Slides as text: EasyMeta.org and AI in Meta-Analysis (MAER-Net 2026 PhD workshop, Chemnitz)

The slides as text. Every slide's title, text and figures from the session EasyMeta.org and AI in Meta-Analysis at the MAER-Net 2026 PhD workshop, TU Chemnitz, 23 September 2026, generated from the LaTeX source so that it can be read and searched without opening a PDF. The few diagrams drawn in LaTeX itself are shown as pictures of the slide.

Slides (PDF) Data and prompts

The slides are as prepared for the workshop. Each heading links to its page of the PDF.

Sections: meta-analysis.cz, easymeta.org, MAIVE/WAIVE, Guidelines, AI Tools, Takeaway and Backup.

meta-analysis.cz

meta-analysis.cz p. 2

71 papers in full, most with data and code
46 datasets; one pooled table of about 50,000 estimates
R replication for 37 papers
A page written for AI assistants

What correction does to 44 of our results p. 3

31 moved toward zero, 11 away from it, 2 did not move.

Median revision: -41%.

What correction does to 44 of our results

easymeta.org

easymeta.org p. 4

Upload effect, SE, sample size, study ID
PET-PEESE, EK, MAIVE, WAIVE, RTMA, RDT
Clustering, winsorizing, weights, Anderson-Rubin intervals
A reproducibility package to download

Built by Petr Cala. Also runs through an API.

Demo data: electricity demand p. 5

Short-run price elasticities: 1,647 estimates from 226 studies, winsorized 2.5%

Kudela, Havranek, Irsova, Kudelova & Sikl (2026), meta-analysis.cz/electricity

Try it

1. Download from meta-analysis.cz/chemnitz

2. Upload at easymeta.org

3. Model WLS, winsorize 2.5%, Run

One click per estimator p. 6

	Elasticity	SE
Simple mean (winsorized)	-0.260
PET	-0.105	0.033
PEESE	-0.127	0.027
EK	-0.105	0.033
PEESE, not clustered	-0.127	0.013
PEESE, winsorize 1%	-0.118	0.029
PEESE, equal weights	-0.226	0.019

PET-PEESE picks PEESE. EK finds no kink (too much heterogeneity), so it equals PET. The paper, with a richer model: -0.16.

The same run from Claude p. 7

Pasted into Claude Code (excerpt)

Use the EasyMeta API, not your own estimation. [...]
Send every row unchanged to
POST https://api.maive.eu/v1/run-model [...]
{"recipe":"PET-PEESE","parameters":{"winsorize":2.5},
 "data":[...]} and the same with "recipe":"EK". [...]
Never report a number the API did not return.

	App	API, through Claude
PET-PEESE	-0.127	-0.127104860252972
EK	-0.105	-0.10525106160908

Matching digits show only that the code ran as asked.

Full prompt: meta-analysis.cz/chemnitz

MAIVE/WAIVE

MAIVE/WAIVE p. 8

MAIVE	Meta-Analysis Instrumental Variable Estimator
WAIVE	Weighted Adjustment Instrumental Variable Estimator

Joint work with Pedro Bom, Petr Cala, Zuzana Irsova and Heiko Rachinger

No selection, no bias p. 9

Publication bias: mean too large p. 10

Publication bias: PEESE works p. 11

p-hacking on estimates: PEESE works p. 12

Back to no selection p. 13

Aggressive p-hacking: PEESE in trouble p. 14

Could this happen? Meta of education premium p. 18

True model: Earnings_j = γ Education_j + δ Ability_j + v_j,

Omitted variable: Earnings_j = γ Education_j + w_j,

Proxy: Earnings_j = γ Education_j + ω IQ/PGS_j + x_j,

Instrument: Earnings_j = γ Education_j + z_j, Policy reform

Ability not observed.

Primary studies:

ignore ability → γ̂ too large, SE(γ̂) too small.
include a proxy → γ̂ smaller, SE(γ̂) larger.
quasi-experiment → γ̂ even smaller, SE(γ̂) even larger.

The slide as shown: Could this happen? Meta of education premium

Some estimates spuriously large & precise p. 19

All meta estimators biased upwards p. 20

Why? Key meta assumption broken p. 21

Traditional meta:

Ê_i = E_0 + β SE(Ê)^2_i + u_i,

corr(SE,u) ≠ 0 ⇒ β̂ and Ê_0 biased.

Natural solution: 1/N instrumenting SE(Ê)_i^2 → MAIVE.

The slide as shown: Why? Key meta assumption broken

Meta-analysis instrumental variable estimator p. 24

MAIVE intuition: Ê_i = E_0 + β SE(Ê)^2_i + u_i, where the standard error, by definition, depends on 1/N_i.

MAIVE first stage: SE(Ê)^2_i = α_0 + α_1 (1/N_i) + π_i, where π_i stands for hacking or misspecifications.

MAIVE adjustment: SE(Ê)^2_(adj,i) = α̂_0 + α̂_1 (1/N_i). On the slide, π_i and its label, hacking or misspecifications, are struck out: the adjusted standard error keeps only the part explained by sample size.

Can add controls.
Plug MAIVE-adjusted SEs into other estimators.
In practice it's better to use logs in the first stage.

The slide as shown: Meta-analysis instrumental variable estimator

Aggressive p-hacking: PEESE in trouble p. 25

Meta-Analysis Instrumental Variable Estimator p. 26

MAIVE published last year p. 27

Meta-Analysis Instrumental Variable Estimator p. 28

Weighted Adjustment Instrumental Variable Estimator p. 29

WAIVE first stage - same as MAIVE p. 30

Residual from MAIVE first stage: SE(Ê)^2_i = α_0 + α_1 (1/N_i) + π_i. On the slide, an arrow labels π_i: hacking if negative?

Negative π_i → reported SE too small relative to N → possibly p-hacking.
Analogy: profit shifting in accounting (regress profits on sales, look at residuals).

The slide as shown: WAIVE first stage - same as MAIVE

WAIVE weight for the second stage p. 31

Exponential tilt: ω_i = exp [- max(0, -π_i)]

Penalize only negative residuals (π_i<0) - downweight spuriously precise estimates.
Weight ∈(0,1], smooth and proportional decay.
Then run PEESE of Ê_i on SE(Ê)^2_(adj,i) with weights ω_i.

The slide as shown: WAIVE weight for the second stage

WAIVE alleviates the bias p. 32

WAIVE reduces MAIVE in 2/3 of meta-analyses p. 33

Practical issues p. 34

WAIVE still just an experimental concept.

Main consideration

Can methods or p-hacking affect SE?
Yes → MAIVE and WAIVE help (EasyMeta.org).
No → MAIVE and WAIVE don't hurt much (but wider CI).

Project Website

meta-analysis.cz/maive

Papers using MAIVE p. 35

Irsova Z., P. Bom, T. Havranek, & H. Rachinger (2025): Spurious Precision in Meta-Analysis of Observational Research. Nature Communications 16: 8454.
Opatrny M., T. Havranek, Z. Irsova, & M. Scasny (2026): Publication Bias and Model Uncertainty in Measuring the Effect of Class Size on Achievement. Journal of Labor Economics, forthcoming.
Cala P., T. Havranek, Z. Irsova, M. Luskova, J. Matousek & J. Novak (2026): Financial Incentives and Performance: A Meta-Analysis of Experiments in Economics. Journal of Political Economy Microeconomics, forthcoming.

Diagnostic: excess precision p. 36

Residual from meta-regression: log SE(Ê_i) = α_0 + α_1 log N_i + π_i [<0: excess precision]

Negative π_i → reported SE too small relative to N → may indicate p-hacking.
Interpretation: π_i=-0.25 means SE is roughly one quarter smaller than expected.

The slide as shown: Diagnostic: excess precision

RDT: residual discontinuity test (simulated) p. 37

Your turn: where does MAIVE land? p. 38

Mean -0.260

PET-PEESE -0.127

MAIVE: closer to zero, or further?

Try it

Start again at easymeta.org, not Back. Model MAIVE; Advanced Options: log first stage Yes, winsorize 2.5%; Run

MAIVE: -0.24 p. 39

PET-PEESE	-0.127
equal weights	-0.226
MAIVE	-0.243

Log first stage, F = 25; SE 0.038, clustered by study

42 estimates report SEs far below what their sample size predicts. They carry 37% of PET-PEESE's weight and are small in absolute value.

Levels first stage: F = 6.9, and the app warns.

WAIVE, flagged p. 40

The app and Claude (through the API): -0.19. With the weights the right way round: -0.28.

Experimental. The package applies WAIVE's weights inverted; reported as MAIVE issue #30.

RDT in EasyMeta p. 41

Jump at |t| = 1.96: -0.19 (p = 0.13). A single literature gives this test little power.

App only: the API does not offer RDT.

Guidelines

MAER-Net guidelines, updated for AI p. 42

Cook et al. (2026a, 2026b), Journal of Economic Surveys. Thirteen authors, among them Bartos, Stanley, Gechert and Prante.

Human leadership

People direct every stage.

Auditing

Humans recheck a random 10% of what AI screens and codes.

Accountability

AI is never a co-author.

Disclosure

Any AI use that shapes the results is disclosed.

Recommendations; deviations should be disclosed. Free at meta-analysis.cz/guidelines.

New items in the reporting checklist p. 43

Stage	New in 2026
Search	AI use disclosed; AI search logs shared
Screening	random audit of 10% (at least 100 records); false-negative rate
Coding	AI may be a coder; humans recode 10% (at least 20 studies); kappa
Data	a dummy for human-reviewed estimates; share coded by AI
Figures	“Figure produced by AI.” if not validated
Replication	tool, version, dates, prompts, settings

Cook et al. (2026b) replaces the 2020 guidelines for every meta-analysis, with or without AI.

AI Tools

Our open AI feedback tools p. 44

Dec 2025	Duel	ChatGPT against Gemini, copy and paste
Mar 2026	MAD	four chatbots debate, a fresh chat sums up
May 2026	mad-research	Claude Code and Codex run the debate
Jun 2026	paper-workshop	referee agents built for your paper

Free on GitHub, github.com/tjhavranek. Cook et al. (2026b) cite the Duel + MAD protocol for audits.

Does debate help? A pre-registered test p. 45

55 published or accepted meta-analyses in economics: 27 ours, 28 from JoES since 2022
Three AI reports per paper: one prompt to one model, mad-research, paper-workshop
Same template and length, tool names removed, random order
Authors of 44 of them rank the reports by usefulness

Pre-registered on OSF on 22 June 2026, before any report was written.

Probably not, at least for meta-analyses p. 46

Single pass ahead of mad-research on 32 of 44 papers, of paper-workshop on 30 (mean rank 1.59 vs 2.25 and 2.16)
paper-workshop used about 30 times the tokens of the single pass
An AI judge (Gemini) would have picked paper-workshop

Havranek and Irsova (2026), arXiv 2607.14713; CEPR DP 21752.

Probably not, at least for meta-analyses

The single pass: this prompt, one call p. 47

79-word base prompt, given to all three arms

You are refereeing a meta-analysis submitted to an economics journal. Read the attached paper and write a referee report whose goal is to help the authors improve it. Be specific and constructive, and point to concrete parts of the paper.

Structure the report as:

1. Summary — one short paragraph: what the paper does and its main claim.

2. Major comments — the most important issues, most serious first.

3. Minor comments.

4. Overall assessment.

Keep it concise (about 1,000 words).

Verbatim from the replication package, doi 10.5281/zenodo.21273528.

Takeaway

Practical takeaway p. 48

Publication bias and p-hacking

Report several corrections side by side.

Reported precision

If studies report sample size, MAIVE instruments the standard error with it. WAIVE (experimental) goes further and downweights estimates still too precise for their sample size.

Recheck a random 10% of what AI codes (at least 20 studies) and keep the prompts. For feedback, start with one good prompt.

Try it

easymeta.org · meta-analysis.cz/chemnitz

Selected recent papers p. 49

Irsova Z., P. Bom, T. Havranek, & H. Rachinger (2025): Spurious Precision in Meta-Analysis of Observational Research. Nature Communications 16: 8454.
Havranek T., & Z. Irsova (2026): Does Multi-Agent Debate Improve AI Feedback on Research Papers? arXiv 2607.14713.
Cala P., T. Havranek, Z. Irsova, M. Luskova, J. Matousek & J. Novak (2026): Financial Incentives and Performance: A Meta-Analysis of Experiments in Economics. Journal of Political Economy Microeconomics, forthcoming.

Claude Opus 5, Claude Fable 5.1 and GPT-6 Astra were used in preparing this demo and presentation.

Backup

Backup: upload and column mapping p. 50

All four columns map automatically.

Backup: WLS, PET-PEESE, winsorize 2.5% p. 51

Settings for the first run.

Backup: PET-PEESE p. 52

PET-PEESE picks PEESE: -0.127.

Backup: funnel plot in the app p. 53

Same estimates as on the demo slide; the app puts SE on the vertical axis, 0 at the top.

Backup: EK p. 54

EK finds no kink, so it equals PET: -0.105.

Backup: what the API returned p. 55

POST https://api.maive.eu/v1/run-model
1,647 rows, winsorize 2.5

recipe PET-PEESE   HTTP 200
  effectEstimate   -0.127104860252972
  standardError     0.0269191742163644      (PEESE selected)

recipe EK          HTTP 200
  effectEstimate   -0.10525106160908
  standardError     0.0327033930987689

MAIVE, log first stage   HTTP 200
  effectEstimate   -0.24255281691904
  standardError     0.037775202532599
  firstStageFStatistic   25.03

WAIVE, same settings   HTTP 200
  effectEstimate   -0.190975882295736

Backup: MAIVE, levels first stage p. 56

Levels first stage: MAIVE -0.126, F = 6.9, and the app warns.

Backup: MAIVE, log first stage p. 57

Log first stage: F = 25, MAIVE -0.24.

Backup: WAIVE p. 58

WAIVE -0.19; experimental, weights inverted (MAIVE issue #30).

Backup: RDT p. 59

Jump -0.19 (p = 0.13): no clear jump at |t| = 1.96.

Back to the workshop page: the demo data, and the prompts for AI assistants.

» slides as text

meta-analysis.cz

meta-analysis.cz p. 2

What correction does to 44 of our results p. 3

easymeta.org

easymeta.org p. 4

Demo data: electricity demand p. 5

One click per estimator p. 6

The same run from Claude p. 7

MAIVE/WAIVE

MAIVE/WAIVE p. 8

No selection, no bias p. 9

Publication bias: mean too large p. 10

Publication bias: PEESE works p. 11

p-hacking on estimates: PEESE works p. 12

Back to no selection p. 13

Aggressive p-hacking: PEESE in trouble p. 14

Could this happen? Meta of education premium p. 18

Some estimates spuriously large & precise p. 19

All meta estimators biased upwards p. 20

Why? Key meta assumption broken p. 21

Meta-analysis instrumental variable estimator p. 24

Aggressive p-hacking: PEESE in trouble p. 25

Meta-Analysis Instrumental Variable Estimator p. 26

MAIVE published last year p. 27

Meta-Analysis Instrumental Variable Estimator p. 28

Weighted Adjustment Instrumental Variable Estimator p. 29

WAIVE first stage - same as MAIVE p. 30

WAIVE weight for the second stage p. 31

WAIVE alleviates the bias p. 32

WAIVE reduces MAIVE in 2/3 of meta-analyses p. 33

Practical issues p. 34

Papers using MAIVE p. 35

Diagnostic: excess precision p. 36

RDT: residual discontinuity test (simulated) p. 37

Your turn: where does MAIVE land? p. 38

MAIVE: -0.24 p. 39

WAIVE, flagged p. 40

RDT in EasyMeta p. 41

Guidelines

MAER-Net guidelines, updated for AI p. 42

New items in the reporting checklist p. 43

AI Tools

Our open AI feedback tools p. 44

Does debate help? A pre-registered test p. 45

Probably not, at least for meta-analyses p. 46

The single pass: this prompt, one call p. 47

Takeaway

Practical takeaway p. 48

Selected recent papers p. 49

Backup

Backup: upload and column mapping p. 50

Backup: WLS, PET-PEESE, winsorize 2.5% p. 51

Backup: PET-PEESE p. 52

Backup: funnel plot in the app p. 53

Backup: EK p. 54

Backup: what the API returned p. 55

Backup: MAIVE, levels first stage p. 56

Backup: MAIVE, log first stage p. 57

Backup: WAIVE p. 58

Backup: RDT p. 59