Structural Change in Production Networks and Economic Growth

The paper studies two networks — the input-output network, the sectoral distribution of intermediate input production and purchases, and the investment network, the distribution over the production and purchases of new capital — and documents that in each, goods sectors systematically produce a smaller share of output over time, offset by increased production by service sectors. The measurement uses the BEA Input Output Database Make and Use Tables from 1947 through 2020, supplemented for the investment network by data from vom Lehn and Winberry (2022), extended by the authors through 2020. The dataset covers 40 NAICS-defined sectors including agriculture and government, with “goods” defined as the 22 agriculture, mining, construction and manufacturing sectors and “services” as the remaining 18. Three energy-intensive sectors — oil and gas extraction, utilities, and petroleum and coal manufacturing — are excluded because price growth in them can be large enough to overshadow price growth in other intermediate input producers; the authors report in an appendix that these sectors play no role in structural change in production networks. The authors describe these changes as large, similar in magnitude to structural change observed in consumption expenditures, and widespread, occurring in nearly all sectors and in many countries. They also note that which specific goods and services sectors are rising and falling varies by network: most sectors buy investment from a common set of hubs — construction, machinery and motor vehicles manufacturing, and information and professional/technical services — whereas intermediates suppliers are much more sector-specific.

Because the goods and services sectors producing consumption, investment, and intermediates are distinct, the authors measure the price of goods and services separately for each final use, and they argue that not doing so induces an aggregation bias that reverses the sign of the investment price trend. They map BEA final expenditure data on 68 consumption and 30 investment commodities in the NIPA to the goods or services sectors that produce them. Intermediates are not in GDP and therefore not in NIPA expenditure data, so their prices are inferred as residuals from the BEA’s sectoral gross output prices, which are a weighted average of consumption, investment, and intermediates prices. The authors report that this approach is internally consistent with published aggregate price data and that the resulting intermediates price series align almost perfectly with BLS producer price indices for the select years those are available. The stated key to their result is that since only around 5% of services output is used for investment, the price of investment produced by services sectors is not accurately represented in the aggregate price index for the services sector as a whole; they show this aggregation bias persists even when using more disaggregated sectoral output prices.

While the relative prices of services that produce intermediates and consumption are rising, the authors find that the relative price of services that produce investment is falling over time. They locate this against recent work allowing for structural change in investment — Herrendorf, Rogerson and Valentinyi (2021), García-Santana, Pijoan-Mas and Villacorta (2021), and Sposi, Yi and Zhang (2021) — which measures a single price for goods and services across uses and finds the price of services rising faster than that of goods. The authors reproduce that result for consumption and intermediates, but obtain the opposite result for investment. The examples they give are that consumption services include education and health care, intermediate services include financial services and wholesale trade, and investment services include software development and R&D.

The paper extends the multi-sector neoclassical growth model to allow for production networks in both intermediates and investment, with each sector producing gross output from capital, labor, and intermediate inputs, and with each sector’s intermediates and investment being a CES bundle of inputs purchased from many sectors. This setup means changes in relative prices across sectors, induced by changes in technology, can generate changes in the composition of intermediates and investment production — structural change in production networks. In a special case that admits a balanced growth path, the authors show that the higher the elasticity of substitution, the faster growth will be in either network, and that the composition of aggregate growth endogenously shifts toward the network with the higher elasticity of substitution. The full model does not admit a balanced growth path, so the quantitative work solves for a transition path between two steady states, informed by work on stable transformation paths by Buera, Kaboski, Mestieri and O’Connor (2024). The baseline calibration studies six sectors, with each of the consumption, investment, and intermediates sectors split into a goods and a services subsector, which the authors describe as a natural extension of Greenwood, Hercowitz and Krusell (1997) and Ngai and Samaniego (2009).

The calibrated elasticity parameters confirm that goods and services are complements in the bundling of intermediates in all sectors (elasticities less than one), but substitutes in investment bundling in all sectors (elasticities greater than one). Elasticities are calibrated by minimizing least-squares differences between the model and data series for structural change in each of the six sectors over 1947-2020, given share parameters set to match 1947 expenditure shares. For consumption, the calibration yields a goods share of 0.13 and an elasticity of substitution of 0.28, with a non-homotheticity parameter of 0.54; the authors note the consumption and intermediates elasticities imply strong complementarity, consistent with the existing literature they cite. The economic content of the split is that complementarity in intermediates and consumption means structural change reallocates expenditure toward producers with the slowest TFP growth — what the authors call the “bottlenecks” of growth — while substitutability in investment means expenditures reallocate toward the producers with the fastest TFP growth, the “frontiers” of growth. Because these forces push in opposite directions, the authors state that the net impact on aggregate growth is a quantitative question.

Over the full 1947-2019 sample, investment-specific technical change accounts for 52% of aggregate GDP-per-worker growth, consumption-specific for 35%, and intermediates-specific for 31%, with the contribution of bundling technical change alone reported as -1%. The decomposition is computed from counterfactual transition paths in which only a subset of TFP series is allowed to grow, so the components are not constrained to sum to 100% given the nonlinear relationships between individual technology series and the aggregates; the authors flag this explicitly, and also note the bundling series are already contained in the investment- and intermediates-specific counterfactuals. The shares move sharply over time: for 2000-2019, investment-specific technical change accounts for 94% of growth, consumption-specific for 54%, and intermediates-specific for -4%. Their first observation from the exercise is that all three sources of technical change significantly contribute to aggregate growth.

Replacing CES investment aggregation with Cobb-Douglas, which rules out structural change in investment, implies 5% lower total growth in GDP per worker over the full sample, and 20% lower growth since 2000. When either intermediates or consumption aggregation is made Cobb-Douglas, growth is instead faster, because resources no longer reallocate toward the slowest-growing producer; over the entire sample the effect of complementarity in either of these products is to reduce GDP growth by about 3-5%, though the effect attributable to consumption since 2000 is 9%. The authors draw three observations: investment-specific technical change is an increasingly important force in generating economic growth and structural change in investment has played a substantial role in driving that growth since 2000; structural change in consumption and intermediates has exerted a modest drag on aggregate growth, with the consumption drag larger since 2000; and intermediates-specific technical change has stagnated, with a negative total contribution to growth since 2000.

Intermediates-specific technical change appears in the model as a significant contributor to the recent slowdown in economic growth, having stagnated with a negative total contribution since 2000. The authors add a timing observation: although significant attention has been paid to bottlenecks in supply chains since the COVID-19 pandemic, they find substantial sluggishness along this network in the years preceding the pandemic. In the data, GDP per worker grows at a fairly steady rate over most of the sample but shows a pronounced slowdown since 2010, which the calibrated transition path matches by construction.

In the authors’ projection it does not, because substitutability in investment inputs offsets the drag: forecast growth in GDP per worker starts at 0.66% in 2032 and rises to 0.87% in 2070. The forecast assumes all TFP series grow at a constant rate from 2021-2070, with growth rates set to the average growth in each series over 2010-2020, and the authors discard the first ten years of the window because growth there partially reflects delayed capital accumulation dynamics. In counterfactual paths where TFP grows only among consumption producers or only among intermediates producers, aggregate growth slows over time, consistent with Baumol’s cost disease; when TFP grows only at investment producers, growth rises over time. When the elasticity of substitution in investment inputs is instead set to 1 (Cobb-Douglas) or 0 (Leontief) — assumptions the authors describe as commonly made in the literature — aggregate growth is substantially lower and declining over time. The authors’ stated reading is that investment reallocation serves as a critical counterbalance to the forces driving long-run stagnation from Baumol’s cost disease, and that this mechanism appears sufficient to fully offset the growth drag caused by structural change in consumption and intermediate production. In the conclusion they characterize the finding as providing “some optimism” for the impact of sectoral reallocation on aggregate economic growth, while noting the intermediates network in their framework does indeed appear to suffer from the cost disease.

The results are conditional on the calibrated elasticities, on the transition-path solution between two steady states, and — for the projections — on the assumption that sectoral TFP grows at constant 2010-2020 average rates through 2070. On data, the authors note that the BEA reports only total spending on imports and does not differentiate how imports are used, so imports cannot be removed from the networks within their measurement framework; they argue this cannot account for the patterns they document because imports are never more than 8% of total consumption, intermediates, and investment in any given year. The energy-intensive sector exclusion is a deliberate measurement choice defended in an appendix. The authors report in an appendix that they also calibrate the investment elasticities to different sub-periods within 1947-2020, given the novelty of those estimates. All data and computer code necessary to reproduce the figures and tables are stated to be available in a replication package.