Convergence masking and the innovation paradox in OECD panels

This repository contains the reproduction package for the paper "Convergence masking and the innovation paradox in OECD panels" (S. Kim and Y.-W. Sawng). Research hypothesis. In two-way fixed-effects panel regressions of technological change on R&D intensity, omitting the lagged technology level biases the R&D coefficient toward zero. Proximity to the technology frontier raises both R&D intensity and the technology level, while a higher technology level is followed by smaller subsequent technological gains; when the lagged level is omitted, this negative convergence association loads onto the R&D coefficient. We term this mechanism "convergence masking" and hypothesize that adding a single convergence control recovers the within-country R&D-technology association. Data. Two country-year panels in long format covering the same 24 OECD countries: (1) main_panel_P1.csv, 2002-2019 (432 rows), where the technology variable is a principal-components composite of PCT international applications and triadic patent families; (2) wipo1b_panel_P1.csv, 2002-2022 (504 rows), where it is the WIPO 1b national-phase-entry patent count, used for robustness. All series derive from public sources: OECD Main Science and Technology Indicators (R&D intensity, gross R&D expenditure over GDP) and WIPO patent statistics. Variables are documented in README.md; lags were constructed within country. Notable findings. With country and year fixed effects and country-clustered standard errors (N = 408), the lagged R&D coefficient rises from 1.161 (p = 0.068) to 6.982 (p = 0.026) once the lagged technology level is added; the within R-squared rises from 0.001 to 0.104. The direction of this shift is robust to the alternative patent measure, to R&D lags of one to three years, and to leave-one-out country exclusion. Its magnitude is sample- and measurement-dependent (excluding Korea lowers the coefficient by about 73 percent) and is not part of the claim. Interpretation and use. Results should be read as a convergence-corrected within-country association, not a causal parameter, and as evidence on the direction of the bias rather than its size. Running P1_reproduce.R in R (4.0 or later) with the fixest and dplyr packages reproduces Table 1 and all robustness checks in the paper; expected outputs are annotated in the script and listed in README.md. Reported p-values come from fixest; other software may apply slightly different small-sample corrections to clustered standard errors, shifting p-values marginally while leaving coefficients unchanged.本仓库包含论文《OECD面板中的收敛遮蔽与创新悖论》（S. Kim与Y.-W. Sawng）的复现包。 研究假设。在以技术变革为被解释变量、研发强度为解释变量的双向固定效应面板回归中，遗漏滞后技术水平会使研发系数向零偏移。技术前沿距离同时提升研发强度与技术水平，而更高的技术水平会伴随后续技术增益的缩减；当遗漏滞后技术水平项时，这种负向收敛关联会被加载到研发系数上。我们将这一机制命名为「收敛遮蔽（convergence masking）」，并提出假设：仅加入一项收敛控制变量即可恢复国家内部的研发-技术关联。 数据。包含两份长格式的国家-年度面板数据，覆盖相同的24个OECD成员国：(1) main_panel_P1.csv，时间跨度为2002-2019年，共432条观测值，其中技术变量为PCT国际申请与三方专利家族的主成分合成指标；(2) wipo1b_panel_P1.csv，时间跨度为2002-2022年，共504条观测值，其中技术变量为WIPO 1b类国家阶段进入专利数量，用于稳健性检验。所有序列均取自公开数据源：OECD主要科技指标（研发强度，即研发总支出占GDP的比重）与WIPO专利统计数据。变量说明详见README.md；滞后项均按国家分组构建。 主要研究发现。在控制国家与年份固定效应、并采用国家聚类标准误的设定下（N=408），加入滞后技术水平变量后，滞后研发系数从1.161（p=0.068）提升至6.982（p=0.026）；组内R²从0.001上升至0.104。这一系数变动的方向在替换专利衡量指标、采用1至3阶研发滞后项、以及逐一剔除单个国家的稳健性检验中均保持一致。其幅度则依赖于样本与衡量方式（排除韩国后系数约下降73%），且不属于本文的核心主张范畴。 解读与使用说明。本研究结果应被解读为经收敛校正后的国家内部关联，而非因果参数，同时仅能作为偏差方向而非偏差大小的证据。通过在R（4.0及以上版本）中运行P1_reproduce.R脚本（需安装fixest与dplyr包），即可复现论文中的表1及所有稳健性检验结果；脚本中已对预期输出进行注释，且输出文件列表详见README.md。报告的p值由fixest包计算得出；其他软件可能对聚类标准误采用略有不同的小样本校正方法，仅会小幅改变p值，而不影响系数结果。