Thursday, 9 July 2026

The canonical entry point for the macro-level view is Besley, Burgess, Khan & Xu's NBER working paper on bureaucracy and development¹, which surveys how administrative data and field experiments connect to the older, more theoretical literature on state-capacity origins. It's a review paper, so the value is architectural: it maps what the profession knows about bureaucrat incentives, political interference, and relationships to citizens and firms, and flags where the frontier is still open. Worth reading in full if you want a single orienting document for the literature.

For a concrete data point on what state investment in citizens can actually do, Bruhn, Garber, Zia & Koyama's VoxDev summary² of a nine-year RCT follow-up in Brazil shows that a school financial-education programme embedded in standard curricula produced lasting reductions in expensive credit use and a measurable shift toward entrepreneurship — effects that only become visible well past the typical one-year evaluation window. The mechanism is a public-delivery question as much as a financial-literacy one.

Gaggl, Gorry & vom Lehn's Review of Economic Studies paper on structural change in production networks¹ is the more tractable of the two pieces here: it documents a rising services share in both input-output and investment networks, builds a multi-sector growth model to match the divergent relative-price trends, and lands on a clean headline — investment-specific technical change accounts for 20% of US aggregate growth since 2000, and structural change within investment networks alone can offset Baumol's cost disease in other sectors. The complement is Farboodi & Veldkamp's Restud paper on the data economy², which formalises data as a depreciating, tradeable intangible asset and calibrates the model to suggest US GDP was mismeasured by up to 6% in 2018. Together they sketch a picture of an economy where the locus of growth is increasingly in assets national accounts were not designed to see.

Scott Cunningham's Mixtape Substack post on conditional parallel trends and covariate inclusion¹ tackles a specific practitioner confusion head-on: the belief that a diff-in-diff estimate that shifts when you add covariates is evidence the design is broken. Cunningham works through the algebra carefully — parallel trends is itself a 2×2 expression on the missing counterfactual, and covariate adjustment is warranted whenever the unconditional version of that assumption is implausible. It's pedagogical in register but the substance is real; worth bookmarking as a reference for the next referee report that flags coefficient instability as a validity concern.

Di Matteo & Sacco's working paper on musical ecosystems¹ nests superstar economics, rational addiction, and Bayesian social learning in a single agent-based framework where preferences are not fixed but co-evolve with the consumption environment. The key prediction — algorithmic curation suppresses diversity past a sharp nonlinear threshold — is tested against four national cases (Italy's Sanremo system, Brazil, South Korea, the UK). The welfare argument is the sharpest part: because listeners adapt their preferences to impoverished environments, revealed preference cannot evaluate the damage, which grounds the case for intervention without resorting to paternalism. The active-inference machinery borrowed from cognitive science is a commitment, but the economic question being asked is the right one.