Evidence table builder online: model-ready, audit-proof evidence tables for HEOR and HTA

Model-ready means every extracted value has a defined field, a unit, a timepoint, and a source location, so a modeller can map it to a parameter without re-reading the paper. Audit-proof means a reviewer can trace any dossier number back to the page it came from and to the person who extracted or verified it. In health economics and health technology assessment, that chain is what carries a hazard ratio, a utility, or a unit cost from a trial table into a model and then into a submission table. One worked row is below, including the field that blocks a hazard ratio from the wrong analysis population.

This page is informational. It is not regulatory advice. Use the current methods guide for the jurisdiction, and your own standard operating procedures, before a number is locked for a submission.

One row: source, parameter, dossier

The numbers in this row are synthetic. They are not results from a published trial. They show the fields a modeller needs on a single relative-effect parameter.

Step	What is recorded
Trial arm	Example RCT. Intention-to-treat (ITT). Drug A 10 mg once daily versus placebo, randomised 1:1 (n = 240 vs 238).
Outcome	Progression-free survival, investigator-assessed. Time origin is randomisation. A higher hazard is worse.
Effect estimate	Hazard ratio 0.72 (95% CI 0.58 to 0.89). Cox model, stratified by prior therapy, with no further covariate adjustment.
Source location	Example trial report, Table 2, page 8. The cell stores that location with the number, not in a separate note that can drift.
Model parameter	Natural log of the hazard ratio = −0.3285. Standard error of the log hazard ratio = (ln 0.89 − ln 0.58) / 3.92 = 0.1092. For probabilistic analysis the draw is normal on the log scale, then exponentiated.
HTA dossier table	Clinical-effectiveness relative-effect row for progression-free survival: ITT hazard ratio, confidence interval, analysis population, and a footnote to Table 2, page 8. In a PBAC submission this trial result is the starting point of the stepped economic table (Guidelines Table 3A.8.1), before any translation step.
Who checked it	Extracted value and page recorded on the row. Reviewer B marked the cell verified on 12 March 2026 after reading Table 2.

The standard-error step follows the Cochrane Handbook, Chapter 6 (Higgins, Li, and Deeks). Section 6.3.2 says to take the natural logs of a ratio measure and its confidence limits first. Section 6.3.1 then uses 3.92 as the divisor for a 95% interval (3.92 = 2 × 1.96). Rounded to four decimal places:

• ln(0.72) = −0.3285
• ln(0.89) = −0.1165
• ln(0.58) = −0.5447
• (−0.1165 − (−0.5447)) / 3.92 = 0.4282 / 3.92 = 0.1092

Exporting the hazard ratio 0.72 into a cell that the model reads as a log hazard ratio is a different error. The parameter field has to name the scale.

Failure mode: hazard ratio from the wrong population

The same synthetic paper also reports a per-protocol hazard ratio of 0.61 (95% CI 0.44 to 0.84) in a footnote to Table 2. The model's baseline progression risk is the ITT Kaplan–Meier curve. Pairing 0.61 with that baseline applies a relative effect to a different set of patients from the set that generated the baseline.

The field that catches it is analysis_population, stored on both the baseline-risk row and the relative-effect row. The check is equality: both must say ITT, or both must name the same prespecified subgroup, before either number is marked verified. A second field, effect_adjustment, records whether the hazard ratio is unadjusted or adjusted and lists the covariates. Chapter 6 of the Cochrane Handbook treats an adjusted estimate as a different extracted result, and says to record the variables used for adjustment.

If two extractors disagree — one copied 0.72 (ITT) and the other copied 0.61 (per protocol) — both numbers are in the paper. That is an interpretation mismatch, not a typo. Use the ladder in Resolving dual-extraction disagreements: keep the value that matches the prespecified population, and keep both source pointers on the audit row.

What an evidence table is here

An evidence table is the study-by-study grid of the variables you will compare or enter into a model; column design is covered in What is an evidence table.

Fields to lock before a number enters the model

Record these on the row. A blank is a failed row, not a default.

• Analysis population (ITT, modified ITT, per protocol, or a named subgroup) and the arm definition (regimen, dose, line of therapy).
• Outcome construct, measurement method, and direction (which way is worse).
• Time origin and timepoint, or follow-up window. Trial follow-up and the model horizon are different fields.
• Effect measure and scale (hazard ratio, or log hazard ratio plus standard error).
• Confidence interval or standard error, and the sample size that belongs to that estimate.
• Adjusted or unadjusted, with the adjustment set if adjusted.
• For costs: unit, currency, and price year. For utilities: instrument, value set, and any mapping method.
• Source location: table, figure, or page.
• Who verified the cell, and when.

Where AI is doing a first pass, the operating rules for which fields a person must still check are in AI as a second extractor in HEOR systematic reviews. Reusing the same field list across runs is what extraction templates are for. Do not rebuild the population and scale fields from memory on each model.

What the jurisdiction manuals require of that row

These are reference-case points from the current public manuals. They are not a complete submission checklist.

NICE. The health technology evaluations manual (PMG36) reference case takes costs from the NHS and personal social services perspective, prefers EQ-5D for health-related quality of life in adults (section 4.3.6), and discounts costs and health effects at 3.5% per year (section 4.5). A non-reference-case analysis has to be justified and shown separately. When effects are extrapolated past the studies, PMG36 section 4.2.24 says that comparing several alternative scenarios is desirable, including a scenario in which benefit stops when treatment stops.

Canada (CDA-AMC). The fourth-edition Canadian guidelines are published on the Canada's Drug Agency site. The working group named on that page includes CADTH health economists. The reference-case perspective is the publicly funded health care payer. Costs and outcomes beyond one year are discounted at 1.5% per year in the reference case, with 0% and 3% as sensitivity analyses.

PBAC. The PBAC Guidelines, section 3A.1, discount costs and outcomes at 5% per year in the base case, and ask for sensitivity analyses at 3.5% and at 0%. Section 3A.8 asks for a stepped table from the clinical data to the modelled incremental cost-effectiveness ratio. A translation of population, time horizon, or outcome has to appear as its own step. The ISPOR-SMDM Modeling Good Research Practices overview (Caro, Briggs, Siebert, and colleagues, Value in Health, 2012) is the cross-jurisdiction good-practice statement: a model input should be transparent enough that someone else can see what was used.

Reporting of the economic evaluation itself follows CHEERS 2022 (Husereau and colleagues, BMJ). The evidence table is the input log behind that report, not a substitute for it.

What Evidence Table Builder stores

Builder takes uploaded PDFs and writes answers into the fields you set for that run. A run accepts up to 10 active extraction fields. Split a wide HEOR schema into passes (for example efficacy, then utilities and costs) or drop fields until the active set fits.

The Excel download has a Results sheet (one row per file, one column per field) and an Audit Trail sheet. Each audit row can carry the verbatim quote, location type, location label, page, confidence, and the flags not_found, ambiguous, and needs_human_review. It also carries verification status, reviewer name, and reviewed-at. Yellow highlights in the PDF mark the quoted passage. They are not a sign that a person has checked it. An answer is verified only when a reviewer sets the status to verified, or to edited and verified. The column list and those status meanings are described in the audit trail feature.

Builder does not calculate the log-hazard-ratio standard error above, does not choose ITT versus per protocol for you, and does not apply a jurisdiction's discount rate. Those stay in the parameter workbook. Formal risk-of-bias tools (RoB 2, ROBINS-I, Newcastle–Ottawa) run in the separate quality assessment workflow, not as columns invented inside an extraction export.

Screening and PRISMA flow stay upstream of this table. Study Screener is one place to keep inclusion decisions. Systematic Review Tools lists search and appraisal tools. Builder starts once the included full texts are the files you upload.

References

• Higgins JP, Li T, Deeks JJ. Chapter 6: Choosing effect measures and computing estimates of effect. In: Cochrane Handbook for Systematic Reviews of Interventions, version 6.5. Chapter 6.
• NICE. NICE health technology evaluations: the manual. Process and methods PMG36, chapter 4.
• Canada's Drug Agency. Guidelines for the Economic Evaluation of Health Technologies: Canada — 4th Edition.
• Pharmaceutical Benefits Advisory Committee. Guidelines, section 3A.1 and section 3A.8.
• Caro JJ, Briggs AH, Siebert U, et al. Modeling good research practices — overview: a report of the ISPOR-SMDM Modeling Good Research Practices Task Force-1. Value in Health. 2012;15(5):796-803. ISPOR.
• Husereau D, Drummond M, Augustovski F, et al. Consolidated Health Economic Evaluation Reporting Standards 2022 (CHEERS 2022). BMJ. 2022;376:e067975. Full text.