Long-term Effect Analysis
Let \(X \in \mathbb{R}^{p}\) be baseline covariates. Let \(D \in \{0, 1\}\) indicate treatment assignment. Let \(M \in \mathbb{R}\) be an intermediate/short-term outcome and \(Y \in \mathbb{R}\) be a long-term outcome. An analyst may wish to measure the effect of \(D\) on \(Y\) (a long-term outcome), yet the experimental sample only includes \(M\) (a short-term outcome).
If the analyst has access to an additional observational sample that includes the long-term outcome, then long-term causal inference is still possible. Specifically, assume the analyst has access to (i) an experimental sample, indicated by \(G=0\), where \((D, M, X)\) are observed; and (ii) an observational sample, indicated by \(G=1\), where \((M, X, Y)\) are observed, and \(D\) is either observed or not. Depending on whether \(D\) is also revealed in the observational sample will give rise to different assumptions that identify the long-term treatment effect. Specifically, the key identifying assumption when we do not observe \(D\) is that the short-term outcome is a statistical surrogate for the long-term outcome, while the identifying assumption for the case when we observe \(D\) is that unobserved confounding is mediated through the short-term outcome in the observational sample. Following Athey et al., 2016, we refer to these models as Surrogacy model or Latent unconfounded model, respectively (Athey et al., 2020).
Long-term effect
Formally, define the long-term counterfactual \(\mathbb{E}\left[Y^{(d)}\right]\) as the counterfactual mean outcome for the full population in the thought experiment in which everyone is assigned treatment value \(D=d\).
Effects for the experimental or observational subpopulation use target-specific influence-function weights. Select them with sample_G="G=0" or sample_G="G=1", respectively. Subgroup targeting is not obtained merely by multiplying the pooled-population score by a group indicator: the leading regression term and the residual-correction weights change with the target population.
Note
In addition to the population ATE, the implementation supports conditional long‑term ATEs within the experimental or observational subpopulations:
\(\mathbb{E}[Y^{(1)}-Y^{(0)} \mid G=0]\) and \(\mathbb{E}[Y^{(1)}-Y^{(0)} \mid G=1]\).
Set sample_G="G=0" or sample_G="G=1" to target these effects; use sample_G="all" for the pooled-population target. Identification and influence‑function estimators (with the associated nuisance components) for the Surrogacy and Latent‑Unconfounded models are given in Theorem 3.1 and Theorem B.2 of Chen & Ritzwoller (2023).
Project STAR application
The plotted proposed estimators use sample_G="all" and therefore target the pooled STAR and NYC population; the oracle remains STAR-only.
In the heterogeneous analysis, prior ability defines the grid and kernel weights only (include_V=False); it is excluded from the nuisance models and oracle NPIV inputs. The proposed estimators use the pooled nonmissing-prior sample for their grid, bandwidth, and kernel normalization, while the STAR oracle uses the same grid and bandwidth with STAR normalization. These are finite-bandwidth localizations of the chosen score, not automatically the usual conditional causal effect given prior ability.
Surrogacy Model
Define the regression and the conditional distribution
For the pooled target, sample_G="all", the four nuisances associated with the model are
and the long-term counterfactual is
Latent Unconfounded Model
When we observe \(D\) in the observational sample, the regression becomes
For the pooled target, sample_G="all", the nuisances under this model are given by
The long-term counterfactual is
Experimental-population target
Let \(\pi_0=\mathbb{P}(G=0)\) and
For the Surrogacy model, write \(e_d(x)=\mathbb{P}(D=d\mid X=x,G=0)\), \(e_d(m,x)=\mathbb{P}(D=d\mid M=m,X=x,G=0)\), and \(s(m,x)=\mathbb{P}(G=1\mid M=m,X=x)\). The experimental-target residual weights are
For the Latent-Unconfounded model, let \(\gamma(x)=\mathbb{P}(G=1\mid X=x)\) and define the counterfactual treatment propensity
Although \(M(d)\) is not jointly observed with both treatment states, this propensity is identified by the observable Bayes-rule representation
This is the propensity estimated by the corresponding sample_G="G=0"
branch; it is generally not the ordinary observed-data propensity
\(\mathbb{P}(D=d\mid M=m,X=x,G=1)\). Then
For either model, the uncentered score for treatment arm \(d\) is
Target-specific localization
For a localization variable \(V\), evaluation value \(v\), and bandwidth \(\lambda\), the experimental-target weight is
The finite-bandwidth target is a ratio. Consequently, if \(H_{d,\lambda}^{(0)}=\ell_{\lambda,v}^{(0)}H_d^{(0)}\), its centered influence value is
not \(H_{d,\lambda}^{(0)}-\theta_{d,\lambda}^{(0)}\). The implementation uses this ratio centering for pointwise and uniform covariance estimation. For a treatment contrast, subtract the corresponding expressions for \(d=0\) from those for \(d=1\).
The observational-population construction is analogous, with \(q_1(W)=\mathbb{1}_{G=1}/\mathbb{P}(G=1)\) and kernel normalization conditional on \(G=1\). For the pooled target, \(q=1\) and the kernel is normalized over the pooled population.
When sample_G selects a subgroup, automatic bandwidth selection and kernel normalization use that subgroup.
If v_values is omitted, the implementation localizes at the target-subgroup mean; explicitly supplied values should be chosen from the intended target distribution.
Setting include_V=True appends \(V\) to the covariates used by the nuisance models in addition to using it for localization.
|
Debiased Machine Learning for long-term causal analysis (DML-longterm) class with joint/sequential model fitting. |
References
Athey, S., Chetty, R., Imbens, G., 2020. Using experiments to correct for selection in observational studies.
Athey, S., Chetty, R., Imbens, G., Kang, H., 2016. Estimating treatment effects using multiple surrogates: the role of the surrogate score and the surrogate index.
Chen, J., Ritzwoller, D. M., 2023. Semiparametric estimation of long-term treatment effects. Journal of Econometrics, Volume 237, Issue 2, Part A.