# Sources for cir/panel.csv Retrieved 2026-10-03. Raw downloads are kept unchanged in `cir/raw/`; `cir/build.ml` rebuilds `panel.csv` and `panel_flags.csv` from them with no network access (`dune exec data/cir/build.exe`, from the repository root). No value in the panel was typed by hand. Panel: 38 OECD member countries (AUS AUT BEL CAN CHL COL CRI CZE DNK EST FIN FRA DEU GRC HUN ISL IRL ISR ITA JPN KOR LVA LTU LUX MEX NLD NZL NOR POL PRT SVK SVN ESP SWE CHE TUR GBR USA) × 1995–2024 = 1,140 rows. An empty cell means the source publishes no value (or flags it `O`, missing). `panel_flags.csv` has the same shape and holds the source observation-status codes for each cell: `B` time-series break, `D` definition differs, `E` estimated, `P` provisional. For `researchers_per_1000` the flag is the union of the flags on its two inputs. ## 1. OECD Main Science and Technology Indicators (MSTI) - Dataflow `OECD.STI.STP:DSD_MSTI@DF_MSTI(1.3)`, "Main Science and Technology Indicators". DSD dimensions: REF_AREA, FREQ, MEASURE, UNIT_MEASURE, PRICE_BASE, TRANSFORMATION. - Download (all countries): `https://sdmx.oecd.org/public/rest/data/OECD.STI.STP,DSD_MSTI@DF_MSTI,/.A.B+G+B_RS+TOT_EMP+B_FG+GDP+TOT_POP.PT_B1GQ+USD_PPP+FTE+PS+PT_BERD..?startPeriod=1995&format=csvfilewithlabels` saved as `raw/msti_all.csv` (6.3 MB, 22,198 rows, latest year 2025). - Download (France, every MSTI measure, for inspection): `https://sdmx.oecd.org/public/rest/data/OECD.STI.STP,DSD_MSTI@DF_MSTI,/FRA.A....?startPeriod=1995&format=csvfilewithlabels` saved as `raw/msti_FRA.csv`. - The SDMX API publishes no release date for this dataflow. The vintage is the one served on 2026-10-03. - Dataflow description as published: "The OECD Main Science and Technology Indicators (MSTI) provide a set of indicators that compare the Science and Technology (S&T) performance of OECD member countries and selected non-member economies. The MSTI database focuses principally on tracking financial and human resources devoted to research and experimental development (R&D), as defined in the OECD Frascati Manual, [...]" | panel column | MSTI key (MEASURE.UNIT_MEASURE.PRICE_BASE.TRANSFORMATION) | label as published | |---|---|---| | `berd_gdp` | `B.PT_B1GQ._Z._Z` | Business Enterprise Expenditure on R&D (BERD), Percentage of GDP | | `berd_ppp` | `B.USD_PPP.Q._Z`, UNIT_MULT 6, BASE_PER 2020 | BERD, US dollars, PPP converted, constant prices; millions of 2020 USD PPP | | `gerd_gdp` | `G.PT_B1GQ._Z._Z` | Gross Domestic Expenditure on R&D (GERD), Percentage of GDP | | `researchers_per_1000` | derived: `B_RS.FTE._Z._Z` ÷ `TOT_EMP.PS._Z._Z` (TOT_EMP in thousands) | Business enterprise sector researchers (FTE) per 1 000 total employment | | `direct_funding_berd` (extra) | `B_FG.PT_BERD._Z._Z` | BERD financed by government, Percentage of BERD | `researchers_per_1000` is the only derived column from MSTI. MSTI publishes a per-1,000-employment ratio for total researchers (`T_RS.10P3EMP`) but not for the business sector. The derivation reproduces OECD's own method: for France 2020, `T_RS.FTE` 321,398 ÷ `TOT_EMP` 28,645 thousand = 11.2200, equal to the published `T_RS.10P3EMP` value of 11.2200384. ## 2. OECD R&D Tax Incentives database ### Implied tax subsidy rates (1 − B-index) - Dataflow `OECD.STI.STP:DSD_RDTAX@DF_RDSUB(1.0)`, "Implied tax subsidy rates on R&D expenditures". Dimensions: REF_AREA, FREQ, MEASURE (`RDSUB`), UNIT_MEASURE (`IX`), SIZE (`LARGE`, `SME`), PROFIT_SCENARIO (`PROFITABLE`, `LOSS_MAKING`). - Download: `https://sdmx.oecd.org/public/rest/data/OECD.STI.STP,DSD_RDTAX@DF_RDSUB,/all?format=csvfilewithlabels` saved as `raw/rdsub_all.csv` (6,016 rows, 2000–2025). The series starts in 2000, so 1995–1999 are empty for every country. - Columns: `tax_subsidy_large` = LARGE × PROFITABLE, `tax_subsidy_sme` = SME × PROFITABLE; extras `tax_subsidy_large_loss`, `tax_subsidy_sme_loss` are the LOSS_MAKING scenario. - Definition as published: "Implied R&D tax subsidy rates are defined as 1 minus the B-Index, a measure of the before-tax income needed by a "representative" firm to break even on one additional monetary unit of R&D outlay (Warda, 2001; OECD, 2023). [...] The OECD time-series estimates of implied R&D tax subsidy rates is based on headline tax credit and allowance rates. Due to limited historical data availability, the estimates are not adjusted for provisions that bound the tax benefits received by firms (e.g. ceilings, thresholds). They therefore provide an upper bound for the marginal tax subsidy implied by R&D tax relief measures at central government level across countries over time." - Consequence for France: the EUR 100m ceiling of the post-2008 CIR (above which the rate falls to 5%) is not reflected. For a large firm whose R&D exceeds EUR 100m the marginal subsidy is overstated. ### Government tax relief for business R&D (GTARD) - Dataflow `OECD.STI.STP:DSD_RDTAX@DF_RDTAX(1.0)`, "R&D tax expenditure and direct government funding of BERD". - Download: `https://sdmx.oecd.org/public/rest/data/OECD.STI.STP,DSD_RDTAX@DF_RDTAX,/.A.RDTAX+DF+RDTAXSUB.PT_B1GQ+PT_BERD..?format=csvfilewithlabels` saved as `raw/rdtax_all.csv` (8,700 rows, 2000–2024). France, every measure: `.../DSD_RDTAX@DF_RDTAX,/FRA.A....?format=csvfilewithlabels` saved as `raw/rdtax_FRA.csv`. - Columns: `gtard_gdp` = `RDTAX.PT_B1GQ`, "Indirect government support through R&D tax incentives (GTARD)", % of GDP; extras `gtard_sub_gdp` = `RDTAXSUB.PT_B1GQ` (subnational part) and `direct_funding_gdp` = `DF.PT_B1GQ`, "Government-financed BERD", % of GDP. - `RDTAX` already includes subnational relief. Check in `raw/rdtax_CAN_HUN_2010_identity_check.csv` (`.../CAN+HUN.A.DF_GTARD+RDTAX+RDTAXSUB+DF.PT_B1GQ..?startPeriod=2010&endPeriod=2010`): Canada 2010 `DF_GTARD` 0.2194 = `RDTAX` 0.184 + `DF` 0.0354, and Hungary 0.2576 = 0.1631 + 0.0945. `RDTAXSUB` is therefore a component of `RDTAX`. Do not add the two. - Description as published: "The OECD R&D Tax Incentives database provides a set of indicators that reflect the level and structure of central and subnational government support for business R&D in form of R&D tax incentives and direct funding [...] These estimates of the cost of central and subnational R&D tax relief have been combined with data on direct government funding of business expenditure on R&D (BERD), as compiled by National Statistical Offices based on reports from firms [...]" - France `gtard_gdp` for 2020–2023 is flagged `P` (provisional), and 2024 is `O` (missing). ## 3. Covariates ### GDP per capita, PPP: World Bank WDI - `NY.GDP.PCAP.PP.KD`, "GDP per capita, PPP (constant 2021 international $)". WDI "Last Updated Date" 2026-07-13. - Download: `https://api.worldbank.org/v2/en/indicator/NY.GDP.PCAP.PP.KD?downloadformat=csv` saved as `raw/wb_NY.GDP.PCAP.PP.KD.zip` (unzipped alongside). ### Manufacturing share of value added: OECD national accounts - Dataflow `OECD.SDD.NAD:DSD_NAMAIN10@DF_TABLE1_OUTPUT(2.0)`, "Annual GDP and components - output approach". - Download: `https://sdmx.oecd.org/public/rest/data/OECD.SDD.NAD,DSD_NAMAIN10@DF_TABLE1_OUTPUT,/A..S1.S1.B1G._Z.C+_T._Z.XDC.V.N.T0101?startPeriod=1995&format=csvfilewithlabels` saved as `raw/oecd_va_C_T.csv`. - `manuf_share` = 100 × B1G (gross value added), ACTIVITY `C` (Manufacturing) ÷ B1G ACTIVITY `_T` (Total, all activities), national currency, current prices. This is a share of total gross value added, not of GDP. Gaps: Canada has C only for 1997–2022, the US from 1997, Colombia from 2005, Chile from 1996, New Zealand to 2023. OECD status `P` (provisional) is carried into the flags file, and status `A` (normal) is dropped. - France: the share jumps from 10.42% (2022) to 11.28% (2023) because B1G C rises from EUR 246.9bn to 287.2bn in the source. The World Bank manufacturing series also rises in 2023 (9.30 to 9.77% of GDP). This is recorded as published and was not investigated further. ### Extra covariates: World Bank WDI (WDI last updated 2026-07-13) - `manuf_gdp_wb` = `NV.IND.MANF.ZS`, "Manufacturing, value added (% of GDP)". `https://api.worldbank.org/v2/en/indicator/NV.IND.MANF.ZS?downloadformat=csv` - `mht_manuf_share` = `NV.MNF.TECH.ZS.UN`, "Medium and high-tech manufacturing value added (% manufacturing value added)", source UNIDO Competitive Industrial Performance database. Covers 1995–2022 only. `https://api.worldbank.org/v2/en/indicator/NV.MNF.TECH.ZS.UN?downloadformat=csv` ## 4. Verification - **No OECD web page could be read.** www.oecd.org answered HTTP 403 to every fetch, including the R&D tax incentives topic page, the indicator pages and the OECD iLibrary. The old R&D tax country-profile URL answered 410, and the oe.cd/rdtax short link redirects to the blocked domain. No series could therefore be checked against a number quoted on an OECD web page. The substitute checks below are all against named downloads. - **BERD % GDP against Eurostat.** Eurostat `rd_e_gerdtot`, sector BES, unit PC_GDP, last update 29/09/26: `https://ec.europa.eu/eurostat/api/dissemination/sdmx/2.1/data/rd_e_gerdtot/A.BES.PC_GDP.?format=SDMX-CSV` saved as `raw/eurostat_rd_e_gerdtot_BES_PC_GDP.csv`. - Overall: of 925 country-years in common, 844 agree within rounding (0.005) and 862 within 0.02. - France: every year from 2004 to 2023 agrees to rounding. For example, 2008 is 1.29 in Eurostat and 1.2945 in MSTI, and 2022 is 1.47 against 1.4682. - United Kingdom 2014–2019: Eurostat is lower by 0.6–0.7 points of GDP. MSTI flags a UK BERD break in 2014. The likely cause is ONS's later upward BERD revision, which Eurostat's post-Brexit UK data do not carry. This was not checked against an ONS download. Prefer MSTI for the UK. - Japan: the two sources differ by about 0.05–0.11 throughout. - 2024: the gaps are larger because the 2024 values are provisional. - **Direct funding consistency inside OECD.** For France, MSTI `B_FG.PT_BERD` equals the DF_RDTAX `DF.PT_BERD` to rounding (2022: 9.2131 against 9.21). - **Researchers ratio**: reproduces OECD's published `T_RS.10P3EMP` (section 1). - **Implied subsidy rates and GTARD** were not checked against any outside figure. The checks above found no problem with these downloads, but they rest on the OECD SDMX download alone. ## 5. Series breaks ### France (source flags in `panel_flags.csv`) | year | `berd_gdp` / `berd_ppp` | `researchers_per_1000` (B_RS FTE) | `gerd_gdp` | `direct_funding_berd` | |---|---|---|---|---| | 1997 | B | B | B | B | | 2000 | – | – | B | – | | 2001 | B | B | – | B | | 2004 | B | – | B | B | | 2006 | B | B | – | B | | 2010 | – | – | B | – | | 2014 | – | – | B | – | | 2023–2024 | P | P (2024) | P | – | Eurostat flags the same France BES breaks as MSTI (b in 1992, 1997, 2001, 2004, 2006). - **2004 and 2006.** Both are BERD breaks. Eurostat's France national metadata for the business enterprise sector (`https://ec.europa.eu/eurostat/cache/metadata/EN/rd_simsbe2_fr.htm`, metadata updated 31 October 2025) explains 2004: "In 2007, the sampling method in the BE sector was modified and the 2004 data revised according to the new methodology." It explains 2001: "Coverage of the BE sector was expanded, and the data communicated by the Ministry of Defense now cover research that was not considered as R&D in earlier years." It gives the comparable series lengths as "R&D expenditure: 2001 onwards" and "R&D personnel (FTE): 2004 onwards". The 2006 BERD break is flagged in both MSTI and Eurostat, but no explanation was found in the sources consulted. - **2010 and 2014: not BERD breaks in these data.** MSTI flags them on GERD only, which points to the other performing sectors rather than BERD. Neither `berd_gdp` nor `berd_ppp` is flagged in those years, and the Eurostat BE metadata does not list them. If the 2010 and 2014 breaks you have in mind come from MESR/SIES documentation, they are not reflected in OECD or Eurostat BERD flags. - **2020 (Covid).** No break flag on any France R&D series. `berd_gdp` rises from 1.448 (2019) to 1.494 (2020) because GDP fell. BERD in constant PPP fell from 52,333m to 49,960m (−4.5%), then 51,998m in 2021. Treat 2020 as a denominator shock, or end the post-period at 2019 as a robustness check. The 2020 reform of the data-collection regulation (EU 2020/1197, applied from reference year 2021) is noted in the Eurostat metadata. - **Pre-treatment window.** The 2004 and 2006 BERD breaks fall inside the natural 1995–2007 pre-period. A 2004–2007 pre-period is short, and the 2004 and 2006 CIR reforms (which added a volume component) are visible in the implied subsidy rate: 0.08 in 2000–03, 0.15 in 2004–05, 0.21 in 2006–07, 0.43 from 2008. The treatment is therefore not a single step in 2008. ### Other countries (break flags on `berd_gdp`) MSTI `B` flags on `berd_gdp` or `berd_ppp`, 1995–2024: AUT 2016; BEL 2009; CAN 2014; CHL 2009, 2021; CRI 2022, 2023; CZE 1995, 2005; DNK 2007; ESP 2002, 2008; FRA 1997, 2001, 2004, 2006; GBR 2001, 2014, 2022; GRC 2008; HUN 2018; ISL 2010, 2011, 2013; ISR 2000, 2017; ITA 2016; JPN 1996; KOR 2007; LTU 2023; LUX 2012; MEX 2004; NLD 1996, 2011, 2012, 2013, 2022; NOR 1995; NZL 2001; POL 1995, 2016; PRT 1995; SVK 1997; SVN 2008, 2011; SWE 1995, 2005, 2023; USA 2015, 2016, 2021, 2023. Biennial or irregular BERD reporting in the pre-period is a separate issue, listed in NOTES.md. ## CNEPI 2021 opinion (for the study page) - CNEPI (France Stratégie), *Évaluation du crédit d'impôt recherche*, avis 2021, June 2021: `https://www.strategie-plan.gouv.fr/files/files/Publications/2021/0601%20CNEPI/fs-2021-rapport-cnepi-cir-juin.pdf`, from the page `https://www.strategie-plan.gouv.fr/publications/evaluation-credit-dimpot-recherche-rapport-cnepi-2021`. Retrieved 2026-10-14. The text of PDF pages 3, 11 and 12 (printed pages 3, 12 and 13) is in `raw/cnepi-2021-excerpt.txt`. - p. 3: the 2019 studies found the 2008 reform raised firms' R&D "d'un montant à peu près équivalent à la dépense fiscale additionnelle résultant de la réforme"; "pour les entreprises déjà bénéficiaires du CIR avant 2008, les effets identifiés sont positifs sur les PME, mais non significatifs sur les grandes entreprises". - p. 12: "pour les relativement plus grandes entreprises (y compris les ETI), l'étude n'identifie pas d'effet positif significatif des créances du CIR". - p. 13, quoting OECD (2020): "les petites entreprises (moins de 50 salariés) investissent plus de 1,4 euro dans la R-D, les moyennes entreprises (50-249 salariés) 1 euro et les grandes entreprises (250 salariés ou plus) seulement 0,4 euro". ## Tests written by analysis/cir.ml - `benchmark.csv`: for each main design and each multiple c of the rise in France's tax relief for business R&D since 2007 (`gtard_gdp`), France's series less c times that rise from 2008, its mean gap after 2008 and the share of units (France and each donor treated in turn) with a fit ratio at least as large (`p_ratio`) or a scaled mean gap at least as low (`p_lower`) or high (`p_upper`), after Firpo and Possebom (2018). With 12 donors no p-value can fall below 1/13. - `loo.csv`: the main designs refitted leaving out, in turn, each donor with a weight. - `fit-2001` in `summary.csv` and `results.csv`: the 2008 design fitted on 2001-2007 only, after the 1997 and 2001 breaks in France's series; the 2004 and 2006 breaks fall inside every possible pre-period. ## Patents (`patents.csv`) Added 2026-10-15 for the credit's innovation aim. Source: OECD, "Patents by technology", dataflow `OECD.STI.PIE:DSD_PATENTS@DF_PATENTS(1.0)`, downloaded on 2026-10-15 with `https://sdmx.oecd.org/public/rest/data/OECD.STI.PIE,DSD_PATENTS@DF_PATENTS,/all?startPeriod=1995&format=csvfilewithlabels` (1.5 GB, kept outside the repository). `dune exec data/cir/patents.exe -- DF_PATENTS.csv data/cir/patents.csv` streams it and keeps IP5 patent families (authority `9P50_3`, measure `PF`) by priority date, inventor country (agent role `INVENTOR`), every technology (`WIPO=_T`, `OECD_TECHNOLOGY_PATENT=_Z`), no counterpart and no co-operation type, for countries only (three-letter `REF_AREA` codes; regional aggregates are dropped): `country_iso3`, `year`, `ip5_families`, fractional counts as published, rounded to one decimal (the precision the file's `DECIMALS` column gives; `OBS_VALUE` carries four), 1995-2022 (recent priority years are incomplete, which is why the designs stop in 2019).