Thread by @WomenInStat

Missing Data, a thread ⬇️
Missing values are everywhere! We have listed more than 150 R packages in cran.r-project.org/web/views/Miss…
So let us give few pointers:
The method of handling missing data depends on the purpose of the analysis: estimation, completion, prediction, etc.

1) For inference with missing values, estimating as well as possible a parameter and giving a confidence interval, consider likelihood approaches (using EM algorithms) or multiple imputation

2) Single Imputation/Matrix completion aims at completing (predicting the missing entries) a dataset as best as possible. Multiple imputation aims at estimating parameters and their variability, taking into account the uncertainty due to missing values

3) Imputing by the mean is the worst thing you can for inference but can be acceptable for prediction with missing data in the covariates (with lots of data, powerful learner and imputing train and test by the same constant) - see, arxiv.org/abs/1902.06931

4.1) Should I delete a variable with >40% missing values? NO! Many of my clinical colleagues use this rule. It is not only the percentage that counts, but also the structure of the data

4.2) Imagine data with all variables perfectly correlated, even with 80% missing data, you can perfectly predict the missing values from the observed values. On the other hand, with very poorly correlated variables, even a small percentage of missing data can be problematic

5) Give me an incomplete data, I easily output a completed data (> 100 imputation methods); properties depend on the imputation model used: Impute with random forest, you can not extrapolate, Impute with linear models, you can not consider non linear relationships, etc..

6) What confidence should be given to an analysis performed from incomplete data ? Multiple Imputation can help you assessing the variability+ As in any data analysis, you should also consider descriptive statistics

7) Missing data in causal inference: be careful, results very sensitive to the identifiability assumptions and the distribution of missing values. For pipelines in R to estimate ATE with IPW, AIPW see my former PhD student @imkemay ‘s notebooks on Rmistatic

@imkemay 8) Missing data in linear and logistic regression: misaem R package implements estimation of the regression coefficients and prediction with missing values - for now only with continuous covariates and MAR, MCAR data for those familiar.

@imkemay 9) Missing data in random forests: Everyone says that trees handle missing data, but not all methods are equivalent! We recommend for prediction missing incorporated in attributes, implemented in the R package grf.

@imkemay 10) Missing data in PCA: visualisation with the R package missMDA, also for categorical variables, group of variables, etc.

@imkemay Amazing women developing methods for missing values: @madeleineudel (low rank methods), @raziehnabi (identifiability with graphical models), Shu Wang (weighting, multiple imputation, fractional imputation),@LeyClem (causal inference),

@imkemay @raziehnabi @LeyClem @N_Erler (longitudinal data, bayesian approach and amazing teaching resources), Marine Le Morvan (Supervised learning, neural nets) and of course so many others.

@imkemay @raziehnabi @LeyClem @N_Erler Citations:
“One of the ironie of big data is that missing values play an even more significant role” (arxiv.org/abs/1906.12125).
“The idea of imputation is both seductive and dangerous (Dempster and Rubin, 1983)”

@imkemay @raziehnabi @LeyClem @N_Erler It was first elements, questions I am often asked and methods I use. Feel free to reach out as we develop tools for users and we are glad if it may help other scientists analyzing their data.

@imkemay @raziehnabi @LeyClem @N_Erler Today I present my project of a joint team, PreMeDICaL (Precision Medicine by Data Integration and Causal Learning), between @inria_sophia and @Inserm (IDESP). We are keeping our fingers crossed. See you tomorrow.

@imkemay @raziehnabi @LeyClem sorry typo: @madeleineudell