FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteEvidence version preferred94 / 162 · 9728e36ee848f98c

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

1 Introduction

… behavioral sciences are causal in nature. For example, what is the mortality risk difference amongst patients offered a surgical procedure versus those who did not receive it in a given population? [ 1 ] Causal inference methods may be used to answer this scientific question where a clinical trial is unfeasible or unethical (e.g. a surgical procedure). [ 2 ] Causal inference methods are based on the counterfactual framework introduced by Neyman in the randomized experiment setting, then extended to time-fixed and time-varying observational studies by Rubin and Robins, respectively. ( 3 , 4 )

Over the years, rapid on-going advances in the field causal inference have created a set of different approaches to estimate the causal effect of a treatment (or exposure) on an outcome (i.e., methods that incorporate propensity scores, the G-computation, or a combination of both, namely double-robust methods). Overall, double-robust methods now are preferred over naive regression approaches because the latter are biased under misspecification of a parametric outcome model when the research question is causal in nature. [ 5 , 6 ]

↓ 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

Double-robust estimators combine an outcome regression model with a treatment (or exposure) model, so that the estimated causal effect remains consistent if either model—but not necessarily both—is correctly specified. [ 5 , 6 ] This property makes them attractive in observational settings where the true form of the outcome or treatment mechanism is unknown.

Passages supplied to the Evidence version

Robust semiparametric estimators: missing data and causal inference ↗

Semiparametric inference with missing outcome data (including causal inference) is based on partially specified models which are not of direct interest (e.g., model for missingness/treatment assignment mechanism). Different class of estimators exist, which are more or less robust to misspecification…

Read full passage excerpt

Semiparametric inference with missing outcome data (including causal inference) is based on partially specified models which are not of direct interest (e.g., model for missingness/treatment assignment mechanism). Different class of estimators exist, which are more or less robust to misspecification of these models. Another type of threat to the validity of the inference occur in situations where some observations are contaminated (generated by some nuisance distribution). Classical semiparametric inference is not robust to such contamination, and a single observation may have an arbitrary large effect on bias as measured by the influence function. We introduce inverse probability weighted, double robust and outcome regression estimators of location and scale parameters, which are robust to contamination in the sense that their influence function is bounded.

No Evidence

Same model and position · no retrieved passages · first output

Not preferred in this comparison

Despite these methodological advances, applied researchers face practical barriers to implementing causal inference methods, including the need to navigate multiple software environments and reconcile differences in estimation and reporting across them. This tutorial addresses that gap by providing a unified, reproducible workflow for computational causal inference in Stata, R and Python.

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 = Evidence; Output B = No Evidence.

Output A provides a coherent, grounded continuation that elaborates on double-robust methods—the topic just introduced in the draft's final sentence. It explains what makes double-robust estimators 'preferred' (consistent if either model is correct) and connects this to 'observational settings,' which aligns with the draft's focus on causal inference when trials are unfeasible. The source snippet supports this by discussing 'double robust' estimators and their robustness properties. No unsupported claims: the explanation of double-robustness is a standard characterization, and the source mentions this class of estimators. Output B introduces an entirely new topic (software barriers, a tutorial, Stata/R/Python workflow) with no grounding in the draft or source. The draft never mentions software, tutorials, or reproducible workflows; this is a non-sequitur that changes the manuscript's direction. The source also does not mention software or tutorials. B's claim that 'this tutorial addresses that gap' is unsupported—the draft does not establish that this is a tutorial paper, and no source supports this framing.

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.