CausalELM enables estimation of causal effects in settings where a randomized control trial would be impossible or infeasible. Estimation of the average treatment effect (ATE), intent to treat effect (ITE), and average treatment effect on the treated (ATT) can be estimated via G-computation or double machine learning (DML) while the ATE or cumulative treatment effect (CTE) can be estimated from an interrupted time series analysis. CausalELM also supports estimation of individual treatment effects or conditional average treatment effects (CATE) via S-learning, T-learning, X-learning, and R-learning. The underlying machine learning model for all these estimators is an extreme learning machine or L2 regularized extreme learning machine. Furthermore, once a model has been estimated, CausalELM can summarize the model, including computing p-values via randomization inference, and conduct sensitivity analysis. All of this can be done in foru lines of code.
In some cases we would like to know the causal effect of some intervention but we do not have the counterfactual, making conventional methods of statistical analysis infeasible. However, it may still be possible to get an unbiased estimate of the causal effect (ATE, ATE, or ITT) by predicting the counterfactual and comparing it to the observed outcomes. This is the approach CausalELM takes to conduct interrupted time series analysis, G-Computation, DML, and meatlearning via S-Learners, T-Learners, X-Learners, and R-learners. In interrupted time series analysis, we want to estimate the effect of some intervention on the outcome of a single unit that we observe during multiple time periods. For example, we might want to know how the announcement of a merger affected the price of Stock A. To do this, we need to know what the price of stock A would have been if the merger had not been announced, which we can predict with machine learning methods. Then, we can compare this predicted counterfactual to the observed price data to estimate the effect of the merger announcement. In another case, we might want to know the effect of medicine X on disease Y but the administration of X was not random and it might have also been administered at mulitiple time periods, which would produce biased estimates. To overcome this, G-computation models the observed data, uses the model to predict the outcomes if all patients recieved the treatment, and compares it to the predictions of the outcomes if none of the patients recieved the treatment. Double machine learning (DML) takes a similar approach but also models the treatment mechanism and uses it to adjust the initial estimates. This approach has three advantages. First, it is more efficient with high dimensional data than conventional methods. Second, it allows one to model complex, nonlinear relationships between the treatment and the outcome. Finally, it is a doubly robust estimator, meaning that only the model of the outcome OR the model of the treatment mechanism has to be correctly specified to yield unbiased estimates. The DML implementation in CausalELM. also overcomes bias from regularization by employing cross fitting. Furthermore, we might be more interested in how much an individual can benefit from a treatment, as opposed to the average treatment effect. Depending on the characteristics of our data, we can use metalearning methods such as S-Learning, T-Learning, X-Learning, or R-Learning to do so. In all of these scenarios, how well we estimate the treatment effect depends on how well we can predict the counterfactual. The most common approaches to getting accurate predictions of the counterfactual are to use a super learner, which combines multiple machine learning methods and requires extensive tuning, or tree-based methods, which also have large hyperparameter spaces. In these cases hyperparameter tuning can be computationally expensive and requires researchers to make arbitrary decisions about how many and what models to use, how much regularization to apply, the depth of trees, interaction effects, etc. On the other hands, ELMs are able to achieve good accuracy on a variety of regression and classification tasks and generalize well. Moreover, they have a much smaller hyperparameter space to tune and are fast to train becasue they do not use backpropagation to update their weights like conventional neural networks.
- Simple interface enables estimating causal effects in only a few lines of code
- Validate and test sensitivity of a model to possible violations of its assumption in one line of code
- Estimation of p-values and standard errors via asymptotic randomization inference
- Incorporates latest research from statistics, machine learning, econometrics, and biostatistics
- Analytically derived L2 penalty reduces cross validation time and multicollinearity
- Fast automatic cross validation works with longitudinal, panel, and time series data
- Includes 13 activation functions and allows user-defined activation functions
- Single interface for continous, binary, and categorical outcome variables
- No dependencies outside of the Julia standard library
- Added support for dataframes
- Estimators can handle any array whose values are <:Real
- Estimator constructors are now called with model(X, T, Y) instead of model(X, Y, T)
- Eliminated unnecessary type constraints on many methods
- Improved documentation
- Permutation of continuous treatments draws from a continuous, instead of discrete uniform distribution during randomization inference
The priority for v0.6 will be on implementing doubly robust estimation for the average effect (ATE/ATT) as well as the CATE. We will also make minor improvements for speed, maintainability, etc.
CausalELM is extensively tested and almost every function or method has multiple tests. That being said, CausalELM is still in the early/ish stages of development and may have some bugs. Also, expect breaking releases for now.All contributions are welcome. Before submitting a pull request please read the contribution guidlines.