Scribd
Upload
4 views2 pages

05 Practical Issues Positive Inference

The document discusses targeted learning in causal inference, emphasizing the importance of positivity and the challenges posed by extreme weights due to sparse data. It outlines diagnostic methods and mitigation strategies to address issues related to positivity and variance in estimates. Additionally, it provides a standard inference workflow and common extensions for various types of outcomes.

Uploaded by

marco vitturini
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
4 views2 pages

05 Practical Issues Positive Inference

The document discusses targeted learning in causal inference, emphasizing the importance of positivity and the challenges posed by extreme weights due to sparse data. It outlines diagnostic methods and mitigation strategies to address issues related to positivity and variance in estimates. Additionally, it provides a standard inference workflow and common extensions for various types of outcomes.

Uploaded by

marco vitturini
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd

Targeted Learning — Practical Issues

(positivity, diagnostics, inference)


Summary notes (non-exhaustive) — generated 2026-05-29

Positivity, sparsity, and why weights can go feral


Identification of many causal estimands requires positivity: within strata of confounders W,
treatment assignment must have nonzero probability:
0 < P(A=1|W=w) < 1.

In practice, if ĝ(W) is very close to 0 or 1 for some w, clever covariates H(A,W) contain 1/ĝ
or 1/(1−ĝ), creating extreme weights. This can induce:
• high variance,
• finite-sample bias (“sparse data bias”),
• unstable confidence intervals.

Diagnostics and mitigation strategies


 Diagnose: inspect propensity distributions, effective sample size, and instability of
H(A,W). Some approaches use bootstrap diagnostics to detect sparse-data bias.
 Mitigate by design: restrict to overlap region, improve measurement of confounders, or
change study design.
 Mitigate analytically: truncation of ĝ (e.g., cap at [0.01,0.99]), targeted selection of ĝ
(C-TMLE), or use alternative estimands that respect limited overlap.
 Prefer bounded fluctuation models for Q when Y is bounded; substitution estimators
avoid impossible predictions.

Finite-sample performance: the “bias–variance budget”


Targeted learning treats estimation as budgeting error: some nuisance misspecification is
acceptable if it doesn’t move Ψ̂ much, but positivity problems can move it a lot.

A practical mindset:
• reduce extreme weights first (stability),
• then use flexible Q models (bias),
• then validate inference via EIF and, when needed, cross-fitting.
In other words: don’t let one near-zero propensity value turn your analysis into
performance art.

Inference workflow (standard template)


1. Estimate nuisances (Q̂ , ĝ, censoring if needed) with ML + CV.
2. Compute TMLE (or CV-TMLE / C-TMLE).
3. Compute EIF values D*(ᵖP)(O_i).
4. SE = sqrt( Var_n(D*) / n ).
5. CI = Ψ̂ ± z_{0.975}·SE (or use t-approx in small samples).
6. Report diagnostics: overlap/positivity, sensitivity to truncation, robustness checks.

Common extensions (so you can recognize them in the wild)


 Survival outcomes: target survival curves; nuisance includes hazards and censoring
mechanisms.
 Case-control: weighting schemes to correct biased sampling; still TMLE structure.
 Longitudinal regimes: g-computation / sequential regression; multiple clever covariates
for time-varying treatment and censoring.
 Mediation/direct effects: careful identification assumptions; targeted estimation of a
nontrivial functional.

You might also like