Causal bounds and instruments
Instrumental variables have proven useful, in particular within the social sciences and economics, for making inference about the causal effect of a random variable, B, on another random variable, C, in the presence of unobserved confounders. In the case where relationships are linear, causal effect...
| 主要作者: | |
|---|---|
| 格式: | Journal article |
| 語言: | English |
| 出版: |
Association for Uncertainty in Artificial Intelligence (AUAI) Press
2007
|
| 主題: |
| _version_ | 1826301056498270208 |
|---|---|
| author | Ramsahai, R |
| author_facet | Ramsahai, R |
| author_sort | Ramsahai, R |
| collection | OXFORD |
| description | Instrumental variables have proven useful, in particular within the social sciences and economics, for making inference about the causal effect of a random variable, B, on another random variable, C, in the presence of unobserved confounders. In the case where relationships are linear, causal effects can be identified exactly from studying the regression of C on A and the regression of B on A, where A is the instrument. In the more general case, bounds have been developed in the literature for the causal effect of B on C, given observational data on the joint distribution of C, B, and A. Using an approach based on the analysis of convex polytopes, we develop bounds for the same causal effect when given data on (C, A) and (B, A) only. The bounds developed are thus in direct analogy to the standard use of instruments in econometrics, but we make no assumption of linearity. Use of the bounds is illustrated for experiments with partial compliance. The bounds are, for example, relevant in genetic epidemiology, where the 'Mendelian instrument' S represents a genotype, and where joint data on all of C, B, and A may rarely be available but studies involving pairs of these may be abundant. Other examples of bounding causal effects are considered to show that the method applies to DAGs in general, subject to certain conditions. |
| first_indexed | 2024-03-07T05:26:35Z |
| format | Journal article |
| id | oxford-uuid:e0c52783-3f19-4aa9-b101-5e22a4527e9d |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-03-07T05:26:35Z |
| publishDate | 2007 |
| publisher | Association for Uncertainty in Artificial Intelligence (AUAI) Press |
| record_format | dspace |
| spelling | oxford-uuid:e0c52783-3f19-4aa9-b101-5e22a4527e9d2022-03-27T09:49:43ZCausal bounds and instrumentsJournal articlehttp://purl.org/coar/resource_type/c_dcae04bcuuid:e0c52783-3f19-4aa9-b101-5e22a4527e9dArtificial IntelligenceStatisticsEnglishOxford University Research Archive - ValetAssociation for Uncertainty in Artificial Intelligence (AUAI) Press2007Ramsahai, RInstrumental variables have proven useful, in particular within the social sciences and economics, for making inference about the causal effect of a random variable, B, on another random variable, C, in the presence of unobserved confounders. In the case where relationships are linear, causal effects can be identified exactly from studying the regression of C on A and the regression of B on A, where A is the instrument. In the more general case, bounds have been developed in the literature for the causal effect of B on C, given observational data on the joint distribution of C, B, and A. Using an approach based on the analysis of convex polytopes, we develop bounds for the same causal effect when given data on (C, A) and (B, A) only. The bounds developed are thus in direct analogy to the standard use of instruments in econometrics, but we make no assumption of linearity. Use of the bounds is illustrated for experiments with partial compliance. The bounds are, for example, relevant in genetic epidemiology, where the 'Mendelian instrument' S represents a genotype, and where joint data on all of C, B, and A may rarely be available but studies involving pairs of these may be abundant. Other examples of bounding causal effects are considered to show that the method applies to DAGs in general, subject to certain conditions. |
| spellingShingle | Artificial Intelligence Statistics Ramsahai, R Causal bounds and instruments |
| title | Causal bounds and instruments |
| title_full | Causal bounds and instruments |
| title_fullStr | Causal bounds and instruments |
| title_full_unstemmed | Causal bounds and instruments |
| title_short | Causal bounds and instruments |
| title_sort | causal bounds and instruments |
| topic | Artificial Intelligence Statistics |
| work_keys_str_mv | AT ramsahair causalboundsandinstruments |