Résultats de recherche pour « Data Imputation »

By revisiting imputation from the modern perspective of missing data graphs, we correct common guidance about which auxiliary variables should be included in an imputation model. We propose a generalized definition of missingness at random (MAR), called “z-MAR'', which makes explicit the set of variables to be analyzed and a separate set of auxiliary variables that are included in the imputation model, but not analyzed. We provide a graphical equivalent of z-MAR that we call the “m-backdoor criterion”. In a sense we formalize, a standard imputation model trained on complete cases is valid for a given analysis if and only if the m-backdoor criterion holds. As the criterion indicates, the standard recommendations to use all available auxiliary variables, or all that are associated with missingness status, can lead to invalid imputation models and biased estimates. This bias arises from collider stratification and can occur even with non-causal estimands. Instead, the set of auxiliary variables should be restricted to those that affect incomplete variables or missingness indicators. These auxiliary variables will always suffice to have a valid imputation model, if such a set does exist among the candidate auxiliary variables. Applying this result does not require full knowledge of the graph.

By revisiting imputation from the modern perspective of missing data graphs, we correct common guidance about which auxiliary variables should be included in an imputation model. We propose a generalized definition of missing at random (MAR), termed z-MAR, which distinguishes analysis variables from auxiliary variables included only in the imputation model. We derive a corresponding graphical condition, the m-backdoor criterion, which characterizes when a broad class of MAR-based imputation models yields valid inferences. In a sense we formalize, the m-backdoor criterion is both necessary and sufficient for valid imputation. Whereas common guidance is to include all available auxiliary variables in the imputation model, our results show that some auxiliary variables (e.g., colliders) can exacerbate bias. Thus, the modeled auxiliary variables should be restricted, at minimum, to those that causally affect either the incomplete variables or the missingness indicators. This set will always suffice to ensure valid imputation, if any sufficient set does exist among the candidate auxiliary variables. Applying this result does not require full knowledge of the graph. We demonstrate these results through simulations and an empirical application to the Stockholm Public Health Cohort, where including a presumed collider in the imputation model substantially altered estimates.

Abstract This paper presents a stochastic imputation approach for large datasets using a correlation selection methodology when preferred commercial packages struggle to iterate due to numerical problems. A variable range-based guard rail modification is proposed that benefits the convergence rate of data elements while simultaneously providing increased confidence in the plausibility of the imputations. A large country conflict dataset motivates the search to impute missing values well over a common threshold of 20% missingness. The Multicollinearity Applied Stepwise Stochastic imputation methodology (MASS-impute) capitalizes on correlation between variables within the dataset and uses model residuals to estimate unknown values. Examination of the methodology provides insight toward choosing linear or nonlinear modeling terms. Tailorable tolerances exploit residual information to fit each data element. The methodology evaluation includes observing computation time, model fit, and the comparison of known values to replaced values created through imputation. Overall, the methodology provides useable and defendable results in imputing missing elements of a country conflict dataset.