Seminar by Dr Kevin Credit (Maynooth University) – May 21, 2025

We kindly invite you to attend a lecture titled “Modelling spatial treatment effects using causal machine learning".

The event will take place at 15:00 in Room F at the Faculty of Economic Sciences and online via the Zoom platform. (The speaker will present remotely.)

Link to the meeting: https://uw-edu-pl.zoom.us/j/91231110629?pwd=2QG0mL4tgyya17LMgAAs07wErC3hlx.1

[Meeting ID: 912 3111 0629
Passcode: 545309]

Below, we present the abstract of the presentation.

---
Over the past ten years, we have seen rapid advancement in the study of causal inference in two areas that are particularly interesting for spatial data scientists, econometricians, and geographers: 1) the development of methods of causal inference that include and account for spatial data structures (“spatial-causal inference”) and 2) leveraging machine learning (ML) algorithms for estimating heterogenous causal effects (“causal ML”). Building on prior work (Credit & Lehnert, 2023), this paper seeks to further integrate spatial ways of thinking into causal ML by testing the use of different operationalisations of spatial treatment effects in two ML algorithms, the causal forest and the T-learner. The basic idea is that the kind of “treatments” often analysed in urban planning and policy contexts – such as location within a given distance buffer of a new infrastructure investment (Credit, 2018) – likely carry significant spatial spillovers that are not accounted for in standard (non-spatial) causal modelling frameworks, including non-ML approaches such as difference-in-differences (DID). The paper’s hypothesis – which is confirmed by its analysis of the impact of developing a new urban pedestrian/cycling trail in Chicago on nearby neighbourhood-level residential construction – is that accounting for these treatment spillovers reduces the size (and standard error) of the estimated treatment effects, which may have important implications for the conclusions that can be drawn from studies like this. Overall, the use of a continuous distance-decay treatment parameter – rather than a binary treatment/control cutoff – appears to perform best. To conduct this test for both methods, the paper also develops a novel approach for estimating conditional average treatment effects (CATE) from a continuous treatment parameter using the T-learner.