thomvolker

set.seed(123)

# Number of predictors
P <- 10
# Sample size
N <- 10000
# Generate variance covariance matrix as toeplitz
V <- toeplitz(P:1/P)
# Relative size of coefficients

thomvolker

A slightly modified estimice() and .norm.draw()

Thom Benjamin Volker

04-03-2025

An attempt at a faster estimice() function for the R-package mice.

devtools::load_all(path = "C:/Users/5868777/surfdrive/Documents/mice")

thomvolker

TabPFN in R with reticulate

Generate some example data.

X1 <- runif(200, 0, 10)
X2 <- sin(X1) + rnorm(200, 0, 0.5)
Y <- 3 + 0.5 * X1 + X2 + rnorm(200, 0, 1)

thomvolker

Prediction intervals with missing data

authors: Florian van Leeuwen, Thom Benjamin Volker, Gerko Vink and Stef van Buuren

The development and application of (clinical) prediction models is complicated by missing data, as most analysis techniques do not readily allow for incorporating missing values. Consequently, model parameters cannot be estimated, and predictions cannot be calculated. Ad-hoc fixes to deal with missing data, such as listwise deletion or mean imputation, work only under limited circumstances, such as MCAR, which are unlikely to hold in practice. A more principled approach dealing with missing data is multiple imputation (MI). Many studies confirmed that MI allows one to obtain unbiased and efficient estimates of model parameters under fairly general conditions. Practitioners in (clinical) prediction commonly conceive single imputation to be sufficient. The present study compares single versus multiple imputation for making point estimates (predictions) and prediction interval

thomvolker

thomvolker

Can we get away with single imputation when the goal is to obtain well-calibrated prediction intervals

Of course not! 😁

library(foreach)

set.seed(123)

nsim <- 200

thomvolker

Density ratio estimation is invariant to one-to-one and onto transformations

Thom Benjamin Volker

At least in the univariate case, the multivariate case must be considered in a future gist still.

N <- 10000000
x1 <- rexp(N, 1)

thomvolker

The following code shows how to obtain density ratio estimates that are regularized to $1$ instead of $0$ (or an estimated intercept).

pred_adapt <- function(nu, de, ce, sigma, lambda) {
  Knu <- densityratio::distance(as.matrix(nu), as.matrix(ce), TRUE) |> kernel_gaussian(sigma)
  Kde <- densityratio::distance(as.matrix(de), as.matrix(ce), TRUE) |> kernel_gaussian(sigma)

  Kdede <- crossprod(Kde) / nrow(Kde)
  Knunu <- colMeans(Knu)
 alpha &lt;- solve(Kdede[-1, -1] + lambda * diag(ncol(Kde)-1), Knunu[-1] - Kdede[1, -1])

thomvolker

Divergence-based testing using density ratio estimation techniques

Thom Benjamin Volker

In this notebook, we explore properties of divergence-based two-sample testing, using the methods employed in the densityratio package (Volker et al., 2023). We consider three settings: two groups are drawn from the same distribution, two groups are drawn from distributions with different means, but with the same variance-covariance matrix, and two groups are drawn from distributions with the same means but different covariances. Subsequently, we compare the performance of the

thomvolker

mice.impute.pmm() can give counterintuitive results when using type 1 matching, see for example the following scenario.

The following code is provided by Stef van Buuren, inspired by Templ 2023, Visualization and Imputation of Missing Values.

library(mice)
#> 
#> Attaching package: 'mice'
#> The following object is masked from 'package:stats':
#> 
#>     filter