{
 "_readme": "One row per paper on meta-analysis.cz whose own headline states a SINGLE bias-corrected or best-practice value in the SAME units as that literature's estimate-level `effect` column -- or, for a methods paper admitted on its worked example under `rule.illustration`, in the units of the external literature that example is drawn from. `corrected` is that value, read off the paper; `corrected_quote` is the sentence it comes from, verbatim from estimates.csv `source_quote`. `mean` is what the reported estimates average to: the paper's own stated simple mean where it gives one (preferred -- it is the published comparator and the authors chose it), otherwise the 1%-winsorised mean of that literature's `effect` column in site/data/v1/estimates_harmonised.csv, computed by build_correction_figure.py and never typed here. The revision is (|corrected| - |mean|) / |mean|, the same relative index as Table 3 of Gechert et al. (2025), which is his own paper. The rule is DIRECTION-BLIND and was rewritten to be so after an adversarial review; see `rule.direction_blind`, and `rule.sign_change` for the related point that a reversal of sign is disclosed rather than excluded. `rule.comparator` and `rule.corrected` are ordered hierarchies: the first option that exists in the paper wins, and anything below the first rung is drawn as a ring. Exclusions are listed with their reason and are part of the record: the figure's caption states how many papers qualified and why the rest did not.",
 "rule": {
  "include": "the paper states, in its own words, a single corrected or best-practice value for ONE clearly defined estimand -- the whole literature where it gives one, otherwise a named horizon or subsample -- AND a comparator for that same estimand in the same units and the same transform: either the paper's own uncorrected/reported mean, or a value commensurate with its estimate-level effect column. Anything narrower than the whole literature, or read off a table rather than the headline, or taken as the central value of a range or an upper bound, is admitted under `tier` and drawn as a ring with a note saying so.",
  "comparator": "a fixed order, first one that exists wins. (1) the paper's own uncorrected mean for the same estimand, units, transform and horizon -- the simple mean where it prints one, otherwise the mean it differences its own best-practice estimate against in a printed Diff column, matched weighting for weighting. It is the published comparator and the authors chose it; (2) failing that, the paper's own uncorrected POOLED estimate of exactly that estimand, quoted; (3) failing that, the 1%-winsorised mean of the estimate-level data released here, or the mean of the released source column read from that dataset's codebook where the harmonised table does not carry the literature; (4) only where a paper reports none of those, the canonical value the paper ITSELF names as the number its field had been working with. Never import a benchmark from outside the paper merely to create a row. Anything but (1) is drawn as a ring and has to quote the sentence its comparator comes from. As of this writing no paper reaches rung 4: every one that looked like a benchmark case turned out to print a mean of its own.",
  "exclude": "ranges, several values for different horizons or subsamples, narrative answers, a different estimand, units that do not match, a headline value that is not itself a correction, a second row on a literature already represented, and any case where two defensible comparators disagree about the DIRECTION. \"Defensible comparator\" means one of the two the comparator clause names, the paper's own stated mean or the winsorised mean of the estimate-level data -- not any statistic that can be computed from the column. Reading it loosely kept `incentives` out on a median and an inverse-variance mean, while the two comparators the rule actually names agreed to three decimals. Note what is NOT excluded: a best-practice figure computed as a synthetic study at the literature's preferred values. That is how these papers construct best practice and how Gechert et al. build their corrected column; excluding it would remove most papers' own headline conclusions. The rows concerned say so in their notes.",
  "direction_blind": "The rule was rewritten after an adversarial review found that every paper whose correction moved the number AWAY from zero had landed in the excluded set. Nothing in the rule refers to the sign of the revision, and `migrant` (+69%) and `armington` (+144%) are in the figure because the rule admits them, not despite it. If a future paper moves up, it appears.",
  "verbal_zero": "A paper that says its corrected effect is 'about zero', 'negligible' or 'insignificant' without giving a number does NOT enter as 0. Coding a verbal hedge as an exact zero forces exactly -100% and would have put four such papers at the extreme of the figure. Insignificance is not a point estimate.",
  "corrected": "a fixed order too. (1) the corrected or best-practice point estimate the paper says it prefers; (2) a paper-stated corrected or best-practice value for the whole literature, or for the horizon or subsample the paper leads with; (3) where a paper reports several correction methods, prefers none of them, and they agree on the substantive conclusion, the MEDIAN of those methods, with every method listed in the row and the median computed by the builder rather than typed; (4) where the methods disagree about the sign or the substantive conclusion, exclude. Never convert 'insignificant', 'negligible', 'about zero' or 'unbiased' into a number the paper did not report.",
  "sign_change": "A reversal of sign is NOT an exclusion. The plotted index is defined on absolute magnitudes and exists either way; deleting such rows would be a second way of letting the rule decide the finding, which is the error the direction-blind rewrite was meant to end. Instead the builder detects `mean * corrected < 0`, draws that dot outlined, and names it in the caption, the tooltip and the SVG description, because the one thing a magnitude axis cannot say is that the number changed sign. `forward` is the case that matters: its magnitude moves 1.5% while its sign goes from -0.602 to +0.611.",
  "small_base": "Where both the reported and the corrected level are economically negligible in the paper's own terms, a large percentage is a large relative change from a small base and says less than it looks. Such a row carries `small_base` and says why in `approximation`; the builder refuses to plot it otherwise, and the caption names them. This is disclosure, not exclusion. It is NOT a licence to divide by any small number: bma is plotted as no change because its own test finds no bias to correct and its authors decline a best-practice definition, where competition and students each print a corrected value their authors stand behind and keep their computed ratios.",
  "illustration": "A methods paper hosted here may enter on its own worked example: the literature it applies its correction to, even where that literature is external and this site holds no data for it. Conditions, all of them: the paper is one of the 55; both sides are quoted from it; the external literature appears in no other row; and the paper stands behind the illustration's finding rather than disclaiming it. That last condition is what separates `correlations`, which concludes from its example that 'the most reliable and informative studies in this area of research find no evidence of a positive correlation', from `pcc_survey`'s ICT example, whose own footnote says 'We make no inference about the true unconditional average PCC of ICT.' Such a row is always drawn as a ring."
 },
 "rows": [
  {
   "project": "dst",
   "corrected": -0.01,
   "mean": -0.34,
   "mean_from": "paper",
   "corrected_quote": "The resulting global estimate is -0.01%, quite distant from -0.34%, the simple average effect reported in the literature.",
   "note": "the paper states both numbers in one sentence Its best-practice figure is a synthetic study: the paper evaluates its model at the literature's preferred values: recent data, more observations, better-cited outlets. That is the construction these authors use and the one behind Gechert et al.'s corrected column, and it is the paper's own headline conclusion."
  },
  {
   "project": "excess_sensitivity",
   "corrected": 0.11,
   "mean": 0.37,
   "mean_from": "paper",
   "corrected_quote": "When corrected for the bias due to aggregation and the bias due to publication selection, the literature yields a mean excess sensitivity of merely 0.11.",
   "note": "the headline itself carries the comparator: '0.11, against a mean reported estimate of 0.37'"
  },
  {
   "project": "sigma",
   "corrected": 0.3,
   "mean": 0.9,
   "mean_from": "paper",
   "corrected_quote": "The mean elasticity conditional on the absence of these issues is 0.3.",
   "note": "the caveat column states the comparator: 'the mean reported estimate is 0.9'. The winsorised mean of the shipped data is 0.66; the paper's own number is used.",
   "quote_source": "pdf",
   "corrected_locator": "sigma2.pdf p.1 (abstract); 'these issues' are the publication bias, aggregation and first-order-condition problems named in the preceding sentence"
  },
  {
   "project": "discrate",
   "corrected": 0.33,
   "mean": 0.8,
   "mean_from": "paper",
   "corrected_quote": "The corrected mean annual discount rate is 0.33.",
   "mean_quote": "publication bias exaggerates the mean reported discount rate more than twofold, from 0.33 to 0.80 (the simple uncorrected mean)",
   "note": "the paper states both numbers in one sentence. The site's harmonised column holds only the 539 of 927 estimates that report a standard error, and its mean of 1.02 appears nowhere in the paper."
  },
  {
   "project": "esg",
   "corrected": 0.125,
   "mean": 0.28,
   "mean_from": "paper",
   "corrected_quote": "Correcting for publication bias alone brings the typical effect down to between 0.08 and 0.17 points, depending on the method.",
   "tier": "range",
   "approximation": "plotted at 0.125, the midpoint of the 0.08 to 0.17 the paper reports for the whole sample; both ends are its own figures. The 0.12 the page quotes is the best-practice estimate for most of the world, and the paper reports 0.276 for the Middle East and near −0.11 for Southeast Asia, so it does not cover the same firms as an average over every estimate.",
   "quote_source": "pdf",
   "corrected_locator": "Havránek, Iršová et al., abstract: the sentence beginning 'Much of that reflects selective reporting'",
   "mean_quote": "diversity raises ESG scores by about 0.28 points"
  },
  {
   "project": "water",
   "corrected": 0.15,
   "mean": null,
   "mean_from": "data",
   "corrected_quote": "Attempting to correct simultaneously for publication and endogeneity biases, we find that the mean underlying elasticity is approximately 0.15 or less.",
   "note": "'0.15 or less' is an upper bound; taking 0.15 understates the correction, which is the conservative direction",
   "tier": "range",
   "approximation": "The paper says the elasticity is 'approximately 0.15 or less', an upper bound rather than a point, so 0.15 understates the correction."
  },
  {
   "project": "frisch",
   "corrected": 0.25,
   "mean": 0.5,
   "mean_from": "paper",
   "corrected_quote": "the two panels suggest that 0.25 is a reasonable estimate for the mean Frisch elasticity at the extensive margin",
   "mean_quote": "publication bias in the literature is substantial and likely to exaggerate the mean reported elasticity approximately twofold",
   "note": "0.25 is the paper's estimate for the EXTENSIVE margin, which is the margin the shipped column holds; it also happens to be the paper's total-hours headline. The earlier version of this row quoted the total-hours sentence against a per-margin mean, which crossed estimands. The pair and the -50% are unchanged.",
   "quote_source": "pdf",
   "corrected_locator": "frisch2.pdf p.8; the same page concludes that publication bias 'is likely to exaggerate the mean reported elasticity approximately twofold', i.e. 0.25 against the reported 0.5"
  },
  {
   "project": "beauty",
   "corrected": 3,
   "mean": 5,
   "mean_from": "paper",
   "corrected_quote": "Conservative corrections for publication bias reduce the mean premium from about 5% to about 3%, while other techniques suggest a more aggressive reduction.",
   "tier": "range",
   "approximation": "both ends are the paper's own approximate figures, and 3% is the conservative end: it says other techniques cut the premium further. The 1.1% the page quotes is a different quantity -- it controls for cognitive ability and reweights occupations towards US employment shares, and the paper says its subgroups do not average to the overall mean.",
   "quote_source": "pdf",
   "corrected_locator": "Iršová, Havránek et al., abstract: the sentence after 'The mean premium reported in the literature is exaggerated by publication bias.'",
   "mean_quote": "reduce the mean premium from about 5%"
  },
  {
   "project": "learning",
   "corrected": -0.12,
   "mean": null,
   "mean_from": "data",
   "corrected_quote": "Our preferred specifications, RoBMA and MAIVE, rely on different assumptions yet converge on an effect size of approximately -0.12 SD, equivalent to a learning loss of about 30% of a school year."
  },
  {
   "project": "migrant",
   "corrected": 22,
   "mean": 13,
   "mean_from": "paper",
   "corrected_quote": "After correcting for this bias, the implied elasticity rises from an uncorrected average of 13 to approximately 22.",
   "note": "both ends are the paper's own whole-literature figures. Its best-practice estimate of about 17, which the page quotes, conditions on the baseline regional specification -- Table 5 gives about 8 for national models -- so it is not commensurate with an uncorrected mean taken across every design.",
   "quote_source": "pdf",
   "corrected_locator": "Iršová et al., section 1: the sentence follows the precision-selection result and precedes the relative-wage implication",
   "mean_quote": "the implied elasticity rises from an uncorrected average of 13"
  },
  {
   "project": "armington",
   "corrected": 3.8,
   "mean": 1.56,
   "mean_from": "paper",
   "corrected_quote": "Third, the elasticity implied by the literature after accounting for both publication bias and study quality lies in the range 2.5--5.1 with a median of 3.8.",
   "note": "the paper corrects long-run elasticities only, so the comparator is its own long-run uncorrected mean rather than the mean of every estimate in the file. Table 1 splits the sample: 556 short-run estimates average 0.88, 2,968 long-run estimates 1.56.",
   "mean_quote": "Long-run effect ... 1.56 (Table 1, 2,968 estimates)",
   "mean_quote_kind": "table"
  },
  {
   "project": "skill",
   "corrected": 4,
   "mean": 1.8,
   "mean_from": "paper",
   "corrected_quote": "The implied mean elasticity is 4, with a lower bound of 2.",
   "note": "An upward revision, and the paper names both numbers. It works in INVERSE elasticities: publication bias exaggerates the reported inverse, correcting pulls it toward zero, and the implied elasticity rises from 1.8 to 4. The corrected 4 is the inversion of a corrected mean inverse elasticity, so it is not commensurable with a mean of the raw elasticities in the harmonised column (median 1.42, maximum 1000) -- which is why the paper's own pair is used, the same treatment as migrant, sigma and dst.",
   "mean_quote": "even our preferred estimate of 4 is much larger than the uncorrected mean implied elasticity of 1.8, a difference which shows that publication bias dominates attenuation bias"
  },
  {
   "project": "size",
   "corrected": 1.72,
   "mean": 5.08,
   "mean_from": "paper",
   "quote_source": "pdf",
   "units": "annualised size premium, % per year",
   "corrected_quote": "The implied difference in percentage returns between the 10th and the 90th percentile of NYSE stocks is 0.142%, or 1.72% in annualized terms, which is roughly three times lower than the unadjusted value.",
   "corrected_locator": "size2.pdf p.14 (JoES 33(5), p.1476), end of Section 4; tabulated in Table 6, row '10th'",
   "mean_quote": "As a benchmark for the unadjusted size premium, we use the simple mean reported coefficient of -0.092 reported in Table 1 ... We obtain a benchmark monthly size premium of 0.415%, or 5.08% annualized.",
   "note": "The two numbers are the SAME transform of the same quantity, applied to two slopes: the corrected FAT-PET intercept and the mean reported coefficient. Table 6 prints both at 19 percentiles and their ratio holds at 0.337-0.340 throughout, so the choice of the 10th/90th cut does not drive it. Previously excluded for pairing 1.72% with the raw slope mean of -0.10%, which crosses a sign-flipping transform -- true of that pairing, but the paper states its own benchmark in the headline units and that was never tested."
  },
  {
   "project": "border",
   "corrected": 1.76,
   "mean": 3.03,
   "mean_from": "paper",
   "quote_source": "pdf",
   "units": "semi-elasticity of trade w.r.t. the within-country dummy",
   "corrected_quote": "The overall mean semi-elasticity is 1.76, which translates into a border effect of 5.8-almost four times smaller than the border effect based on the sample mean of the semi-elasticities reported in the literature.",
   "corrected_locator": "border2.pdf p.22 (IMF Economic Review 65(2), p.386), Section 5 'Best Practice'; same value in the 'All countries' row",
   "mean_quote": "Table 1, 'Border Effects Differ Across Countries', p.372, last row: All countries | 1,271 | 3.03 (unweighted mean)",
   "note": "Same units as the shipped column (semi-elasticity, the same 1,271 estimates). Excluded before because the site quotes the paper's 'one-third reduction in trade' sentence, which is a different transform -- but the paper states the semi-elasticity pair itself. Its best-practice figure is a synthetic study: the paper evaluates its model at the literature's preferred values: recent data, more observations, better-cited outlets. That is the construction these authors use and the one behind Gechert et al.'s corrected column, and it is the paper's own headline conclusion.",
   "mean_quote_kind": "table"
  },
  {
   "project": "climate",
   "corrected": 1.6,
   "mean": 3.27,
   "mean_from": "paper",
   "quote_source": "pdf",
   "units": "degrees Celsius, equilibrium climate sensitivity",
   "corrected_quote": "After correction for publication bias, the best estimate assumes that the mean climate sensitivity equals 1.6 with a 95% confidence interval (1.246, 1.989). This is one half of the simple uncorrected average, 3.27.",
   "corrected_locator": "climate2.pdf p.6 (Energy & Environment 26(5), p.858), Section 5, final paragraph; restated in the Conclusion",
   "mean_quote": "The 48 CS estimates collected from 16 studies range from 0.7 to 10.4, with a mean of 3.27.",
   "note": "Excluded before as 'a range, 1.4-2.3 degrees', which is the abstract's phrasing; the body states a single best estimate of 1.6 and its own uncorrected mean in the same sentence.",
   "tier": "range",
   "approximation": "The site's headline for this paper is a range, 1.4-2.3 C, which is the span across specifications; the figure uses the paper's own stated best estimate of 1.6 C."
  },
  {
   "project": "education",
   "corrected": -0.037,
   "mean": -0.19,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "table",
   "units": "partial correlation coefficient, on both sides of the ratio",
   "approximation": "The paper's headline answer is verbal - the elasticity is 'close to zero' - so the number here is its best-practice estimate from Table 6 rather than a figure it puts in the abstract. The Table 6 figure is a synthetic study: publication characteristics are set at their sample maxima, the standard error and OLS at their minima, and the endogeneity control at its maximum.",
   "corrected_quote": "The 'best-practice' estimation in Table 6 yields a partial coefficient of −0.037 with a 95% confidence interval of (−0.055; −0.019).",
   "corrected_locator": "education2.pdf Table 6 ('All estimates', BMA column), p.21; restated in prose on p.22",
   "mean_quote": "the mean reported value of the tuition-enrolment partial correlation coefficient, −0.19 (shown in Table 2), is significantly exaggerated by the presence of publication bias",
   "note": "This is the right way to handle a verbal-zero headline: use the number the paper actually reports, never an exact 0. -0.037 against -0.19 is -80%, not the -100% that coding 'about zero' as zero would have produced."
  },
  {
   "project": "risk",
   "corrected": 1,
   "mean": 7.5,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "subsample",
   "units": "coefficient of relative risk aversion",
   "approximation": "The paper gives economics and finance separately. This is economics -- 590 estimates from 58 studies, the value its abstract leads with -- against the paper's own uncorrected mean for the same subsample. Finance moves further: 1-3 corrected against 45 uncorrected.",
   "corrected_quote": "The corrected mean coefficient of relative risk aversion is around 1 for economics and 1-3 for finance, compared with uncorrected means of 7.5 and 45, respectively.",
   "corrected_locator": "risk2.pdf p.2320 (JoES 39, Section 3, discussion of Table 2); Table 1 Panel A p.2318 gives Economics 590 estimates, mean 7.50",
   "mean_quote": "compared with uncorrected means of 7.5 and 45, respectively"
  },
  {
   "project": "eis",
   "corrected": 0.0145,
   "mean": 0.5,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "table",
   "units": "elasticity of intertemporal substitution",
   "approximation": "The site's headline, 0.3-0.4, is the corrected value for a narrower subsample: micro estimates for asset holders. What is plotted is Table 1's corrected mean for the WHOLE literature, which is the quantity this figure measures and which the paper reports as essentially zero. The two ends also treat the tails differently: the mean of 0.5 excludes estimates above 10 in absolute value, while the regression keeps all 2,735 because precision weighting already gives the extremes little weight.",
   "corrected_quote": "The first column of Table 1 reports the baseline result. The estimated β is approximately two and the constant equals zero, suggesting strong selective reporting and zero underlying elasticity on average.",
   "corrected_locator": "Havranek 2015, JEEA 13(6), Table 1 column 'FE', the paper's preferred specification: Constant 0.0145, s.e. 0.00881, 2,735 estimates",
   "mean_quote": "The mean reported estimate from all studies is 0.5."
  },
  {
   "project": "scc",
   "corrected": 134,
   "mean": 411,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "range",
   "units": "USD per ton of carbon, 2010 prices, emission year 2015",
   "approximation": "134 is the LARGEST of the paper's corrected means, and 411 the uncorrected mean of the same estimates -- the pair the paper itself puts in one sentence. Taking the largest corrected value understates the correction, which is the conservative direction.",
   "corrected_quote": "The largest corrected mean SCC we get for estimates with uncertainty is USD 134 per ton of carbon at 2010 prices for emission year 2015; because the uncorrected mean of these estimates is 411, our results indicate that the reported estimates of the SCC are exaggerated",
   "corrected_locator": "Havranek, Irsova, Janda & Zilberman 2015, Energy Economics 51: 394-406, p.405, concluding remarks",
   "mean_quote": "the uncorrected mean of these estimates is 411"
  },
  {
   "project": "activism",
   "corrected": 0.75,
   "mean": 1.49,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "range",
   "units": "percent, short-window cumulative abnormal return",
   "approximation": "The paper gives a corrected range, 0% to 1.5%, and names no central value, so the midpoint is plotted. The ten estimators behind that range average 0.60%, so the midpoint understates the correction rather than overstating it.",
   "corrected_quote": "We find that, after correcting for this bias, the value created by shareholder activism is positive but much smaller than commonly proposed. Our estimates range from 0% to 1.5%, depending on the estimation technique.",
   "corrected_locator": "Bajzik, Havranek, Irsova & Novak 2025, Corporate Governance 33: 1039-1061, p.1041; the ten corrected values are in Table A2 of the appendix",
   "mean_quote": "Table 2, row 'All': Nobs 1,973, Mean 1.49; the note to Table A2 reads 'The uncorrected mean value creation by shareholder activism is 1.49%'",
   "mean_quote_kind": "table"
  },
  {
   "project": "gasoline_price",
   "corrected": -0.31,
   "mean": -0.691,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "horizon",
   "units": "price elasticity of gasoline demand, long run",
   "approximation": "The paper gives two horizons. This is the long run, the one its abstract names first, against the paper's own simple mean for the same horizon; the short run moves the same way.",
   "corrected_quote": "Using mixed-effects multilevel meta-regression, we show that after correction for publication bias the average long-run elasticity reaches - 0.31 and the average short-run elasticity only - 0.09.",
   "corrected_locator": "Havranek, Irsova & Janda 2012, Energy Economics 34, Abstract p.201; body p.205",
   "mean_quote": "This sharply contrasts to the simple uncorrected averages amounting to -0.23 and -0.69: publication bias exaggerates the average reported elasticity more than twofold."
  },
  {
   "project": "gasoline",
   "corrected": 0.1,
   "mean": 0.28,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "horizon",
   "units": "income elasticity of gasoline demand, short run",
   "approximation": "The paper gives two horizons. This is the short run, the one its abstract leads with, and the only one whose corrected value and comparator come from the same whole sample.",
   "corrected_quote": "Our results suggest that the income elasticity of gasoline demand is on average much smaller than reported in previous surveys: the mean corrected for publication bias is 0.1 for the short run and 0.23 for the long run.",
   "corrected_locator": "Havranek & Kokes 2015, Energy Economics 47, Abstract p.77; Table 5 p.83, row 'Preferred estimate'",
   "mean_quote": "The studies cover many countries and report a mean elasticity of 0.28 for the short run and 0.66 for the long run."
  },
  {
   "project": "electricity",
   "corrected": -0.16,
   "mean": -0.267,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "horizon",
   "units": "own-price elasticity of electricity demand, short run",
   "approximation": "The paper gives two horizons. This is the short run, the value its abstract leads with, against the paper's own raw mean for the same row; the long run moves the same way.",
   "corrected_quote": "the publication-bias-corrected short-run elasticity is about -0.16 (a 10% rise in the electricity price cuts consumption by under 2%)",
   "corrected_locator": "Kudela et al. 2026, electricity.pdf, Abstract p.1; PET column of the 'Short-run elasticities (full sample)' row of Table 2",
   "mean_quote": "every corrector pulls the raw mean of -0.27 toward zero, to -0.16 (PET), -0.23 (PEESE), -0.11 (WAAP)"
  },
  {
   "project": "hedge",
   "corrected": 0.335,
   "mean": 0.36,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "range",
   "units": "percent per month, hedge fund alpha",
   "approximation": "The paper's headline is a 30-40 basis point range; plotted here is the mean of its own twelve corrected estimates, 0.335%, which it states in Table 7, against its own unconditional sample mean of 0.36%.",
   "corrected_quote": "The table shows that, for the full sample, our estimates based on various techniques of the “representative” alpha coefficient corrected for the publication selection bias range from 0.274 and 0.386. The mean and median values of 0.335 and 0.338, respectively",
   "corrected_locator": "Yang, Havranek, Irsova & Novak 2024, JoES 38: 1085-1131, p.1120 and Table 7, row 'Table 2 / Full sample'",
   "mean_quote": "The vertical line in Figure 3 denotes the unconditional sample mean of monthly alphas of 0.36%"
  },
  {
   "project": "finance_growth",
   "corrected": 0.199,
   "mean": 0.15,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "table",
   "units": "partial correlation coefficient",
   "approximation": "The paper's headline is verbal, an 'authentic positive link', so the number here is the FAT-PET effect from its Table 2 rather than a figure it puts in the abstract.",
   "corrected_quote": "The statistically significant estimate of β 0 , however, indicates that the literature identifies, on average, an authentic link between financial development and economic growth.",
   "corrected_locator": "finance_growth2.pdf p.8, Table 2: 1/SE_r (Effect) 0.199*** (0.018), 1,334 estimates from 67 studies",
   "mean_quote": "Simple average r 0.15 (0.095, 0.20)",
   "note": "An UPWARD row, and the same construction as the included eis row: a FAT-PET intercept against the literature's own simple average. The earlier exclusion called the answer narrative, which was wrong: both numbers are tabulated."
  },
  {
   "project": "class",
   "corrected": -0.21,
   "mean": -0.25,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "table",
   "units": "normalised class-size effect, hundredths of a standard deviation per one-student increase, on both sides of the ratio",
   "approximation": "The paper's headline answer is verbal -- the effect is negligible except in Tennessee's STAR experiment and in classes under 15 -- so the number here is Table 8's overall corrected mean rather than a figure from the abstract. Table 8 is a synthetic study evaluated at a zero standard error and the preferred treatment of endogeneity. Its confidence interval, (-1.38, 0.97), is about twelve times the size of the move.",
   "corrected_quote": "Table 8, Overall corrected mean: -0.21, 95% confidence interval (-1.38, 0.97).",
   "mean_quote": "Table 2, All estimates: mean -0.25, 95% confidence interval (-0.34, -0.17), 2,434 estimates.",
   "corrected_locator": "class.pdf Table 8; the comparator is Table 2, row 'All estimates'",
   "note": "Excluded before with 'one reviewer accepted it and a second refuted it; it stays out until that is settled'. A third reading settled it: both numbers are stated, in the same normalisation, over the same 2,434 estimates. The width of the interval is a caveat for the tooltip, and uncertainty is not an eligibility test anywhere else in this figure.",
   "quote_kind": "table",
   "mean_quote_kind": "table"
  },
  {
   "project": "exercise",
   "corrected": 0.227,
   "mean": 0.474,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "subsample",
   "units": "standardised mean difference, general cognition, on both sides of the ratio",
   "approximation": "General cognition is the first of the paper's three outcomes and the one its abstract leads with; memory and executive function are corrected further, by about -94% and -97%, so leading with general cognition understates the correction rather than choosing the friendly number. The paper adds that the corrected pooled value 'is of little value due to the extreme between-study heterogeneity' -- its prediction interval runs from -1.052 to 1.500.",
   "corrected_quote": "general cognition: standardized mean difference, SMD, = 0.227, 95% credible interval 0.116 to 0.330.",
   "mean_quote": "A three-level publication bias-unadjusted Bayesian model-averaged (BMA) meta-analysis reproduced an overall effect size estimate (SMD = 0.474 [0.410, 0.537]).",
   "corrected_locator": "exercise.pdf, abstract and Results; both values are for the same 835 effect sizes from 82 meta-analyses",
   "note": "The correction is what this figure measures, and the authors' warning is about how much weight to put on either pooled number, not about whether the correction happened. Every dot here is a pooled estimate; the warning belongs in the tooltip, which carries it."
  },
  {
   "project": "reforms",
   "corrected": -0.38,
   "mean": -0.052,
   "mean_from": "paper",
   "quote_source": "pdf",
   "tier": "horizon",
   "units": "partial correlation coefficient, short run, on both sides of the ratio",
   "approximation": "The paper's headline pairs a short-run cost with a long-run gain. The short run is plotted because it is the horizon where the paper finds significant publication bias: its long-run corrected effect is close to the simple average, and the long-run rise to 0.27 comes from best-practice conditioning alone. That pairing, 0.146 to 0.27, is +85%, so the horizon does not decide the direction. The comparator, -0.052, is itself small, which is why a move to -0.38 reads as +631%.",
   "corrected_quote": "The improved estimate of the reform effect for the short run reaches −0.38, which means virtually no change compared with the case when we only corrected the simple average for publication bias. In contrast, the improved estimate of the long-run effect reaches 0.27, which is almost thrice more than the estimate in the previous section.",
   "mean_quote": "Table 2, Simple average: short run -0.052, 95% confidence interval (-0.084, -0.021); long run 0.146, (0.118, 0.173).",
   "corrected_locator": "reforms2.pdf Table 2 for the simple averages and the best-practice discussion that follows it",
   "note": "Excluded before because 'the two horizons move in OPPOSITE directions'. They do not, under the index this figure plots: in absolute magnitude the short run moves +631% and the long run +85%, both away from zero. The old reason applied signed direction to a rule that is direction-blind, and it is corrected here. Re-read for the off-scale check: the move is not an artefact of the best-practice construction. Table 3 puts the short-run effect beyond publication bias at -0.39 under all three estimators, -0.394, -0.395 and -0.394, and the paper says the best-practice figure of -0.38 is 'virtually no change compared with the case when we only corrected the simple average for publication bias'. Correcting for selection alone moves -0.052 to about -0.39; conditioning on best practice moves it back a little. The paper's own reading is that this turns the short-run cost of reform from negligible into 'strong' on Doucouliagos's guidelines. An earlier version of this row quoted a sentence for the corrected value that appears nowhere in the paper. The quote above is the paper's own, and the builder now checks prose quotes against the full text on this site so that cannot recur. One thing a sceptical reader will notice: the paper's own words for this move are that the corrected estimates 'reach -0.39, which is approximately four times more than the simple averages'. Four times fits the fixed-effects average, -0.081; against the simple average the paper names first, -0.052, it is seven times. The comparator rule takes the simple average, and the looser multiplier is the paper's phrasing rather than a different number.",
   "mean_quote_kind": "table"
  },
  {
   "project": "lags",
   "corrected": 29.2,
   "mean": 33.5,
   "mean_from": "paper",
   "corrected_quote": "The average transmission lag implied by our definition of the ideal study is 29.2 months, which is less than the simple average by approximately 4 months.",
   "note": "the paper states both sides in one sentence. Its best-practice figure is the synthetic ideal study the rule admits: the BMA results evaluated at best-practice methodology and maximum publication characteristics, everything else at sample means. The comparator, 33.5 months, is the simple average of all 198 collected transmission lags. This was recorded as excluded for 'no estimate-level data on this site and no stated pair'. The first half is true and stops mattering once the comparator comes from the paper rather than from the data; the second half was simply wrong.",
   "quote_source": "pdf",
   "corrected_locator": "lags2.pdf p.18, the ideal-study paragraph after Table 5; the comparator 33.5 is on p.6 and in Table 2 on p.7, row 'Estimates from all Impulse Responses', 198 observations"
  },
  {
   "project": "house_prices",
   "corrected": -1.154,
   "mean": -1.2,
   "mean_from": "paper",
   "corrected_quote": "The mean maximum corrected semi-elasticity is −1.2, which suggests, in practical terms, insufficiently strong transmission of monetary policy to house prices",
   "note": "both numbers are the two-year peak of the impulse response, so the pair is commensurate: this was excluded once as 'a peak against an average', which was wrong, because the paper states its own simple uncorrected mean OF THAT PEAK, -1.2%. It then sat in the figure at exactly 0%, because the paper rounds its corrected peak to -1.2 as well. That was an artefact of the rounding. Table 3's baseline row prints the corrected response at every horizon, -0.710, -0.716, -0.935, -1.154, -1.000, -0.714, so the maximum is -1.154 at eight quarters and the correction moves the number toward zero. The released data agrees and settles the size: house.do multiplies the effect column by 100 and keeps the level responses, which leaves the paper's own 237 impulse responses from 37 studies, and their mean at the eight-quarter horizon is -1.216. Against that the move is -5.1%; against the paper's stated -1.2, which the comparator rule prefers, it is -3.8%. An earlier version of this note said the released data could not settle it and quoted -0.31 as its eight-quarter mean. That was the mean of estimateraw, the un-normalised column, and it was wrong. What the paper is doing is worth keeping in view: correcting for publication bias alone moves the two-year response to -0.23%, an 81% move toward zero, and conditioning on best practice moves it nearly all the way back.",
   "quote_source": "pdf",
   "corrected_locator": "house_prices2.pdf p.25, Table 3, first row 'Agnostic on control variables (baseline)', eight-quarter column, -1.154, and the paragraph on p.26 that reads it; the comparator is on p.14, '-0.23% after two years compared to the simple uncorrected mean estimate of -1.2%', and again in the note to Table 1, 'The mean uncorrected effect at the 8-quarter horizon was -1.2'",
   "tier": "table",
   "approximation": "the corrected value is Table 3's baseline row at the eight-quarter horizon, -1.154, which the paper's prose rounds to -1.2; the comparator is printed to one decimal, and the released data pins it at -1.216, so the move is about -5%"
  },
  {
   "project": "spillovers",
   "corrected": 0.94,
   "mean": 0.88,
   "mean_from": "paper",
   "tier": "subsample",
   "approximation": "backward spillovers, the channel the paper's best-practice exercise is built on, rather than all three channels its abstract reports",
   "corrected_quote": "Such defined best-practice estimate of the underlying semi-elasticity, e0, reaches 0.94 and is significant at the 1% level with the 95% confidence interval (0.66, 1.21).",
   "note": "this was excluded on the ground that '9% higher productivity per 10-percentage-point rise in foreign presence' is not the elasticity the effect column holds. The units objection dissolves once the comparator comes from the paper: 9% per 10 percentage points IS a semi-elasticity of 0.9, and the paper prints the precise best-practice value, 0.94, against its own arithmetic average of all published backward-spillover estimates, 0.88. The same paper also reports a publication-bias-corrected 0.178 against that same 0.88, a fivefold exaggeration and an 80% move toward zero. The row plots the best-practice value because that is the paper's headline and what the rule prefers, and because house_prices is read the same way: publication-bias correction shrinks the number and best practice moves it back.",
   "quote_source": "pdf",
   "corrected_locator": "spillovers2.pdf p.7, the best-practice paragraph in Section 5, 'Results of the multivariate meta-regression'; the comparator 0.88 is on p.4, 'The arithmetic average of all published estimates of backward spillovers is 0.88'"
  },
  {
   "project": "incentives",
   "corrected": 0.031,
   "mean": 0.051,
   "mean_from": "paper",
   "tier": "table",
   "approximation": "both numbers are read off tables rather than the abstract: best practice from Table 5, the comparator from Table 1's unweighted mean of all 1,252 estimates. The running text rounds the corrected value to 0.03",
   "corrected_quote": "The overall mean incentive-performance effect based on our definition of “best practice” described above is 0.03.",
   "note": "this was excluded on the direction clause, on the ground that the literature's median (0.0221) and its inverse-variance mean (0.0174) put the corrected 0.031 ABOVE the comparator while the paper's own mean puts it below. That was a misreading of the rule. The comparator clause names exactly two candidates, the paper's own stated mean and the 1%-winsorised mean of the estimate-level data, and here they agree to three decimals: the paper states 0.051 and the shipped data winsorise to 0.0508. A median and an inverse-variance mean are neither of those. Nothing disagrees, and the row goes in at -39%.",
   "quote_source": "pdf",
   "corrected_locator": "incentives.pdf p.31 for the prose and p.30 Table 5, row 'Mean best practice', 0.031 with 95% confidence interval (-0.034, 0.095); the comparator 0.051 is on p.11 and in Table 1 on p.14, row 'All estimates', unweighted, 1,252 estimates"
  },
  {
   "project": "competition",
   "corrected": 0.022,
   "mean": -0.012,
   "mean_from": "paper",
   "tier": "table",
   "small_base": true,
   "approximation": "both sides are the paper's weighted all-country construction, read from Table 1 and Table 6, and both levels are far below the boundary between a zero and a small partial correlation",
   "corrected_quote": "Table 6, Best-Practice Estimates of The Competition Coefficient, weighted: All countries 0.022 (95% CI -0.022, 0.066), Diff. 0.034",
   "note": "excluded before against a comparator of -0.0002, the winsorised mean of the released column. That broke the comparator rule, which prefers the paper's own stated mean: Table 1 gives a weighted all-country mean of -0.012 and Table 6 prints the difference from it, 0.034, explicitly. The pairing is weighted against weighted. The unweighted pair, -0.001 against 0.038, is the same story with a denominator that is a statistical zero, which is why the weighted construction is the one used. The sign changes and both levels are negligible; the row says so twice. The comparator's own confidence interval, (-0.035, 0.011), spans zero, which is a further reason the row is flagged for a small base. The weighted panel is the paper's baseline: it reports the unweighted regressions as a robustness check in the right-hand part of each table.",
   "quote_source": "pdf",
   "corrected_locator": "competition2.pdf p.21 Table 6, row 'All countries', weighted column; the comparator -0.012 is p.7 Table 1, row 'All', weighted mean, 598 estimates",
   "quote_kind": "table"
  },
  {
   "project": "students",
   "corrected": -0.038,
   "mean": -0.017,
   "mean_from": "paper",
   "tier": "table",
   "small_base": true,
   "approximation": "the best-practice value is Table 6's overall row rather than an abstract scalar, and both levels sit below the paper's own boundary between a zero and a small partial correlation",
   "corrected_quote": "Table 6, Overall effect -0.038 (95% CI -0.079, 0.003); the table's note reads: the table presents the mean partial correlation coefficients implied by the Bayesian model averaging exercise and our definition of best practice for various contexts",
   "note": "excluded before as 'corrected about zero against a mean of -0.017, already negligible'. That is an interpretation of the levels, not a clause of the rule: the pair is the paper's own stated simple mean against its own stated best-practice value, in the same units, for the whole literature. It is a useful row precisely because better methodology made the estimate larger rather than smaller. Both levels remain negligible and the row is flagged for it. The best-practice interval spans zero as well, and the paper names the boundary the row is flagged against: 'The boundary between zero and small effects is 0.07, so the bulk of the literature is consistent with the notion that working while in school has no material effect.'",
   "quote_source": "pdf",
   "corrected_locator": "students2.pdf p.14 Table 6, row 'Overall effect'; the comparator -0.017 is on p.4, 'The mean partial correlation coefficient is -0.017'",
   "quote_kind": "table"
  },
  {
   "project": "price_puzzle",
   "corrected": -0.331,
   "mean": -0.067,
   "mean_from": "paper",
   "tier": "horizon",
   "approximation": "six months, the horizon at which the paper's best-practice response bottoms out and the one its headline conclusion names; both numbers are that horizon. At thirty-six months the same comparison runs the other way, -0.116 against -0.561, so the horizon decides the direction here as it does for no other row. The enlargement at six months comes from filtering the misspecifications that create the puzzle, not from the publication-bias correction, which by itself leaves a maximum decrease the paper calls negligible, 0.02%. Peak against peak the paper's corrected response is smaller and faster than the reported average, -41%",
   "corrected_quote": "Table 5, Consequences of misspecifications: Best practice -0.157 (3 months), -0.331 (6 months), -0.225 (12), -0.155 (18), -0.116 (36)",
   "note": "excluded before on the ground that the corrected six-month peak could only be paired with the uncorrected maximum at thirty-six months. That was wrong: Table 3 prints the raw mean response at every horizon, including -0.067 at six months, so a same-horizon pair exists. Six months is the horizon the paper's headline conclusion names, because that is where its best-practice response bottoms out; the raw average response never bottoms out inside the window at all. An earlier draft of this note said the abstract names that horizon. It does not: the abstract is qualitative throughout. The body's conclusion is the anchor. The honest weakness of this row, which no other horizon row shares, is that the direction is a property of the horizon: at thirty-six months the same pair gives -79% and peak against peak gives -41%. All three are in the paper, the ring says the horizon is a choice, and the tooltip now names the reversal.",
   "quote_source": "pdf",
   "corrected_locator": "price_puzzle2.pdf p.24 Table 5, row 'Best practice', 6-month column; the comparator -0.067 is p.17 Table 3, row 'Response (6M)', mean",
   "quote_kind": "table"
  },
  {
   "project": "forward",
   "corrected": 0.611,
   "mean": -0.602,
   "mean_from": "paper",
   "tier": "table",
   "approximation": "the fixed-effects funnel-asymmetry estimate the paper says it prefers, read from Table 2 rather than from the abstract, which reports intervals by currency group",
   "corrected_quote": "Table 2, Panel A: FAT-PET, Mean beyond bias (1/SE) FE 0.611, WLS 0.605, IV 0.332; and: We prefer including dummy variables for each study, which controls for unobserved study-level characteristics that can be related to quality; the resulting specification is labeled “FE” for fixed effects",
   "note": "excluded before because the correction flips the sign. The rule is direction-blind and the index is defined on magnitudes, so a sign reversal is information to disclose rather than grounds to delete a row; the dot is drawn outlined and the reversal is named in the caption, the tooltip and the description. This row is the reason that matters: the magnitude barely moves, +1.5%, while the sign goes from -0.602 to +0.611 and the forward premium puzzle stops being a puzzle. The plotted percentage is the least interesting thing about it, and the figure now says so. The paper states a weighted mean too, -0.840, under which the revision would be -27%. The unweighted mean is the comparator because the paper itself compares its corrected values against it: 'all these estimates are clearly larger than the simple mean estimates of beta, which are reported in Table 1'.",
   "quote_source": "pdf",
   "corrected_locator": "forward2.pdf p.10 Table 2 Panel A, FE column, 'Mean beyond bias', and the preceding paragraph; the comparator -0.602 is p.8 Table 1, row 'All estimates', unweighted mean over 2,989 estimates",
   "quote_kind": "table"
  },
  {
   "project": "inflation",
   "corrected": 0.61,
   "mean": 0.74,
   "mean_from": "paper",
   "tier": "table",
   "approximation": "the comparator is Table 3's pooled mean of all 777 estimates, and the corrected side is the paper's preferred synthesis of the same literature: one row per study, at each study's preferred specification",
   "corrected_quote": "Our preferred headline is 0.61 percent per year, taken from the simple mean of authors' preferred specifications (one row per study, paper-clustered standard errors, 116 clusters, 95% confidence interval [0.25, 0.97]).",
   "note": "first coded against the two-percent target, on the ground that 0.61 is the literature's own mean and not a correction of it. An adversarial review showed the comparator hierarchy forbids that: rung four is reachable only where a paper reports no mean of its own, and Table 3 prints one, 0.74 over all 777 estimates. So the row is the move the paper actually makes, from the pooled mean of every estimate to one row per study at each study's preferred specification, which is the whole of the 18%. The paper reads its result against the two-percent target and finds no bunching there, and its own caveat belongs with the dot: measurement error in published price indices could close, widen or reverse that gap.",
   "quote_source": "pdf",
   "corrected_locator": "inflation.pdf, Discussion; the comparator 0.74 is Table 3 on p.21, row 'Optimal inflation (pp/year)', 777 estimates"
  },
  {
   "project": "euro",
   "corrected": 0.00899,
   "corrected_from": "method_median",
   "methods": [
    {
     "name": "FAT-PET",
     "value": 0.000667
    },
    {
     "name": "ROBUST",
     "value": 0.0265
    },
    {
     "name": "RIM",
     "value": 0.00899
    }
   ],
   "mean": 0.038,
   "mean_from": "pooled",
   "tier": "method_median",
   "mean_quote": "Using fixed effects, the pooled estimate of the euro's γ is very low: a mere 0.038 (β j = 3.87%) with the 95% confidence interval CI = (3.36, 4.39%), although it is very significant (z-stat. = 14.9).",
   "approximation": "the comparator is the paper's own uncorrected fixed-effects pool, because it prints no simple mean of euro-only estimates, and the corrected value is the median of the three correction methods it reports, none of which it prefers",
   "corrected_quote": "Table 1, Tests of publication bias and the true effect, eurozone studies: prec (effect) FAT-PET 0.000667 (0.05), ROBUST 0.0265 (1.52), RIM 0.00899 (0.90)",
   "note": "excluded before as a verbal zero. Table 1 does print corrected point values, three of them, so the reason had to be the comparator instead, and the released column here mixes eurozone with other currency unions and so measures a different thing. The paper's own uncorrected pooled estimate for exactly the eurozone, 0.038, is the honest comparator. The median of the three methods is 0.00899, which is one of them, RIM, and not an interpolation. All three are statistically insignificant and all three carry the same conclusion, which is what makes a median admissible here; the paper contrasts them with the 5 to 10 percent and 10 to 15 percent figures the literature had been quoting.",
   "quote_source": "pdf",
   "corrected_locator": "euro2.pdf p.8 Table 1; the comparator 0.038 is on p.4",
   "quote_kind": "table"
  },
  {
   "project": "remittances",
   "corrected": 0.039,
   "corrected_from": "method_median",
   "methods": [
    {
     "name": "Top10",
     "value": 0.025
    },
    {
     "name": "WAAP",
     "value": 0.042
    },
    {
     "name": "Andrews and Kasy",
     "value": 0.121
    },
    {
     "name": "stem-based",
     "value": 0.036
    }
   ],
   "mean": 0.103,
   "mean_from": "paper",
   "tier": "method_median",
   "approximation": "the median of the four correction methods the paper reports, none of which it prefers; the Andrews and Kasy estimate lies above the uncorrected mean and the other three below it",
   "corrected_quote": "Table 3, Alternative approaches to correcting for publication bias: Top 10 0.025, WAAP 0.042, A&K 0.121, Stem-based bias correction model 0.036, Uncorrected mean 0.103",
   "note": "excluded before because the four corrected values straddle the comparator and the paper names no preferred one, so a midpoint would invent a central value the authors declined to give. It enters now under a stated rule rather than a judgement: where a paper reports several correction methods, prefers none, and they agree on the substantive conclusion, the median of them is used and every one is listed. All four here are positive and all four are described by the paper as economically small, which is the agreement the rule requires. The tooltip and this note carry all four. The paper certifies the agreement the median rule requires: 'once the correction for publication bias is performed, the underlying effect of remittances on economic growth is small in all of the methodological approaches: none passes Doucouliagos's bar for a medium effect.'",
   "quote_source": "pdf",
   "corrected_locator": "remittances2.pdf, Table 3 and the paragraph that reads it; the comparator 0.103 is Table 1, row 'Simple Average', long-term column, 469 estimates",
   "quote_kind": "table"
  },
  {
   "project": "correlations",
   "corrected": 0.05,
   "mean": 0.128,
   "mean_from": "pooled",
   "tier": "illustration",
   "approximation": "the worked example the paper applies its estimator to, an external psychology meta-analysis of earnings potential and romantic evaluations, rather than a literature of the paper's own; Hunter and Schmidt corroborates it at 0.047",
   "mean_quote": "Conventional random-effects estimate the correlation of the earnings potential of the target on women's romantic evaluations as: 0.128; 95% CI (0.092, 0.164), k = 73.",
   "corrected_quote": "On the other hand, UWLS+3 reduces the RE estimate by over 60%: 0.050, 95% CI (0.022, 0.078). That is, UWLS+3 reduces a small correlation to a trivial one by Cohen's benchmarks.",
   "note": "excluded before as 'a methods paper about converting effects to partial correlations; no literature effect to correct'. Too broad: section 3.5 is a worked example with a conventional estimate and the paper's own focal correction for the same 73 correlations, and the paper states the revision itself, 'over 60%'. UWLS+3 is the estimator the paper introduces and recommends, not the smallest one available. The literature is external and appears nowhere else on this site, so nothing is counted twice; the ring says the dot is an illustration rather than a literature of the authors' own. The comparator is the conventional random-effects pooled estimate rather than a simple mean, so the row sits on the second rung of the comparator hierarchy and quotes the sentence it comes from.",
   "quote_source": "pdf",
   "corrected_locator": "correlations/paper/, section 3.5 'An illustration'; both numbers are in the same two paragraphs, and the source meta-analysis is Eastwick et al. (2014)"
  },
  {
   "project": "bma",
   "corrected": 0.021,
   "mean": -0.002,
   "mean_from": "paper",
   "verbal_zero": true,
   "small_base": true,
   "tier": "unmoved",
   "quote_source": "pdf",
   "quote_kind": "table",
   "corrected_locator": "Table 4, Test of publication bias, study fixed effects column: Constant 0.021, standard error 0.015, p = 0.150",
   "corrected_quote": "Constant 0.021 0.015 0.150",
   "mean_quote": "The simple mean of the remaining estimates is -0.002, not significantly different from zero at any conventional level.",
   "mean_quote_kind": "sentence",
   "approximation": "Neither number is distinguishable from zero: the simple mean of -0.002 is not significant at any conventional level, and the bias-corrected constant of 0.021 has p = 0.150, as does the publication-bias term itself. The paper's own conclusion is that horizontal spillovers are on average zero, and correcting for selection does not change that",
   "note": "the owner's ruling: the correction leaves this literature where it was, at zero. Dividing 0.021 by 0.002 would report +950%, which is a property of the denominator rather than a finding, so the row is plotted as no change and marked small_base. The paper declines a best-practice definition on principle: 'If we excluded studies that do not correspond to a particular definition of best practice, we would greatly increase the subjectivity of our analysis.'"
  },
  {
   "project": "habits",
   "corrected": 0.4,
   "mean": 0.4,
   "mean_from": "paper",
   "verbal_zero": true,
   "tier": "unmoved",
   "quote_source": "pdf",
   "quote_kind": "sentence",
   "corrected_locator": "Appendix A.4, Publication Bias, closing paragraph, and the abstract for the mean of 0.4",
   "corrected_quote": "we find little evidence of any systematic bias resulting from this selection",
   "mean_quote": "The mean reported strength of habit formation equals 0.4",
   "mean_quote_kind": "sentence",
   "approximation": "The paper tests for publication bias and reports finding none that matters: the funnel plots 'show little signs of asymmetry (especially compared to other fields of empirical economics)' and the formal tests find 'little evidence of any systematic bias'. Its corrected value is therefore the 0.4 it reports, and the correction moves nothing",
   "note": "the owner's ruling: the difference is zero. Recorded here rather than left out, because a literature the correction leaves alone is a result this figure should show, not a gap in it. The 0.4 is the mean reported estimate; the paper states no separate corrected scalar because its own tests say none is needed."
  },
  {
   "project": "resource_curse",
   "corrected": 0.026,
   "mean": -0.078,
   "mean_from": "paper",
   "tier": "table",
   "quote_source": "pdf",
   "quote_kind": "table",
   "corrected_locator": "Table 3, Tests of the true effect and publication selection, Panel A, clustered OLS column: Constant (true effect) 0.026, t = 1.69, p = 0.099",
   "corrected_quote": "Constant (true effect) 0.026",
   "mean_quote": "The arithmetic mean yields a partial correlation coefficient of -0.078 with a 95% confidence interval",
   "mean_quote_kind": "sentence",
   "approximation": "The paper reports four constants across four estimators and names no preferred one: 0.026 (clustered OLS), 0.038 (instrumental variables), and two from its fixed and mixed effects specifications. This row plots the first, the clustered OLS column of Table 3, which is the specification the paper leads with; the instrumental-variables column would give a similar move and the other two would not",
   "note": "the owner's ruling on which of the four constants to plot. The comparator is the simple average partial correlation of -0.078 from Table 2. The correction removes the curse and reverses its sign, which is the paper's own conclusion in words: support for the resource curse is weak once publication bias and method heterogeneity are taken into account."
  },
  {
   "project": "fdi",
   "corrected": 1.1,
   "mean": -0.118804,
   "mean_from": "codebook",
   "mean_column": "e_w",
   "tier": "projection",
   "quote_source": "pdf",
   "quote_kind": "sentence",
   "corrected_locator": "Section VI, Concluding remarks: 'we compute the spillover value implied by the best practice methodology in the literature for the year 2018. The result is 1.1 overall'",
   "corrected_quote": "we compute the spillover value implied by the best practice methodology in the literature for the year 2018. The result is 1.1 overall",
   "approximation": "The corrected value is the paper's best-practice construction, but it is evaluated at a data year of 2018, later than any vintage in its own sample: 'for the year of the data, we plug in 2018 in order to estimate current effects, assuming the trend that we see in the literature has continued to this day'. The comparator pools every data year the sample actually contains, so part of this move is the projection forward rather than the correction itself. The paper's own summary of the literature points the other way: 'on average, the reported spillovers seem to be zero, even after controlling for potential publication selection bias'",
   "note": "the owner's ruling. The paper states no single simple mean over its 332 estimates, only category means of -0.1, -0.16 and -0.09 for horizontal, backward and forward spillovers, so the comparator is the mean of the winsorised effect column the paper itself released, e_w, over all 332 rows. This literature is not in the harmonised table because the file carries no standard error, so the comparator is read from the published codebook instead. Data year is one of the variables the paper finds matters, with a coefficient of 0.150 in the general model and 0.135 in the specific one; its posterior inclusion probability in the Bayesian model averaging is 0.144, and the paper says of it and three others that 'the posterior inclusion probabilities fall short of 50%'."
  }
 ],
 "excluded": [
  {
   "project": "alphas",
   "why": "the same hedge fund performance literature is already in the figure as `hedge`, against the identical 0.36 comparator, and one literature gets one dot. Two rows on it at -7% and -78% would be incoherent. (Earlier reasons here rested on its best-practice value conditioning on the sample maximum data year. That is close to the ground on which fdi is excluded, but it is not the ground here: the duplicate literature decides it, and dst, border and lags are admitted on within-sample constructions of the same family.)"
  },
  {
   "project": "conventional_wisdom",
   "why": "a review of 24 other meta-analyses, not a meta-analysis of one literature; its own revision figure is on its page"
  },
  {
   "project": "debate",
   "why": "a randomised experiment on feedback, not a meta-analysis with a corrected mean"
  },
  {
   "project": "guidelines",
   "why": "the practitioner's guide: a methods text, not a meta-analysis"
  },
  {
   "project": "maive",
   "why": "the estimator paper itself; it has no literature of its own to correct"
  },
  {
   "project": "outliers",
   "why": "a methods paper on outlier treatment"
  },
  {
   "project": "pcc",
   "why": "a methods paper on partial correlation coefficients"
  },
  {
   "project": "pcc_survey",
   "why": "a cross-literature aggregate that demonstrably counts this figure's own dots twice. Its Table 1 does give a pair, a median random-effects estimate of 0.087 against a median UWLS of 0.061 across 172 economics meta-analyses, and the paper states the revision itself: 'UWLS's median is 30% smaller than RE's.' But its released data names those 172, and two of them are rows in this figure: Valickova, Havranek and Horvath on financial development and growth, plotted here as finance_growth, and Zigraiova and Havranek on bank competition and stability, plotted here as competition. Plotting it would be plotting a summary of the figure inside the figure, which is the ground on which `conventional_wisdom` is excluded, and it has to apply to both or neither. Its own worked illustration, on ICT and growth, is out on a further ground: the paper refuses the inference, saying 'We make no inference about the true unconditional average PCC of ICT.' Both medians are also medians across meta-analyses rather than means within one, which no other row is."
  },
  {
   "project": "spillovers_bias",
   "why": "the same 3,626 estimates are already in the figure as `spillovers`, and one literature gets one dot. This paper does state a pair -- a publication-bias-corrected backward spillover of 0.12 against a simple average of 1.14 for estimates in high-quality journals, an 89% move toward zero -- but plotting it beside `spillovers` at +7% would put the same literature in the figure twice with opposite answers, which is the ground on which `alphas` is excluded. The two differ because this paper corrects for publication bias alone while its companion also conditions on best practice; the `spillovers` note records both."
  },
  {
   "project": "substitution",
   "why": "the same 2,735 estimates are already in the figure as `eis`, and one literature gets one dot. Independently, this paper states no corrected scalar: its headline is which country characteristics explain the heterogeneity, not a level."
  },
  {
   "project": "trust",
   "why": "the paper declines to estimate the quantity this figure plots, and says so: 'We do not fit a full selection model, because the target is the institutional moderators of reported slopes, not the bias-corrected mean size premium.' Its headline is which institutions move the premium, not a level."
  }
 ]
}