Causal analysis and machine learning
Lead Research Organisation:
University of Bristol
Department Name: Economics
Abstract
Recent developments have integrated techniques from machine learning, like the Lasso, into the econometric analysis of causal effects estimation. This project will analyse the conditions needed for these method to produce reliable estimates, both for models where selection is on observables as for models where selection is on unobservables, including the case where not all instruments are valid. The methods will be applied to causal effect estimation of various exposures on various outcomes, using cross-section and panel data, from e.g. understanding society, but also from Biobank.
Organisations
People |
ORCID iD |
| Frank Windmeijer (Primary Supervisor) | |
| Xiaoran Liang (Student) |
Studentship Projects
| Project Reference | Relationship | Related To | Start | End | Student Name |
|---|---|---|---|---|---|
| ES/P000630/1 | 01/10/2017 | 30/09/2027 | |||
| 1925904 | Studentship | ES/P000630/1 | 01/10/2017 | 30/01/2021 | Xiaoran Liang |
| Description | Instrumental variables (IV) estimation is a well-established method for determining and estimating the causal effect of the exposure/treatment on the outcome when their relationship is confounded by unobserved factors. Our reaearch is concerned with the situation where there are a large number of potential IVs, but some of them might be invalid in the sense that they may have direct effect on the outcome, or affect the outcome through unobservabled factors. Without prior knowledge about which ones of the candidate IVs are valid, we have gained new insights into the problem and developed various new methods to select the valid IVs and estimate the causal effect. Our Confidence Interval method can achieve consistent selection and oracle estimation, and has better finite sample performance than previous methods. We also developed an adaptive Lasso method to deal with the IV selection problem when there are two endogenous variables, which is the first study to do so. We are working on a new method, the clustering method, to deal with the IV selection problem in a more general way that we may be able to take weak instruments and heterogenous treatment effects into account. Our methods have been applied to Mendelian Randomization studies. |
| Exploitation Route | Our methods can be used for general IV selection, especailly Mendelian Randomization studies and demand estimation when there are a large number of candidate IVs whose validity might be in doubt. |
| Sectors | Financial Services, and Management Consultancy,Healthcare |