Does Manufacturing Matter for Economic Growth in the Era of Globalization? Results from Growth Curve Models of Manufacturing Share of Employment (MSE) To formally test trends in manufacturing share of employment (MSE) for different groups of countries, I estimated unconditional growth curve models of manufacturing share of employment for different samples of countries that include random intercepts and slopes. As expected, developed countries exhibit a downward time trend in manufacturing share of employment throughout the period under study (b= -1.395; p<0.001). Similarly, post-socialist transition countries also exhibit a downward time trend (b= -1.915; p<0.001). The same model for LDCs (excluding post-socialist transition countries) does not provide evidence of a trend in manufacturing share of employment because the coefficient of the time trend is negatively signed but not significant (b= -0.062; p=0.643). Therefore, there is no evidence of an increase in average manufacturing share of employment at the country-level among LDCs between 1970 and 2010. Testing for Endogeneity using GMM Instrumental Variable Regression It is possible that the positive association between manufacturing (MSE) and economic growth reported in Tables 1-2 suffers from endogeneity bias. Indeed, faster rates of economic growth could spur investment in profitable sectors like manufacturing and lead to more manufacturing employment. On the demand side, rising incomes could lead to increased demand for manufactured goods and spur manufacturing employment. Studies that seek to explain trends in manufacturing employment almost always include an independent variable that captures the level of GDP per capita (Alderson 1999; Kaya 2010). While the processes that link level of development to manufacturing employment might be different, it is nevertheless important to test for reverse causality. In order to test this, I employ a two-stage generalized method of moments instrumental variable regression (GMMIV) model for the whole sample that includes manufacturing employment and the neoclassical controls along with t-1 time dummies. The use of GMMIV models allow for the estimation of cluster-robust standard errors. In the first stage of GMMIV models, the suspected endogenous variable (manufacturing employment) is regressed on other exogenous variables and two select instruments. In the second stage, the dependent variable (economic growth) is regressed on the exogenous variables and the predicted values from the first stage regression. The instruments used in GMMIV models must be sufficiently correlated with the suspected endogenous variable (i.e. strong instruments) uncorrelated with the error term of the second-stage regression (i.e. valid instruments). Previous research suggests that lagged values and higher moments could be suitable instruments provided they are strong and valid (Bollen 2012 ; Curwin and Mahutga 2014). I use the logged second difference and the lagged second difference of manufacturing share of employment as instruments for the manufacturing share of employment covariate. Tests for the strength and validity of instruments reveal that we can reject the null hypothesis of weak instruments (F = 15.30, p<0.0001) and we are unable to reject the null hypothesis of valid instruments (Hansen s J-statistic = 0.005, p=0.94). In the model, manufacturing has a strong positive effect on economic growth which is entirely consistent with previous findings (b = 0.016; p<0.05).
Most importantly, tests for endogeneity reveal that we cannot reject the null hypothesis that manufacturing is exogenous (Difference-in-Sargan C-statistic = 0.193; p=0.66). In sum, the significant effect of manufacturing on economic growth in the GMMIV model coupled with the endogeneity test indicating that manufacturing (MSE) is in fact exogenous suggests that the main difference models are appropriate. Arellano-Bond Estimator with LDV and Two-Way Fixed Effects Estimator I estimated models (see online supplement) that (1) implement the Arellano-Bond estimator that allows for the inclusion of a lagged-dependent variable in the face of endogeneity bias (Nickell 1981); (2) implement the within version of the fixed estimator by employing indicator variables for each country and time point. These results from these models confirm the substantive findings from the main difference models. See next page for the results in Table S1. The inclusion of a lagged dependent variable (LDV) in a first-difference regression causes endogeneity issues because the error term is correlated with the LDV by construction (Nickell 1981). The Arellano-Bond estimator deals with this issue by instrumenting for the LDV using all lagged levels of ln (where z > 1). In addition, the consistency of the Arellano-Bond estimator depends on the assumption of no second-order serial correlation in the first-differenced error term. The M2 statistic is a test of this assumption and as can be seen, the null hypothesis of no second order serial correlation in the error terms cannot be rejected for all models. The two-way fixed effects models includes country and time dummies and compares the preand post-1990 manufacturing effect using Seemingly Unrelated Regression (SUR). These models also include a lagged level of per capita GDP as a covariate (see equation on next page) which makes it a growth model and helps reduce serial correlation of error terms. As the results show, we can reject the null hypothesis of equal pre- and post-1990 manufacturing coefficients for the LDC subsample but not for the DC subsample. Preliminary Analysis of Disaggregated Manufacturing Industries I conduct a preliminary investigation (available upon request) of trends in and returns to disaggregated manufacturing industries using INDSTAT s 2-digit ISIC D industry dataset. Out of the 23 2-digit industries, I analyze the 18 industries for which data is available for the whole time period considered. Consistent with the trend for the aggregate manufacturing sector, the locus of manufacturing in each of the 18 2-digit industries examined shifts to LDCs. All of the 18 manufacturing industries see increases in the proportion of their worldwide labor being employed in LDCs from the 1970s to the 2000s, due to geographical relocation of these activities and productivity increases. Next, I run models for LDCs where I assess the impact of the share of each of the industries employment in total employment of an economy on national economic growth. Again, consistent with the findings in the main analysis of the aggregate manufacturing sector, I find that the growth effect of employment in almost all (16 out of 18) of the 2-digit manufacturing industries exhibit significant declines for LDCs. The other two industries, Textiles [ISIC 17] and Petroleum Products [ISIC 23], do not exhibit significant declines because they fail to exert significant effects on economic growth throughout the time period under study.
Table S1. Alternative Estimators: Two-Way Fixed Effects & Arellano-Bond LDV Two-Way Fixed Effects a,b Arellano-Bond Estimator LDC DC LDCs Developed Countries (S1) (S2) (S3) (S4) (S5) (S5) (S6) (S7) Lagged DV 0.425 *** 0.366 *** 0.404 *** 0.318 *** 0.421 *** 0.939 *** (0.092) (0.105) (0.092) (0.090) (0.084) (0.082) Manufacturing & Period Effects Manufacturing Share of Employment 0.026 *** 0.058 *** 0.049 *** 0.011 *** -0.004 0.009 * (0.004) (0.012) (0.009) (0.003) (0.005) (0.004) Time (coded 0-7) 0.152 *** 0.150 *** 0.079 *** 0.078 *** (0.018) (0.018) (0.015) (0.016) Manufacturing x Time (coded 0-7) -0.007 ** 0.005 *** (0.002) (0.001) Globalization Period 1990-2010 0.036 ** -0.020 (0.012) (0.014) Manufacturing x Globalization (>1990) -0.032 ** 0.010 (0.010) (0.006) Manufacturing Effect pre-1990 0.024 *** 0.005 (0.006) (0.003) Manufacturing Effect post-1990 0.006 0.010 ** (0.004) (0.004) N 469 158 333 333 333 133 133 133 M2 - - -0.93-0.66-0.29 0.97 0.84-0.02 χ 2 test (pre-1990 = post-1990) 6.70 ** 1.12 - - - - - - Notes: Robust standard errors in parentheses. M2 is a test of second-order serial correlation of error term in A-B estimator. Neoclassical effects and constants are not shown for the sake of brevity. a Two-way Fixed Effects models include t-1 time and country dummies as well as lagged GDP per capita. b The pre- and post-1990 coefficients from the FE models were compared using Seemingly Unrelated Estimation. + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001 (two-tailed tests) Two-Way Fixed Effects Estimator within each time period (pre- and post-1990): ln ln where y is per capita GDP for country i and time t, is vector of country fixed-effects, and is vector of time fixed-effects, x and m are predictors. The chi-squared test from the seemingly-unrelated estimation presented above compares with. Arellano-Bond Estimator: ln ln where y is per capita GDP for country i and time t, and is the non-differenced time variable, and x and m are predictors.
Testing Manufacturing Value Added as % of GDP and Controlling for Manufacturing Productivity I also tested manufacturing share of GDP as a measure of manufacturing activity using data on manufacturing value added data as a percentage of GDP from the United Nations National Accounts Main Aggregates Database (United Nations Statistics Division 2015). Using values for ISIC category D, I ran the same main analyses and found that this measure of manufacturing consistently fails to exert significant effects on economic growth for all samples and that these effects do not change over time. Furthermore, the corresponding t-statistics for these coefficients never exceeded 1.0 in any of the models. The non-significant effects for manufacturing share of GDP, coupled with the strong positive effects of manufacturing employment on growth suggest that manufacturing employment is an operationalization that is more relevant to the theoretical discussion in the paper. In addition, I ran the main models (i.e. with manufacturing share of employment) while controlling for manufacturing sector productivity (manufacturing value added/worker) and obtained substantively identical conclusions as the main results (i.e. declining growth effect of manufacturing share of employment for LDCs but not for developing country subsample). Analyses that Control for Agricultural Employment The results from this analysis are shown in Table S2 on the next page. If we conceptualize national economies as being divided into three sectors (primary, secondary, and tertiary), the inclusion of covariates that capture employment in two of these three sectors in models for my supplemental analysis will help alleviate concerns about omitted variable bias. Secondary employment is already covered with the manufacturing employment covariate. For the other employment covariate, I look to the World Bank s cross-national data on agricultural employment as a percentage of total employment (includes employment in agriculture, hunting, forestry, and fishing). This represents employment in the primary sector. I use agricultural employment (primary) instead of service sector employment (tertiary) because the aforementioned heterogeneity in the service sector means that for example, low-wage employment in the fast-food industry is grouped in with employment in high-end IT industry. Therefore, I employ World Bank s data on agricultural employment (primary) to address concerns about omitted variable bias. Unfortunately, the World Bank s data on employment (for all sectors) only starts in 1980. This means that the first two 5-year time-periods in my main analyses (1970-1975 & 1975-1980) are not available for analysis when I include this agricultural employment covariate in my models. In addition, the cross-national data coverage for this covariate in the 1980s and 1990s is quite sparse. The missing data for the agricultural employment means that the inclusion of this covariate in my models reduces the sample size by about 56% (See Table S2). Results from this supplemental analysis for the LDC subsample are shown in Table S2. In Table S2, the models that are not shaded are the ones with the original sample from the main analysis and they are shown for comparison. As the results indicate, even with a substantially different sample (both in terms of countries and years included/excluded), I obtain very similar substantive results as the main analyses. For the LDC subsample, I observe a statistically significant decline in manufacturing employment s estimated effect on economic growth after the early 1990s (p<0.01; see models S12 and S15). I do not observe a similar decline for the
developed country subsample (results for developed countries available upon request). Furthermore, when I compare models with and without the agricultural employment covariate with the sample constrained to be common between both models (shaded models in Table S2), the coefficient estimates and standard errors for the interaction and other terms do not change appreciably between models and the substantive results remain the same. This suggests that omitted variable bias, even if present, might not be a major factor in the analyses. Consistent with expectations, agricultural employment is negatively associated with growth.
Table S2. Estimates from FD Regression in LDC Subsample Controlling for Agricultural Employment a (S9) (S10) (S11) (S11) (S12) (S13) (S14) (S15) Manufacturing Share of Employment 0.021 *** 0.020 *** 0.060 *** 0.065 *** 0.065 *** 0.051 *** 0.053 *** 0.054 *** (0.005) (0.005) (0.012) (0.017) (0.016) (0.008) (0.013) (0.013) Time (coded 0-7) 0.012 0.013 0.006 + 0.009 0.009 (0.009) (0.009) (0.003) (0.008) (0.008) Manufacturing x Time (coded 0-7) -0.008 *** -0.009 ** -0.009 ** (0.002) (0.003) (0.003) Globalization Period 1990-2010 0.039 ** 0.054 * 0.054 + (0.013) (0.027) (0.027) Manufacturing x Globalization (>1990) -0.034 *** -0.036 * -0.039 ** (0.009) (0.014) (0.014) Agriculture Employment -0.003 + -0.003 * -0.003 + (0.001) (0.001) (0.001) Secondary School Enrollment 0.001 0.001 0.002 * 0.001 0.001 0.002 * 0.001 0.001 (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Trade Openness -0.000-0.000-0.001 * -0.000-0.000-0.001 * -0.001-0.001 (0.001) (0.001) (0.000) (0.001) (0.001) (0.000) (0.001) (0.001) Gross Capital Formation 0.011 *** 0.010 *** 0.008 *** 0.011 *** 0.011 *** 0.008 *** 0.010 *** 0.010 *** (0.002) (0.002) (0.001) (0.002) (0.002) (0.001) (0.002) (0.002) Labor Force Participation 0.002 0.004 0.002 0.001 0.002 0.002 0.002 0.003 (0.006) (0.006) (0.004) (0.006) (0.006) (0.004) (0.006) (0.006) Population (ln) -0.709 ** -0.655 ** -0.854 *** -0.615 ** -0.554 * -0.830 *** -0.644 ** -0.587 ** (0.214) (0.206) (0.149) (0.218) (0.209) (0.139) (0.201) (0.195) Constant 0.126 * 0.115 * 0.153 *** 0.134 * 0.122 * 0.156 *** 0.143 *** 0.133 *** (0.058) (0.057) (0.026) (0.055) (0.054) (0.020) (0.032) (0.031) N adj. R 2 0.387 0.399 382 0.344 0.407 0.421 382 0.350 0.415 0.428 BIC -223.3-222.4-442.1-224.9-224.7-445.8-227.1-226.9 Notes: Heteroskedasticity and serial correlation consistent standard errors in parentheses a Models S7, S8, S10, S11, S13, and S14 are constrained to have the same sample, while Models S9 and S12 are replications of Models 8 and 9 in the main paper. + p < 0.10, * p < 0.05, ** p < 0.01, *** p < 0.001
References for Alderson, Arthur S. 1999. Explaining Deindustrialization: Globalization, Failure, or Success? American Sociological Review 64(5):701. Bollen, Kenneth A. 2012. Instrumental Variables in Sociology and the Social Sciences. Annual Review of Sociology 38(1):37-72. Curwin, Kevin, and Matthew Mahutga. 2014. Foreign Direct Investment and Economic Growth: New Evidence from Post-Socialist Transition Countries. Social Forces 92(3):1159-1187. Kaya, Yunus. 2010. Globalization and Industrialization in 64 Developing Countries, 1980-2003. Social Forces 88(3):1153-1182. Nickell, Stephen. 1981. Biases in Dynamic Models with Fixed Effects. Econometrica 49(6):1417 26.