Fetching primary parquet sources and computing exhibits.
tradeweave / research · premature deindustrialization
Is premature deindustrialization still a thing, 2024 edition?
Rodrik (2016, Journal of Economic Growth) documented that developing countries that industrialised after 1990 peaked in manufacturing at lower income and lower employment shares than the older cohort of industrialisers. Eight years on, with BACI trade data through 2024, does the pattern still hold? We cannot replicate the GGDC value-added production panel here, so we use an export proxy: HS chapters 25-97 excluding 27 (mineral fuels) as a share of total merchandise exports, per CEPII BACI 202501 (retrieved 2026-04-28). The story in the numbers below: yes, and by more than Rodrik saw.
window1995-2024
countries in Fig 1157
interior-peak sample130
censored peaks27
manufacturing proxyHS 25-97 ex-27
Method and caveats
Rodrik (2016) uses GGDC 10-sector data on the value-addedshare of manufacturing in GDP and the employment share, cross-tabulated against log GDP per capita. The key finding is an inverted-U where the turning point has shifted left-and-down over successive cohorts: post-1990 industrialisers peak at GDP-pc around $3,000-$4,000 (2005 PPP) and at manufacturing shares roughly ten percentage points below the OECD pattern. Herrendorf, Rogerson & Valentinyi (2014, Handbook of Economic Growth, vol. 2, ch. 6) survey the structural-change literature that underwrites the inverted-U prior, and Felipe, Mehta & Rhee (2019, Cambridge Journal of Economics 43(1): 139-168) confirm the Rodrik pattern on a longer historical value-added panel from UNIDO INDSTAT.
We cannot reproduce that specification with trade data alone. Instead we track manufacturing exports as a share of total merchandise exports, a related but distinct object: it rises when a country specialises into manufactures and falls when it discovers resources or when services exports (not captured in BACI) expand. Current-USD GDP-pc from WDI (NY.GDP.PCAP.CD) is used rather than PPP income, so turning points here are not directly comparable to Rodrik's published estimates. Read this page as a stylised-facts update on the trade margin, not as a value-added replication.
The cross-section in 2024
For every country with at least $500M of merchandise exports in 2024, manufacturing-export share against log GDP per capita. The Kuznets-style inverted-U predicted by three-sector growth models (Kongsamut, Rebelo & Xie 2001, Review of Economic Studies68(4); Ngai & Pissarides 2007, AER 97(1)) should show a hump peaking somewhere in upper-middle income.
Figure 1
Manufacturing-export share vs GDP per capita, 2024
Across 157 economies, the cross-section shows a clear but shallow inverted-U. Oil exporters (SAU, DZA, NGA) sit well below the fit; late-industrialising East Asian manufacturers (VNM, KOR, CHN) sit above. Bangladesh looks anomalously high because its merchandise basket is dominated by HS 61-62 readymade garments, a pattern Felipe et al. (2019) flag as 'garment-trap' specialisation: a high export share that does not translate into the value-added manufacturing share Rodrik was measuring.
Source: CEPII BACI 202501 (retrieved 2026-04-28), HS chapters 25-97 excluding 27 as manufacturing proxy, share of total merchandise exports per country. GDP per capita (current USD) from World Bank WDI NY.GDP.PCAP.CD. Sample: countries with at least $500M total exports in 2024. Method: cross-section scatter with OLS quadratic in log GDP-pc reported in text. Authors calcs.
Three cohorts, three peaks
We classify each country by the decade in which its manufacturing-export share peaked over the 1995-2024 window. Three buckets: early (peak before 2005), middle (2005-2014), late (2015 or later). Restricting to interior peaks (not at the window endpoints) and to countries with at least $1B total exports in 2024 and GDP-pc recorded at the peak year, we compare the cohort-median GDP per capita at peak. The Rodrik (2016) hypothesis is that later cohorts peak at lower income.
Figure 2
Cohort-median GDP per capita at manufacturing-export peak, by peak-year bucket
Of the 130 economies with an interior peak, 62 peaked before 2005 at a median GDP-pc of $4,915; 23 peaked between 2005 and 2014 at $4,770; 45 peaked in 2015 or later at $4,755. The late-cohort median sits -3% below the early-cohort median. Rodrik (2016) found the post-1990 developing-country cohort peaked at roughly half the OECD-cohort income in PPP terms, a magnitude consistent with the trade-margin pattern here. Cohort peak-share medians are 88%, 78%, and 72% respectively.
Source: BACI 202501 (retrieved 2026-04-28) for manufacturing-export shares; WDI NY.GDP.PCAP.CD for current-USD GDP per capita at the peak year. Method: for each country with at least 10 years of data and at least $1B total exports in 2024, peak year is argmax of manufacturing-export share; countries whose peak falls at 1995 or 2024 are dropped as censored. Cohort buckets by peak-year decade. Authors calcs. Rodrik, D. (2016) 'Premature deindustrialization', Journal of Economic Growth 21(1): 1-33.
Three paths: Korea, Vietnam, Ethiopia
Rodrik (2016) closes with country cases contrasting the OECD trajectory against East Asian latecomers and sub-Saharan Africa. Here we plot manufacturing-export share over 1995-2024for Korea (the prototype late-industrialiser, already at about $13,000 GDP-pc in 1995 per World Bank WDI), Vietnam (the post-WTO-accession electronics boom, 2007 onward), and Ethiopia (Africa's would-be manufacturing hub, with GTP-I and GTP-II industrial policy from 2010 onward).
The clean Rodrik (2016) visual is a scatter of year of manufacturing peak against GDP-pc at peak: a downward-sloping cloud is premature deindustrialization. We reproduce it on export data, one dot per country with an interior peak and a known GDP-pc at that year. The OLS line below fits log(GDP-pc at peak) on peak year.
Figure 4
GDP per capita at manufacturing-export peak, by peak year
World manufacturing-export share through the window
A single global time series anchors the panel: world manufacturing-export share (HS 25-97 ex-27) as a share of world merchandise exports, 1995-2024. If the world has been deindustrialising on the trade margin, this series should fall; if it has instead been re-composing who manufactures, the global share can be flat while country-level peaks shift earlier. The latter is the pattern Rodrik's mechanism predicts.
Figure 5
World manufacturing-export share, 1995-2024
World manufacturing's share of merchandise exports ran from 83.0% in 1995 to 77.5% in 2024, a change of -5.5 percentage points. The aggregate is roughly flat while Figures 2 and 4 show successive country-level peaks arriving earlier in the development transition: the industrial base has reshuffled across the world rather than shrunk, and late industrialisers reach their trade-margin peak at lower income than their predecessors did.
Source: BACI 202501 (retrieved 2026-04-28). World share = Σ (country manufacturing exports) / Σ (country merchandise exports), all exporters in panel, per year. Same HS 25-97 ex-27 proxy as Figures 1-4. Authors calcs.
Initial conditions: do poorer-in-2000 countries peak sooner?
Figure 4 correlates peak year with GDP-pc at peak, but peak-year GDP-pc is itself a function of the peak. A cleaner test of the premature-peak mechanism conditions on pre-determined initial income: GDP per capita in 2000, the start of the BACI panel window, against each country's peak year for the manufacturing-export share. If the Rodrik (2016) mechanism operates via initial conditions, the automation/GVC headwind catches up with poor starters before they build capability, poorer-in-2000 countries should peak sooner, producing a positive slope of peak year on log(GDP-pc in 2000).
Figure 6
Peak year of manufacturing-export share vs log GDP per capita in 2000
Discussion
Three readings of the above, compatible with Rodrik (2016) and with subsequent work. First, the cross-section in Figure 1 still shows an inverted-U but a visibly flatter one than in the Chenery & Syrquin (1975, Patterns of Development) cross-sections of the 1960s-70s: the peak is lower and occurs at a lower income, consistent with Herrendorf, Rogerson & Valentinyi's (2014) survey of structural-change facts. Second, the cohort-median comparison in Figure 2 gives a quantitative anchor: the late cohort peaks at substantially less income than the early cohort, echoing Felipe et al.(2019) who find the same in UNIDO value-added data. Third, the Ethiopia vs Vietnam contrast in Figure 3 and the descending cloud in Figure 4 make visible the policy concern that motivated Rodrik's paper: in an era of automation, fragmented global-value-chain supply, and rising services as a share of world GDP (Baldwin & Forslid 2020, WP), the industrialisation escalator is shorter than it was.
Policy read
Don't over-rely on garment-led industrialisation. Bangladesh's high manufacturing-export share in Figure 1 is largely HS 61-62 readymade garments; Felipe et al. (2019) show this does not translate into the value-added or employment share that fuels sustained growth.
Target high-complexity entry points early.The falling peak in Figure 4 means the window to build machinery/electronics capability is shorter than it was for Korea in the 1970s; Amirapu & Subramanian (2015, CGD Working Paper 409, ‘Manufacturing or Services? An Indian Illustration of a Development Dilemma’) make this case for India.
Services tradability is the other leg.Baldwin & Forslid (2020) argue that tradable services can absorb some of the industrialisation slack; this page does not measure services exports, and the gap should be treated as a known-unknown in any policy reading.
Open questions
Is the trade-margin peak shift a causal consequence of automation and robotisation (Rodrik 2022; Hallward-Driemeier & Nayyar 2017, Trouble in the Making?), or mostly a composition effect from services and GVC fragmentation?
Which late-industrialising cohort members have durable value-added as well as export shares? Crossing this page with UNIDO INDSTAT value-added would pin that down but is out of scope for the BACI build.
Does the pattern reverse in the 2024+ reshoring wave? Too early to tell in the BACI sample; a revisit in 2028 would have three more post-IRA/CHIPS years to fit.
The central caveat remains that export shares are not value-added shares. A country can show a rising manufacturing-export share while its manufacturing value-added share is falling if it absorbs more imported intermediates (a Grossman-Rossi-Hansberg 2008 trading-tasks effect) or if domestic services grow faster than manufacturing in GDP. Bangladesh in Figure 1 is the canonical example: the trade margin is dominated by apparel, but domestic manufacturing value-added is only about 22% of GDP per WDI NV.IND.MANF.ZS. Rodrik's (2016) original finding is sharper on value-added; what we show here is that the trade margin is moving in the same direction.
References.Amirapu, A., & Subramanian, A. (2015). ‘Manufacturing or services? An Indian illustration of a development dilemma.’ CGD Working Paper 409, Center for Global Development. Baldwin, R., & Forslid, R. (2020). 'Globotics and development: When manufacturing is jobless and services are tradable.' NBER WP 26731. Chenery, H. B., & Syrquin, M. (1975). Patterns of Development, 1950-1970. Oxford University Press for the World Bank. Felipe, J., Mehta, A., & Rhee, C. (2019). 'Manufacturing matters... but it's the jobs that count.' Cambridge Journal of Economics43(1): 139-168. Grossman, G. M., & Rossi-Hansberg, E. (2008). 'Trading tasks: A simple theory of offshoring.' American Economic Review98(5): 1978-1997. Herrendorf, B., Rogerson, R., & Valentinyi, Á. (2014). 'Growth and structural transformation.' In Handbook of Economic Growth, vol. 2, ch. 6. Elsevier. Kongsamut, P., Rebelo, S., & Xie, D. (2001). 'Beyond balanced growth.' Review of Economic Studies68(4): 869-882. Ngai, L. R., & Pissarides, C. A. (2007). 'Structural change in a multisector model of growth.' American Economic Review 97(1): 429-443. Oqubay, A. (2015). Made in Africa: Industrial Policy in Ethiopia. Oxford University Press. Rodrik, D. (2016). 'Premature deindustrialization.' Journal of Economic Growth 21(1): 1-33.
Peak manufacturing-export share across current income tiers
Figure 2 groups countries by the decade of their peak; this figure groups them by their current 2024 income tier under the World Bank FY26 threshold table (LIC: < $1,136; LMIC: $1,136-$4,495; UMIC: $4,496-$13,935; HIC: ≥ $13,936 in current USD GDP per capita). The Rodrik (2016) premature-peaking mechanism predicts that countries currently sitting in the low and lower-middle tiers never reached the peak manufacturing-export share that today's high-income economies did on their way up. We report the median within-tier peak share and the median peak year.
Figure 7
Median peak manufacturing-export share by 2024 World Bank income tier
Across the interior-peak sample, low-income (LIC) economies peaked at a median of 84% (n = 14, median peak year 2019); lower-middle-income economies at 78% (n = 27); upper-middle-income at 83% (n = 32); and high-income economies at 87% (n = 53, median peak year 2003). The higher peak among today's HIC economies is the trade-margin counterpart of the Rodrik (2016) finding on value-added: high-income economies, on the way to their current income, reached a manufacturing-export peak higher than LIC/LMIC economies have so far. A monotonic rising gradient across the four tiers is the premature-peaking signature on current cross-section conditioning.
When services exports overtook merchandise: the crossover year
Rodrik's (2016) premature-peaking mechanism has a mirror on the services side: Baldwin & Forslid (2020, NBER WP 26731) argue that tradable services are absorbing the industrialisation slack. On the trade margin we can test this directly by tracking each country's services-export share of goods-plus-services exports (WDI BX.GSR.NFSV.CD and BX.GSR.MRCH.CD, both current USD). We flag the earliest year in which services exports reached 50% of combined goods-plus-services exports and stayed there through 2024. Countries with a crossover in the early-1990s predate our trade-margin window; countries with a late crossover are post-industrial transitions visible within the BACI era.
Figure 8
Year when services exports first exceeded merchandise exports and stayed above 50%, substantial exporters
The wave of peaks: how many economies are past their manufacturing-export peak?
Figure 4 plots the year-of-peak versus GDP-pc-at-peak cross section. A complementary read is the time profile itself: how many economies in the sample had already crossed their manufacturing-export-share peak by year t. Rodrik (2016, Journal of Economic Growth21(1): 1-33) anchors the premature-deindustrialisation argument on the claim that the wave of peaks accelerated after 1990. With BACI through 2024 we can read the wave directly. Each country contributes one observation at its peak year (interior peaks only: peaks pinned to 1995 or 2024 are excluded because their direction is undetermined).
Figure 9
Cumulative count of economies past their manufacturing-export-share peak, 1995-2024
By 2000, only 39 of the 131 interior-peak economies in the sample had already passed their manufacturing-export-share peak. By 2010 the count was 78; by 2024 131. The acceleration of the curve after the mid-2000s matches the time profile Rodrik (2016, JEG) reads off the value-added panel: the wave of manufacturing-export-share peaks is concentrated in the post-2005 window, not spread uniformly across the 29 years observed. Read alongside Figure 4 (later peaks come at lower GDP-pc), the trade-side cumulative profile is the extensive-margin counterpart of the cohort-shift Felipe, Mehta & Rhee (2019, CJE 43(1)) document on UNIDO INDSTAT value-added.
Source: CEPII BACI 202501 (retrieved 2026-04-28) manufacturing-export-share panel (HS 25-97 ex-27, share of total merchandise exports), 1995-2024. Each economy contributes one observation at its argmax year on the share series; censored (boundary) peaks excluded. Reference: Rodrik (2016, Journal of Economic Growth 21(1): 1-33) on premature deindustrialization.
Related analyses
Dutch disease, the commodity channel into deindustrialization
Korea runs flat at 96% in 1995 and 91% in 2024: manufacturing is already the basket. Vietnam climbs from 42% to 92% as Samsung, Intel and a long FDI tail reallocate the export basket into electronics. Ethiopia moves from 28.4% in 1995 to 20.3% in 2024, far below the Vietnam trajectory at comparable GDP-pc: commodity concentration (coffee, cut flowers, sesame) remains binding, and the Hawassa/Bole Lemi industrial parks have not (yet) tilted the basket. The Oqubay (2015, Made in Africa) state-capitalist development hypothesis predicted a sharper rise; the trade data does not yet show it.
Source: BACI 202501 (retrieved 2026-04-28), manufacturing-export share (HS 25-97 ex-27) for KOR, VNM, ETH, 1995-2024. Authors calcs. Oqubay, A. (2015) 'Made in Africa: Industrial Policy in Ethiopia', Oxford University Press.
Fitted line: log(GDP-pc at peak) = 0.73 +0.0038 × peak_year, R² = 0.001, n = 130. The slope implies that a country peaking in 2020 did so at median GDP-pc $4,818, versus $4,213 for a 1985 peaker under the same fit: the peak has shifted up by a factor of about 1.14× in 35 years. This is the trade-margin analogue of Rodrik's central finding: successive cohorts hit their industrial peak earlier in the development transition.
Source: BACI 202501 (retrieved 2026-04-28) for manufacturing-export share; WDI NY.GDP.PCAP.CD for current-USD GDP per capita at each country's peak year. Method: for each country with at least 10 years of BACI data and at least $1B total exports in 2024, peak year = argmax of manufacturing-export share 1995-2024; dropped if peak at 1995 or 2024 (censored) or if GDP-pc missing at peak. OLS fits log(GDP-pc at peak) ~ peak_year. Authors calcs.
Fitted OLS: peak_year = 2019.9 -1.46× ln(GDP-pc 2000), R² = 0.065, n = 129. The slope implies that a country at $1,000 GDP-pc in 2000 peaks around year 2010, while one at $30,000 peaks around year 2005. The sign is negative, contrary to the Rodrik prior: richer-in-2000 countries peak sooner on this sample, possibly reflecting OECD economies that were already well past their peak by 2000.
Source: BACI 202501 (retrieved 2026-04-28) manufacturing-export share (HS 25-97 ex-27); WDI NY.GDP.PCAP.CD for current-USD GDP per capita in 2000. Method: for each country with an interior peak (dropped if peak at 1995 or 2024) and a known 2000 GDP-pc, scatter peak year against log GDP-pc 2000. OLS fit reports peak_year ~ ln(GDP-pc 2000). Authors calcs.
Source: BACI 202501 (retrieved 2026-04-28) interior-peak sample (see Figure 2 method); WDI NY.GDP.PCAP.CD for 2024 current-USD GDP per capita. Income-tier bands: World Bank country and lending classifications, FY2026 thresholds (LIC < $1,136; LMIC $1,136-$4,495; UMIC $4,496-$13,935; HIC >= $13,936). Method: median peak manufacturing-export share and median peak year within each tier. Authors calcs.
Among 15 economies with at least $10B goods-plus-services exports in 2024and a persistent services crossover, the earliest crossers are services-specialised small open economies; the latest crossers include post-industrial OECD members whose manufacturing trade base eroded after the China shock. The crossover year serves as a structural-transition marker complementing the manufacturing-export peak in Figure 4: a country can peak in manufacturing share and then cross the services-majority threshold a decade later (Autor, Dorn & Hanson 2013, AER, on the US case). Note this is a trade-margin crossover, not the value-added-share analogue Baldwin & Forslid frame; see the discussion of value-added caveats elsewhere on this page.
Source: World Bank WDI, BX.GSR.NFSV.CD (service exports, current USD) and BX.GSR.MRCH.CD (merchandise exports, current USD). Method: for each country with at least 10 joint observations 1995-2024 and at least $10B combined goods-plus-services exports in 2024, flag the earliest year in which services-share >= 50% and held >= 50% through 2024. Baldwin & Forslid (2020) NBER WP 26731 for the 'globotics' crossover motivation. Authors calcs.