wytzzka's picture
Upload 11 files
2027f34 verified
|
Raw History Blame Contribute Delete
4.63 kB

Proximal starter v1: mechanisms and parameter priors

Each table is a 500-observation study and is scored at one independent baseline query x*. Target: CATE(x*) = E[Y(1)-Y(0)|X=x*].

Common proxy model

Let X1,X2 iid N(mu, sigma_x^2), c=(.25,.25) and p(x)=1/[1+exp{.125*r_p*(x1+x2)}]. The observation sampler draws A|X ~ Bernoulli(p(X)), then jointly Gaussian (Z,W,U)|A,X with mean

(.25,.25,.25) + A*(.25,.25*r_w,.25) + (c'X)*(1,1,1)

and covariance

[[1,       r_z*r_w, r_z],
 [r_z*r_w, 1,       r_w],
 [r_z,     r_w,     1  ]].

This is a factorization of a joint distribution, not a claim that treatment causally changes pre-treatment latent U or the baseline proxies. The coupled covariance and W-shift preserve the intended Gaussian conditional proxy restrictions. Let

m_W(u,x)=.25+.25(x1+x2)+r_w*(u-.25-.25(x1+x2)),

g_W(x)=E[W|X=x]=.25+.25(x1+x2)+.25*r_w*p(x).

Generate

Y = 2 + f(X) + beta_W*m_W(U,X) + 2*(W-m_W(U,X)) + s*A*delta(X,m_W(U,X)) + epsilon.

The result bridge is h(W,A,X)=2+f(X)+beta_W*W+s*A*delta(X,W). Because each delta is affine or constant in W, the analytic query effect is tau(x)=s*delta(x,g_W(x)). Natural W|X is integrated out; W|A,X is not substituted. U and references never enter the observed table.

Four families

Family f(x) beta_W delta(x,w) Noise
Cui linear .25(x1+x2) 4q 2 N(0,(.25v)^2)
Qi heterogeneous linear .25(x1+x2) 8q .5+3x1-5x2+c4*w Uniform(-v,v)
Qi quadratic x1^2+x2^2 5q .25-6x1*x2 Uniform(-v,v)
Qi trigonometric x1^2+x2^2 4q 2.3+abs(x1-1)-abs(x2+1)+(-2.5+sin(x1)-2cos(x2))*w Uniform(-v,v)

Qi linear first samples L1/L2 equally, with c4=.25 or 0. These scenarios are not counted as extra independent mechanism families. Query x* is a fresh independent draw from the same N_2(mu*1, sigma_x^2 I) covariate population.

Hyperprior and source adaptations

Family probabilities are 1/4 each. Within a task the independent continuous draws are s~U(.5,1.5), v~LogUniform(.5,2), mu~U(0,.5), sigma_x~U(.2,.4), r_p~U(.5,1.5), q~U(.75,1.25), and r_z,r_w~U(.35,.7). Covariance and W mean-shift are then updated jointly. These are benchmark hyperpriors, not a claim of exact source-paper settings.

  • Cui, Pu, Shi, Miao & Tchetgen Tchetgen (2024), Semiparametric Proximal Causal Inference, JASA 119(546),1348-1359, DOI:10.1080/01621459.2023.2191817. Mechanism locator: arXiv:2011.08411v4 §5.1, PDF21-22, equations13-14.
  • Qi, Miao & Zhang (2024), Proximal Learning for Individualized Treatment Regimes Under Unmeasured Confounding, JASA119(546),915-928, DOI:10.1080/01621459.2022.2147841. Mechanism locators: §6.1 Table3 L1/L2; Table4 N2 (quadratic), N1 (trigonometric), and author ProxITR code.

Qi treatment coding is adapted to 0/1 and the target to a query CATE. Extraneous noise covariates and auxiliary source labels are not imported. In particular, the source N2 helper labels with an inconsistent branch call are not used; the reference follows the displayed N2 outcome/bridge equations.

Public starter generation

The shipped starter uses 48 train,16 dev,32 test tables, each from a distinct sampled parameter configuration. Each configuration contributes one table. They have fresh seed namespaces and do not reproduce official observations. The labels in each split's truth.csv are exact analytic query references. This public seed is for the starter only and is not the official collection seed.

python generate.py --output extra --seed 314159 --train 2000 --dev 16 --test 32

Requires Python3.10+ and NumPy1.26 or later. The generator refuses to replace an existing output directory. catalog has public labels, queries are stored separately, and observations use round-trip float64 CSV precision. New training draws should not replace the saved official inputs.

Official design and scoring boundary

The frozen official collection has 2,000 different parameter draws, one table and one query each, without ID/OOD labels. It cannot estimate a within-fixed- configuration sampling variance from paired repeats. Report across-task error variance with N-1 denominator, with raw and optional SD(Y)-normalized scales kept separate. Nominal 95% intervals do not guarantee coverage.

Public illustrative mechanisms, equations and technical consistency checks do not replace scientific review or establish identification for every possible adaptation. The source registry's needs_scientific_review state is retained.