Text and Causal Inference: A Review of Using Text to Remove Confounding\n from Causal Estimates
Katherine A. Keith, David Jensen, Brendan O’Connor
Abstract
Open-access reader
Katherine A. Keith, David Jensen, Brendan O’Connor
Abstract
Open-access reader
Many applications of computational social science aim to infer causal\nconclusions from non-experimental data. Such observational data often contains\nconfounders, variables that influence both potential causes and potential\neffects. Unmeasured or latent confounders can bias causal estimates, and this\nhas motivated interest in measuring potential confounders from observed text.\nFor example, an individual's entire history of social media posts or the\ncontent of a news article could provide a rich measurement of multiple\nconfounders. Yet, methods and applications for this problem are scattered\nacross different communities and evaluation practices are inconsistent. This\nreview is the first to gather and categorize these examples and provide a guide\nto data-processing and evaluation decisions. Despite increased attention on\nadjusting for confounding using text, there are still many open problems, which\nwe highlight in this paper.\n
A significance statement is not available in the OpenAlex record.
A contribution statement is not available in the OpenAlex record.
Method details are not available in the OpenAlex metadata.
Findings are not separately available in the OpenAlex metadata.
Limitations are not available in the OpenAlex metadata.
Application details are not available in the OpenAlex metadata.
Many applications of computational social science aim to infer causal\nconclusions from non-experimental data. Such observational data often contains\nconfounders, variables that influence both potential causes and potential\neffects. Unmeasured or latent confounders can bias causal estimates, and this\nhas motivated interest in measuring potential confounders from observed text.\nFor example, an individual's entire history of social media posts or the\ncontent of a news article could provide a rich measurement of multiple\nconfounders. Yet, methods and applications for this problem are scattered\nacross different communities and evaluation practices are inconsistent. This\nreview is the first to gather and categorize these examples and provide a guide\nto data-processing and evaluation decisions. Despite increased attention on\nadjusting for confounding using text, there are still many open problems, which\nwe highlight in this paper.\n
Key concepts: Confounding, Causal inference, Observational study, Categorization, Inference, Computer science, Causality (physics), Psychology