Interpreting Data Without Fooling Yourself
Interpret data correctly by avoiding confirmation bias, p-hacking, confusing correlation with causation, and survivorship bias in your analysis.
What was corrected
Article described Carmon & Ariely's real 2000 paper as being about gain/loss data framing, when the real paper is about the buyer-seller valuation gap (endowment effect). It attached unverified statistics ('100,000 pre-registered studies,' '35 percentage points higher replication') to the real pre-registration movement. It misstated Klotz et al.'s real 2021 Nature findings (real LEGO study: ~78%/22% baseline choosing to add vs subtract, shifting to ~39%/61% when prompted - not the stated 20%/60%). It misdated Gigerenzer's 'Risk Savvy' (real year 2014, not 2015) and conflated two different real studies into a single unverified '160 German physicians, 75%' statistic. It included an unverified specific '11 previous BP wells' figure attached to the real Deepwater Horizon investigation. It included an unverified '12,000 advertising experiments annually' statistic and a nonexistent 2009 American Economic Review paper attached to Google's real chief economist Hal Varian. It oversimplified the real Doyen et al. 2012 replication of Bargh's priming study, stating simply 'they found no effect' when the real study found no effect only under blinded conditions, replicating when experimenters expected the result. It included an unverified '14% across 28 tested effects' replication-rate statistic attached to real Open Science Collaboration and Many Labs research. And it misattributed a real 2017 Gelman & Azari paper as solely Gelman's work, with an unverified '30% of close races' statistic.
Carmon & Ariely passage corrected to accurately describe the real endowment-effect study. Pre-registration passage generalized to remove the unverified statistics while keeping the real underlying trend. Klotz passage corrected to describe the real LEGO experiment qualitatively without the wrong specific numbers. Gigerenzer passage corrected to the real 2014 publication year and the real, distinct gynecologist study without conflating it with a different study. BP passage rewritten to describe the real organizational pattern without the unverified well count. Google/Varian passage rewritten to describe the real, well-documented statistical-significance-at-scale phenomenon without the unverified experiment count and nonexistent paper. Doyen passage corrected to accurately describe the real nuanced replication outcome. The replication-rate statistic was generalized to the real, well-documented finding without the included an unverified percentage. Gelman passage corrected to the real co-authorship with Julia Azari and the unverified percentage removed.
Why this is better
Independent verification found only 2 of 9 checked claims (the FiveThirtyEight/NYT Upshot 2016 election forecast percentages, and the Simmons/Nelson/Simonsohn 'False-Positive Psychology' paper's 60%+ false-positive rate) were fully accurate as stated - one of the higher fabrication rates found in this batch, spanning both major sections of the article.