multinom_weightit: Fitting (Weighted) Multinomial Regression Models
In WeightIt: Weighting for Covariate Balance in Observational Studies

multinom_weightit

R Documentation

Fitting (Weighted) Multinomial Regression Models

Description

multinom_weightit() fits a multinomial logistic regression model with a covariance matrix that accounts for estimation of weights, if supplied. By default, this function uses M-estimation to construct a robust covariance matrix using the estimating equations for the weighting model and the outcome model when available.

Usage

multinom_weightit(
  formula,
  data,
  link = "logit",
  weightit = NULL,
  vcov = NULL,
  cluster,
  R = 500L,
  offset,
  start = NULL,
  control = list(...),
  x = FALSE,
  y = TRUE,
  contrasts = NULL,
  fwb.args = list(),
  br = FALSE,
  ...
)

Arguments

`formula`	an object of class `formula` (or one that can be coerced to that class): a symbolic description of the model to be fitted.
`data`	a data frame containing the variables in the model. If not found in data, the variables are taken from `environment(formula)`, typically the environment from which the function is called.
`link`	a string corresponding to the desired link function. Currently, only `"logit"` is allowed.
`weightit`	a `weightit` or `weightitMSM` object; the output of a call to `weightit()` or `weightitMSM()`. If not supplied, an unweighted model will be fit.
`vcov`	string; the method used to compute the variance of the estimated parameters. Allowable options include `"asympt"`, which uses the asymptotically correct M-estimation-based method that accounts for estimation of the weights when available; `"const"`, which uses the usual maximum likelihood estimates (only available when `weightit` is not supplied); `"HC0"`, which computes the robust sandwich variance treating weights (if supplied) as fixed; `"BS"`, which uses the traditional bootstrap (including re-estimation of the weights, if supplied); `"FWB"`, which uses the fractional weighted bootstrap as implemented in \pkgfunfwbfwb (including re-estimation of the weights, if supplied); and `"none"` to omit calculation of a variance matrix. If `NULL` (the default), will use `"asympt"` if `weightit` is supplied and M-estimation is available and `"HC0"` otherwise. See the `vcov_type` component of the outcome object to see which was used.
`cluster`	optional; for computing a cluster-robust variance matrix, a variable indicating the clustering of observations, a list (or data frame) thereof, or a one-sided formula specifying which variable(s) from the fitted model should be used. Note the cluster-robust variance matrix uses a correction for small samples, as is done in `sandwich::vcovCL()` by default. Cluster-robust variance calculations are available only when `vcov` is `"asympt"`, `"HC0"`, `"BS"`, or `"FWB"`.
`R`	the number of bootstrap replications when `vcov` is `"BS"` or `"FWB"`. Default is 500. Ignored otherwise.
`offset`	optional; a numeric vector containing the model offset. See `offset()`. An offset can also be present in the model formula.
`start`	optional starting values for the coefficients.
`control`	a list of parameters for controlling the fitting process.
`x`, `y`	logical values indicating whether the response vector and model matrix used in the fitting process should be returned as components of the returned value.
`contrasts`	an optional list defining contrasts for factor variables. See `model.matrix()`.
`fwb.args`	an optional list of further arguments to supply to \pkgfunfwbfwb when `vcov = "FWB"`.
`br`	`logical`; whether to use mean bias reduction, i.e., to solve the bias-reducing adjusted score equations of Firth (1993) rather than the score equations. This yields estimates with smaller asymptotic bias that are always finite, even when the maximum likelihood estimates are not (e.g., under separation). Default is `FALSE`. See Details.
`...`	arguments to be used to form the default control argument if it is not supplied directly.

Details

multinom_weightit() implements multinomial logistic regression using a custom function in WeightIt that optionally computes a coefficient variance matrix that can be adjusted to account for estimation of the weights if a weightit or weightitMSM object is supplied to the weightit argument. This implementation is less robust to failures than other multinomial logistic regression solvers and should be used with caution. Unless br = TRUE, estimation of coefficients should align with that from mlogit::mlogit() and mclogit::mblogit() but might differ from nnet::multinom() due to the relaxed convergence thresholds of the latter.

When no argument is supplied to weightit or there is no "Mparts" attribute in the supplied object, the default variance matrix returned will be the "HC0" sandwich variance matrix, which is robust to misspecification of the outcome family (including heteroscedasticity). Otherwise, the default variance matrix uses M-estimation to additionally adjust for estimation of the weights. When possible, this often yields smaller (and more accurate) standard errors. See the individual methods pages to see whether and when an "Mparts" attribute is included in the supplied object. To request that a variance matrix be computed that doesn't account for estimation of the weights even when a compatible weightit object is supplied, set vcov = "HC0", which treats the weights as fixed.

Bootstrapping can also be used to compute the coefficient variance matrix; when vcov = "BS" or vcov = "FWB", which implement the traditional resampling-based and fractional weighted bootstrap, respectively, the entire process of estimating the weights and fitting the outcome model is repeated in bootstrap samples (if a weightit object is supplied). This accounts for estimation of the weights and can be used with any weighting method. It is important to set a seed using set.seed() to ensure replicability of the results. The fractional weighted bootstrap is more reliable but requires the weighting method to accept sampling weights (which most do, and you'll get an error if it doesn't). Setting vcov = "FWB" and supplying fwb.args = list(wtype = "multinom") also performs the resampling-based bootstrap but with the additional features fwb provides (e.g., a progress bar and parallelization).

Bias reduction

When br = TRUE, the coefficients solve the bias-reducing adjusted score equations of Firth (1993) instead of the score equations, which removes the first-order term in the asymptotic bias of the estimates. Because the multinomial logistic regression model uses the canonical link, the adjusted score function is the gradient of the log-likelihood penalized by the Jeffreys invariant prior, \log|F(\beta)|/2, where F(\beta) is the expected information matrix (Kosmidis & Firth, 2011). Consequently, the estimates are always finite and correspond to the maximum of a penalized likelihood. Estimation should align with that from \pkgfunbrglm2brmultinom with its default type = "AS_mean".

Weights are treated as multinomial totals, as they are by brglm2, which makes the estimates invariant to whether the data are supplied as individual units or as groups of identical units with weights equal to their counts. As for br = TRUE in glm_weightit(), the reported variance matrix uses the information matrix at the estimates rather than the Jacobian of the adjusted score, i.e., the adjustment is treated as fixed; the two differ by a term that vanishes asymptotically. M-estimation and bootstrapping can be used with br = TRUE just as they can without it.

Value

A multinom_weightit object.

Unless vcov = "none", the vcov component contains the covariance matrix adjusted for the estimation of the weights if requested and a compatible weightit object was supplied. The vcov_type component contains the type of variance matrix requested. If cluster is supplied, it will be stored in the "cluster" attribute of the output object, even if not used.

The model component of the output object (also the model.frame() output) will include two extra columns when weightit is supplied: (weights) containing the weights used in the model (the product of the estimated weights and the sampling weights, if any) and (s.weights) containing the sampling weights, which will be 1 if s.weights is not supplied in the original weightit() call.

References

Firth, D. (1993). Bias reduction of maximum likelihood estimates. Biometrika, 80(1), 27–38. \Sexpr[results=rd]{tools:::Rd_expr_doi("10.1093/biomet/80.1.27")}

Kosmidis, I., & Firth, D. (2011). Multinomial logit bias reduction via the Poisson log-linear model. Biometrika, 98(3), 755–759. \Sexpr[results=rd]{tools:::Rd_expr_doi("10.1093/biomet/asr026")}

Examples

data("lalonde", package = "cobalt")

# Logistic regression ATT weights
w.out <- weightit(treat ~ age + educ + married + re74,
                  data = lalonde, method = "glm",
                  estimand = "ATT")

# Multinomial logistic regression outcome model
# that adjusts for estimation of weights
lalonde$re78_3 <- factor(findInterval(lalonde$re78,
                                      c(0, 5e3, 1e4)))

fit <- multinom_weightit(re78_3 ~ treat,
                         data = lalonde,
                         weightit = w.out)

summary(fit)

# Same model using mean bias reduction
fit_br <- multinom_weightit(re78_3 ~ treat,
                            data = lalonde,
                            weightit = w.out,
                            br = TRUE)

summary(fit_br)

WeightIt documentation built on Aug. 4, 2026, 1:09 a.m.