The Twenty Percent Border: Impact of the EU Migration and Asylum Pact
Does EU-wide fast-track asylum processing reduce asylum applications?
Preliminary. Data through 2026-08. Last successful update: 2026-11-01 11:58 UTC.
1 Setting and research idea
Under the EU’s new Asylum Procedure Regulation (2024/1348), applicants’ claims are subject to an accelerated border examination (max 12 weeks) whenever the prior-year recognition rate of applicants from the same origin country is 20% or lower, considerably shortening the proecedure. The rule creates a sharp, numerical, and publicly announced eligibility threshold: origins just below 20% face a faster procedure than origins just above it. This note tracks whether that eligibility change shifts the number of asylum applications from affected countroes. This allows us to study the effect of EU-wide policies on asylum applications, which typically can’t be done; single-EU country policy evaluations on asylum applications can only identify relative effects because of diversion. The rule entered into force June 2026, so everything shown here is still pre-policy: the exercise validates that treated and control origins were not already trending apart before the cutoff bites.
2 Research design
The note runs two complementary designs side by side. The first is a difference-in-differences: origins with a 2025 recognition rate between 12.5% and 20% (treated by the cutoff) are compared to origins between 20% and 27.5% (control), using a Poisson (PPML) event study with origin and quarter fixed effects. The second is a dynamic difference-in-discontinuities: at each quarterly horizon, it compares the change in EU-wide applications for origins just below the 20% cutoff against origins just above it, relative to the most recent (reference) quarter, using a fixed 20-percentage-point bandwidth with triangular-kernel local-linear and local-quadratic fits on each side. The first design is simpler and does not depend on local smoothing; the second uses a wider 0-40% support and treats the cutoff itself, rather than the 12.5-27.5% band, as the source of identification. Agreement between the two is a stronger pre-policy check than either alone. Data on applications and recognition rates originate from eurostat.
Technical details
The running variable is the 2025 EU-wide recognition rate minus 20. The sample keeps origins with a 2025 recognition rate strictly above 12.5% and at or below 27.5%, and at least 200 EU-wide applications in 2025 (12 treated, 12 control origins as of this update). For quarterly horizon l, the outcome is the change in that origin’s applications relative to the reference quarter r = 2026 Q2, normalized by the origin’s average quarterly 2025 applications: (Y_r - Y_l) / (Y_2025 / 4) for horizons before the Pact’s application date, and (Y_l - Y_r) / (Y_2025 / 4) for horizons at or after it – so a pre-policy value above zero and a post-policy value above zero both read as movement in the same, policy-consistent direction. Both Y_l and Y_r are rescaled to a full-quarter-equivalent by 3 / months observed whenever a quarter is incomplete, so a partial reference or partial current quarter is not mechanically under-counted relative to complete quarters. The PPML event study applies the same exposure adjustment as a regression offset, log(months observed / 3), rather than rescaling counts directly. Standard errors are clustered by origin country throughout. The data vintage advances automatically to the latest month for which at least 85% of EU Member States have reported any observations. Member States that have not yet reported that month are left out of all periods rather than counted as zero (currently ES, PL, SI), so application totals are sums over the remaining Member States. The sample window, quarter count, and axis extent all update on their own as new Eurostat releases arrive – nothing here is pinned to a calendar date except the Pact’s own June 2026 entry into force.
3 Preliminary results
Figure 1 shows where the 24 origins in the 12.5-27.5% window actually sit relative to the 20% cutoff – the dashed line marks the threshold, with treated origins (accelerated) below it and control origins above it. Figure 2 plots the raw monthly application totals for the two groups; both have been drifting down over 2025-26 without an obvious break yet, but volumes and composition differ, so this alone is not evidence of parallel trends. Figure 3 is the difference-in-differences design’s quarter-fixed-effects event study: the treated-control gap is close to zero in every pre-reference quarter and never statistically distinguishable from it. Figure 4 is the dynamic difference-in-discontinuities design’s main estimate, plotted on the same y-axis scale as Figure 3 for direct comparison: linear and quadratic fits agree that the gap has been converging toward zero as the reference quarter approaches, with wide, overlapping confidence bands throughout.
Figure 1: origin countries around the 20% cutoff
Explainer: what’s behind Figure 4 (the local RDD fits)
Each panel of the figure below is one quarterly horizon’s cutoff comparison: origin-level normalized changes plotted against the 2025 recognition rate, with application-weighted bins and the triangular-kernel linear/quadratic fits that Figure 4 summarizes into a single treated-minus-control estimate per quarter. It’s the diagnostic view behind the headline numbers, useful for checking that no single origin or outlier is driving a quarter’s estimate.
Figure 6 zooms in on two periods: the last quarter before the Pact (2026 Q1) and the latest data after it (July–August 2026), with linear fits only. Each is compared with 2026 Q2, the quarter in which the Pact took effect; an incomplete quarter is labelled by the months it covers and scaled to a full-quarter equivalent. The outcome is the same as in Figure 5. The dotted lines are the application-weighted mean change on each side of the cutoff across the full 20-point bandwidth, so the gap between them compares all treated with all control origins in the bandwidth, much like a simple first difference. The gap between the two solid lines at 20% is the RDD estimate, which uses only the variation close to the cutoff.
In 2026 Q1, the treated-side mean is -0.07 and the control-side mean -0.16, and the linear RDD estimate is -0.00 (95% CI -0.22 to 0.22). Just before the Pact, the side means differ by 0.09 and the fits show essentially no jump at the cutoff. (For pre-policy quarters the change is signed reference minus horizon, as in Figure 5.) In July–August 2026, the treated-side mean is -0.24 against 0.06 on the control side, so treated origins as a group lost applications relative to control origins. The linear RDD estimate is much smaller, -0.01 (95% CI -0.33 to 0.30). In the data through 2026-08, the treated-side decline is largest among origins with recognition rates far below 20%, and the treated-side fit rises toward the cutoff. The gap in means therefore reflects how treated and control origins differ in general more than a jump at the threshold itself. One post-policy month is too little to tell these apart.
4 Conclusion
With 12 treated and 12 control origins, the event study design has reasonable power: the most recent complete pre-reference quarter’s PPML estimate is -0.037 (SE 0.063). Because PPML is a log-linear model, that estimate reads on the percentage scale as a -3.6% pre-trend (95% CI -14.8% to 9.0%) – small enough to detect a moderate-sized effect, so the PPML design should have real power to pick up a genuine fast-track deterrent once post-policy data arrive. The dynamic difference-in-discontinuities estimates are far more imprecise, so most of that power comes from the PPML side rather than the RDD one, even though the RDD is arguably better identified. However, as every year in January the recognition rates list updates, every january becomes a new natural experiment, which will allow us to gradually build evidence, as some treated countries become control and vice versa. Let the data flow in!
Two extensions are worth flagging even at this pre-policy stage. First, the composition of applicants near the cutoff may itself respond to the rule: an origin just below 20% faces accelerated screening, which could raise its recognition rate the following year if weaker claims are filtered out disproportionately, pushing that origin back above the threshold. Repeated across years, this would produce a zigzag in treatment status – origins alternating between treated and control as their own recognition rate reacts to the prior year’s exposure – which threatens designs that compare a given origin’s treatment status across years, including the difference-in-differences. The RDD is less exposed to this, since it re-identifies the discontinuity within each year’s cutoff rather than tracking a given origin’s status over time. Second, applicants who prefer to avoid the border procedure altogether have an incentive to evade detection at the border and travel onward to register in a different, typically more western, Member State. That kind of diversion would not show up in recognition-rate composition but would show up in the geographic distribution of first registrations, which is worth examining directly once enough countries report application counts by Member State of registration.
How this page updates itself
Once a month, a GitHub Actions workflow checks Eurostat’s bulk data files for a new release, reruns the versioned R analysis script (update.R), and re-renders this Quarto note. If the new data pass validation – enough Member States reporting, no broken joins – the workflow commits the refreshed figures, results, and this page straight to the repository, and GitHub Pages serves the update automatically. Nothing here is edited by hand: the same code that produced the numbers above produces every future update, and the full history of past versions stays visible in the repository’s commit log.
Sources: Eurostat migr_asydec1pc (first-instance decisions) · Eurostat migr_asyappctzm (first-time applications) · Code: update.R · Contact: j.a.h.adema94@gmail.com