Principled methods for causal inference in the presence of violations of overlap and consistency
Abstract
AbstractApproved for public release.Classical methods in causal inference from observational data rely on three fundamental assumptions:no unmeasured confounding, positivity or overlap, and consistency. Our last ONR-funded projectand our ongoing work deals with violations of the unmeasured confounding assumption, which isa perennial difficulty in causal inference. The other assumptions, however, are equally importantfor valid inference, and methods for dealing with their violations have received comparatively littleattention in the literature, and in particular in causal machine learning where such violations arefairly common.Positivity violations often occur in settings with covariate shift, domain adaptation, or high-dimensionalcovariates (e.g. high-dimensional networks). Existing work either estimates causal effects where positivitydoes hold and ignores the covariate regions where it does not (trimming, truncating) or developstheory for slower rates of convergence where positivity is nearly violated. In Thrust 1, we propose insteadto use alternative identification strategies: extrapolation of the outcome regression functional,adaption of identification theory methods in the presence of hidden variables, as well combining thesenew approaches with more traditional strategies applied in the region of overlap. These strategieshave never been rigorously formalized to deal with positivity violations, and the theory describingtheirperformance will be a novel and complementary alternative to existing approaches.One way that consistency can fail is when there isinterference, and this has received a lot of attention.But there are many other more neglected types of consistency failures. Consistency fails for #fathand" interventions where the very act of intervening has consequences beyond the intended effectsof the intervention itself. This is the case for most interventions on network structure: interveningon networks ties, brain networks; for the use case of intervening on algorithms or their input tounderstand their inner workings (e.g. LLMs); and for genetic, genomic, and cellular interventions.Consistency also fails when there are multiple versions of treatments with possibly different effects,and this can be especially pernicious if experimental and real world settings necessarily involve differentversions of treatments. This is closely related to the idea that context can matter for causaleffects: whether you are receiving a treatment as part of a clinical trialor as part of routine standardof care could affect its mechanisms. Finally, the noisy labels problem in ML can be seen as a consistencyviolation, as can the problem of inflated 0s (where small levels of a variable are undetectableand measured to be 0; alternatively where missigness status of a binary outcome corresponding todisease status or condition is coded as 1 its absence, e.g. 0). Theseissues occur frequently in bothmedicine and engineering. No framework currently exists for rigorously dealing with all these kindsof consistency violations, and in Thrust 2 we propose to develop principled new methods, describedin more detail below, that will fill that gap.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 09, 2024
- Source ID
- N000142412701
Entities
People
- Elizabeth L Ogburn
Organizations
- Johns Hopkins University
- Office of Naval Research
- United States Navy