| Title: | Tools for Analysis of Stacked Multiple Imputations |
|---|---|
| Description: | Provides methods for inference using stacked multiple imputations augmented with weights. The vignette provides example R code for implementation in general multiple imputation settings. For additional details about the estimation algorithm, we refer the reader to Beesley, Lauren J and Taylor, Jeremy M G (2020) “A stacked approach for chained equations multiple imputation incorporating the substantive model” <doi:10.1111/biom.13372>, and Beesley, Lauren J and Taylor, Jeremy M G (2021) “Accounting for not-at-random missingness through imputation stacking” <arXiv:2101.07954>. |
| Authors: | Lauren Beesley [aut], Mike Kleinsasser [cre] |
| Maintainer: | Mike Kleinsasser <[email protected]> |
| License: | GPL-2 |
| Version: | 0.1.0 |
| Built: | 2024-11-12 03:40:36 UTC |
| Source: | https://github.com/cran/StackImpute |
This function takes a dataset with stacked multiple imputation and a model fit and applies bootstrap to estimate the covariance matrix accounting for imputation uncertainty.
Bootstrap_Variance(fit, stack, M, n_boot = 100)Bootstrap_Variance(fit, stack, M, n_boot = 100)
fit |
object with corresponding vcov method (e.g. glm, coxph, survreg, etc.) from fitting to the (weighted) stacked dataset |
stack |
data frame containing stacked dataset across multiple imputations. Could have 1 or M rows for each subject with complete data. Should have M rows for each subject with imputed data. Must contain the following named columns: (1) stack$.id, which correspond to a unique identifier for each subject. This column can be easily output from MICE. (2) stack$wt, which corresponds to weights assigned to each row. Standard analysis of stacked multiple imputations should set these weights to 1 over the number of times the subject appears in the stack. (3) stack$.imp, which indicates the multiply imputed dataset (from 1 to M). This column can be easily output from MICE. |
M |
number of multiple imputations |
n_boot |
number of bootstrap samples |
This function implements the bootstrap-based estimation method for stacked multiple imputations proposed by Dr. Paul Bernhardt in “A Comparison of Stacked and Pooled Multiple Imputation" at the Joint Statistical Meetings, 2019.
Variance, estimated covariance matrix accounting for within and between imputation variation
data(stackExample) fit = stackExample$fit stack = stackExample$stack bootcovar = Bootstrap_Variance(fit, stack, M = 5, n_boot = 10) VARIANCE_boot = diag(bootcovar)data(stackExample) fit = stackExample$fit stack = stackExample$stack bootcovar = Bootstrap_Variance(fit, stack, M = 5, n_boot = 10) VARIANCE_boot = diag(bootcovar)
This function is called internal to Bootstrap_Variance and re-estimates glm model parameters
func.boot(data, indices)func.boot(data, indices)
data |
matrix with indices of possible imputed datasets to sample |
indices |
sampled indices |
numeric vector of parameter coefficients
This function is internal to Jackknife_Variance. This estimates model parameters using a subset of the stacked data.
func.jack(leaveout, stack)func.jack(leaveout, stack)
leaveout |
indexes the multiple imputation being excluded from estimation |
stack |
data frame containing stacked dataset across multiple imputations. Could have 1 or M rows for each subject with complete data. Should have M rows for each subject with imputed data. Must contain the following named columns: (1) stack$.id, which correspond to a unique identifier for each subject. This column can be easily output from MICE. (2) stack$wt, which corresponds to weights assigned to each row. Standard analysis of stacked multiple imputations should set these weights to 1 over the number of times the subject appears in the stack. (3) stack$.imp, which indicates the multiply imputed dataset (from 1 to M). This column can be easily output from MICE. |
numeric vector of parameter coefficients
The goal of this function is to estimate the glm dispersion parameter using data across imputed datasets while correctly accounting for the weights.
glm.weighted.dispersion(fit)glm.weighted.dispersion(fit)
fit |
an object of class glm |
an estimate of the glm dispersion parameter
data(stackExample) glm.weighted.dispersion(stackExample$fit)data(stackExample) glm.weighted.dispersion(stackExample$fit)
This function takes a dataset with stacked multiple imputation and a model fit and applies jackknife to estimate the covariance matrix accounting for imputation uncertainty.
Jackknife_Variance(fit, stack, M)Jackknife_Variance(fit, stack, M)
fit |
object with corresponding vcov method (e.g. glm, coxph, survreg, etc.) from fitting to the (weighted) stacked dataset |
stack |
data frame containing stacked dataset across multiple imputations. Could have 1 or M rows for each subject with complete data. Should have M rows for each subject with imputed data. Must contain the following named columns: (1) stack$.id, which correspond to a unique identifier for each subject. This column can be easily output from MICE. (2) stack$wt, which corresponds to weights assigned to each row. Standard analysis of stacked multiple imputations should set these weights to 1 over the number of times the subject appears in the stack. (3) stack$.imp, which indicates the multiply imputed dataset (from 1 to M). This column can be easily output from MICE. |
M |
number of multiple imputations |
This function implements the jackknife-based estimation method for stacked multiple imputations proposed by Beesley and Taylor (2021).
Variance, estimated covariance matrix accounting for within and between imputation variation
data(stackExample) fit = stackExample$fit stack = stackExample$stack jackcovar = Jackknife_Variance(fit, stack, M = 5) VARIANCE_jack = diag(jackcovar)data(stackExample) fit = stackExample$fit stack = stackExample$stack jackcovar = Jackknife_Variance(fit, stack, M = 5) VARIANCE_jack = diag(jackcovar)
This function takes a dataset with stacked multiple imputations and a glm or coxph fit and estimates the corresponding information matrix accounting for the imputation uncertainty.
Louis_Information(fit, stack, M, IMPUTED = NULL)Louis_Information(fit, stack, M, IMPUTED = NULL)
fit |
object of class glm or coxph from fitting to the (weighted) stacked dataset |
stack |
data frame containing stacked dataset across multiple imputations. Could have 1 or M rows for each subject with complete data. Should have M rows for each subject with imputed data. Must contain the following named columns: (1) stack$.id, which correspond to a unique identifier for each subject. This column can be easily output from MICE. (2) stack$wt, which corresponds to weights assigned to each row. Standard analysis of stacked multiple imputations should set these weights to 1 over the number of times the subject appears in the stack. |
M |
number of multiple imputations |
IMPUTED |
deprecated parameter, not used in current version |
This function uses the observed information matrix principle proposed in Louis (1982) and applied to imputations in Wei and Tanner (1990). This estimator is a further extension specifically designed for analyzing stacks of multiply imputed data as proposed in Beesley and Taylor (2019) https://arxiv.org/abs/1910.04625.
Info, estimated information matrix accounting for within and between imputation variation
data(stackExample) Info = Louis_Information(stackExample$fit, stackExample$stack, M = 50) VARIANCE = diag(solve(Info))data(stackExample) Info = Louis_Information(stackExample$fit, stackExample$stack, M = 50) VARIANCE = diag(solve(Info))
This function takes a dataset with stacked multiple imputations and a score matrix and covariance matrix from stacked and weighted analysis as inputs to estimates the corresponding information matrix accounting for the imputation uncertainty.
Louis_Information_Custom(score, covariance_weighted, stack, M)Louis_Information_Custom(score, covariance_weighted, stack, M)
score |
n x p matrix containing the contribution to the outcome model score matrix for each subject (n rows) and each model parameter (p columns). |
covariance_weighted |
p x p matrix containing the estimated covariance matrix from fitting the desired model to the stacked and weighted multiple imputations. Note: For GLM models, use summary(fit)$cov.unscaled*StackImpute::glm.weighted.dispersion(fit) as the default dispersion parameter will be incorrect. |
stack |
data frame containing stacked dataset across multiple imputations. Could have 1 or M rows for each subject with complete data. Should have M rows for each subject with imputed data. Must contain the following named columns: (1) stack$.id, which correspond to a unique identifier for each subject. This column can be easily output from MICE. (2) stack$wt, which corresponds to weights assigned to each row. Standard analysis of stacked multiple imputations should set these weights to 1 over the number of times the subject appears in the stack. |
M |
number of multiple imputations |
This function uses the observed information matrix principle proposed in Louis (1982) and applied to imputations in Wei and Tanner (1990). This estimator is a further extension specifically designed for analyzing stacks of multiply imputed data as proposed in Beesley and Taylor (2019) https://arxiv.org/abs/1910.04625.
Info, estimated information matrix accounting for within and between imputation variation
data(stackExample) fit = stackExample$fit stack = stackExample$stack covariates = as.matrix(cbind(1, stack$X, stack$B)) score = sweep(covariates, 1, stack$Y - covariates %*% matrix(coef(fit)), '*') / glm.weighted.dispersion(fit) covariance_weighted = summary(fit)$cov.unscaled * glm.weighted.dispersion(fit) Info = Louis_Information_Custom(score, covariance_weighted, stack, M = 50) VARIANCE_custom = diag(solve(Info))data(stackExample) fit = stackExample$fit stack = stackExample$stack covariates = as.matrix(cbind(1, stack$X, stack$B)) score = sweep(covariates, 1, stack$Y - covariates %*% matrix(coef(fit)), '*') / glm.weighted.dispersion(fit) covariance_weighted = summary(fit)$cov.unscaled * glm.weighted.dispersion(fit) Info = Louis_Information_Custom(score, covariance_weighted, stack, M = 50) VARIANCE_custom = diag(solve(Info))
Function for updating a model fit using either new data or a new model structure
my_update(mod, formula = NULL, data = NULL, weights = NULL)my_update(mod, formula = NULL, data = NULL, weights = NULL)
mod |
object of class 'glm' or 'coxph' |
formula |
formula for updated model fit, default = no change |
data |
data used for updated model fit, default = no change |
weights |
weights used for updated model fit, default = no change |
the updated model fit object of the same class as the given model
Example data set for Louis_Information()
a list with
fit glm fit from vignette example
stack stacked imputed data sets from vignette example