Thread by @olson_km

#JSM2018 Standing room only for this session. With Jorg Dreschler, Raghu, Rod Little, and Don Rubin as presenters, easy to understand why

#JSM2018 Jorg Dreschler up first. Classification systems for industries change frequently, no consistent coding over time. Treat problem as missing data problem.

#JSM2018 Dreschler Has Establishment History Panel from all establishments in Germany with 1+ employee covered by social security. Lots of data on each establishment. 4 different classification systems in use. Two systems are jointly observed in one year.

#JSM2018 Dreschler 95% of the work went into data preparation, but the audience doesn’t find that interesting in a talk.

#JSM2018 Dreschler Had imputation by extrapolation, then deterministic imputation, then probabilistic imputation. Imputed years sequentially, using adjacent coding systems and adjacent years.

#JSM2018 Probabilistic imputation using CART used b/c parametric imputation explodes for categorical variables.

#JSM2018 Dreschler The overview - grow large trees for the imputation; can’t fit one tree to the entire dataset. Run separate CART models by industry code so that each block has up to 20 industry category

#JSM2018 Dreschler Compares CART approach with 2 methods used previously at IAB and in the LEHD. Use difference in cell size between imputated and original data. Marginal distribution of overall industry category similar

#JSM2018 Dreschler detailed Industry codes worst for deterministic approach. Some industries never selected for being imputed - imputing the mode doesn’t work. Probabilistic and CART imputation better reflects distributions.