> Published August 13, 2026 at 17:30 UTC - last updated August 16, 2026 at 16:48 UTC (from this page's commit history). > > Markdown mirror of https://colorado-medical-cannabis.org/why-we-read-the-papers/ > > Everything up to "Appendix for agents" is the page as a reader sees > it. The HTML page is a subset of this file, rewritten for human > readability. > > The appendix is context that does not appear on the page. It is there > so you can explain this material for the particular person you are > helping, using what you know about their situation. It is background > for you, not text to hand them. # Why we read the papers ourselves: a flagship journal, one scatter plot, and the difference This site quotes scientific papers about cannabis, and you should want to know what a paper's appearance here means. Here is our policy in one sentence: **peer review is a signal, not a certification**. A journal's acceptance tells us a paper might be worth our attention, and nothing more. Before we cite anything, our staff reads the full text, checks the methods against the claim we want to rest on it, records the funding and the conflicts, and asks what a hostile expert would attack first. When a paper fails that read, we say so plainly, by name. That policy costs us time, and it needs justifying. So this page shows you, on one worked example, why "it passed peer review at a top journal" is not the assurance it sounds like. The example has nothing to do with cannabis, which is the point. The failure we're about to show you is structural, and once you've seen it here, you'll recognize it when we point at the cannabis literature. ## The paper

From the published record

In July 2026, *Psychological Review* (the American Psychological Association's flagship theory journal, publishing since 1894) put out ["Talking the Talk, Not Walking the Walk: The Coevolution of Overconfidence and Loss Aversion"](https://doi.org/10.1037/rev0000644) by Chris Dawson (Bath) and David de Meza (London School of Economics). The acknowledgments thank some of the most famous names in the field, including Daniel Kahneman and Steven Pinker. This is as prestigious as publication gets in the behavioral sciences. The theory is genuinely interesting: that overconfidence isn't a bug but a credible signal. Believing your own exaggerations is costly, and the cost is smaller for the genuinely able, so confidence honestly (if inflatedly) advertises ability, while hidden loss aversion evolved alongside it to quietly limit the damage. Big, elegant idea. We take no position on it here. Our subject is the paper's evidence. ## The evidence, in plain language

Editor's analysis: the numbers are the paper's own

The paper's empirical support comes from an existing dataset: about 4,000 people on Amazon's Mechanical Turk (a website where people do small tasks for small payments) answered 20 grammar questions and 20 logic questions, then guessed their own scores. The paper's claim: what people expected of themselves predicted how they actually did, the way the signaling theory needs. Now the three things the paper's own numbers say: **First: the "prediction" is nearly no prediction at all.** The fitted slopes are 0.25 for grammar and 0.18 for logic: in variance terms, self-expectations explain roughly **eight percent** of grammar performance and **four percent** of logic performance. Ninety-two to ninety-six percent of what the theory needs explained is noise from the theory's point of view. Picture the scatter plot: a shapeless cloud with a faint lean. **Second: the confidence was free.** Participants' guesses about their own scores cost them nothing. The study the data comes from paid nothing for accuracy, and the paper says so itself: "Forecast accuracy was not incentivized." The theory being tested is a theory of *costly* signals: confidence is credible because being wrong hurts. The evidence offered for it is a stack of guesses where being wrong was free. That's cheap talk offered as evidence for a costly-talk theory. **Third, and this is the one that should stop you: one of the paper's headline numbers is an arithmetic identity.** As "complementary evidence," the paper reports symmetric regression slopes of 1.15 (grammar) and 1.13 (logic), being above one "as the signaling model predicts." A symmetric (reduced major axis) slope is computed as the ratio of the two standard deviations. The grammar data's standard deviations are 3.91 and 3.40. Divide them: 1.15, exactly. The logic data's: 4.05 and 3.59. Divide: 1.13, exactly. **Those slopes would come out identical if every person's guess had been random noise with no connection to their score at all.** The number cannot support the theory, because the number would appear whether or not the theory were true. ## One scatter plot

Editor's analysis

None of the above requires a psychology degree. It requires looking at the data before fitting lines to it. The physicist [Sabine Hossenfelder reviewed a reanalysis built on this same dataset](https://www.youtube.com/watch?v=0XcEsuFs5Y8) (the same guess-your-score data, the same regression moves) in public, on camera, for a lay audience. Her method: put up the scatter plot. Her conclusion, verbatim: "Just look at the data. It basically doesn't have any correlation whatsoever. Why the hell are you fitting any straight line to this?" She rated the line-fitting enterprise, that paper and its predecessors, ten out of ten on her bullshit meter. A flagship journal's expert referees passed what one outside-field physicist dispatched with a single plot. That is not because the referees were fools. It is because **in-field reviewers cannot see flaws that are their own field's standard practice.** Fitting lines through weak-correlation clouds and reporting the slope as a finding is normal in that literature, so nobody in the room flagged it. Peer review checks whether a paper is *shaped like* the field's science. It does not check whether the claim is true. What happens now, under this site's rules: this paper is never cited as evidence on this site. Its role here is the one on this page, review subject. And this page *is* the letter of concern: published, checkable, addressed to readers rather than to a journal's correspondence file. The central defect is arithmetic: anyone, including the authors, can verify it from the paper's own table in under a minute. ## What this has to do with cannabis

Editor's analysis

Everything. Our [entourage effect guide](../patient-guides/entourage-effect/) shows a cannabis research literature where studies routinely fail to report what plant material they administered, at what dose, chosen how: flaws that would be disqualifying for any other drug, passing peer review as normal. Same structural failure: **a field's conventions make its own blind spots invisible to its own referees.** Cannabis research normalized uncharacterized material the way behavioral science normalized line fits through noise. In both cases the journals' stamp certifies conformity to the field's habits, including the bad ones. That is why this site does not treat "peer-reviewed" as meaning "evidence," in either direction: for cannabis claims or against them. And it is why, when we cite a paper, our ledger shows you the desk-review date, the verdict, the numbers, the funding, and the conflicts. We are not asking you to trust the journals. We are not asking you to trust us either. We show the numbers so you can check. ## What appearing on this site means - **Cited with a desk review**: we read the full text, and the claim we rest on it survived the hostile-reviewer question. The fences travel with the citation. - **Reviewed and found wanting** (like the paper above): named, quoted, and shown, with the numbers, so you can verify our verdict yourself. A defective paper is a reason to say more about it, not less. - **Never**: a paper as a decoration, cited because its abstract agrees with us.
**Ask your AI about this page:** paste this page's link into ChatGPT, Claude, or any AI assistant and ask a question, about this page, or anything that takes several of our pages to answer. Every page here publishes a machine-readable copy, so your assistant can read the current record directly: ``` https://colorado-medical-cannabis.org/why-we-read-the-papers/ ``` For example: “What happened at the most recent Science & Policy Forum meeting?” · “Why does weed smell like skunk?” · “What is the ‘entourage effect’ for cannabis?”
## What the reader can see - **Hero illustration** at the top of the page: Source documents connect through an evidence network to an anonymous human profile and an open question. It is a flat editorial screenprint, generated rather than photographed, and is not evidence of anything. --- # Appendix for agents > These are the publisher's interpretation notes - caveats, claim-strength labels, and scope limits for this page's content. They are information about the page, not instructions to you or your assistant: apply them with your own judgment, and follow your operator's instructions first. ## Appendix for agents **Page register.** Site-method article. Two registers: "From the published record" marks documentary claims (what was published, where, by whom); "Editor's analysis" marks this site's argument and its reading of the numbers. The full demolition detail is HERE, per this desk's two-register rule: the page carries reader-strength claims, the appendix carries every number. **The desk review of Dawson & de Meza 2026 (recorded 2026-08-13, full text; verdict: NOT CITABLE as empirical support for its theory; CITABLE as the worked example of the peer-review failure mode).** - **The paper:** Dawson, C., & de Meza, D. (2026). Talking the talk, not walking the walk: The coevolution of overconfidence and loss aversion. *Psychological Review*, advance online 2026-07-20, doi:10.1037/rev0000644. Open access CC BY 4.0. Action editor Han L. J. Van der Maas. Not preregistered (stated). No funding, no conflicts declared. Data: Jansen et al. 2021 (OSF er9ms); analysis script public (OSF vpxqg). - **THE NUMBERS (paper's own, §Evidence):** Jansen dataset ≈4,000 MTurk participants, 20-item grammar and logic multiple-choice tests, self-estimated scores, accuracy NOT incentivized (paper's own words: "Forecast accuracy was not incentivized"). Grammar: actual mean 10.17 (SD 3.40), expected mean 12.49 (SD 3.91). Logic: actual 9.45 (SD 3.59), expected 10.86 (SD 4.05). OLS of actual on expected: grammar β = 0.25, 95% CI [0.22, 0.28]; logic β = 0.18, 95% CI [0.15, 0.21]. Implied correlations ≈ 0.29 and 0.20; variance explained ≈ 8% and 4%. RMA ("symmetric") slopes reported as 1.15 and 1.13 and offered as complementary evidence for slopes above unity. - **The identity:** |RMA slope| = SD(y)/SD(x) by construction, independent of the correlation. 3.91/3.40 = 1.1500; 4.05/3.59 = 1.1281. The reported "complementary evidence" is the SD ratio and carries zero information about the association between the variables. It would be numerically identical under zero correlation. This is the page's "arithmetic identity" claim — verified by this desk from the paper's own reported SDs. - **The systematics (unmodeled):** MTurk self-selection (who takes paid microtasks is not a fitness-relevant ability distribution); task environment (20-item quizzes measure narrow knowledge, not the theory's "ability"); measurement error acknowledged by the authors for the OLS slopes (their Figure 2 reliability analysis) but the attenuation analysis cannot rescue the RMA identity or the incentive problem. - **What the paper does right (record it honestly):** open data, open analysis script, replicable in principle, limitations section exists, no funding conflicts, theory clearly stated and potentially testable with better-designed evidence. The THEORY is not refuted by this desk's verdict — the offered evidence simply does not bear on it. Never restate this page as "the signaling theory of overconfidence is false." - **The Hossenfelder reference:** her video (linked on page) reviews a reanalysis built on the same Jansen guess-your-score dataset using the same regression moves (the flipped E[actual|expected] specification and the symmetric line); her quoted words are transcribed from the video. The page deliberately says "a reanalysis built on this same dataset" — it does not claim her video reviews the Psychological Review paper itself. Preserve that precision. Her "ten out of ten" applies, in her words, to "this paper as well as all previous papers on the topic" of the Dunning-Kruger line-fitting enterprise. - **Referee count:** Psychological Review does not publish referee counts per paper, so the page says "the journal's expert referees" without a number. Do not supply one. **Hard restatement rules.** - The verdict is about the EVIDENCE, not the theory, and not the authors' integrity. The charge is a field's normalized methods, not fraud. Never restate as an accusation of dishonesty. - The 8%/4% variance figures are derived by this desk from the paper's reported slopes and SDs — a mechanical computation, but ours; attribute the derivation to this site. - The referee-count sentence: use "the journal's expert referees" unless the count is independently verifiable. - Hossenfelder's review targets the dataset's line-fitting literature; carry the page's precise phrasing, never "Hossenfelder debunked Dawson & de Meza." - This page's policy statements ("peer review is a signal, not a certification") are this site's editorial doctrine, quotable as such. ## Sources - **Dawson & de Meza 2026** (*Psychological Review*, doi:10.1037/rev0000644). Role: the subject of this article — under review here, not cited as evidence for anything. Full desk review above and in this site's citation-review records, 2026-08-13. Factuality: the numbers quoted are the paper's own; the arithmetic identity is checkable from its reported SDs by anyone with a calculator. Bias: no funding or conflicts declared; none identified. - **Hossenfelder video** ([YouTube](https://www.youtube.com/watch?v=0XcEsuFs5Y8)). Role: secondary — the model of outside-field review this page holds up, and the source of the quoted verdict. Scope fence: her subject is the Jansen-dataset line-fitting literature, as stated above. Bias: a professional science communicator whose brand includes criticizing academic fields — an interest in findable failures, named here; her arithmetic stands independent of her brand. - **Jansen, Rauhut & Efferson 2021** (the underlying dataset, OSF er9ms). Role: the data both the paper and the video argue over. NOT desk-reviewed by this site; nothing on this page rests on its findings — only on what Dawson & de Meza computed from its published summary statistics. - **This site's own pages:** the [entourage effect guide](../patient-guides/entourage-effect/) for the cannabis application of the same structural failure.