ImmigrationDBVerified immigration data

More foreign residents, more crime? The answer changes at every border

By · Published · Updated

Question: in cities with more foreign residents, is more crime recorded?

Answer: it depends on the country, and the differences between countries are larger than the pattern itself. In Germany and the United States the two barely move together. In the Netherlands, Spain and Belgium they move together substantially. Any single figure quoted for Europe is an average of countries that disagree with each other.

The numbers: 0.06 in Germany across 133 cities and 0.06 in the United States across 3,818; 0.50 in Spain across 303 and in Belgium across 316; 0.59 in the Netherlands across 40. Pooling all 4,703 cities gives 0.28 — a figure that describes which countries happen to be in the sample, not any city in it.

What we measured

We measured whether cities with a higher foreign-population share also tend to rank higher on police-recorded offences per 100,000 residents. The dataset contains one row for each of 4,703 cities in 19 countries that publish both measures.

The foreign-population measure is a percentage of residents, not a count of arrivals and not an immigration-flow measure. The crime measure is recorded crime, not all crime and not a universal safety score. It covers offences recorded by a police force and published by a statistical or public authority. Reporting, detection, classification and publication practices can differ between cities and countries, so the rate must retain its source and observation-period caveat.

The observation periods are not uniform. Across the full pooled sample, recorded-crime values run from 2022 through 2025, while foreign-share observations run from 2014 through 2026 Q2. The six country samples large enough to report are more concentrated but still not perfectly aligned: Germany combines 2022 and 2024 crime observations with foreign-share observations from 2014 to 2024; Spain combines 2024 crime with 2023–2024 foreign-share data; the Netherlands combines 2024 crime with 2024 foreign-share data except for one observation dated 1 January 2025; and the United States, Belgium and Canada each pair their most recent published crime year with the nearest available foreign-share year.

The unit of analysis is a city. A city-level association says nothing about the conduct of any individual resident. Reading an aggregate city coefficient as evidence about people is the ecological fallacy, and this analysis does not make that inference in either direction.

We report a country's own coefficient only when at least 20 of its cities publish both measures. Six countries qualify: the United States with 3,818, Belgium with 316, Spain with 303, Germany with 133, the Netherlands with 40 and Canada with 36. The other 13 countries contribute the remaining cities to the pooled dataset and the scatter figure, but none is reported with its own coefficient.

What the data shows

The pooled coefficient of 0.28 describes the rank ordering created after all 4,703 cities and all 19 national source systems are mixed together. It does not describe a relationship that repeats consistently within countries. Separated by country, and reporting only samples of at least 20 cities, the coefficients cover a range from 0.06 to 0.59.

Germany’s coefficient is 0.06 for 133 cities, and the United States sits at the same 0.06 across 3,818 — small positive numbers, not zeroes, but city rankings that are only slightly aligned. Spain and Belgium are both 0.50, across 303 and 316 cities, and the Netherlands is 0.59 across 40. Those values answer a limited question about within-country city ranks; they do not identify a cause and cannot be translated into a statement about residents.

The pooled value is also sensitive to which countries are included. All 4,703 cities give 0.28; dropping the United States, which alone supplies 3,818 of them, moves the pooled figure substantially without changing a single city's underlying value. Neither pooled coefficient is the substantive answer — the country figures are.

This matters because the pooled calculation combines two types of ordering: differences among cities within the same national reporting system and differences among rows drawn from different systems. The first can be examined with a country coefficient. The second cannot be treated as a comparable crime-level ordering because national recording and publication rules differ. A single pooled number blends the two.

Where it breaks down

The relationship breaks down at the country boundary, and the six reportable samples split into two groups that have almost nothing in common.

Belgium, Spain and the Netherlands sit at 0.50, 0.50 and 0.59 — a substantial alignment between a city's foreign-population share and where it ranks on recorded crime. Germany, the United States and Canada sit at 0.06, 0.06 and 0.10, which is close enough to nothing that the ranking on one measure tells you almost nothing about the other. The gap between the two groups is larger than the pooled coefficient itself.

The medians in that table show why the levels cannot be compared across the border either. Germany's median city records 7,356 offences per 100,000 residents and the United States' 1,721 — a fourfold difference that reflects what each system counts and publishes, not four times the crime. The same table shows Germany with the highest median foreign share, 17.5%, alongside one of the weakest coefficients.

The scatter makes the sample problem visible. The United States contributes 3,818 of the 4,703 paired cities, so a pooled figure is largely a statement about American places measured under one national system. Thirteen further countries contribute the remaining cities without reaching the twenty needed to report separately. Pooling gives every city one rank; it does not make the source systems equivalent.

What this means if you are moving

This analysis cannot tell you whether a city will be safe for you, and foreign-population share should not be used as a shortcut for that decision. The practical measure is the city’s own recorded-crime rate, read together with its observation year, definition and national source context.

For the largest-city examples in the reportable samples — Berlin, Madrid and Amsterdam — the relevant reading is the same: use the published city rate as a source-specific indicator, not the city’s foreign-share percentage as a proxy. Comparisons are more defensible between cities covered by the same national reporting system and a similar period than between cities in different countries.

ImmigrationDB can surface both measures on a city page, but their presence together does not make one an explanation for the other. The foreign-share figure describes population composition under the source’s definition. The recorded-crime figure describes what the relevant authority recorded and published. A relocation decision should keep those meanings separate.

Density does not solve the safety question either. Its coefficients show that another city characteristic can align with recorded-crime ranks differently by country. The useful lesson is methodological: inspect the city-level rate and its metadata, compare like with like, and reject any single cross-country correlation that claims to summarise a person’s likely experience.

Limits

The strongest limit is measurement comparability. Recorded crime is not total offending or victimisation. It is affected by what reaches or is detected by police, how offences are classified and what an authority publishes. These processes differ, so the article does not compare crime levels between countries.

The second limit is the mismatch in observation periods. The pooled analysis combines crime observations from 2022–2025 with foreign-share observations from 2014–2026 Q2. Germany contains the widest spread: 101 of its 133 crime observations are from 2022 and 32 are from 2024, while 118 foreign-share observations are from 2024 and 15 are older. Spain and the Netherlands are more closely aligned, but not perfectly synchronous.

Third, this is a publication-based sample rather than a random sample of cities. A city appears only when both measures are available at city grain. Thirteen of the 19 countries have fewer than 20 such cities, leaving six for country coefficients. The results therefore describe the available rows, not every city or country.

Fourth, every coefficient is bivariate. Density is a comparison, not a control, and the analysis does not estimate the independent contribution of foreign share, density, unemployment, population or any unmeasured factor. Correlation does not establish direction, mechanism or causation.

Finally, the ecological fallacy remains decisive. City averages cannot identify individual behaviour, exposure or risk, and the data have no neighbourhood-level grain. Even a larger within-country coefficient would still be an aggregate association, not evidence about any person or group.

Method

The analysis covers every city in our database that publishes both a foreign-population share and a recorded-crime rate per 100,000 residents at city level — 4,703 of them, across 19 countries. Each city was weighted equally; population weighting was not applied. The pooled Spearman coefficient was calculated across all complete pairs.

Country coefficients were calculated only for groups with at least 20 complete foreign-share and crime pairs. Three countries passed: Germany with 133 cities, Spain with 97 and the Netherlands with 40. Thirteen countries, containing 63 cities in total, remained in the pooled analysis and scatter but were excluded from individual coefficient reporting. No below-floor country result is inferred or described.

Spearman rank correlation is the primary coefficient because it tests whether the ordering of one variable tends to move with the ordering of the other without requiring a linear relationship in raw units. It is also less dependent than Pearson correlation on an extreme city value. Pearson values are retained in the country-detail table as diagnostics, but the article’s findings use Spearman throughout.

Density correlations use pairwise-complete rows and no imputation: 129 cities in Germany, 75 in Spain and 40 in the Netherlands. The foreign-share correlations use all 133, 97 and 40 cities respectively. Coefficients are reported to two decimals, percentages to one decimal and rates in the precision the source publishes.

The analysis is descriptive. It tests the pooled and within-country rank associations, compares them with density in the same country samples, and stops where the data stop. It does not fit a causal model, compare national crime levels or infer individual behaviour.

Sources

Frequently asked questions

Does immigration increase crime according to these data?

These data cannot answer whether immigration increases crime. Across 4,703 cities, the pooled Spearman coefficient is 0.28, but the six reportable country coefficients range from 0.06 to 0.59. The unit is a city, so the analysis cannot identify what any resident does, establish causation or measure an immigration flow.

Why can’t recorded-crime rates be compared directly across countries?

Police-recorded crime depends on what is reported, detected, classified and published under each national system. A rate from one country is therefore not directly equivalent to a rate from another. This analysis compares how cities rank within countries; it does not rank countries by crime level.

Which country has the largest city-level correlation?

Among countries with at least 20 cities, the Netherlands has the largest Spearman coefficient at 0.59 for 40 cities. Spain and Belgium are both 0.50, for 303 and 316 cities. Germany is 0.06 for 133 and the United States 0.06 for 3,818. These are city-level associations, not causal estimates or statements about foreign residents.

What should someone moving to a city use instead?

Use the city’s own recorded-crime rate, observation year and source definition, and compare it with cities covered by the same national reporting system. Do not treat foreign-population share as a shortcut for safety. The analysis shows that its association with recorded crime changes substantially by country.

Why use Spearman correlation for this analysis?

Spearman correlation compares ranks rather than raw distances between values. That makes the result less dependent on extreme city values and more appropriate when variables are skewed or not linearly related. Each city receives equal weight, and the coefficient describes association only.

More research

Clean air barely shows up in the rent

Do you pay more rent for cleaner air? Across 392 German housing markets the link is weak — Spearman −0.12 — and 31 districts sit below the median on both.

The cities gaining migrants are not the cities that are growing

Where are foreign populations growing fastest? Across 423 cities, growth is almost unrelated to whether the city is growing at all: 0.11 pooled, near zero by country.