fuyao11's picture
Streamline dataset card and starter documentation to match Mediation layout
6de515d verified
|
Raw History Blame Contribute Delete
3.19 kB

Public starter mechanisms

The starter contains 96 newly generated studies: 48 train, 16 dev and 32 test. Each split uses an equal number of the following two presets. One configuration fixes the mechanism parameters and an independently sampled CATE query; two independent studies of n=500 share that configuration. Configurations, queries and dataset seeds are separate across splits and from the official test. The public generation_plan.json describes only these starter studies.

Split Configurations Tables Intended use
Train 24 48 Fit or adapt a method using known effects
Dev / validation 8 16 Choose method settings and check interval construction
Public test 16 32 Score a fixed method locally

Each observed table has a population ATE label and a CATE label at its supplied query. The two replicate studies share that query. Labels are provided in truth_ate.csv and truth_cate.csv within each split.

nie_wager_a: continuous outcome, five covariates

Source: Nie and Wager (2021), Quasi-oracle estimation of heterogeneous treatment effects, Section 6.1, Setup A of the author manuscript. Published paper.

Draw five independent Uniform(0,1) covariates. For one configuration draw s ~ Uniform(0.5,1.5) and sigma ~ Uniform(0.5,1.5). Set

e(x) = clip(sin(pi*x1*x2), 0.1, 0.9)
b(x) = sin(pi*x1*x2) + 2*(x3-0.5)^2 + x4 + 0.5*x5
tau(x) = s*(x1+x2)/2
T | X ~ Bernoulli(e(X))
Y = b(X) + T*tau(X) + sigma*epsilon, epsilon ~ Normal(0,1)

The population ATE is s/2. CATE at the supplied query is tau(query). This is an adaptation: dimension is fixed at five, the effect and noise scales vary over the stated benchmark distributions, and treatment is uncentered. The cited source uses centered treatment; this changes the baseline outcome surface while preserving the stated treatment-effect contrast.

benkeser_binary: binary outcome, two covariates

Source: Benkeser, Carone, van der Laan and Gilbert (2017), Doubly robust nonparametric inference on the average treatment effect, Section 5.1. Published paper.

Draw X1 ~ Uniform(-2,2) and X2 ~ Bernoulli(0.5), independently. For one configuration draw a ~ Uniform(0.1,0.3). With sigmoid(z)=1/(1+exp(-z)), set

z(x) = -x1 + 2*x1*x2
e(x) = sigmoid(z(x))
mu_t(x) = sigmoid(z(x) + a*t)
T | X ~ Bernoulli(e(X))
Y | X,T ~ Bernoulli(mu_T(X))

CATE is mu_1(query)-mu_0(query), a probability difference, not a log-odds effect. With softplus(z)=log(1+exp(z)), the population ATE is

[softplus(a+2)-softplus(a-2)-softplus(2)+softplus(-2)] / 4.

The source coefficient a=0.2 is varied over the stated benchmark distribution.

Generate more studies

Run python generate.py --output regenerated to regenerate the supplied starter with the default seed. NumPy's version used for generation is recorded in the plan. Use another seed to generate further example studies into a new directory. All probabilities, outcomes, effect labels and calculations use float64 where applicable. Treatment is binary and all observed covariates precede treatment.