[
  {
    "date": "2026-08-10",
    "datetime": "2026-08-10 08:22:32",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7492465389231206400",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-08-10_p1_1.png"
    ],
    "image_size": [
      [
        696,
        626
      ]
    ],
    "image_alt": [
      "Title block and abstract of the paper “Financial Incentives and Performance: A Meta-Analysis of Experiments in Economics” by Petr Cala, Tomas Havranek, Zuzana Irsova, Martina Luskova, Jindrich Matousek and Jiri Novak, in the Journal of Political Economy Microeconomics."
    ],
    "comment_links": [
      "https://www.journals.uchicago.edu/doi/10.1086/743543",
      "https://meta-analysis.cz/incentives/"
    ],
    "slug": "2026-08-10-financial-incentives-and-performance",
    "anchor": "2026-08-10",
    "text": "🔎 New in the Journal of Political Economy Microeconomics: a meta-analysis of the effect of financial incentives on performance.\n\nI teach Econ 101, and the first thing I tell students is that economics is all about incentives. Well, financial incentives do not automatically improve performance, at least in economics experiments.\n\nIn fact, the mean effect of financial incentives, corrected for publication bias and p-hacking, is close to zero in most field settings. Why could that be?\n\n1️⃣ First, money could crowd out the enjoyment we otherwise derive from a task (\"intrinsic motivation\"), though our results don't fully support this explanation.\n\n2️⃣ Second, economists like to be original, so they often focus on \"funny,\" unintended effects of incentives in weird settings and at small stakes. That's what we see in a meta-analysis.\n\nDoes that mean financial incentives don't work in practice? One of our referees was late with his report. He (yes, he signed his report) assured us he would have submitted it on time if the journal had offered him a million dollars.\n\nWe shouldn't assume that financial incentives automatically make people more productive. As a rule of thumb, I like Uri Gneezy and Aldo Rustichini's \"pay enough or don't pay at all.\" I also highly recommend Uri's book on incentives, Mixed Signals, for the full picture.\n\n📄 Paper available here: https://www.journals.uchicago.edu/doi/10.1086/743543\n🖥️ One-click replication package available at meta-analysis.cz: https://meta-analysis.cz/incentives/\n\nJoint work with Petr Čala, Tomas Havranek, Martina Lušková, Jindřich Matoušek, and Jiri Novak."
  },
  {
    "date": "2026-07-16",
    "datetime": "2026-07-16 16:15:33",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3AugcPost%3A7483554931325542401",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-07-16_p2_1.jpeg",
      "2026-07-16_p2_2.jpeg"
    ],
    "text": "Does multi-agent debate improve AI feedback on research papers?\n\nNot in our experiment.\n\nWe just posted a pre-registered study in which authors ranked three AI reports on their own paper. The reports were blinded and came from three setups: a single prompt and two multi-agent tools. We expected the multi-agent tools to win.\n\nWe find:\n\n1) Multi-agent debate does not help here. The single prompt beat both multi-agent tools, even though one of them spent about 30x the tokens.\n\n2) If an independent AI model ranked the reports in the authors' place, it would put the most expensive multi-agent tool first.\n\n3) Authors who recalled their real journal referee feedback usually ranked it above all the AI reports. In contrast, the AI judges almost always ranked the human feedback last.\n\nI was also surprised by how little the author and AI rankings agree (correlation 0.14). Authors seem to have actually read the AI reports (!) and thought about them, not just used a chatbot to rank them.\n\nThanks to all 47 authors who participated!\n\nPaper: https://lnkd.in/d7YsNxWc\n\nBoth multi-agent tools are open source:\nhttps://lnkd.in/dtGSi8zP\nhttps://lnkd.in/dk5NQpcE",
    "image_size": [
      [
        800,
        1035
      ],
      [
        800,
        678
      ]
    ],
    "image_alt": [
      "Title page of the working paper “Does Multi-Agent Debate Improve AI Feedback on Research Papers?” by Tomas Havranek and Zuzana Irsova, dated 16 July 2026.",
      "Chart: author mean rank plotted against tokens per paper on a log scale. The single pass ranks best at roughly 25k tokens, while mad-research and paper-workshop rank worse despite spending about 200k and 800k tokens."
    ],
    "comment_links": [
      "https://meta-analysis.cz/debate",
      "https://cepr.org/publications/dp21752",
      "https://github.com/tjhavranek/mad-research",
      "https://github.com/tjhavranek/paper-workshop",
      "https://doi.org/10.17605/OSF.IO/E6XGW",
      "https://osf.io/7nfyb",
      "https://doi.org/10.5281/zenodo.21273528",
      "https://arxiv.org/abs/2607.14713"
    ],
    "slug": "2026-07-16-does-multi-agent-debate-improve-ai-feedback",
    "link_map": {
      "https://lnkd.in/d7YsNxWc": "https://meta-analysis.cz/debate",
      "https://lnkd.in/dtGSi8zP": "https://github.com/tjhavranek/mad-research",
      "https://lnkd.in/dk5NQpcE": "https://github.com/tjhavranek/paper-workshop"
    },
    "anchor": "2026-07-16"
  },
  {
    "date": "2026-07-08",
    "datetime": "2026-07-08 11:39:51",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7480586444739375104",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-07-08_p3_1.jpeg"
    ],
    "text": "This week I have been at Stanford, spending my days working and my evenings at Nitoboxing in Palo Alto. Honestly, it was exactly what I needed!\n\nResearch is intense, travel tiring, and sometimes the best way to get your head straight is not another coffee, but gloves and sweat.\n\nHuge thanks to Mark and the whole team for the warm welcome, great energy, and for helping me reset after long days.",
    "image_size": [
      [
        800,
        481
      ]
    ],
    "image_alt": [
      "Two photographs from the Nitoboxing gym in Palo Alto: Zuzana Havránková standing with a trainer beside the ring, and the gym's chalkboard sign reading “Boxing — it's cheaper than therapy”."
    ],
    "comment_links": [
      "https://www.instagram.com/nito_boxing_paloalto/"
    ],
    "slug": "2026-07-08-stanford-and-nitoboxing",
    "anchor": "2026-07-08"
  },
  {
    "date": "2026-07-04",
    "datetime": "2026-07-04 16:54:11",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7479216000241033216",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-07-04_p4_1.jpeg"
    ],
    "text": "Proud to be affiliated with Stanford METRICS. Few groups have done more for reproducibility and research integrity than John Ioannidis and his team, and it's a pleasure to meet in person some of the postdocs behind so much of the work: Alejandro Sandoval Lentisco, Quentin Loisel, and Sarah Tanveer.\n\nHappy Fourth of July, and happy 250th, America!!",
    "image_size": [
      [
        800,
        450
      ]
    ],
    "image_alt": [
      "Group selfie of six people around an outdoor table on a sunny campus terrace, lunch containers and drinks on the table, trees and a lawn behind them."
    ],
    "slug": "2026-07-04-affiliated-with-stanford-metrics",
    "anchor": "2026-07-04"
  },
  {
    "date": "2026-06-23",
    "datetime": "2026-06-23 12:20:50",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7475160943409135616",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-06-23_p5_1.jpeg"
    ],
    "text": "If you plan to work on your ERC proposal this summer, the following tool might help you:\n\nhttps://lnkd.in/gAN6S747\n\nIt checks your draft at various stages against the official ERC evaluation criteria and rules, flags the routine weak spots, gives you tips, etc. The idea is to clear the easy problems before you ask humans, not to replace them.\n\nJust read the privacy note first: use a paid model with training turned off, or don't paste anything sensitive.\n\nWe built this with Tomas Havranek, who was on the ERC Advanced economics panel in 2020 and is involved in the expert group supporting ERC applicants in Czechia.\n\nYou might also check our related tools: mad-research (https://lnkd.in/dtGSi8zP), which stress-tests your paper or proposal with a Claude/Codex debate, and paper-workshop (https://lnkd.in/dk5NQpcE), which simulates an expert workshop on your stuff.\n\nIf you have any comments or feedback, we'll be happy to incorporate them into the tools.",
    "image_size": [
      [
        800,
        161
      ]
    ],
    "image_alt": [
      "Screenshot of a repository page headed “erc-ai-feedback”, describing a small package that gives ERC Starting and Consolidator Grant applicants a rubric-based pre-review of a draft proposal in one chat session with a frontier model, to clear routine structural problems before workshop time is spent on them. It notes that it does not replace human review."
    ],
    "slug": "2026-06-23-a-tool-for-your-erc-proposal",
    "link_map": {
      "https://lnkd.in/gAN6S747": "https://github.com/tjhavranek/erc-ai-feedback",
      "https://lnkd.in/dtGSi8zP": "https://github.com/tjhavranek/mad-research",
      "https://lnkd.in/dk5NQpcE": "https://github.com/tjhavranek/paper-workshop"
    },
    "anchor": "2026-06-23"
  },
  {
    "date": "2026-06-10",
    "datetime": "2026-06-10 12:01:57",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7470445147063803904",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-06-10_p6_1.jpeg"
    ],
    "text": "Imagine a panel of the world's leading experts, assembled for your research paper, arguing it out from rival schools and then revising it themselves.\n\nWell, that's not possible. But we built an approximation with Claude Code, and it now uses the full power of Anthropic's Mythos-class model, Claude Fable 5.\n\nBelow is the skill, open and free. You give it your paper, ideally with the data and draft code. You get a stress test of your paper + a revision in track changes + a replication package. We ran it end to end on colleagues' papers and our own paper accepted at JPE Micro, Stata and R re-runs included. Enjoy:\n\nhttps://lnkd.in/dk5NQpcE",
    "image_size": [
      [
        800,
        189
      ]
    ],
    "image_alt": [
      "Screenshot of a page headed “CRUCIBLE — the paper-workshop skill”: a panel of leading experts assembled for one specific paper, arguing it out from rival schools and then rebuilding it themselves, re-running the author's code. It promises a tracked redline, a clean draft and a replication package."
    ],
    "slug": "2026-06-10-a-panel-of-experts-for-your-paper",
    "link_map": {
      "https://lnkd.in/dk5NQpcE": "https://github.com/tjhavranek/paper-workshop"
    },
    "anchor": "2026-06-10"
  },
  {
    "date": "2026-06-04",
    "datetime": "2026-06-04 07:30:20",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3AugcPost%3A7468202463888609280",
    "reshare_with_comment": false,
    "lang": "cs",
    "images": [
      "2026-06-04_game.png"
    ],
    "text": "Nesehnali jste už lístky na Smetanova Litomyšl, ale přesto chcete zakusit genius loci perly východních Čech? Snadná pomoc!\n\nDěti připravily festivalové vydání své závodní hry z Litomyšle:\n\nhttps://lnkd.in/gvEF-zD7\n\nZávodíte v litomyšlských ulicích s dalším hráčem na počítači, nebo na mobilu s botem. Můžete střílet, sbírat lepší zbraně, ničit budovy (např. školu).\n\nPozor na bomby a zombie Bedřichy!\n\nHru postavily děti samy v zimě s Claude Code za několik hodin (ač pak trvaly na mnoha hodinách testování 😉). Jasně, vypadá jako z 80. let. Ale je to jen malá ukázka toho, co současné agentní AI systémy dokážou, a není to zdaleka jen o kódování.\n\nPokud jste Claude Code nebo Codex ještě nezkoušeli, určitě stojí za to nainstalovat, ať už se živíte čímkoli. Nejtěžší je překonat počáteční strach. Ono je to nakonec jednoduché.\n\nDetaily hry tady:\n\nhttps://lnkd.in/gVmACuyi",
    "image_size": [
      [
        1280,
        715
      ]
    ],
    "image_alt": [
      "Snímek závodní hry: mapa Litomyšle se zámkem, Smetanovým náměstím a gymnáziem, dva závodící hráči a hořící budovy v apokalyptickém režimu hry."
    ],
    "slug": "2026-06-04-smetanova-litomysl-genius-loci",
    "link_map": {
      "https://lnkd.in/gvEF-zD7": "https://tjhavranek.github.io/race/",
      "https://lnkd.in/gVmACuyi": "https://github.com/tjhavranek/race"
    },
    "anchor": "2026-06-04"
  },
  {
    "date": "2026-05-30",
    "datetime": "2026-05-30 06:57:28",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7466382253208645634",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-05-30_p8_1.jpeg"
    ],
    "text": "Do you know that Claude Code can use Codex to stress test its work?\n\nI had both installed and used them interchangeably: for some tasks, one seems to be better, and vice versa. Also, sometimes I hit usage limit in one, so I continue in the other.\n\nThis can be automated easily: work from Claude Code and call Codex when needed. Here is the Claude skill, applied to stress-testing research:\n\nhttps://lnkd.in/dtGSi8zP\n\nThis builds on our previous manual protocols for using different AI models to test research ideas or papers:\n\nhttps://lnkd.in/dRwKz63g\n\nDoes the skill work for you? What should we change? Should we add Gemini?",
    "image_size": [
      [
        800,
        602
      ]
    ],
    "image_alt": [
      "Screenshot showing Claude Code invoking Codex to stress-test its own work, as described in the post."
    ],
    "slug": "2026-05-30-claude-code-with-codex-stress-test",
    "link_map": {
      "https://lnkd.in/dtGSi8zP": "https://github.com/tjhavranek/mad-research",
      "https://lnkd.in/dRwKz63g": "https://github.com/tjhavranek/research-audit-duel-protocol"
    },
    "anchor": "2026-05-30"
  },
  {
    "date": "2026-05-13",
    "datetime": "2026-05-13 05:39:56",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7460202147213791232",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-05-13_p9_1.jpeg"
    ],
    "text": "Reporting Guidelines for Meta-Analysis in Economics, updated for AI, just published in the Journal of Economic Surveys:\nhttps://lnkd.in/d_BciGAd\n\nTwo practical points I would emphasize (my personal opinion), beyond the reporting checklist itself:\n\n1️⃣ If you use AI for searching, screening, or coding, don't rely on a single model. Use meta-analysis thinking: each model is trained differently and on different data (think Claude vs. Grok). Even if one model strictly dominates, there will be useful information in the others, and you need to stress-test your favorite model brutally regardless. We have developed a simple Research Audit Protocol based on Multi-Agent Debate (MAD) for exactly this:\nhttps://lnkd.in/d63DFbPz\n\n2️⃣ These guidelines intentionally do not recommend any particular methodology. We do so in our 2024 method guidelines (https://lnkd.in/dPcDR7Gr). Brief update: I think the baseline meta-analysis technique is now Robust Bayesian Meta-Analysis (RoBMA) by František Bartoš, Maximilian Maier, and Eric-Jan Wagenmakers -- a principled way to average over various bias-correction methods. But these methods don't address p-hacking, so RoBMA should be complemented with MAIVE (easy to apply via https://easymeta.org) and RTMA (Maya Mathur).\n\nThe updated reporting guidelines were led by Nikolai Cook and co-authored with František Bartoš, Pedro Bom, Sebastian Gechert, Klára Kantová, Jerome Geyer-Klingeberg, Dr.-Ing., Tomas Havranek, Martina Lušková, Matej Opatrny, Franz Prante, Heiko Rachinger, and Tom Stanley.",
    "image_size": [
      [
        800,
        344
      ]
    ],
    "image_alt": [
      "Journal of Economic Surveys article header, open access: “Reporting Guidelines for Meta-Analysis in Economics — Updated for AI”, by Nikolai Cook, František Bartoš, Pedro R. D. Bom, Sebastian Gechert, Klára Kantová, Jerome Geyer-Klingeberg, Tomáš Havránek, Zuzana Irsova, Martina Luskova and others. First published 12 May 2026."
    ],
    "comment_links": [
      "https://github.com/tjhavranek/research-audit-duel-protocol"
    ],
    "slug": "2026-05-13-reporting-guidelines-updated-for-ai",
    "link_map": {
      "https://lnkd.in/d_BciGAd": "https://onlinelibrary.wiley.com/doi/10.1111/joes.70116",
      "https://lnkd.in/d63DFbPz": "https://github.com/tjhavranek/research-audit-duel-protocol/",
      "https://lnkd.in/dPcDR7Gr": "https://onlinelibrary.wiley.com/doi/full/10.1111/joes.12595"
    },
    "anchor": "2026-05-13"
  },
  {
    "date": "2026-04-28",
    "datetime": "2026-04-28 15:29:43",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7454914753266692097",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-04-28_p10_1.jpeg"
    ],
    "text": "Will our results change if we redo the meta-analysis from scratch?\n\nWe just pre-registered a big revision of our meta-analysis on beauty and professional success:\n\nhttps://lnkd.in/dBSAB5N2\n\nThe current paper follows standards commonly used in economics meta-analysis. For the revision we decided to do the search and data collection differently, following multidisciplinary systematic-review practice: a librarian-designed multi-database search, dual screening, coding reliability checks, PRISMA documentation, and a full audit trail.\n\nA nice side effect is that this becomes a natural experiment on our own work. Honestly, I'm curious how much it will move the results.\n\nThe amazing Martina Lušková joined the team to help lead study selection and coding, and we are working with a librarian at the University of Amsterdam on the search strategy.\n\nIn the current version we find that the effect of beauty on earnings is smaller than commonly thought once you correct for publication bias and p-hacking, and smaller still when more weight is given to studies that control for cognitive ability. The one clear exception is sex workers. For politicians, the beauty premium mostly goes away after correction.\n\nTo make the comparison fully transparent, we also uploaded the current paper, data, and code to the registration.\n\nWe should do much more pre-registration in observational research. It doesn't fully prevent p-hacking, but it helps a lot and the cost is low.\n\nCo-authored with Tomas Havranek, František Bartoš, Xenia Bortnikova, and Martina Lušková",
    "image_size": [
      [
        800,
        520
      ]
    ],
    "image_alt": [
      "First page of a pre-registered protocol headed “Systematic review — update protocol: Meta-Analysis of Field Studies on Beauty and Professional Success”. A table gives the protocol type, describing substantial revisions to the literature search, screening, coding, analysis and reporting of the previous version, and records the registration on the Open Science Framework."
    ],
    "comment_links": [
      "https://doi.org/10.17605/OSF.IO/B3D7W"
    ],
    "slug": "2026-04-28-redoing-a-meta-analysis-from-scratch",
    "link_map": {
      "https://lnkd.in/dBSAB5N2": "https://doi.org/10.17605/OSF.IO/B3D7W"
    },
    "anchor": "2026-04-28"
  },
  {
    "date": "2026-04-22",
    "datetime": "2026-04-22 05:32:10",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7452590050799853568",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-04-22_p11_1.jpeg"
    ],
    "text": "Not an easy task, but in a new note just out in the Journal of Economic Surveys we try to set a basic floor on the use of AI in meta-analysis.\n\nhttps://lnkd.in/dZNGRfJb\n\nShort version:\n\n🔹 Human leadership. Humans direct the search, coding, and analysis, and record where they override the AI.\n\n🔹 Human accountability. AI cannot be a co-author. If your name is on the paper, the errors are yours.\n\n🔹 Human auditing. AI can serve as one of the coders, as long as humans audit at least 10% of screening records and 10% of coded studies (or 100 and 20, whichever is larger), and report a measure of agreement.\n\n🔹 Human disclosure. Anything that shapes search, screening, coding, analysis, or conclusions should be disclosed, with prompts and model versions saved.\n\nThe effort was led by the amazing Nikolai Cook. It will be periodically updated at maer-net.org.\n\nFor the full discussion of how we agreed on these guidelines (scroll down to comments):\nhttps://lnkd.in/dME6se-W",
    "image_size": [
      [
        759,
        334
      ]
    ],
    "image_alt": [
      "Journal of Economic Surveys article header, open access: “Guidance for the Use of AI in the Meta-Analysis of Economics Research”, by Nikolai Cook, František Bartoš, Pedro R. D. Bom, Sebastian Gechert, Klára Kantová, Jerome Geyer-Klingeberg, Tomáš Havránek, Zuzana Irsova, Martina Luskova, Matěj Opatrný, Franz Prante, Heiko J. Rachinger and T. D. Stanley. First published 21 April 2026."
    ],
    "slug": "2026-04-22-note-in-journal-of-economic-surveys",
    "link_map": {
      "https://lnkd.in/dZNGRfJb": "https://onlinelibrary.wiley.com/doi/10.1111/joes.70105",
      "https://lnkd.in/dME6se-W": "https://www.maer-net.org/post/developing-guidelines-for-the-use-of-ai-in-meta-analysis-of-economics-research-guai-maer-and"
    },
    "anchor": "2026-04-22"
  },
  {
    "date": "2026-04-02",
    "datetime": "2026-04-02 08:52:45",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7445392769788968960",
    "reshare_with_comment": true,
    "lang": "en",
    "images": [],
    "image_size": [],
    "image_alt": [],
    "reshare_of": {
      "label": "She reshared the IES announcement of this Nature study, which she co-authored",
      "url": "https://doi.org/10.1038/s41586-026-10251-x",
      "url_label": "Brodeur, A., Mikola, D., Cook, N. et al., “Reproducibility and robustness of economics and political science research”, Nature 652, 151–156 (2026)"
    },
    "text": "It was fun to be (a very small) part of the amazing team led by Abel Brodeur. The numbers look great, but note that we replicate papers in journals that have mandatory data and code sharing. So: share your data and code! And while you're at it, pre-register your papers and upload pre-analysis plans. This is the way!",
    "slug": "2026-04-02-brodeur-team-reproducibility-numbers",
    "anchor": "2026-04-02"
  },
  {
    "date": "2026-03-28",
    "datetime": "2026-03-28 12:52:04",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7443641057592033280",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [],
    "text": "Top 7 non-fiction books I recommend. 📚\n\nNot an endorsement of everything in these books, but I learned a great deal about economics, history, and politics.\n\n1️⃣ Bourgeois Equality (McCloskey): How the West got rich.\n2️⃣ In the Garden of Beasts (Larson): An American family in 1930s Berlin, watching a society die.\n3️⃣ Recession (Tyler Goodspeed): Why economies shrink. Excellent new book.\n4️⃣ Cicero (Everitt): The life of arguably the greatest statesman the West ever produced.\n5️⃣ Red Dawn Over China (Dikötter): How China became communist.\n6️⃣ The Lives of the Caesars (Suetonius): The most readable classical text I know. Reads like a novel.\n7️⃣ The Fiscal Theory of the Price Level (John H. Cochrane): A bold, rigorous book on modern macroeconomics.\n\nWhat's the best non-fiction book you've read this year?",
    "image_alt": [],
    "slug": "2026-03-28-top-7-non-fiction-books",
    "anchor": "2026-03-28"
  },
  {
    "date": "2026-03-19",
    "datetime": "2026-03-19 07:22:32",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7440296636783960064",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [],
    "text": "Two AI models dueling worked. Four models debating works better.\n\nWe updated our research audit protocol. The new version (MAD v2.0) uses ChatGPT, Claude, Gemini, and Grok in structured adversarial rounds:\n\n1️⃣ Independent critique — no model sees the others, every claim grounded in the document.\n2️⃣ Cross-examination — each model attacks the weakest peer arguments.\n3️⃣ Final arbiter synthesizes what survived.\n\nNo code needed. Copy-paste prompts. Free model versions work (just register for each model).\n\nAdvanced users: the entire workflow can be automated via the models' APIs using frameworks like AutoGen, LangGraph, or CrewAI.\n\nUse it for high-stakes documents: stress-testing your papers, grant proposals, referee reports.\n\nProtocol (GitHub): 👉 https://lnkd.in/dRwKz63g",
    "image_alt": [],
    "slug": "2026-03-19-four-models-debating-works-better",
    "link_map": {
      "https://lnkd.in/dRwKz63g": "https://github.com/tjhavranek/research-audit-duel-protocol"
    },
    "anchor": "2026-03-19"
  },
  {
    "date": "2026-03-11",
    "datetime": "2026-03-11 08:36:32",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7437416155155177472",
    "reshare_with_comment": false,
    "lang": "cs",
    "images": [
      "2026-03-11_screening.jpg"
    ],
    "text": "Doporučuji všem rodičům včas testovat děti na cukrovku. Nám to bohužel vyšlo jednou pozitivní, ale díky tomu testu + moderní medicíně můžete nástup nemoci oddálit o několik let.\n\nVčasný test = šance na roky navíc bez inzulinu.\n\nDíky Barbora Berka a Natálie Chrástecká za skvělý přístup v Motole.\n\nDíky Jan Hrušovský za hezký rozhovor s mým manželem, který poprvé na kameru mluvil o něčem jiném než o ekonomii (což je většinou velmi ... poutavé 😉 ).\n\nhttps://lnkd.in/duWtcSbA",
    "image_alt": [
      "Úvodní záběr z videa: Tomáš Havránek a moderátor Jan Hrušovský stojí v nahrávacím studiu podcastu, přes obrázek je nápis „Screening cukrovky — je lepší, když o nemoci víte“."
    ],
    "image_size": [
      [
        1280,
        720
      ]
    ],
    "reshare_of": {
      "label": "Odkaz z příspěvku vede na rozhovor",
      "url": "https://www.youtube.com/watch?v=FQMmJGOw_Ik",
      "url_label": "Screening cukrovky — je lepší, když o nemoci víte (YouTube)"
    },
    "slug": "2026-03-11-testujte-deti-na-cukrovku",
    "link_map": {
      "https://lnkd.in/duWtcSbA": "https://www.youtube.com/watch?v=FQMmJGOw_Ik"
    },
    "anchor": "2026-03-11"
  },
  {
    "date": "2026-03-09",
    "datetime": "2026-03-09 15:51:18",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7436800791929151488",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-03-09_p16_1.jpeg"
    ],
    "text": "Meta-analyses should try to correct not just for publication bias, but also for p-hacking.\n\nSome estimates are more likely to be reported than others, so every good summary of research should correct for this publication bias. In case you're wondering — yes, this can be done, there are dozens of methods and decades of research on this. There is even a great way to put these different correction techniques together: see RoBMA by František Bartoš and colleagues.\n\nThe problem is that these techniques assume that the reported estimates are individually unbiased. This is a strong assumption, as researchers can tweak models (consciously or unconsciously) to get more \"sensible\" results. This is called p-hacking. Most of us do it.\n\nIn our recent Nature Communications paper we show that under some forms of p-hacking, classical models correcting for publication bias can actually be more biased than a simple average of published estimates.\n\nAs far as I know, there are only two meta-analysis corrections for (some forms of) p-hacking: our MAIVE (from the Nature Comms paper) and RTMA by Maya Mathur. If you know of more, please let me know in the comments!\n\nIf you want to see MAIVE and RTMA applied, take a look at our meta-analysis of the beauty premium. Spoiler: apart from the sex industry, beauty doesn't matter much in the labor market.",
    "image_size": [
      [
        800,
        572
      ]
    ],
    "image_alt": [
      "Funnel plot of estimates of the beauty effect on earnings. The bulk of estimates cluster near zero; a red line marks the mean of 4.3 per cent, and estimates for sex workers, shown separately in red, sit well to the right of it."
    ],
    "comment_links": [
      "https://www.easymeta.org/",
      "https://www.nature.com/articles/s41467-025-63261-0",
      "https://meta-analysis.cz/beauty",
      "https://fbartos.github.io/RoBMA/",
      "https://onlinelibrary.wiley.com/doi/full/10.1002/jrsm.1701"
    ],
    "slug": "2026-03-09-correcting-for-p-hacking-too",
    "anchor": "2026-03-09"
  },
  {
    "date": "2026-01-31",
    "datetime": "2026-01-31 13:55:50",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3AugcPost%3A7423363383665508352",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-01-31_game.png"
    ],
    "text": "A small experiment at home: our kids (8-12) were sick and bored, so we introduced them to Claude Code.\n\nTwo hours later, they had a working browser racing game set in our hometown. They did almost everything themselves; we helped with publishing it on GitHub.\n\nThe graphics are… charmingly 1980s. But it runs smoothly and it’s fun.\n\nTakeaway for me: if you’re a researcher, it’s worth setting aside an hour to play with these tools -- you might be surprised what you can do now.\n\n(Link to the game in the first comment!)\n\n#ResearchTools #ClaudeCode #Litomysl",
    "image_size": [
      [
        1280,
        715
      ]
    ],
    "image_alt": [
      "Screenshot of the racing game the children built: a top-down map of Litomysl with the chateau, Smetana square and the grammar school as landmarks, two players racing, and several buildings on fire in the game's apocalyptic mode."
    ],
    "comment_links": [
      "https://tjhavranek.github.io/race/"
    ],
    "slug": "2026-01-31-kids-and-claude-at-home",
    "anchor": "2026-01-31"
  },
  {
    "date": "2026-01-27",
    "datetime": "2026-01-27 07:18:44",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3AugcPost%3A7421813900372918272",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2026-01-27_easymeta.png"
    ],
    "text": "A browser-based tool for correcting publication bias and p-hacking.\n\nWe built EasyMeta.org to make bias-corrected meta-analysis easier to run. Upload your dataset and benchmark your conclusions against modern corrections -- directly in your browser.\n\nWhy: Selective reporting can distort what enters the literature and how results are reported, but applying corrections often requires specialized software or a complex workflow.\n\nWhat it offers:\n• Bias corrections: MAIVE (our Nature Communications method) plus benchmarks like PET-PEESE.\n• Robust inference: options for clustering and different data structures.\n• Minimal setup: free, open, no installation, no coding.\n• Reproducibility: export R code for the results you generate.\n\nSwipe through the four slides to see the workflow.\n\nTry it with your own data: 👉 easymeta.org (https://www.easymeta.org/)\n\nWith Pedro Bom, Tomas Havranek, Heiko Rachinger, and Petr Čala\n\n#MetaAnalysis #OpenScience #ResearchMethods #Econometrics #Statistics",
    "image_size": [
      [
        1200,
        1500
      ]
    ],
    "image_alt": [
      "First slide of the EasyMeta walkthrough: “Seamless Meta-Analysis with MAIVE — adjust your data for publication bias, p-hacking, and spurious precision”, with buttons to upload data or run a demo."
    ],
    "slug": "2026-01-27-browser-tool-for-publication-bias",
    "anchor": "2026-01-27"
  },
  {
    "date": "2025-12-29",
    "datetime": "2025-12-29 12:27:07",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7411382256684060672",
    "reshare_with_comment": false,
    "lang": "cs",
    "images": [
      "2025-12-29_p22_1.jpeg"
    ],
    "text": "Jak vytěžit z AI maximum?\nNechte modely bojovat mezi sebou. 🧠🤖\n\nV dnešních Hospodářské noviny ukazujeme jednoduchý, ale účinný postup, jak pracovat efektivněji s umělou inteligencí.\n\nMísto jedné odpovědi: duel AI modelů.\n\nChatGPT vs. Gemini (nebo jiná kombinace).\n🔹 Jeden model dostane roli kreativního vizionáře, který nápad rozvíjí.\n🔹 Druhý je ďáblův advokát – systematicky hledá faktické i logické chyby.\n🔹 Modely si navzájem napadají výstupy a opravují se.\n\nVýsledek?\n👉 Méně halucinací.\n👉 Tvrdší „stress test“ nápadů.\n👉 Kvalitnější výstup než od jednoho modelu.\n\nNemáte předplatné ChatGPT?\nŽádný problém. Duel zvládnete i manuálně – stačí otevřít vedle sebe dva zdarma dostupné AI modely, dát jim stejná data a protichůdné role a jejich odpovědi mezi sebou křížit. Po několika kolech už dává smysl výstupy číst.\n\n🔗 Odemčený článek v HN:\nhttps://lnkd.in/dBbUetrP\n🔗 Duelový protokol (prompt + návod):\nhttps://lnkd.in/dRwKz63g",
    "image_size": [
      [
        800,
        533
      ]
    ],
    "image_alt": [
      "Ilustrace: dva roboti, jeden se znakem ChatGPT a druhý s logem Gemini, proti sobě zaráží pěsti a mezi nimi šlehají jiskry. Pod nimi sedí u notebooků dva lidé obklopení hromadami papírů, v pozadí grafy a binární kód."
    ],
    "slug": "2025-12-29-jak-vytezit-z-ai-maximum",
    "link_map": {
      "https://lnkd.in/dBbUetrP": "https://nazory.hn.cz/c1-67828260-jedna-ai-nestaci-nechte-modely-spolu-bojovat",
      "https://lnkd.in/dRwKz63g": "https://github.com/tjhavranek/research-audit-duel-protocol"
    },
    "anchor": "2025-12-29"
  },
  {
    "date": "2025-12-23",
    "datetime": "2025-12-23 12:00:01",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7409201113406603264",
    "reshare_with_comment": true,
    "lang": "en",
    "images": [],
    "image_size": [],
    "image_alt": [],
    "reshare_of": {
      "label": "Commenting on her own interview for Startitup.sk",
      "url": "https://meta-analysis.cz/komentare/zi-iv-startitup-liahen-talentov/",
      "url_label": "„Slovensko zostane len liahňou talentov a vidiekom Prahy“ — Startitup.sk, 18 December 2025"
    },
    "text": "I still love Slovakia and Gemer. When our kids were born, one of the first things my husband did was go to the Slovak embassy in Prague and arrange Slovak citizenship for them. I hope I’m wrong in this interview.",
    "slug": "2025-12-23-slovakia-and-gemer",
    "anchor": "2025-12-23"
  },
  {
    "date": "2025-12-12",
    "datetime": "2025-12-12 08:39:03",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7405164270306504704",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [],
    "text": "For deep analytical work, I don’t use one AI model. I make them duel.\n\nFor thinking-intensive tasks (research, strategy, due diligence) a single AI model often converges too fast. It sounds convincing but skips edge cases and hidden assumptions.\n\nSo Tomas Havranek and I formalized a workflow to create structured disagreement:\n\n🔹 Anchor (ChatGPT): Writes a first-pass assessment grounded in the files. 🔹 Audit (Gemini): Actively tries to break it (logic gaps, counterexamples, failure modes). ChatGPT defends. The duel iterates. 🔹 Synthesis (You): You analyze the conflict to see what survives.\n\nThe output is not “AI approval.” It is a clearer map of risks, boundary conditions, and what you must verify.\n\nAdvanced users: You can swap the auditor for Claude or automate this via API (MAD).\n\nFor everyone else: no code needed. You can copy-paste the protocol and try it via ChatGPT Agent.\n\nProtocol (GitHub): 👉 https://lnkd.in/dRwKz63g\nResearch example: 👉 https://lnkd.in/dw98RKuW\n\n#AI #DecisionMaking #CriticalThinking #Productivity #Research",
    "image_alt": [],
    "slug": "2025-12-12-i-make-ai-models-duel",
    "link_map": {
      "https://lnkd.in/dRwKz63g": "https://github.com/tjhavranek/research-audit-duel-protocol",
      "https://lnkd.in/dw98RKuW": "https://www.maer-net.org/post/ai_duel"
    },
    "anchor": "2025-12-12"
  },
  {
    "date": "2025-12-10",
    "datetime": "2025-12-10 07:23:06",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7404420382004539392",
    "reshare_with_comment": true,
    "lang": "en",
    "images": [],
    "text": "🔎 MAIVE is now on CRAN: Better tools for trustworthy meta-analysis\n\nMeta-analysis is one of the key pillars of evidence-based research. It combines results from many studies to give us a clearer, more reliable answer.\n\nBut in observational research, studies can sometimes appear too precise --not because the evidence is strong, but because of method choices, selective reporting, or p-hacking. This can distort the final meta-analytic conclusion.\nMAIVE helps address this.\n\nYou can now install MAIVE directly in R:\n\ninstall.packages(\"MAIVE\")\n\nOr try it instantly in your browser (no coding needed):\n 🖥️ https://www.easymeta.org\nMore details:\n 📦 CRAN package: https://lnkd.in/dQ7XjiAq\n\n 📄 Nature Communications article: https://lnkd.in/eFfrP6H2\n\n 📰 Blog overview: https://lnkd.in/ejkgbU6X\n\nA big thank-you to Petr Čala, who built the MAIVE web app and prepared the CRAN release. His work makes high-quality meta-analysis methods accessible.\n\n#metaanalysis #OpenScience #researchmethods #RStats",
    "image_alt": [],
    "slug": "2025-12-10-maive-is-now-on-cran",
    "link_map": {
      "https://lnkd.in/dQ7XjiAq": "https://CRAN.R-project.org/package=MAIVE",
      "https://lnkd.in/eFfrP6H2": "https://www.nature.com/articles/s41467-025-63261-0",
      "https://lnkd.in/ejkgbU6X": "https://communities.springernature.com/posts/spurious-precision-in-meta-analysis-of-observational-research"
    },
    "anchor": "2025-12-10"
  },
  {
    "date": "2025-10-26",
    "datetime": "2025-10-26 11:18:13",
    "url": "https://www.linkedin.com/feed/update/urn%3Ali%3Ashare%3A7388172095639265282",
    "reshare_with_comment": false,
    "lang": "en",
    "images": [
      "2025-10-26_spurious.png"
    ],
    "text": "🔎 New in Nature Communications: “Spurious precision in meta-analysis of observational research.”\n\nSometimes studies appear too precise because their reported uncertainty reflects method choices rather than real evidence strength. This can mislead meta-analyses.\n\nWe introduce MAIVE, a simple way to detect and correct such bias, including publication bias and p-hacking.\n\n🖥️ Try it in your browser (free, no coding): https://www.easymeta.org\n\n 📄 Paper: https://lnkd.in/eFfrP6H2\n\n 📰 Blog: https://lnkd.in/ejkgbU6X\n\n— Kudos to my great co-authors Pedro Bom, Tomas Havranek, and Heiko Rachinger\n\n #metaanalysis #researchmethods #OpenScience #NatureCommunications",
    "image_size": [
      [
        866,
        459
      ]
    ],
    "image_alt": [
      "Two funnel plots side by side from the Nature Communications paper. In (a), selection on estimates, the surviving studies sit above the t=1.96 line and the effect size is inflated. In (b), selection on standard errors, the funnel is equally asymmetric but the effect size is not inflated. Both plot standard errors against estimates."
    ],
    "slug": "2025-10-26-spurious-precision-nature-communications",
    "link_map": {
      "https://lnkd.in/eFfrP6H2": "https://www.nature.com/articles/s41467-025-63261-0",
      "https://lnkd.in/ejkgbU6X": "https://communities.springernature.com/posts/spurious-precision-in-meta-analysis-of-observational-research"
    },
    "anchor": "2025-10-26"
  },
  {
    "date": "2026-07-25",
    "lang": "en",
    "images": [
      "2026-07-25_outliers_title.png",
      "2026-07-25_outliers_table.png"
    ],
    "text": "So, how much do different outlier treatments matter for meta-analysis results?\n\nWe have just preprinted a study that recomputes 358 behavioral science meta-analyses under five different outlier treatments.\n\nThe mean effect barely moves: the median absolute change in Cohen’s d is at most 0.047 and usually much smaller. But the interpretation moves more. In 11.5% of the meta-analyses, at least one treatment changes statistical significance; in 15.9%, it changes whether the effect reaches a smallest effect size of interest.\n\nWinsorizing changes conclusions least often; DFBETAS changes them most often.\n\nPaper, online appendix, and data:\nhttps://lnkd.in/dhtNgTMx\nPre-registration:\nhttps://lnkd.in/du4_6YDT\nReplication package:\nhttps://lnkd.in/dQrKGiw2\n\nJoint work with Tomas Havranek, Martina Lušková, and T. D. Stanley",
    "image_size": [
      [
        1670,
        1570
      ],
      [
        1730,
        915
      ]
    ],
    "image_alt": [
      "Title page of the working paper “Do decisions about outliers and influential effects matter? Evidence from 358 behavioral science meta-analyses” by Tomas Havranek, Zuzana Irsova, Martina Luskova and T. D. Stanley, dated July 2026, with the full abstract below the title.",
      "Table 3 of the paper, “Changes in statistical significance and smallest effect size of interest”. Two panels compare four outlier treatments — drop-extreme, absolute studentized residual above 3, Winsorize 5/95, and absolute DFBETAS above 2/√k — against the do-nothing baseline, each under random-effects and unrestricted weighted least squares estimators. Winsorizing changes the fewest conclusions and DFBETAS the most."
    ],
    "slug": "2026-07-25-do-outlier-decisions-matter",
    "url": "https://www.linkedin.com/posts/zuzanairsova_so-how-much-do-different-outlier-treatments-activity-7486767472440827905-HD5D",
    "datetime": "2026-07-25 13:01:02",
    "link_map": {
      "https://lnkd.in/dhtNgTMx": "https://meta-analysis.cz/outliers",
      "https://lnkd.in/du4_6YDT": "https://doi.org/10.17605/OSF.IO/97CMV",
      "https://lnkd.in/dQrKGiw2": "https://doi.org/10.5281/zenodo.21216506"
    },
    "anchor": "2026-07-25"
  }
]