- NA
impute(surv[c("x", "sex", "status", "H0")], m = 5, imputer = "pmm", seed = 1)
#> mimar imputation
#> Completed datasets: 5
#> Rows x columns: 200 x 4
#> Imputer: pmm maxit: 5 ncore: 1 stochastic: TRUE
#> Missing cells imputed: 40 / 40
#> Variables imputed: x
#> Use complete(x) to extract completed data.
ipcw() returns subject-level inverse-probability-of-censoring weights
Delta_i / G(T_i-), where G is the Kaplan-Meier estimate of the
censoring survivor function. Use them in weighted post-imputation
analyses. aux allows covariate-dependent censoring (stratified
Kaplan-Meier, or method = "cox"); type = "all" keeps a positive
weight for censored subjects, and stabilized and truncate control
the weight range.
w_event <- ipcw(surv, "time", "status")
w_all <- ipcw(surv, time, status, aux = "sex", type = "all", truncate = 0.95)
summary(w_event[w_event > 0])
#> Min. 1st Qu. Median Mean 3rd Qu. Max.
#> 1.000 1.116 1.278 1.544 1.523 11.213
summary(w_all)
#> Min. 1st Qu. Median Mean 3rd Qu. Max.
#> 1.000 1.111 1.317 1.580 1.680 3.982
The Cox-based paths use the survival package.
Pooling
pool() combines post-fit quantities estimated separately in each
completed dataset. The statistical target is the quantity itself, not a
data frame. A quantity can be a scalar, coefficient vector,
covariance-aware parameter vector, matrix of survival probabilities, or
a scalar metric. Data frames are accepted only as a tidy adapter for
scalar model output.
For survival-probability matrices, pool_survmat() applies the
complementary log-log transform internally, pools on that scale, and
back-transforms the result.
For scalar quantities with complete-data variance estimates, pool()
applies Rubin-style pooling:
Q_bar = mean(Q_k)
U_bar = mean(U_k)
B = sample variance of Q_k
T = U_bar + (1 + 1/m) * B
results <- data.frame(
term = rep(c("age", "bmi"), each = 3),
estimate = c(0.10, 0.11, 0.09, 0.30, 0.32, 0.29),
std.error = c(0.04, 0.05, 0.04, 0.08, 0.09, 0.08),
imputation = rep(1:3, times = 2)
)
pool(results)
#> mimar pooled results
#> # basetable: 2 x 14
#> term estimate std.error statistic df p.value conf.low conf.high
#> 1 age 0.1000000 0.04509250 2.217664 465.125 0.027060256 0.01138975 0.1886102
#> 2 bmi 0.3033333 0.08530989 3.555664 1094.452 0.000393134 0.13594390 0.4707228
#> m within_variance between_variance total_variance relative_increase_variance
#> 1 3 0.001900000 0.0001000000 0.002033333 0.07017544
#> 2 3 0.006966667 0.0002333333 0.007277778 0.04465710
#> rule
#> 1 rubin
#> 2 rubin
Direct quantity inputs are preferred when available:
pool(c(0.10, 0.11, 0.09), std.error = c(0.04, 0.05, 0.04), name = "age")
#> mimar pooled results
#> # basetable: 1 x 14
#> term estimate std.error statistic df p.value conf.low conf.high m
#> 1 age 0.1 0.0450925 2.217664 465.125 0.02706026 0.01138975 0.1886102 3
#> within_variance between_variance total_variance relative_increase_variance
#> 1 0.0019 1e-04 0.002033333 0.07017544
#> rule
#> 1 rubin
betas <- list(
c(age = 0.10, bmi = 0.30),
c(age = 0.11, bmi = 0.32),
c(age = 0.09, bmi = 0.29)
)
covariances <- list(
diag(c(0.04, 0.08)^2),
diag(c(0.05, 0.09)^2),
diag(c(0.04, 0.08)^2)
)
pool(betas, covariance = covariances)
#> mimar pooled results
#> # basetable: 2 x 14
#> term estimate std.error statistic df p.value conf.low conf.high
#> 1 age 0.1000000 0.04509250 2.217664 465.125 0.027060256 0.01138975 0.1886102
#> 2 bmi 0.3033333 0.08530989 3.555664 1094.452 0.000393134 0.13594390 0.4707228
#> m within_variance between_variance total_variance relative_increase_variance
#> 1 3 0.001900000 0.0001000000 0.002033333 0.07017544
#> 2 3 0.006966667 0.0002333333 0.007277778 0.04465710
#> rule
#> 1 rubin
#> 2 rubin
When no reliable complete-data variance is supplied, as is common for
some performance metrics, pool() reports robust summaries by default:
median, interquartile range, and range across imputations.
Bounded or skewed quantities (ratios, probabilities, AUC, C-index,
correlations) are better pooled on a transformed scale. transform
accepts "log", "logit", "cloglog", or "fisherz"; results are
reported on the original scale. dfcom, the complete-data degrees of
freedom, switches Rubin pooling to the Barnard and Rubin (1999)
small-sample degrees of freedom.
auc <- c(0.78, 0.81, 0.79, 0.80, 0.77)
auc_se <- c(0.030, 0.028, 0.031, 0.029, 0.032)
pool(auc, std.error = auc_se, transform = "logit", name = "AUC", dfcom = 115)
#> mimar pooled results
#> # basetable: 1 x 14
#> term estimate std.error statistic df p.value conf.low conf.high m
#> 1 AUC 0.7903508 0.03466437 6.343285 36.4344 2.30192e-07 0.7115515 0.852098 5
#> within_variance between_variance total_variance relative_increase_variance
#> 1 0.03281624 0.009125116 0.04376638 0.3336805
#> rule
#> 1 rubin
Pooling fitted models
pool_lm(), pool_glm(), pool_coxph(), pool_survreg(),
pool_clogit(), and pool_panglm() pool a list of fitted models, one
per completed dataset, with Rubin’s rules and Barnard-Rubin degrees of
freedom. They print a coefficient table laid out like the matching
summary() output.
d <- mtcars
d$cyl <- factor(d$cyl)
set.seed(1)
d$hp[sample(32, 6)] <- NA
imp <- impute(d, m = 5, imputer = "pmm", seed = 1)
comp <- complete(imp, "all")
fits <- lapply(comp, function(x) lm(mpg ~ cyl + hp + wt, data = x))
pool_lm(fits)
#> Pooled linear model
#> Pooled across m = 5 imputations (Rubin's rules)
#>
#> Estimate Std. Error df t value Pr(>|t|)
#> (Intercept) 35.93629 2.07206 24.60 17.343 2.64e-15 ***
#> cyl6 -3.11550 1.47296 24.55 -2.115 0.044746 *
#> cyl8 -3.19972 2.21279 23.10 -1.446 0.161606
#> hp -0.02217 0.01220 18.98 -1.818 0.084878 .
#> wt -3.25512 0.71952 25.08 -4.524 0.000127 ***
#> ---
#> Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
pool_test() gives pooled multi-parameter tests, for example all levels
of a factor or a model against a nested one: D1 (multivariate Wald), D2
(combined Wald or likelihood-ratio statistics), and D3 (Meng and Rubin
likelihood ratio for lm and glm), each with an F reference
distribution.
fits0 <- lapply(comp, function(x) lm(mpg ~ hp + wt, data = x))
pool_test(fits, fits0, method = "D1")
#> Pooled multi-parameter test: D1 (multivariate Wald)
#> Terms: cyl6, cyl8
#>
#> statistic df1 df2 p.value riv m
#> 2.06 2 244.6 0.1296 0.111 5
pool_test(fits, terms = c("cyl6", "cyl8"), method = "D2")
#> Pooled multi-parameter test: D2 (combined Wald chi-square)
#> Terms: cyl6, cyl8
#>
#> statistic df1 df2 p.value riv m
#> 2.12 2 435 0.1212 0.08447 5
Installation notes
Learner backends are hard dependencies. Installing mimar installs the
packages needed by the registered learner-backed imputers, including
ranger, rpart, naivebayes, e1071, BART, glmnet, fastgbm,
xgboost, densemlp, missknn, and missMDA.
Citation
citation("mimar")
#> To cite mimar in publications, use:
#>
#> El Badisy I (2026). _mimar: Compact Multiple Imputation, Assessment,
#> and Reporting_. R package version 1.4.2,
#> <https://CRAN.R-project.org/package=mimar>.
#>
#> A BibTeX entry for LaTeX users is
#>
#> @Manual{,
#> title = {mimar: Compact Multiple Imputation, Assessment, and Reporting},
#> author = {Imad {El Badisy}},
#> year = {2026},
#> note = {R package version 1.4.2},
#> url = {https://CRAN.R-project.org/package=mimar},
#> }