Statistics seminar - Dr Margherita Silan - University of Padua

A seminar by Dr Margherita Silan from University of Padua

Title: Matching on Partially Ordered Sets for Multiple Treatments to Study Multidimensional Neighbourhood Effects on Health

Abstract: Estimating neighbourhood effects on health from observational data runs into residential self-selection. Individuals sort into areas according to characteristics that also affect their health, so areas are not directly comparable. If each area is treated as a level of a multivalued treatment, the causal problem involves tens to hundreds of groups. Standard propensity score approaches scale poorly in this setting: generalised propensity score models are hard to specify, common support across all groups is hard to establish, and the computational cost grows quickly.

We propose Matching on Poset-based Average Rank for Multiple Treatments (MARMoT), a non-parametric balancing procedure based on partially ordered set (poset) theory. Individuals’ confounder profiles are arranged in a poset according to the relations among them. Each profile is summarised by its average rank across the linear extensions of the poset, which we approximate for computational efficiency. This balancing score is built from covariates alone: it needs neither the outcome nor a model for how individuals end up in each area. Matching is then done on the average rank across all groups at once. A common reference frequency is set for each rank value, and individuals are resampled within a caliper-based tolerance interval. The procedure is implemented in the R package MARMoT, available on CRAN.

A simulation study based on Medicare data comparing 41 hospitals shows that MARMoT gives better covariate balance and lower estimation bias than template matching. The advantage is largest when the number of groups is high. The study also shows how the choice of reference frequency and caliper affects balance.

We further embed MARMoT in a two-stage framework that pairs individual-level balancing with spatial scan statistics. Contiguous clusters of excess risk are identified on the balanced data, with and without area-level covariates. This separates compositional from contextual components of geographic health variation. Applications to hospitalised fractures in Turin neighbourhoods and dementia incidence in Finnish municipalities show that balancing reshapes the identified risk clusters.

Once compositional differences are removed, the remaining question is which features of the area drive the neighbourhood effects. Ongoing work (DIAMOND project) addresses this question by building a multisource, multidimensional indicator of neighbourhood deprivation.

For further information, please contact RSFAS Seminars.

All information collected by the University is governed by the ANU Privacy Policy.