Convergence masking and the innovation paradox in OECD panels

Abstract

This dataset is the replication package for "Convergence masking and the innovation paradox in OECD panels" (S. Kim and Y.-W. Sawng, Economics Letters). Version 2 accompanies the revised manuscript; Version 1 accompanied the original submission. Research hypothesis. In two-way fixed-effects regressions of technological change on lagged R&D intensity, omitting the lagged technology level biases the R&D coefficient downward. Countries whose R&D intensity rises also reach higher technology levels, and a higher level is followed by smaller gains (convergence). The bias equals the convergence coefficient times the within-country co-movement of R&D and the technology level (−γδ). We call this mechanism "convergence masking". Data. panel_long.csv stores the downloaded series for 41 economies (38 OECD members, China, Chinese Taipei and Singapore) over 1990–2024 (1,435 rows): R&D intensity (OECD Main Science and Technology Indicators), PCT applications and triadic patent families (OECD, Patents by WIPO technology domains), population (OECD) and PCT national-phase entries (WIPO IP Statistics Data Center). pct_accession.csv gives each economy's PCT accession date. panel24.csv and panel24_wipo1b.csv are the two 24-country panels of Version 1, kept for comparison. All variables are constructed in the code; README.md documents the sources, dataflows and columns. Notable findings. In the main sample (37 OECD countries, 1997–2019, years at least three years after PCT accession; N = 812), adding the lagged technology level moves the R&D coefficient from −4.90 (wild cluster bootstrap p = 0.100) to 1.61 (p = 0.796). The convergence coefficient is −0.137 and the co-movement 47.61; the resulting shift of 6.51 is matched by only 3 of 2,000 permutations of R&D paths across countries (p = 0.002). Noise placebos do not reproduce the shift. When the control is built from a patent series other than the dependent variable, the shift falls from 5.85 to 3.94 (p = 0.012) for triadic families and disappears for PCT applications, so part of it is mechanical mean reversion. Korea accounts for 55.8% of the within-country cross-product. In the 24-country panel of Version 1, the coefficient moves from 1.16 to 6.98 (shift 5.82, p = 0.006). Interpretation and use. The estimates are within-country associations, not causal effects. They establish the existence and direction of the bias; its size is partly mechanical, and the controlled coefficient is imprecise. Running replication_main.R and then replication_supplement.R in R (4.0 or later; plm for the GMM estimates, fixest optional) reproduces every estimate in the paper and the Online Appendix; reference_output/ holds the expected output. The main script takes about two hours with the default settings (set FAST <- TRUE for a quick check). Random seeds are fixed, so bootstrap and permutation p-values reproduce exactly when the script is run in full.