maddison-project

The Maddison Project Database (MPD): long-run estimates of real GDP per capita and population by country, from year 1 to the present, kept by the Groningen Growth and Development Centre (GGDC), University of Groningen, continuing the work of Angus Maddison. Three releases, 2018, 2020 and 2023, each as published, with a typed Parquet of each table beside the original files.

CC BY 4.0, with the citations each release asks for (see Licence).

Releases

subset	rows	countries	years	values	from
2023.full_data	131,144	169	1 to 2022	gdppc 21,366, pop 17,875	raw/2023/mpd2023_web.xlsx, sheet Full data
2020.full_data	21,682	169	1 to 2018	gdppc 19,706, pop 17,199	raw/2020/mpd2020.xlsx, sheet Full data
2018.full_data	19,873	169	1 to 2016	cgdppc 17,441, rgdpnapc 17,712, pop 16,859	raw/2018/mpd2018.xlsx, sheet Full data
2018.regional_data	1,039	7 regions and the World	1870 to 2016	761 each	raw/2018/mpd2018_region_data.xlsx, sheet Long data
2018.region_composition	147	147			raw/2018/mpd2018_region_data.xlsx, sheet Composition
2018.benchmark_1990	17,872	169	1 to 2016	gdppc 17,872	raw/2018/mpd2018_1990bm.xlsx, sheet LongForm

"values" counts the rows where the column is not empty. The 2023 table has a row for every country and every year that any country has (776 years), so most of its rows are empty; the 2018 and 2020 tables have rows only where there is a value.

The releases cannot be stacked into one series. Each is a full revision made with its own method, so the same country and year has other values in another release. The 2020 release page says it "uses a different methodology compared to the 2018 update", having "to some extent gone back to the original Maddison approach". Of the 19,213 country-years with gdppc in both 2020 and 2023, 6,601 differ. Japan in 1950:

release	column	value
2018	cgdppc (2011 US$, multiple benchmarks)	2,519
2018	rgdpnapc (2011 US$, 2011 benchmark)	3,023
2018	benchmark_1990.gdppc (1990 US$)	1,921
2020	gdppc (2011 US$)	3,062
2023	gdppc (2011 US$)	3,062

and in 2016: 36,452 and 37,465 (2018), 37,465 (2020), 37,384.56 (2023). Use one release, and say which.

The xlsx and dta files of each release and the documentation papers of 2018 and 2020 are under raw/<release>/, byte for byte as served; each release's MANIFEST.json gives their URLs, sizes, sha256 and Last-Modified, and for 2023 the SHA-1 DataverseNL states. MPD 2013 and the Maddison Database 2010 are not here: their pages state no licence.

Files and columns

The Parquet is made from the xlsx with DuckDB's excel extension. The dta files were not read and are only carried unchanged.

Every cell is read as its text and typed by the column's name, never by inference:

columns	type
year	INTEGER
gdppc, cgdppc, rgdpnapc, pop	DOUBLE
everything else	VARCHAR, the text of the cell

The columns of each table, with the definitions in the workbook's Legend or Notes sheet:

table	column	meaning
all country tables	countrycode	3-letter ISO country code; includes former states such as CSK, SUN, YUG
	country	country name
	year	year
2023, 2020 full_data	gdppc	real GDP per capita in 2011 US$
	pop	population, mid-year (thousands)
2023 full_data	region	one of 8 regions, as the release groups countries
2018 full_data	cgdppc	real GDP per capita in 2011 US$, multiple benchmarks (for cross-country income comparisons)
	rgdpnapc	real GDP per capita in 2011 US$, 2011 benchmark (for cross-country growth comparisons)
	pop	population, mid-year (thousands)
	i_cig	whether the observation is extrapolated, a benchmark, or interpolated
	i_bm	for benchmarks, the kind of benchmark
2018 benchmark_1990	gdppc	GDP per capita in 1990 US$, Maddison benchmark, updated time series
2018 regional_data	region	two-letter code: af, ea, ee, la, wa, we, wo, wd (World)
	region_name	region name
	year, cgdppc, rgdpnapc, pop	as in 2018 full_data
2018 region_composition	countrycode	country
	region1870	its region for the years before 1950 (40 countries)
	region1950	its region from 1950 (all 147)

-- DuckDB, straight off the Hub: GDP per capita of Japan, release 2023
SELECT year, gdppc, pop
FROM 'hf://datasets/yuiseki/maddison-project/parquet/2023/full_data.parquet'
WHERE countrycode = 'JPN' AND gdppc IS NOT NULL
ORDER BY year;

Note that a chart like this one falls under the project's request to cite the original sources (see Licence).

Things to know

• In 2018 full_data, the Legend describes i_cig as "0/1/2" and i_bm as codes 1 to 5, but the xlsx holds the labels: i_cig is Extrapolated (10,318 rows), Interpolated (6,032), Benchmark (1,364) or an empty string (2,159); i_bm is ICP PPP estimates (1,144), Multiple of subsistence (122), Historical income benchmarks (68), Real wages and urbanization (21), Braithwaite (1968) PPPs (9) or an empty string (18,509). The empty strings are kept as empty strings, as in the sheet, not turned into NULL.
• The 2018 sheet Full data has 126 formatted but empty rows at its foot. They are not in the Parquet. No other row was dropped.
• In 2018, GDP per capita is rounded to the nearest dollar and population to the nearest thousand (the Legend says so). In 2020 and 2023 the values are not rounded; 2023 has 8,841 non-integer gdppc and 9,565 non-integer pop.
• 2018 regional_data: before 1950 the regions are made of fewer countries (region1870 in region_composition) and linked to the 1950 regions. The Legend: "Data for the World in 1870 do not include Eastern Europe. Data for the World before 1950 do not include Africa and Western Asia." The years are not continuous before 1950; 278 of the 1,039 rows have no values.
• 2018 region_composition lists 147 countries, not all 169.
• The 2018 regions (7 two-letter codes and wd for the World, with Africa and Western Asia) are not the 2023 region values (8 names, with Middle East and North Africa, and Sub Saharan Africa).
• The 2020 and 2023 workbooks also have a Regional data sheet, wide, with two header rows and unlabelled columns, and the country values again in wide form (GDP pc, Population). They are not made into Parquet; they are in the xlsx under raw/. So are the 2018 sheet Partial countries (England from 1252 and Holland, in wide form with two header rows), the 2023 sheet Maddison original sources, and the Sources sheets.
• The 2018 release has no Sources sheet; its page points to Appendix A of Bolt et al. (2018), the paper in raw/2018/gd174.pdf.
• The xlsx and the dta of a release were not compared with each other.

Checked before upload

scripts/02_verify.py in the source repository checks:

• every raw file against its release's MANIFEST.json (size, sha256, and for 2023 the SHA-1 DataverseNL states) and that it opens as its type
• each sheet is read again with the Python standard library alone (zipfile and xml.etree, not DuckDB), and the Parquet has the same number of rows, equal to the pinned counts above
• every value of every row equals the cell it came from, typed the same way, in sheet order (sha256 over all values printed for both sides)
• the columns are the sheet's header, each of the type its name gives
• (countrycode, year) is unique in the country tables, (region, year) in regional_data, countrycode in region_composition
• a DuckDB with no extension loaded opens every file
• no row group has a bloom filter

Licence

Each release page states CC BY 4.0. Read on 2026-10-05:

2023, https://www.rug.nl/ggdc/historicaldevelopment/maddison/releases/maddison-project-database-2023 :

Maddison Project Database, version 2023 by Jutta Bolt and Jan Luiten van Zanden is licensed under a Creative Commons Attribution 4.0 International License.

DataverseNL (https://doi.org/10.34894/INZBF2) gives the 2023 dataset the licence CC-BY-4.0.

2020, https://www.rug.nl/ggdc/historicaldevelopment/maddison/releases/maddison-project-database-2020 :

Maddison Project Database, version 2020 by Jutta Bolt and Jan Luiten van Zanden is licensed under a Creative Commons Attribution 4.0 International License.

2018, https://www.rug.nl/ggdc/historicaldevelopment/maddison/releases/maddison-project-database-2018 :

Maddison Project Database, version 2018 by Jutta Bolt, Robert Inklaar, Herman de Jong and Jan Luiten van Zanden is licensed under a Creative Commons Attribution 4.0 International License.

The 2020 and 2023 pages also ask, under "Attribution requirement":

All original papers must be cited when: the data is shown in any graphical form subsets of the full dataset that include less than a dozen (12) countries are used for statistical analysis or any other purposes A list of original papers can be found in the source sheet of the database. When neither a) or b) apply, then the MPD as a whole should be cited.

(the two conditions are a list on the page). The Sources sheet is in raw/2020/mpd2020.xlsx and raw/2023/mpd2023_web.xlsx. For 2018 the page says: "For the references to the original research on individual countries, see Appendix A of Bolt et al. (2018)."

What was changed: the xlsx, dta and pdf files are unchanged. The Parquet holds six sheets of those xlsx files, typed by column name, with blank rows left out; no value was changed. LICENSE beside this file says the same.

Citation

Release 2023, as its page asks:

Bolt, Jutta and Jan Luiten van Zanden (2024), "Maddison style estimates of the evolution of the world economy: A new 2023 update", Journal of Economic Surveys, 1-41. DOI: 10.1111/joes.12618

Release 2020, as its page asks:

Maddison Project Database, version 2020. Bolt, Jutta and Jan Luiten van Zanden (2020), "Maddison style estimates of the evolution of the world economy. A new 2020 update".

(the paper is raw/2020/wp15.pdf)

Release 2018, including its regional and 1990 benchmark data, as its page and the workbooks' Legend ask:

Maddison Project Database, version 2018. Bolt, Jutta, Robert Inklaar, Herman de Jong and Jan Luiten van Zanden (2018), "Rebasing 'Maddison': new income comparisons and the shape of long-run economic development", Maddison Project Working paper 10

(the paper is raw/2018/gd174.pdf; the 1990 benchmark workbook writes "Maddison Project Database, version 2018; 1990$ benchmark.")

When the data is charted, or fewer than 12 countries are used, also cite the original papers in the release's Sources sheet (2018: Appendix A of Bolt et al. 2018).