Manufacturing revolutions: Industrial policy and industrialization in South Korea2025

There is a famous Ariel Rubinstein motto that goes: I have not seen any paper in Economics which deserves more than 15 pages. Unfortunately most papers are around 30 pages, and this one in particular is about 50 pages with tables and figures, this is just too much. Excessive robustness sections, repetition of arguments, and so on. This paper compares changes in outcomes between targeted (treated) and non-targeted (control) manufacturing industries. Targeted HCI (heavily-chemical industries) expanded output by 100%, labor productivity was 15% higher, both in comparison to non-targeted manufacturing sectors. There is evidence for dynamic learning by doing and longer term dynamic comaprative advantage of targeted export industries.

There is presence of bourgeois ideology on page 1688, where it is said that the Nixon Doctrine unfortunately coincided on growing military antagonism from the DPRK. It is argued that the 1973 ROK industrial drive was security driven and catalized by both of these crises. The president at the time was Park Chung Hee, a corporatist anti-communist conservative. According to the New Yorker, Park´s background included a period of flirtation with a group of South Korean Communists, whom he betrayed. We can make a counter-point to the paper´s sepculative ideation and posit an ideological motivation for Park´s industrial policy, which is that there was a genuine paranoia surrounding the Juche idea, and a dirigist desire to see ROK grow. Industrialization is thus, in this view, an ideological project, and we can describe both parties as military antagonic. Historically this episode is preceded by the Korean DMZ Conflict (Second Korean War) in 1966-1969.

The HCI drive targeted steel, nonferrous metals, shipbuilding, machinery, electronics, and petrochemicals. The state targeted upstream heavy industries instead of final weapons production. There was some early skepticism from inventors, and it is argued that this is why targeting was not endogeneous of sectors that would have grown regardless. This drive in reality consisted of a bundle of policies: investmet and trade. The investment part was essentialy subsidized credit, around half of domestic credit was policy loans. For trade it mainly consistes of exceptions aimed at HCI, such as import duties being waived. After Park was assasinated in 1979, the incentive structure was dismantled and liberalization was pursued. Lane makes use of a short panel (1970-1986) and a log panel (1967-1986), with a trade-off between industry observations and time periods.

There is a discussion of two externalities that justify industrial policy under neoclassical ideology. The first is dynamic economies of scale, where production generates learning-by-doing which raises future productivity. When learning spills over across firms within an industry, individual firms do not internalize the full social return to production, leading to underinvestment. Even without learning-by-doing spillovers, firms can still underinvest when faced with capital constraints that prevent them from financing early losses despite long-run returns. The second is linkage effects, basically upstream effects on inputs producers, and downstream forward linkages. The main specification is the following linear TWFE difference-in-differences (DD) regression: \[\begin{split} \ln(y_{it}) = \underbrace{\alpha_i + \tau_t}_{\text{Fixed and Time Effects}} & + \sum_{j \neq 1972} \beta_j \big(\text{Targeted}_i \times \text{Year}_t^j\big) \\ & + \sum_{j \neq 1972} \mathbf{X}_i^{\top} \times \text{Year}_t^j \Omega_j + \varepsilon_{it} \end{split}\] with industry \(i\) and year \(t\). Figure II is the DD estimates for the positive effect on output. The paralell trends seem credible despite the short pre-trend. Figure III shows point estimates (yearly) on labor productivity (up), prices (down), employment (up), output and labor share (up), and number of plants (up). It also provides estimates for export devlopment variables, all positive. The results are checked for robustness using Sant'Anna and Zhao (2020) and Callaway and Sant'Anna (2021) double-robust estimators. The ATT estimates are fairly similar to TWFE ones, amd coefficient differences seem to come from distinction between short and long panels. The long panel has lower coefficient estimates for the variables of interest. We then switch to a triple-differences (DDD) setup, where ROK is compared with other countries, while keeping the targeted and non-targeted manufacturing industries as treatment and control. Here we need strong paralell trends: the differential evolution between HCI and non-HCI must have also been the same in the absence of treatment across countries. \[\begin{split} Y_{ict} = \underbrace{\alpha_{it}}_{\text{industry-year}} & + \underbrace{\tau_{ct}}_{\text{country-year}} + \underbrace{\sigma_{ci}}_{\text{country-industry}} \\ & + \sum_{j \neq 1972} \beta_j \big(\text{Korea}_c \times \text{HCI}_i \times \text{Year}_t^j\big) + \varepsilon_{ict} \end{split} \] The results show higher point estimates during for the export development variables, but is even higher following the 1979 liberalization. Lane mentions that there have been no shortage of failures in industrial policty, while here we have arguably a success. I would take it a step further, it is not because an alternative has failed in \(x\) amount of historical settings that disqualifies it from consideration. The author also estimates outcome variables for the policy mechanisms: 81.4% increase in total investment, and 109% increase in input outlays. An attempto to assess the presence of learning by-doing forces the following equation is estimated: \[\begin{split} Y_{it} = \beta_1 \text{Experience}_{it} & + \beta_2 \text{Experience}_{it} \times \text{Targeted}_i \\ & + \theta \text{Size}_{it} + \alpha_i + \tau_t + \mathbf{X}_{it}^{\top} + \varepsilon_{it} \end{split}\] Experience is defined as cumulative production. The author is clear that this is not causal and is purely indicative and correlational. This is also true of all following estimations. \(beta_2\) is of interest, as it identifies the differential effect for the treatment group. The effects show a negative relationship between experience and unit cost reduction, and a positive relationship between experience and TFP. The sign persists for the treatment and experience interaction, suggesting HCI sectors have steeper experience-TFP relationship. It would seem more accumulated output is associated with better performance. Again, this is not causal. Finally, the author uses IO tables to construct backward and forward linkages (how much industry supplies or is supplied by HCI sectors). The following regression is estimated: \[\begin{split} \ln(y_{it}) = \alpha_i + \tau_t & + \sum_{j \neq 1972} \gamma_j \big(\text{Backward Linkage}_i \times \text{Year}_t^j\big) \\ & + \sum_{j} \delta_j \big(\text{Forward Linkage}_i \times \text{Year}_t^j\big) + \varepsilon_{it} \end{split}\] Here the treatment is continuous. The interactions \(\gamma_j\) and \(\delta_j\) capture the differential evolution of industries with strong versus weak exposure to HCI sectors. The results show that industries with strong forward linkages to HCI sectors had a more robust development. Essentially industries that were using HCI inputs grew more, which I can imagine was one of the mechanisms through which the military industrial complex was to be benefitted, as intended by Park. The effect of backward linkages was more limited. There is then a SUTVA section that reads like a response to a referee report.

That is the end of the paper's scope. There is no mention of social cost (Onsan illness) and so on, something that Choi and Levchenko (2025) attempt. Maybe I can read it in the future. For me the main weakness, and it´s not the author´s fault (he did as much as he could), is that there is no clear way of distinguishing what worked, insofar as there is no way of separating the effect of the elements of the policy bundle to the bundle itself, as it is the combination of all measures that is defined as the treatment. It is hard then to prescribe policy based on this, it could have been that one particular element was responsible for all the effects and we simply do not know. However, the paper is very strong, and it shows strong evidence for re-examining knowledge that is considered to be canonical or of common sense. I like that it demystifies the idea of economic miracles, and shows that interventionism can steer entire industries to incredible improvements.