3434</p >
3535
3636<p >
37- CausalELM enables estimation of causal effects in settings where a randomized control trial
38- or traditional statistical models would be infeasible or unacceptable. It enables estimation
39- of the average treatment effect (ATE)/intent to treat effect (ITE) with interrupted time
40- series analysis, G-computation, and double machine learning; average treatment effect on the
41- treated (ATT) with G-computation; cumulative treatment effect with interrupted time series
42- analysis; and the conditional average treatment effect (CATE) via S-learning, T-learning,
43- X-learning, R-learning, and doubly robust estimation. Underlying all of these estimators are
44- ensembles of extreme learning machines, a simple neural network that uses randomized weights
45- and least squares optimization instead of gradient descent. Once a model has been estimated,
46- CausalELM can summarize the model and conduct sensitivity analysis to validate the
47- plausibility of modeling assumptions. Furthermore, all of this can be done in four lines of
48- code.
37+ CausalELM provides easy-to-use implementations of modern causal inference methods. While
38+ CausalELM implements a variety of estimators, they all have one thing in common—the use of
39+ machine learning models to flexibly estimate causal effects. This is where the ELM in
40+ CausalELM comes from—the machine learning model underlying all the estimators is an extreme
41+ learning machine (ELM). ELMs are a simple neural network that use randomized weights and
42+ offer a good tradeoff between learning non-linear dependencies and simplicity. Furthermore,
43+ CausalELM implements bagged ensembles of ELMs to reduce the variance resulting from
44+ randomized weights.
4945</p >
5046
51- <h2 >Extreme Learning Machines and Causal Inference </h2 >
47+ <h2 >Estimators </h2 >
5248<p >
53- In some cases we would like to know the causal effect of some intervention but we do not
54- have the counterfactual, making conventional methods of statistical analysis infeasible.
55- However, it may still be possible to get an unbiased estimate of the causal effect (ATE,
56- ATE, or ITT) by predicting the counterfactual and comparing it to the observed outcomes.
57- This is the approach CausalELM takes to conduct interrupted time series analysis,
58- G-Computation, double machine learning, and metalearning via S-Learners, T-Learners,
59- X-Learners, R-learners, and doubly robust estimation. In interrupted time series analysis,
60- we want to estimate the effect of some intervention on the outcome of a single unit that we
61- observe during multiple time periods. For example, we might want to know how the
62- announcement of a merger affected the price of Stock A. To do this, we need to know what the
63- price of stock A would have been if the merger had not been announced, which we can predict
64- with machine learning methods. Then, we can compare this predicted counterfactual to the
65- observed price data to estimate the effect of the merger announcement. In another case, we
66- might want to know the effect of medicine X on disease Y but the administration of X was not
67- random and it might have also been administered at mulitiple time periods, which would
68- produce biased estimates. To overcome this, G-computation models the observed data, uses the
69- model to predict the outcomes if all patients recieved the treatment, and compares it to the
70- predictions of the outcomes if none of the patients recieved the treatment. Double machine
71- learning (DML) takes a similar approach but also models the treatment mechanism and uses it
72- to adjust the initial estimates. This approach has three advantages. First, it is more
73- efficient with high dimensional data than conventional methods. Metalearners take a similar
74- approach to estimate the CATE. While all of these models are different, they have one thing
75- in common: how well they perform depends on the underlying model they fit to the data. To
76- that end, CausalELMs use bagged ensembles of extreme learning machines because they are
77- simple yet flexible enough to be universal function approximators with lower varaince than
78- single extreme learning machines.
49+ CausalELM implements estimators for aggreate e.g. average treatment effect (ATE) and
50+ individualized e.g. conditional average treatment effect (CATE) quantities of interest.
7951</p >
8052
81- <h2 >CausalELM Features</h2 >
53+ <h3 >Estimators for Aggregate Effects</h3 >
54+ <ul >
55+ <li>Interrupted Time Series Estimator</li>
56+ <li>G-computation</li>
57+ <li>Double machine Learning</li>
58+ </ul >
59+
60+ <h3 >Individualized Treatment Effect (CATE) Estimators</h3 >
61+ <ul >
62+ <li>S-learner</li>
63+ <li>T-learner</li>
64+ <li>X-learner</li>
65+ <li>R-learner</li>
66+ <li>Doubly Robust Estimator</li>
67+ </ul >
68+
69+ <h2 >Features</h2 >
8270<ul >
8371 <li >Estimate a causal effect, get a summary, and validate assumptions in just four lines of code</li >
8472 <li >Bagging improves performance and reduces variance without the need to tune a regularization parameter</li >
@@ -87,25 +75,28 @@ single extreme learning machines.
8775 <li >Most inference and validation tests do not assume functional or distributional forms</li >
8876 <li >Implements the latest techniques form statistics, econometrics, and biostatistics</li >
8977 <li >Works out of the box with arrays or any data structure that implements the Tables.jl interface</li >
78+ <li >Works out of the box with AbstractArrays or any data structure that implements the Tables.jl interface</li >
79+ <li >Works with CuArrays, ROCArrays, and any other GPU-specific arrays that are AbstractArrays</li >
80+ <li >CausalELM is lightweight—its only dependency is Tables.jl</li >
9081 <li >Codebase is high-quality, well tested, and regularly updated</li >
9182</ul >
9283
9384<h2 >What's New?</h2 >
9485<ul >
95- <li >Now includes doubly robust estimator for CATE estimation</li >
96- <li >All estimators now implement bagging to reduce predictive performance and reduce variance</li >
97- <li >Counterfactual consistency validation simulates more realistic violations of the counterfactual consistency assumption</li >
86+ <li >See the JuliaCon 2024 CausalELM demonstration <a href =" https://www.youtube.com/watch?v=hh_cyj8feu8&t=26s " >here.
87+ <li >Includes support for GPU-specific arrays and data structures that implement the Tables.jl API<li >
88+ <li >Only performs randomization inference when the inference argument is set to true in summarize methods</li >
89+ <li >Summaries support calculating marginal effects and confidence intervals</li >
90+ <li >Randomization inference now uses multithreading</li >
91+ <li >Refactored code to be easier to extend and understand</li >
9892 <li >Uses a simple heuristic to choose the number of neurons, which reduces training time and still works well in practice</li >
9993 <li >Probability clipping for classifier predictions and residuals is no longer necessary due to the bagging procedure</li >
100- <li >CausalELM talk has been accepted to JuliaCon 2024!</li >
10194</ul >
10295
10396<h2 >What's Next?</h2 >
10497<p >
105- Newer versions of CausalELM will hopefully support using GPUs and provide interpretations of
106- the results of calling validate on a model that has been estimated. In addition, some
107- estimators will also support using instrumental variables. However, these priorities could
108- also change depending on feedback recieved at JuliaCon.
98+ Efforts for the next version of CausalELM will focus on providing interpreteations for the results of callin validate as well
99+ as fixing any bugs and eliciting feedback.
109100</p >
110101
111102<h2 >Disclaimer</h2 >
0 commit comments