FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteEvidence version preferred134 / 162 · 4cc323b5bc070189

Tutorial: Introduction to computational causal inference using reproducible Stata, R and Python code

Statistics · 2012.09920v2

FLOWING EVIDENCE BENCHMARK

How do we tell whether Evidence helps?

At the same writing position, with the same model and task, how does supplying retrieved paper passages change the first output? We compare matched versions and retain ties, unusable outputs and incomplete reviews.

SAME DRAFT · EVIDENCE ON OR OFF

01 · FIXED WRITING POSITION

MANUSCRIPT

Same writing position ▌

Matched autocomplete at the same draft position

02 · TWO MATCHED INPUTS

SHARED BY BOTH

Manuscript context, model, task and prompt

A · With Evidence

Retrieved paper passages supplied

B · Without Evidence

No retrieved passages supplied

LLM

Same model and version

A → first output

B → first output

Blind judge agent

First outputs are anonymized as X and Y

Continuation review: accuracy, fit to the writing task and usability
Returns: X preferred / tie / Y preferred / both unusable

Order check: X / Y → Y / X

HOW DOES BLIND JUDGING WORK?

① Anonymize both outputs
The judge sees the same draft and both first outputs without knowing which received Evidence.

② Compare and swap order
The judge applies task-specific criteria in X/Y and then Y/X order.

③ Review disagreements
A third pass resolves disagreements. Incomplete reviews remain in the denominator.

How are autocomplete positions stratified?

Before seeing generation outcomes, we check whether a retrieved passage contains a specific proposition that directly supports the next writing move. Those positions appear in the left opportunity group; the rest are ordinary positions on the right. We select a balanced sample from admitted papers in each field. The 50/50 split is experimental, not a measure of how often either type occurs in writing.

AUTOCOMPLETE · 110

Biology, statistics and astrophysics

55 positions on each side; statistics uses the ten-paper rerun.

AUTOCOMPLETE · 52

Psychology and climate science

26 positions on each side; psychology includes nine papers and climate science four.

How are the table percentages calculated?

Across five fields there are 81 source-opportunity positions. The original blind review preferred Evidence in 51; a task check moved one empty Evidence continuation to both unusable, leaving 50 in public counts. In the ordinary group, another pair of empty outputs moved from no winner to both unusable. Original verdicts remain visible on case pages.

50Evidence version preferred
÷
81All positions in this group
=
62%Evidence preference in this group

Source contribution is a separate review: 16 of 21 Evidence wins entered into source review directly used retrieved papers; another 29 wins await review.

These are development-stage model judgments pending independent human review. They are not formal Benchmark conclusions and do not, on their own, establish causality.

Manuscript writing position

Text excerpt · not a PDF page

Research manuscript · excerpt

Tutorial: Introduction to computational causal inference using reproducible Stata, R and Python code

8 Conclusion

It is clear to see that, compared to the true ATE of 17.9%, all of the methods (RA [13.8%], IPTW [55.3%], IPTW-RA [14.8%], AIPTW [13.6%]) produce a biased estimate, but ELTMLE produces an estimate that is unbiased relative to the true ATE. Note that the relative bias of IPTW is very large; IPTW relies on the positivity assumption, which in this simulation is violated because there was a low number of individuals with a higher comorbidity value. Without correcting for this imbalance in the data the methods that rely on this assumption will be vulnerable to bias.

Overall, all the methods introduced here include the estimation of the G-formula (non-parametrically or parametrically), which is a generalization of standardization, and the inverse probability of treatment weighting (IPTW). [ 4 ] However, there are other estimators based on matching strategies that we did not discuss in our tutorial. ( 18 )

↓ Next paragraph was generated here

The highlighted text is near the writing position. This is a layout of the evaluation input, not a PDF screenshot.

A matched comparison at the same writing positionJudgment: Evidence version preferred

Both arms use the same model, task and manuscript context. Only the retrieved passages differ. The first recorded output from each arm is shown below.

Evidence

Same model and position · retrieved passages supplied · first output

Preferred by blind review

Matching-based estimators, such as Kernel Optimal Matching, offer an alternative route to estimating sample average treatment effects when positivity is violated and the treatment assignment model is unknown, as in the spine surgery setting that motivated their development.

Passages supplied to the Evidence version

More robust estimation of sample average treatment effects using Kernel Optimal Matching in an observational study of spine surgical interventions ↗

Inverse probability of treatment weighting (IPTW), which has been used to estimate sample average treatment effects (SATE) using observational data, tenuously relies on the positivity assumption and the correct specification of the treatment assignment model, both of which are problematic assumption…

Read full passage excerpt

Inverse probability of treatment weighting (IPTW), which has been used to estimate sample average treatment effects (SATE) using observational data, tenuously relies on the positivity assumption and the correct specification of the treatment assignment model, both of which are problematic assumptions in many observational studies. Various methods have been proposed to overcome these challenges, including truncation, covariate-balancing propensity scores, and stable balancing weights. Motivated by an observational study in spine surgery, in which positivity is violated and the true treatment assignment model is unknown, we present the use of optimal balancing by Kernel Optimal Matching (KOM) to estimate SATE.

No Evidence

Same model and position · no retrieved passages · first output

Not preferred in this comparison

This tutorial has demonstrated that implementing these estimators requires careful attention to the assumptions underlying each method, particularly the positivity assumption, which was violated in our simulation and led to substantial bias in the IPTW estimate. We encourage researchers to use the reproducible code provided to explore how these methods perform under different data-generating scenarios and to assess the sensitivity of their own causal estimates to violations of these assumptions.

Judgment and task check

The blinded model review preferred the Evidence version. Its original rationale appears below.

Read the original model judgment

In this rationale, Output A = No Evidence; Output B = Evidence.

Output B provides a coherent continuation that bridges the draft's discussion of methods with the contextual note about matching strategies not discussed in the tutorial. It is directly supported by the source about Kernel Optimal Matching (KOM) as an alternative when positivity is violated and the treatment assignment model is unknown—exactly matching the source's stated purpose and the draft's mention of matching strategies. Output A, while topically relevant, contains unsupported claims: it asserts that 'our simulation' violated positivity (the draft discusses 'this simulation' but doesn't establish the authors ran it), claims 'we encourage researchers' and 'reproducible code provided' (no source or draft supports that code was provided or that this is a tutorial with such materials), and makes an unsupported exhortation about exploring 'different data-generating scenarios.' These are material factual claims about the paper's own content and offerings that lack closed-book support. While Output A's prose is fluent, its claims about the manuscript's structure and resources are invented. Output B stays strictly within supported territory, connecting the draft's mention of matching strategies to the sourced information about KOM as an alternative when positivity is violated.

The passages are input material; their presence does not mean an output used them or that they support every claim. Draft and source passages are excerpted; no simulated scores or PDF appear here.