FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteNo-Evidence version preferred86 / 162 · ed1b16ef56464da0

Causal inference methods for combining randomized trials and observational studies: a review

Statistics · 2011.08047v4

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

Causal inference methods for combining randomized trials and observational studies: a review

8 Conclusion

Overall this data analysis highlights the interest of combining two different data sets, but also some challenges : the need for a good understanding of the common covariates, exposure, and outcome of interest before combining the data sets, different missing data patterns, and poor overlap when considering specific target (sub-)populations.

…izability and transportability of RCT from one population to another. The corresponding rich and prolific literature answers a real practical concern: external validity. Indeed, questions about external validity arise as soon as there are treatment effect heterogeneities in the populations under study. We find that, as any growing scientific field, the ideas are in flux: notations differ, implementations are scattered, and the proposed methods proposed still lack real-world benchmarks, generated hand in hand with practitioners. In addition, many open questions still remain as detailled below.

↓ 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: No-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

Not preferred in this comparison

First, the assumptions underlying these methods deserve closer scrutiny. Positivity, in particular, is not merely a technical condition: when overlap between the trial and the observational sample is limited, generalization is possible only to a restricted population without extrapolation, and deterministic violations must be ruled out on substantive grounds.

Passages supplied to the Evidence version

Integrative analysis of randomized clinical trials with real world evidence studies ↗

The transportability of the outcome mean function from the RCT sample to the RWE sample in Assumption 7 is testable, so potential tests for this assumption can be developed to verify this assumption. Assumption 3 states that all real-world patients have nonzero probabilities of participating the RCT…

Read full passage excerpt

The transportability of the outcome mean function from the RCT sample to the RWE sample in Assumption 7 is testable, so potential tests for this assumption can be developed to verify this assumption. Assumption 3 states that all real-world patients have nonzero probabilities of participating the RCT. This assumption requires the absence of patient characteristics that prohibit participation to the RCT. This assumption holds for the CALGB 9633 trial by its protocol. When the positivity assumption is violated, there is limited overlap in the support of covariate in RCT and RWE samples, the generalization can only be made to a restricted population without extrapolation ( Yang and Ding, 2018a ) . In practice, health scientists should carefully examine the question at hand to eliminate deterministic violations of positivity.

No Evidence

Same model and position · no retrieved passages · first output

Preferred by blind review

Several directions for future work emerge from this synthesis. First, the development of real-world benchmarks, generated in collaboration with practitioners, would allow the competing methods to be compared under conditions that reflect the complexities of actual data sets. Second, further work is needed to clarify the relative merits of the different identification strategies when their underlying assumptions are only partially plausible.

Judgment and task check

The blinded model review preferred the no-Evidence version. Its original rationale appears below.

Read the original model judgment

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

Output A contains an unsupported claim. The source discusses positivity, overlap, and generalization to restricted populations, but only in the specific context of 'transportability of the outcome mean function from the RCT sample to the RWE sample' with particular assumptions (Assumption 3, Assumption 7) and the CALGB 9633 trial. Output A generalizes this to 'the trial and the observational sample' broadly and states that 'deterministic violations must be ruled out on substantive grounds,' but the source says health scientists should 'carefully examine the question at hand to eliminate deterministic violations of positivity'—not that they 'must be ruled out on substantive grounds.' The phrase 'not merely a technical condition' is also unsupported editorializing. Output B, by contrast, directly continues the draft's stated open questions: it echoes the draft's mention of 'lack real-world benchmarks, generated hand in hand with practitioners' and 'many open questions still remain,' without adding unsupported factual claims. Both are coherent paragraphs, but A's unsupported material makes it violate the closed-book constraint.

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.