No evidence of ideological bias in immigration research

Borjas and Breznau (2026) is a goner

Yes, this is an intentionally provocative title. Try not to get upset over it. Substantively, I only want to report that George Borjas’ recent study (with Nate Breznau), which purported to show ideological bias in the production of research on immigration, was faulty. There might be bias in immigration research (just like in the social sciences in general), but the study didn’t manage to document it.¹ Actually, it could be argued it inadvertently showed the opposite: no detectable bias.

Borjas and Breznau’s analysis of 158 researchers working in 71 teams was published in early 2026 and touted by many. They ostensibly showed that pro-immigration researchers were more likely to find normatively positive impacts of immigration (on public support for social programs), while anti-immigration researchers were more likely to come to the opposite conclusion. How so? Well, as a researcher, you’re free to specify whatever model you think will best explain the data. So, if you’re ideologically biased, you’re more likely (consciously or not) to run and report model specifications that will show you what you want to see.

Here’s one key figure from the paper.

But is this really so?

Katrin Auspurg and Josef Brüderl have just published a comment arguing the main result is not at all robust. When you look at relationship in a simple way, teams with more pro-immigration views do not clearly produce more pro-immigration findings. The apparent ideological link appears only after certain statistical controls are added. More specifically, it shows up only after discipline-fixed effects are introduced. That might be okay by itself, but Auspurg and Brüderl immediately note two important things.

First, when there’s no simple correlation and a link shows up only after specific controls are introduced, you have to explicitly justify those controls. In fact, as they say,

A pattern in which the unadjusted association is close to zero, yet the adjusted association is statistically significant is known as suppression. Recent methodological research cautions that suppression patterns are rare. Instead, they often signal model mis-specification if they are not theoretically justified.

However, the original study does not explain why “disciplinary composition should act as a suppressor.”

Second, and perhaps even more importantly, due to how Borjas and Breznau decided to code disciplinary categories some teams are singletons and are thus dropped from the estimation of the ideology effect. If you poke around a bit in how you define disciplines such that the four dropped teams get picked up in the analysis, the reported ideology effect vanishes.

There are other technical issues as well, but I won’t bother you with those. What’s important is what happens to the original study’s central result when you test robustness with alternative model specifications. Model 8 is the original study’s key adjusted model, and it’s the only one which is positive and significant.

The replicators go beyond this and report all of the following as well:

Our Supplement provides an even more extensive specification-curve analysis. We vary control sets, apply alternative discipline coding schemes, and estimate models with and without regression weights.

We also implement different heteroscedasticity adjustments, including precision weights, which are commonly used in meta-analytic and many-analyst settings to account for clustering of estimates within research teams or studies.

Additionally, we estimate models at the team level, an alternative approach recommended when the objective is to assign equal weight to teams contributing different numbers of estimates . These aggregate models directly assess whether pro-immigration teams report more positive average effects. We estimate 84 models.

Across 81 reasonable alternative specifications that avoid the discussed idiosyncratic choices, the ideology coefficient remains statistically indistinguishable from zero.

A positive and statistically significant association appears only in B&B’s preferred specification and in two closely related variants that B&B presented as robustness checks and which retain their key modeling choices.

It seems that Borjas and Breznau’s study of ideological bias in immigration research is itself vulnerable to the standard issues that were uncovered during the replication crisis.

Or as Kevin Grier says without holding back:

1. On bias in immigration research/reporting, see this Substack piece by Alexander Kustov.