64th ISI World Statistics Congress - Ottawa, Canada

64th ISI World Statistics Congress - Ottawa, Canada

Advanced inference on mixed effects models for SAE


Prof. stefan sperlich


  • MP
    Monica Pratesi

  • SS
    PROF. DR. stefan sperlich
  • Simultaneous and Comparative Statistics for Small Areas

  • NT
    PROF. DR. Nikolaos Tzavidis
  • Small area estimation using random forests and mixed effects random forests with applications to poverty mapping

  • TK
    Prof. Tatyana Krivobokova
  • Uniformly valid inference based on the lasso in linear mixed models

  • DN
    David Newhouse

  • Category: International Association of Survey Statisticians (IASS)


    As outlined in the description, we focus on recent advances in complex inference for Mixed Effects Models (MEM) as these are extremely popular in many statistical domains. In this session we concentrate on new challenges for using MEM for small area estimation (SAE, where "small areas" may refer to any kind of clusters, not necessarily geographical ones), a statistics domain in which MEMs are particularly popular due to their prediction power.
    More specifically, we consider inference problems that arise when using MEM for small area predictors (a) to do comparative or joint statistics for various or all areas simultaneously, (b) when looking at estimates resulting from model selection, or (c) predictors resulting from highly nonlinear methods like random forests with mixed effects. It is easy to see that the existing classic, commonly used methods fail to work then, or at least don't offer what the practitioner expects them to do. Examples are that e.g. classical prediction intervals for SAE parameters were not made for any of the above problems: not for uniform or simultaneous infernce, not for post-selection inference, not for highly non-linear predictors. They are neither made for doing comparative analysis between small areas, but SAE-based resource-allocation implies exactly this. Morever, focusing on particularly poor or polluted small areas requires joint conditional inference, a topic that so far has largely been ignored. Likelwise, more complex data structures require highly nonlinear methods like random forests that would in addition allow for the inclusion of qualitative data.
    Over the last four to five years, the invited speakers have been working exactly on these particular inference problems, i.e. uniform, post-selection, conditional, and highly non-linear inference for MEM. This session is intended to give an overview of the recent advances of such complex inference problems that practitioners are more and more frequently facing in small area estimation (in particular in official and environmental statistics, and especially for the new requirements of providing SDG-indicators on highly disaggregated levels).
    Prof Tatyana Krivobokova will talk about her findings on conditional and post-LASSO inference in mixed models, Prof. Nikos Tzavidis will present his research on doing inference with mixed effects Random Forests, and Prof. Stefan Sperlich will talk about uniform and simultaneous inference for SAE. Dr. David Newhouse, Senior Economist at the World Bank, will discuss these advances from the practitioners’ point(s) of view, and Prof. Monica Pratesi, an internationally expert for small area estimation in practice and theory, acts as Session Chair.