2024-07-12 –, For Loop (3.2)
Causal inference has proven invaluable to many organizations, however, it is often infeasible because it requires them to conduct expensive experiments. Fortunately, advances in causal machine learning give us new methods to estimate causal effects from observational data. Accordingly, in this talk, I will introduce CausalELM, a package that implements causal estimators based on extreme learning machines, and discuss how it can be used to estimate average and individualized treatment effects.
Causal inference is a useful tool, not only for those conducting academic research, but also large organizations deciding on how to allocate resources. For example, streaming services use A/B tests to decide whether to roll out a new feature, rideshare companies estimate individual treatment effects and use them to determine how much to charge passengers, and e-commerce companies perform causal inference to decide which products to recommend to users. However, estimating causal effects is not easy. Running A/B tests or randomized control trials are expensive: researchers do not have the ability to change public policies, or the dimensionality of the potential covariates is too high for traditional statistical methods. CasualELM addresses these and other challenges by leveraging recent advances in causal machine learning. To that end, we will cover the following main topics:
• Potential Outcomes Framework: We will briefly discuss the potential outcomes framework,
which CausalELM's estimators fall under.
• Motivating Examples: I will motivate the talk with two examples. First, I will demonstrate
how CausalELM can be used to estimate the effect of 401(k) pension plan eligibility on net
worth using double machine learning. Second, I will use a recent study of the effect of
development aid on Taliban attacks to demonstrate how to estimate individualized treatment
effects in CausalELM . These examples will make up the majority of the talk.
• Extreme Learning Machines: We will discuss what extreme learning machines are, how
they work, and how they are used in ensembles in CausalELM.
• Other Estimators: Besides the motivating examples, we will touch upon the other
estimators implemented in CausalELM.
• Comparison with Other Packages: We will talk about how CausalELM is different from
other causal machine learning libraries and the main tradeoffs it makes.
The focus of the talk will be on the implementation of causal machine learning tasks, rather than the mathematical theory behind them, so those without an extensive background in statistics should not feel intimidated.
Darren Colby is an MS candidate in computational analysis and public policy at the University of Chicago. His public policy focus area is international security and development and he is motivated by the prospect of using machine learning and statistics to answer important questions and solve challenges in international security. He also holds an AB in government from Dartmouth College and previously served as an information systems specialist in the US Army for six years.