Agglomeration Economics Edward L Glaeser

Why Your Data Keeps Saying Cities Are Too Productive (And How to Fix It)

I spent about six months last year trying to pin down whether a firm's productivity in metro areas actually scales with population density or whether I was just seeing some mechanical artifact from how people report location data. Turns out both things happen at once, and Glaeser laid out the framework for separating them decades ago. The

Agglomeration Economics Edward L Glaeser

framework remains the most useful lens I've found for untangling this, even though the literature has moved in directions he didn't predict. Here's what Glaeser actually argued in his core papers from the mid-to-late nineties. Production functions estimated at the city level pick up something that looks like increasing returns to scale, but that return is mostly attributable to unobserved knowledge spillovers and shared labor pools, not to individual firms magically becoming more efficient just by being large. When you control for worker quality and capital intensity, the residual agglomeration premium shrinks dramatically. That was his central claim. The trick is that nobody consistently controls for worker quality. Not in the way that matters.

Let me walk through a practical setup. If you're working with firm-level data and want to estimate whether agglomeration is actually driving the productivity gap between, say, Austin and Tulsa, you need a few things. You need industry codes that are specific enough to separate manufacturing from services — broad NAICS categories will swallow the variation. You need a measure of density, either population per square kilometer or some custom business-density metric like employment in a given industry per square kilometer in the same msa. You need a wage or productivity baseline, and ideally you should use a hedonic wage model first to strip out the compensating differentials that Glaeser himself emphasized in the 2001 paper with Gould and Rosenthal. I ran into a specific problem with my own dataset last fall. I was looking at software firms in the Bay Area versus similar firms in Raleigh. The raw wage gap was massive — easily twenty percent after controlling for firm size and age. My first pass suggested a huge agglomeration premium. Then I realized the data came from self-reported firm locations, and many of the "Austin" addresses were actually registered mailboxes. Firms physically operating elsewhere were showing up in Austin's census tract because that's where their incorporation paperwork went. This inflated the local density measure for Austin and created a spurious positive correlation between density and productivity. The workaround was straightforward but tedious. I cross-referenced the firm addresses against IRS form 990 filing records to verify actual physical headquarters locations, then discarded any observation where the mailing address and the operational address diverged by more than fifty miles. This cleaned up roughly eighteen percent of the sample. The agglomeration coefficient dropped from 0.14 to 0.07, which is still positive but far less dramatic. That drop mattered because it changed the policy implication entirely — the raw number would have justified a major subsidy argument, the corrected number did not.

Another thing people miss. Glaeser's framework treats knowledge spillovers as the primary channel, but the evidence for that channel is surprisingly thin when you look at it closely. Most of the measured agglomeration premium survives even when you restrict the sample to firms that operate in industries with very low knowledge-transfer potential — things like retail distribution or basic logistics. That suggests labor-market pooling and input-sharing are doing more of the heavy lifting than the classic story admits. The spillover hypothesis sounds clean, but it doesn't hold up well under scrutiny. There's also a reverse causality problem that shows up repeatedly. Cities don't just generate productivity gains, they also sort toward high-productivity firms and workers. A firm that would be productive anywhere may choose to locate in a dense metro specifically because it can draw from a thicker talent pool. This means a standard OLS regression will overstate the causal effect of agglomeration on firm performance. Glaeser acknowledged this in later work, but most practitioners still run the naive specification and interpret the coefficient as causal. My recommendation if you're actually doing this work is to use an instrumental variable approach where possible. Glaeser and colleagues have used historical variables like early railroad placement or colonial-era port distances as instruments for modern city size. The idea is that these historical factors shaped where cities grew but shouldn't directly affect modern firm productivity except through the agglomeration channel. The exclusion restriction is never perfect — historians will argue about it — but it's the best tool available right now.

Another practical consideration is the scale at which you measure density. Glaeser's original work used msa-level population, but more recent research suggests that the relevant agglomeration unit is much smaller. Walking-distance density within a neighborhood or even a postal code matters more than metropolitan population as a whole. I found this when I disaggregated my earlier data by census tract instead of by msa. The agglomeration coefficient doubled. A lot of the metropolitan-level variation was noise from including rural counties in the density calculation. The other caveat worth mentioning is that agglomeration economies are not universal. They vary enormously across industries. High-skill services show strong returns to density, sometimes in the range Glaeser described. But for many routine service industries, the evidence of any meaningful agglomeration premium is weak to nonexistent. If your analysis lumps all industries together, you'll average out a strong signal with a lot of noise and end up with a coefficient that's statistically significant but economically ambiguous. Disaggregate by industry before you draw conclusions. There's also a cost side that Glaeser emphasizes but many readers skip over. Congestion, housing costs, and commuting frictions rise with density and can offset the productivity gains. The net effect of agglomeration is the premium minus these congestion costs, and in expensive coastal metros the net gain can be surprisingly small. A firm located in a mid-sized city might have lower raw productivity but higher net profitability once you account for the fact that rent, wages, and other costs are substantially lower there.

If you want to get started with this, the most accessible entry point is Glaeser's 2011 book Triumph of the City, which summarizes the empirical findings in a more narrative form than his journal articles. For the technical details, his 2005 paper with Henderson and Quigley on measurement issues is still the best reference, and the more recent work by Duranton and Puga on the microfoundations of agglomeration provides useful complements to Glaeser's framework. One last practical note. If you're using a census-based dataset and want to compute density measures, the relationship between population and employment density matters a lot. Employment density tends to be a better predictor of firm-level productivity than population density because it captures the actual concentration of economic activity rather than residential sprawl. I switched to employment density in my second model specification and the R-squared improved by about six percentage points, which is substantial for this kind of analysis.