FlowingFlowing
← Back to all development cases
Development case · blind model judgment pending independent human review; not a formal Benchmark conclusion.
AutocompleteEvidence version preferred90 / 162 · 7e6fc27444b1cd70

Correcting for partial verification bias in diagnostic accuracy studies: A tutorial using R

Statistics · 2509.12217v1

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

Correcting for partial verification bias in diagnostic accuracy studies: A tutorial using R

1 Introduction

It is essential to obtain valid estimates of accuracy measures to ensure the clinical validity of the tests. [ 7 ] Biased estimates are misleading, which may result in the premature implementation of the new tests and lead to wrong decision making by the clinicians. [ 15 , 4 ] However, the assessment of index tests in diagnostic accuracy studies often suffer from bias because the verification of the target outcomes by gold standard tests can be clinically infeasible due to cost, ethical and clinical considerations, as well as time-consuming and invasive procedures. [ 4 , 9 , 16 , 5 ]

…more likely to be verified by the gold standard test because they are presumably more likely to be positive for the target outcome. In contrast, those with a negative index test result are less likely to be selected in the verification sample because they are presumably less likely to be positive for the target outcome. Selective sampling may also depend on other factors (covariates), such as gender, age and the presence of clinical symptoms. [ 6 , 17 ] Specifically, this selective sampling creates a partial verification bias, resulting in biased estimates of accuracy measures. [ 15 , 4 , 5 ]

↓ 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

The standard approach to correcting for partial verification bias assumes that the true disease status is missing at random (MAR) given the observed test results and covariates, that is, that selection for verification does not depend on the unobserved disease status once these variables are accounted for.

Passages supplied to the Evidence version

Estimation of the Volume Under the ROC Surface in Presence of Nonignorable Verification Bias ↗

In order to correct for verification bias, the researchers often assume that the selection for disease verification does not depend on the disease status, given the test results and other observed covariates, i.e., they assume that the true disease status, when missing, is missing at random (MAR, Li…

Read full passage excerpt

In order to correct for verification bias, the researchers often assume that the selection for disease verification does not depend on the disease status, given the test results and other observed covariates, i.e., they assume that the true disease status, when missing, is missing at random (MAR, Little and Rubin (2002) ).

No Evidence

Same model and position · no retrieved passages · first output

Not preferred in this comparison

Partial verification bias arises when only a subset of patients undergoes the gold standard test, and the selection of that subset is related to the index test result. Because patients with positive index test results are more likely to be verified, the verified sample is enriched for disease, and the resulting estimates of sensitivity and specificity are distorted.

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 is directly and exactly supported by the supplied source, which states: 'In order to correct for verification bias, the researchers often assume that the selection for disease verification does not depend on the disease status, given the test results and other observed covariates, i.e., they assume that the true disease status, when missing, is missing at random (MAR, Little and Rubin (2002)).' Output B reproduces this claim with proper attribution to the source's content. Output A, while topically related to the draft's discussion of partial verification bias, adds unsupported claims: it states that 'the resulting estimates of sensitivity and specificity are distorted' as a definitive outcome, but the source does not mention sensitivity and specificity being distorted—it discusses MAR assumption for correction. More critically, Output A's claim that 'the verified sample is enriched for disease' is an unsupported inference about disease enrichment; the source does not state this. The draft discusses biased estimates of accuracy measures generally, but Output A's specific mechanism of 'enriched for disease' and the specific distortion of 'sensitivity and specificity' goes beyond what the source supports. Output B maintains strict source fidelity without adding unsupported mechanistic claims.

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.