Datasets
This catalogue collects 46 estimate-level datasets from the empirical meta-analyses on this site, each in a standard format.
There are two products here. The archive publishes each paper's original file with automatic CSV and Parquet conversions, a codebook, and a DOI. The harmonised table pools those datasets into one file with a common set of columns, so they can be compared across literatures, and it is stable as of version 2.0.0. For one paper's data use the archive; to work across literatures use the harmonised table, and read its caveats below.
One file, every literature
52,800 estimate-level rows from 42 literatures in a single table, with the effect, its standard error, the sample size, and the characteristics that recur across literatures.
Where the estimates fall. Absolute t-statistics for the 52,582 pooled rows with a headline-scale t-statistic; the first panel shows the 81% below 6, and the second panel magnifies half a unit around 1.96 on its own vertical scale. In equal 0.05-wide bins ([1.91, 1.96) against [1.96, 2.01), so an estimate reported as exactly 1.96 counts as above) there are 550 estimates just below and 683 just above, 1.24 times as many. The observed step is modest, and it is specific to that one threshold: at 1.645 and at 2.576 the count falls as they are crossed instead of jumping. What the picture shows is the shape of the distribution; it does not measure p-hacking. Selective reporting, specification search, and genuinely large effects all leave marks here, and estimates are clustered within studies. The bin counts are in t_distribution.csv.
Also as Parquet, which is smaller and keeps column types.
df <- read.csv("https://meta-analysis.cz/data/v1/estimates_harmonised.csv")import pandas as pd
df = pd.read_parquet("https://meta-analysis.cz/data/v1/estimates_harmonised.parquet")Raw effect levels are not comparable across literatures. An elasticity, a
partial correlation, and a dollar value per tonne of carbon all sit in the same column;
effect_units records which is which. Analyse within each literature. Comparing
across them needs an explicitly standardised measure, and relative changes are meaningful only
where the baseline is safely away from zero.
The harmonised table is version 2.0.0: 52,800 observations from 42 literatures. Every row records the file and the columns it came from, so any value can be traced back to its published dataset.
All 42 pooled literatures are verified:
22 domain-reviewed against the paper's own replication code or,
where a paper ships none, against its published results, and 20
code-traced. Every dataset publishes an audit_status saying
which it is.
One literature at a time
The converted source files hold 67,606 rows, every one keeping the
variables coded for its original paper. After applying each paper's analysis selection and,
where necessary, reshaping the source data into estimate-level observations, the catalogue
below represents 58,405 estimates. The CSV and Parquet are automatic
conversions carrying one data sheet per source workbook, the one named in
datasets.json; any other sheet is absent, and no column is independently
re-checked. For anything load-bearing, work from the original file.
Each dataset is also published on its own, with all of the variables coded for its paper rather than the shared subset, and a codebook describing every column.
The Estimates column counts what each literature contributes to the pooled
table; for the datasets that are not pooled, it counts the rows in the published file after
the paper's own analysis filters. Neither is always the number in the paper's abstract: a
published file is usually the analysis subset rather than the full collection, a pooled
contribution can be smaller still where estimates overlap another literature (trust
contributes the 212 rows size does not already carry, out of 1,613 in its file),
and two literatures contribute a row per impulse-response horizon.
datasets.json carries both counts and a reconciliation for
every dataset where they differ.
46 datasets, 58,405 estimates in their analysis samples (67,606 rows in the source files). 42 literatures are pooled into the harmonised table.
Published but not pooled
Four datasets are published but stay out of the pooled table: two carry no per-estimate standard error, and two duplicate a dataset already included, which would otherwise be counted twice.
- fdi
- no per-estimate precision exists: the file has no standard error, t-statistic, weight, or inverse-SE column, so SE-based synthesis is impossible
- hedge
- duplicate of alphas: identical estimates, row for row
- lags
- outcome is the transmission lag (mon_bot, months to the price trough); it has no sampling standard error, so it cannot enter an effect/SE table. The catalogue used to name `lags` as the effect column, which is the VAR's lag ORDER (median 4), not the transmission lag (mean 34.9 months) -- the same record's own direction_note said so.
- substitution
- duplicate of eis: identical estimates, row for row
Machine-readable
The whole collection is available as JSON: an index of every dataset with paper, DOI, file URLs, and column roles, a Frictionless data package, and a Croissant record. Full documentation is in the API README.
Who else has used these datasets
Slichter and Tran, who have no connection to this project, ask whether better journals publish better estimates, and answer with 14,387 published estimates from 24 literatures. They name three sources for those literatures, and this site is one of them: “a database of meta-analyses operated by the Institute of Economic Studies at Charles University”, with meta-analysis.cz in the footnote. Eleven of their twenty-four literatures are datasets from this page, and those eleven supply 502 of the 871 primary studies in their sample. Their finding is that within a literature the mean and variance of published estimates have little or no correlation with journal rank: the chance that an estimate drawn from a top journal is closer to the truth than one drawn from a mid-ranked journal is about 51%, a coin flip. The paper is accepted at the Journal of Political Economy Microeconomics; until it appears there, it is IZA Discussion Paper 17960.
That is the use this collection is for. Pooling the estimates was worth doing if someone outside the project can ask a question none of the original papers asked, and answer it from the pooled table. The eleven literatures are tuition and enrolment, individual discount rates, capital-labor substitution, the social cost of carbon, intertemporal substitution, skill substitution, water demand, the Armington elasticity, excess sensitivity of consumption, student employment, and the price elasticity of gasoline.
Licence and citation
Everything here is CC BY 4.0, including for training, provided you give credit. The licence file states the terms, and the citation for the collection is in the footer of every page.
Each dataset was assembled for a specific paper. When you use one, cite that paper. The index carries every paper’s title, authors, and DOI. That is how attribution under CC BY is satisfied here.
The Zenodo identifier in that footer is the concept DOI: it always resolves to the newest version. In a replication package, where the exact files matter, cite this release instead: 10.5281/zenodo.22647394, version 2.0.0.