Graphoids over Counterfactuals

Pearl, Judea

doi:10.1515/jci-2014-0028

Cited by 2 publications

(2 citation statements)

References 12 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Our estimations of individual treatment effects depend on the following four standard assumptions in causal inference [15,16].…”

Section: Assumptionsmentioning

confidence: 99%

Applying Inverse Propensity Score Weighting and Doubly Robust for Treatment Effect Estimation in Sepsis

Xiao,

Xia,

Qian

et al. 2023

ACE

View full text Add to dashboard Cite

Sepsis is a severe disease with a relatively high fatality rate. The treatment assignment of antibiotics is one of the most crucial factors in reducing the mortality rate. Existing machine learning models are designed to estimate the individual time-to-treatment effect of the antibiotic. However, most models need more efficiency and stability in overcoming the selection bias problem. This paper applies two methods of debiasing samples and estimating individual treatment effects. By re-weighting, we create pseudo mini-batches that mimic the corresponding randomized controlled trial (RCT) process. We also adopt Inverse Propensity Score Weighting (IPW) method to assign appropriate weight to each sample in the observation dataset by estimating the propensity scores. We assume Doubly Robust (DR) to combine IPW and outcome regression model. On a real-world dataset, the experiments exhibit the ability of the model to successfully identify effective timing of treatment.

show abstract

“…Our estimations of individual treatment effects depend on the following four standard assumptions in causal inference [15,16].…”

Section: Assumptionsmentioning

confidence: 99%

Applying Inverse Propensity Score Weighting and Doubly Robust for Treatment Effect Estimation in Sepsis

Xiao,

Xia,

Qian

et al. 2023

ACE

View full text Add to dashboard Cite

show abstract

“…In DAGs, we have the d-separation criterion, which translate immediately into tests of compatibility with data. In potential outcomes, testability requires non-trivial derivations (Pearl, 2014a).…”

Section: Causal Diagrams Counterfactuals and Potential Outcomesmentioning

confidence: 99%

Interview with Judea Pearl

Pearl¹

2022

Observational Studies

View full text Add to dashboard Cite

Interviewer: Inferring causal effects from data involves many steps. Where do you think your work fits within this overall process?Pearl: I seek to understand the conditions under which such inference is theoretically possible, allowing of course for partial scientific knowledge to guide the inference. My focus has been on a class of models called "nonparametric" which enjoy two unique features: (1) They capture faithfully the kind of scientific knowledge that is available to empirical researchers and (2) they require no commitment to numerical assumptions of any sort. Leveraging these models, I have focused on the problem of identification, rather than estimation. This calls for transforming the desired causal quantity into an equivalent probabilistic expression (called estimand) that can be estimated from data. Once an estimand is derived, the actual estimation step is no longer causal, and can be accomplished by standard statistical methods. This is indeed where machine learning excels, unlike the identification step in which machine learning and standard statistical methods are almost helpless. It is for this reason that I focus on identification -this is where the novelty of causal thinking lies, and where a new calculus had to be developed. Historical PerspectiveInterviewer: What is your perspective of the history of the causal inference movement, and how the movement came to where it is today?Pearl: My perspective comes through the lens of a computer scientist. I look at this movement as a struggle to develop a mathematical language for capturing cause effect relationships, so that we can express our assumptions faithfully and transparently, derive their logical implications and combine them with data. It's really a wedding between two non-intersecting languages, one is the language of cause and effect, the other is the language of data, namely statistics (Pearl, 2019b).The wedding occurred quite late in the history of science because science had not been very kind to causality. It has revolved around the symmetric equality sign '=' of algebra, and thus deprived us of a language to capture the asymmetry of causal relationships. Such a language was developed in the past three decades, using graphs, and it now enables us to answer causal and counterfactual questions with algorithmic precision.Graphs are new mathematical objects, unfamiliar to most researchers in the statistical sciences, and were of course rejected as "non-scientific ad-hockery" by top leaders in the field (Rubin, 2009). My attempts to introduce causal diagrams to statistics (Pearl, 1995;Pearl et al., 2000) have taught me that inertial forces play at least as strong a role in science as they do in politics. That is the reason that non-causal mediation analysis is still practiced in certain circles of social science (Hayes, 2017), "ignorability" assumptions still dominate large islands of research (Imbens and Rubin, 2015), and graphs are still tabooed in the econometric literature (Angrist and Pischke, 2014). While most re-

show abstract

Graphoids over Counterfactuals

Cited by 2 publications

References 12 publications

Applying Inverse Propensity Score Weighting and Doubly Robust for Treatment Effect Estimation in Sepsis

Applying Inverse Propensity Score Weighting and Doubly Robust for Treatment Effect Estimation in Sepsis

Interview with Judea Pearl

Contact Info

Product

Resources

About