Skilled immigrants and state innovation, 2013-2023: a strong cross-state correlation that does not survive within-state variation

Summary

At the US state level over 2013-2023, states with a larger share of skilled (bachelor's-degree-or-higher) immigrants patent far more than states with a smaller share (cross-state correlation r=0.62; regression coefficient +21.4, p<0.001, controlling for population size, robust to dropping any one state). But when the same relationship is tested within each state over time — does a state's own skilled-immigrant share rising or falling track its own patenting rising or falling, holding fixed everything constant about that state and about the national year — the relationship vanishes (coefficient -2.2, p=0.82, two-way fixed effects). A first-differenced Granger-style test in both directions (lagged change in skilled share predicting patent growth, and lagged patent growth predicting change in skilled share) found neither direction statistically significant. The state-level pattern in this data is consistent with places attracting skilled immigrants at least as much as skilled immigrants making places innovative — the correlation looks like it is carried almost entirely by fixed differences between states (research universities, tech clusters, coastal metros) rather than by a state's own trajectory changing when its immigrant composition changes. This does not contradict the strongest causal evidence in the literature, which operates at the firm/inventor level with sharper identification (see below) and finds a real, if geographically concentrated, contribution from skilled immigrants — it says that effect, whatever its true size, is not visible in a coarse 51-unit, 14-year state panel.

What the published literature says

The literature on this question is large and split roughly into two designs: (1) micro-level causal designs using individual inventors, firms, or discontinuities in visa policy, and (2) macro/regional correlational designs using cities, states, or countries.

• Micro-causal evidence is fairly one-sided toward skilled immigrants raising innovation. Bernstein, Diamond, McQuade & Pousada (NBER Working Paper 30797, revised 2025, forthcoming/published in the American Economic Review) use an identification strategy built on premature deaths of inventor collaborators to isolate immigrant inventors' causal contribution, and find immigrants are 16% of US inventors but author 23% of patents, and are responsible for an estimated 36% of aggregate US innovation once spillovers onto native collaborators are counted — immigrant inventors' externalities on collaborators are found to be substantially larger than natives'. Kerr's earlier NBER synthesis (w27075, "Immigration, Innovation, and Growth") and the NBER Digest summary ("The Outsize Role of Immigrants in US Innovation") report consistent findings using patent and inventor microdata.
• Quasi-experimental country/region studies using instruments for endogenous location choice also find positive effects — the IZA working paper (Ozgen/Peri-style shift-share designs, dp15693) and a French department-level study (CEPII Lettre 438, 2023) find that a 1-percentage-point rise in the skilled-immigrant share of a local workforce raises local patenting by several percent, using a Card (2001)-style shift-share instrument for immigrant location to address the reverse-causality/selection concern directly — the same concern this analysis is built around.
• Regional/agglomeration studies flag the identification problem explicitly. The NBER working paper "High-Skilled Migration and Agglomeration" (w22926) and a Springer systematic review/meta-analysis of the immigration-innovation literature both note that innovative regions (Silicon Valley, Boston, Seattle) draw skilled migrants precisely because of their existing innovation ecosystems, so a naive cross-sectional correlation cannot separate "immigrants cause innovation" from "innovation attracts immigrants" — exactly the pattern this analysis surfaces empirically in the state panel.
• The World Economic Forum summary and Cato Institute research brief both restate the Bernstein et al. and Kerr-style findings for a policy audience, and a 2026 replication in Economic Inquiry (Wright, "Replication of 'How much does immigration boost innovation?'") checks a widely-cited estimate of immigration's innovation boost and reports whether it reproduces — direct evidence that even the causal literature's headline numbers are actively being stress-tested, not settled.

The consensus that emerges: at the firm/inventor grain, with designs built to break the selection problem (death shocks, visa lottery discontinuities, shift-share instruments), skilled immigrants show a genuine positive causal contribution to innovation, concentrated in STEM fields and disproportionately large through spillovers onto native collaborators. At the same time, every serious treatment of this literature flags that the SAME innovative places are also the places skilled immigrants are drawn to, so any analysis that does not explicitly separate the two forces will overstate the causal share.

What the state-level data shows here

Data: Patent counts by state come from USPTO PatentsView grant data (patents.patent_activity_by_county, rolled up to state x grant-year, primary-assignee location, 2010-2023). The skilled-immigrant measure is the foreign-born population 25+ with a bachelor's degree or higher, as a share of the total population 25+, from Census ACS 5-Year table B06009 (census.acs_nativity_by_education, 2010-2023). Grain is state x year, n=51 units (50 states + DC) x up to 14 years.

Cross-state (between) relationship: Regressing log(1+patents) on the state's average skilled-immigrant share and log(population) across all 51 units (2013-2023 averages) gives a skilled-share coefficient of 21.4 (SE 5.2, p<0.001, R²=0.74). A leave-one-state-out sensitivity check found no sign flip and no state whose removal crosses p=0.05 (max standardized influence 0.62, from Hawaii). New Jersey, California, DC, New York and Maryland have both the highest skilled-immigrant shares and (adjusting for size) high patenting; low-skilled-immigrant-share states (Mississippi, West Virginia, Montana) patent little. This is the correlation most casual treatments of this question would report, and by itself it says nothing about direction.

Within-state (over time) relationship: A two-way fixed-effects panel (state and year fixed effects absorbed, cluster-robust SEs on state, n=714 state-years, 2010-2023) regressing log(1+patents) on skilled-immigrant share finds a coefficient of -2.2 (SE 9.9, p=0.82) — statistically indistinguishable from zero. This specification asks a different, sharper question than the cross-state one: holding fixed everything permanent about a state (its universities, its industry mix, its coastlines) and everything common to all states in a given year (a national patent-office backlog, a recession), does a state's own change in skilled-immigrant composition track its own change in patenting? Here, no.

Granger-style directional test: Using year-over-year first differences (which remove the trend/level collinearity that made a levels-based lead-lag panel numerically unstable — reported here for transparency: a levels-based attempt returned collinearity warnings between the share and its own lag/lead, r>0.99, so it was discarded in favor of differencing), lagged change in skilled-immigrant share does not significantly predict next-year patent growth (coefficient -19.1, p=0.11 — wrong-signed if anything), and lagged patent growth does not significantly predict next-year change in skilled-immigrant share (coefficient +0.0001, p=0.68). Neither direction of Granger-style causality is detectable in this data.

Interpretation: the pattern — strong between-state correlation, null within-state relationship, no detectable lead-lag causality either direction — is the classic empirical signature of a confound: a state's fixed characteristics (research university density, existing tech clusters, coastal-metro status, cost of living) plausibly drive both its long-run skilled-immigrant share and its long-run patenting level, without either one causing changes in the other on a year-to-year basis within a given state.

Caveats — why this does not settle the question

• State grain is coarse. With 51 units, this panel cannot separate more than one or two confounds, and it averages over enormous internal variation — nearly all US patenting activity concentrates in a handful of metro areas within any given state (Silicon Valley within California, the Route 128 corridor within Massachusetts). The literature's strongest causal designs work at the inventor or firm level for exactly this reason; a state panel dilutes any real effect that operates through specific labs, universities, or firms.
• ACS 5-Year estimates are heavily smoothed. Each year's B06009 estimate is a rolling 5-year average, so consecutive years share 4 of 5 underlying survey years. This attenuates true year-to-year variation and could bias the within-state fixed-effects coefficient toward zero even if a real effect exists — the null result here is therefore a lower bound on statistical power, not necessarily proof of no effect. A true test would want annual (not 5-year) nativity data, which is not available at state grain in ACS.
• Patents measure output at the assignee's location, not the inventor's. A multinational headquartered in Delaware with all its R&D and inventors elsewhere would count its patents toward Delaware; this is a standard limitation of assignee-based patent geography and could dilute a true state-level relationship in either direction.
• This analysis cannot rule in or rule out the specific causal channel the strongest literature identifies (individual immigrant inventors raising collaborators' output) — that operates below the grain this data can resolve. It only shows that, at state grain over the last decade, the visible correlation is a between-state phenomenon and does not show up as a within-state, over-time relationship in either direction.