Init

library(kirkegaard)
load_packages(
  dplyr, tidyr, readr, stringr, purrr, tibble,
  sf, ggplot2, ggrepel, ggcorrplot, patchwork,
  lavaan, modelsummary, gt
)

theme_set(theme_bw(base_size = 11))
options(digits = 3)

knitr::opts_chunk$set(
  echo = TRUE, message = FALSE, warning = FALSE,
  fig.width = 9, fig.height = 6, dpi = 130
)

Project background

A debate in Danish politics about what it means to be Danish motivated this analysis. A politician from the Social Democrats, Frederik Vad, posted commune-level results from the Ministry of Justice’s Tryghedsundersøgelse (safety/tryghed survey), and it was obvious that commune-level ethnic composition correlated strongly with reported safety — but the correlation had not been quantified and the underlying appendix was missing from the printed version of the main alternative source (Realdania’s Vores Livskvalitet, 2. oplag 2025). Realdania later supplied the full Appendix 4A via email.

This notebook documents the data assembly and the analyses we ran. All source scripts live in R/ and all intermediate datasets in data/.

Data sources

Measure	Source	Unit	Time
MoJ Tryghedsundersøgelsen	Ministry of Justice / Rigspolitiet annual reports; commune-level values transcribed from the district maps.	98 communes	2019, 2021, 2022, 2023, 2024
Realdania safety & noise (Appendix 4A)	Vores Livskvalitet (Realdania 2025), received by email.	98 communes	2024/25 wave
Realdania six commune indexes	Vores Livskvalitet appendices 3A, 5B, 6A, 10B, 11A (livskvalitet, socioøkon, naboskab, mental sundhed, tillid).	98 communes	2024/25 wave
FOLK1C (ancestry × origin country)	Danmarks Statistik, direct API pull.	98 communes × 4 HERKOMST × 242 IELAND	2022Q2
STRAF11 (reported offences)	Danmarks Statistik, direct API pull.	98 communes × offense category × quarter	2021-2024
BY2 (population by settlement size)	Danmarks Statistik, direct API pull.	98 communes × BYST	2024
Commune polygons	Dataforsyningen DAWA GeoJSON.	98 commune polygons	current
Postnummer polygons	Dataforsyningen DAWA GeoJSON.	1,089 postnumre	current
Postnummer demographics 2014	A4.csv (DST pull, geometry-indexed) in `data/`.	592 postnumre	2014-01-01
Postnummer demographics 2016	`postnumre dst 2016.xlsx` in `data/`.	457 postnumre	2016-10-01

The R/extract_*.R scripts produce the tidy datasets read below.

Load precomputed datasets

tryghed_long <- readRDS("data/tryghed_MoJ_long.rds")          # long panel 2019-2024
tryghed_mean <- readRDS("data/tryghed_mean_2021_2024.rds")    # commune mean 2021-24
tryghed_rd   <- readRDS("data/tryghed_realdania.rds")         # Realdania Appendix 4A
tryghed_pn   <- readRDS("data/tryghed_postnumre_top_bund.rds")# top/bund 20 postnumre

menapta      <- readRDS("data/commune_menapta_2022Q2.rds")    # commune MENAPTA share
crime        <- readRDS("data/straf11_commune_2021_2024.rds") # reported offences
urban        <- readRDS("data/urbanisering_2024.rds")         # % in byområder ≥200
popdens      <- readRDS("data/popdensity_2022Q2.rds")         # pop / km²
indexes      <- readRDS("data/commune_indexes.rds")           # Realdania book indexes

pn14         <- readRDS("data/postnumre_2014.rds")            # 2014 postnummer demog
pn16         <- readRDS("data/postnumre_2016.rds")            # 2016 postnummer demog
pn_nbs       <- readRDS("data/tryghed_postnumre_neighbours.rds")
path_fits    <- readRDS("data/path_models.rds")

geo_kom <- st_read("data/geo/kommuner.geojson", quiet = TRUE) |>
  filter(!udenforkommuneinddeling) |>
  transmute(kommune = navn, geometry)

MoJ tryghed: inter-year reliability

Five years of map-based commune values (2019, 2021, 2022, 2023, 2024). Rank-order stability and the reliability of a multi-year average.

w <- tryghed_long |>
  select(kommune, year, tryghed_pct) |>
  pivot_wider(names_from = year, values_from = tryghed_pct, names_prefix = "y") |>
  select(-kommune)

pearson_mat  <- cor(w, method = "pearson")
spearman_mat <- cor(w, method = "spearman")
round(pearson_mat, 2)

##       y2019 y2021 y2022 y2023 y2024
## y2019  1.00  0.70  0.54  0.60  0.66
## y2021  0.70  1.00  0.72  0.75  0.75
## y2022  0.54  0.72  1.00  0.72  0.69
## y2023  0.60  0.75  0.72  1.00  0.70
## y2024  0.66  0.75  0.69  0.70  1.00

round(spearman_mat, 2)

##       y2019 y2021 y2022 y2023 y2024
## y2019  1.00  0.61  0.47  0.45  0.56
## y2021  0.61  1.00  0.66  0.64  0.65
## y2022  0.47  0.66  1.00  0.65  0.63
## y2023  0.45  0.64  0.65  1.00  0.59
## y2024  0.56  0.65  0.63  0.59  1.00

z <- scale(w)
alpha <- function(M) {
  k <- ncol(M); v <- apply(M, 2, var); tot <- var(rowSums(M))
  (k/(k-1)) * (1 - sum(v)/tot)
}
sb <- function(k, rbar) k*rbar / (1 + (k-1)*rbar)
r_avg <- function(M) { R <- cor(M); mean(R[lower.tri(R)]) }

tibble(
  model = c("all 5 years", "drop 2019 (4 map-era years only)"),
  alpha = c(alpha(z), alpha(z[, -1])),
  mean_r = c(r_avg(z), r_avg(z[, -1])),
  reliability_of_mean = c(sb(5, r_avg(z)), sb(4, r_avg(z[, -1])))
) |> mutate(across(where(is.numeric), round, 3))

Adjacent-year Pearson correlations are ~0.70–0.75 and adjacent Spearmans ~0.60–0.66, indicating a stable between-commune ranking with meaningful sampling noise in any single year. Cronbach α across 5 years is 0.915; dropping 2019 (different survey format) barely changes α (0.912). We use the mean of 2021-2024 as the primary MoJ safety variable.

Choropleth of mean tryghed 2021-2024

m <- geo_kom |> left_join(tryghed_mean, by = "kommune")

ggplot(m) +
  geom_sf(aes(fill = tryghed_mean_21_24), colour = "white", linewidth = 0.15) +
  scale_fill_viridis_c(option = "mako", direction = -1,
                       name = "% feeling safe\nin neighbourhood\n(2021-2024)",
                       breaks = seq(75, 95, 5)) +
  coord_sf(crs = 25832) +
  labs(title = "Feeling safe in neighbourhood, mean 2021-2024",
       subtitle = "Ministry of Justice Tryghedsundersøgelsen (share answering 1-3 on 1-7 scale)") +
  theme_void(base_size = 11) +
  theme(plot.title = element_text(face = "bold"),
        legend.position = c(0.88, 0.55))

Low-tryghed cluster is concentrated on Copenhagen’s western ring (Vestegn): Ishøj, Brøndby, Albertslund, Høje-Taastrup, Herlev. Islands and rural western Jutland are high.

MENAPTA definition and construction

MENAPTA: share of commune residents (1st + 2nd generation) with origin in the Middle East, North Africa, Pakistan, Turkey, Afghanistan, or the rest of Africa (from Statistics Denmark FOLK1C, 2022Q2, using the midpoint of the 2021-2024 tryghed window).

Excluded: South Africa; Israel; European overseas territories in Africa (Réunion, St Helena, Spanish Africa — all exclaves with European populations). Stateless persons included (the majority in DK are Palestinians).

menapta |>
  arrange(desc(menapta_pct)) |>
  select(kommune, pop_total, menapta_n, menapta_pct) |>
  head(10)

National MENAPTA share is 5.48 %. The top communes are all in the Copenhagen Vestegn ring (Ishøj 29%, Brøndby 23%, Albertslund 22%, Høje-Taastrup 19%).

MENAPTA vs tryghed — main scatter

df_moj <- tryghed_mean |> inner_join(menapta |> select(kommune, menapta_pct), by = "kommune")
r_p <- cor(df_moj$menapta_pct, df_moj$tryghed_mean_21_24)
r_s <- cor(df_moj$menapta_pct, df_moj$tryghed_mean_21_24, method = "spearman")

ggplot(df_moj, aes(menapta_pct, tryghed_mean_21_24)) +
  geom_smooth(method = "lm", se = TRUE, colour = "steelblue",
              fill = "steelblue", alpha = 0.12, linewidth = 0.6) +
  geom_point(alpha = 0.75, size = 1.8, colour = "grey25") +
  geom_text_repel(aes(label = kommune), size = 2.8, max.overlaps = 30,
                  segment.alpha = 0.4, min.segment.length = 0.2) +
  scale_x_continuous(breaks = seq(0, 30, 5), labels = \(x) paste0(x, "%")) +
  labs(x = "MENAPTA share of commune population (2022Q2)",
       y = "Share feeling safe in neighbourhood, mean 2021-2024 (%)",
       title = "Commune safety vs MENAPTA share",
       subtitle = sprintf("Pearson r = %.2f   Spearman ρ = %.2f   n = 98",
                          r_p, r_s)) +
  theme_minimal(base_size = 11)

Comparing the two safety measures

Realdania’s Appendix 4A asks a different question:

MoJ: “On a scale from 1–7 […], how safe or unsafe do you feel? By ‘neighbourhood’ we mean the area immediately surrounding your residence.” Safety % = share answering 1–3.
Realdania: “Are there places in your local area where you avoid going after nightfall because you feel unsafe?” Safety % = share answering “No, I feel safe in the whole local area”.

df_cmp <- inner_join(tryghed_mean, tryghed_rd |> select(kommune, tryg_lokalomraade),
                     by = "kommune")
r_p2 <- cor(df_cmp$tryghed_mean_21_24, df_cmp$tryg_lokalomraade)
r_s2 <- cor(df_cmp$tryghed_mean_21_24, df_cmp$tryg_lokalomraade, method = "spearman")
fit  <- lm(tryg_lokalomraade ~ tryghed_mean_21_24, df_cmp)

df_cmp <- df_cmp |>
  mutate(resid = tryg_lokalomraade - predict(fit),
         label = ifelse(rank(-abs(resid)) <= 15, kommune, NA_character_))

ggplot(df_cmp, aes(tryghed_mean_21_24, tryg_lokalomraade)) +
  geom_abline(slope = 1, intercept = 0, colour = "grey70", linetype = "dashed") +
  geom_smooth(method = "lm", se = TRUE, colour = "steelblue",
              fill = "steelblue", alpha = 0.12, linewidth = 0.6) +
  geom_point(alpha = 0.8, size = 1.8, colour = "grey25") +
  geom_text_repel(aes(label = label), size = 2.9, min.segment.length = 0.2) +
  labs(x = "MoJ: safe in immediate neighbourhood, mean 2021-2024",
       y = "Realdania: safe in whole local area, 2024/25",
       title = "Two commune-level safety measures",
       subtitle = sprintf(
         "Pearson r = %.2f   Spearman ρ = %.2f   n = 98   |   Realdania = %.2f + %.2f · MoJ",
         r_p2, r_s2, coef(fit)[1], coef(fit)[2])) +
  theme_minimal(base_size = 11)

The two measures correlate strongly (r = 0.87) but Realdania’s range is 2.4× wider. All 98 points sit below the 45° line — Realdania values are systematically lower. The largest negative residuals are all urban Vestegn-adjacent communes (Herlev, Vallensbæk, Hvidovre, Slagelse, Glostrup, Ishøj, Albertslund, Frederikssund), consistent with the “whole lokalområde” framing picking up visible nearby problem areas.

Correlation matrix of all commune-level variables

full <- tryghed_mean |>
  left_join(tryghed_rd,  by = "kommune") |>
  left_join(menapta |> select(kommune, menapta_pct), by = "kommune") |>
  left_join(crime   |> select(kommune, rate_straffelov_per1000), by = "kommune") |>
  left_join(indexes, by = "kommune")

m_eng <- full |>
  transmute(
    `Safety MoJ (neighbourhood, 2021-24)` = tryghed_mean_21_24,
    `Safety Realdania (local area)`       = tryg_lokalomraade,
    `% reporting neighbour noise`         = pct_nabostoej,
    `% reporting traffic noise`           = pct_trafikstoej,
    `MENAPTA %`                           = menapta_pct,
    `Crime rate (criminal code /1000/yr)` = rate_straffelov_per1000,
    `Life quality (raw)`                  = livskvalitet_raw,
    `Life quality (adj.)`                 = livskvalitet_adj,
    `Socioeconomic index`                 = socioekon_indeks,
    `% with lack of space at home`        = pladsmangel_pct,
    `Neighbourhood satisfaction`          = naboskab_tilfreds,
    `Neighbour activities`                = nabo_aktiviteter,
    `Mental health (ABC index)`           = mental_sundhed_abc,
    `Trust in others`                     = tillid_andre,
    `Trust in authorities`                = tillid_myndigheder
  )

cm <- cor(m_eng, method = "pearson")
ggcorrplot(cm, type = "lower", hc.order = TRUE, hc.method = "ward.D2",
           lab = TRUE, lab_size = 2.6, outline.col = "white",
           colors = c("#c6362e", "white", "#1f6f8b"), tl.cex = 10,
           legend.title = "Pearson r") +
  labs(title = "Commune-level correlations (n = 98)") +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))

Two clean clusters: a “local wellbeing” bundle (safety, neighbourhood satisfaction, life quality, trust in others) and an “urban stress” bundle (MENAPTA, crime, noise, socioeconomic problems) which anti-correlate. Trust in authorities is the clear outlier — its variation across communes is driven by individual-level demographics (education, age, politics), not the local environment gradient.

Crime rates

crime |>
  arrange(desc(rate_straffelov_per1000)) |>
  select(kommune, rate_tot_per1000, rate_straffelov_per1000, rate_person_per1000) |>
  head(10)

Reported offences sum 2021-2024, expressed as annual rate per 1,000 inhabitants using the 2022Q2 FOLK1C population denominator. We use “straffelov i alt” (all criminal code offences) as the default headline rate; narrower subcategories like violence or sexual offences produce single-digit annual counts in small communes and become unstable.

Urbanization and population density

bind_rows(
  urban   |> arrange(desc(urban_pct))   |> head(3) |>
    transmute(kommune, measure = "Urban %", value = urban_pct),
  popdens |> arrange(desc(pop_density)) |> head(3) |>
    transmute(kommune, measure = "Density /km²", value = pop_density)
)

Plain population density (pop / total commune area) is distorted by empty land (forests, uninhabited coastal zones); residents living in settlements of ≥200 (DST’s definition of a byområde) is a cleaner measure of “how urban the average resident’s environment is”.

Regression table: safety ~ MENAPTA + crime + (urban or density)

dz <- full |>
  left_join(urban   |> select(kommune, urban_pct), by = "kommune") |>
  left_join(popdens |> select(kommune, log_pop_density), by = "kommune") |>
  rename(moj = tryghed_mean_21_24, rd = tryg_lokalomraade,
         menapta = menapta_pct, crime = rate_straffelov_per1000,
         urban = urban_pct, logdens = log_pop_density) |>
  mutate(across(c(moj, rd, menapta, crime, urban, logdens),
                \(x) as.numeric(scale(x))))

models <- list(
  "(1) MoJ + Urban%"               = lm(moj ~ menapta + crime + urban,   dz),
  "(2) MoJ + log10(density)"       = lm(moj ~ menapta + crime + logdens, dz),
  "(3) Realdania + Urban%"         = lm(rd  ~ menapta + crime + urban,   dz),
  "(4) Realdania + log10(density)" = lm(rd  ~ menapta + crime + logdens, dz)
)

modelsummary(
  models,
  coef_map = c(
    "menapta" = "MENAPTA %",
    "crime"   = "Crime rate (criminal code /1000/yr)",
    "urban"   = "Urban % (residents in settlements ≥ 200)",
    "logdens" = "log10(population density, /km²)"
  ),
  gof_map   = tribble(~raw, ~clean, ~fmt,
                      "adj.r.squared", "adj. R²", 3),
  stars     = c("*" = .05, "**" = .01, "***" = .001),
  statistic = "({std.error})",
  coef_omit = "(Intercept)",
  title     = "Commune-level determinants of subjective safety (n = 98) — standardised β",
  notes = c(
    "Standardised β coefficients (all variables z-scored); SE in parentheses. * p<0.05, ** p<0.01, *** p<0.001. n = 98.",
    "MENAPTA %: 1st+2nd-gen population share from Middle East, North Africa, Pakistan, Turkey, Afghanistan, rest of Africa. Excl. South Africa, Israel, EU overseas territories; incl. stateless. Source: DST FOLK1C, 2022Q2.",
    "Crime rate: DST STRAF11 criminal code offences / 1,000 inhabitants / year, 2021-2024.",
    "Urban %: DST BY2 (2024), residents in settlements ≥ 200. Density: FOLK1C pop / commune land area (EPSG:25832)."
  ),
  output = "html"
)

Commune-level determinants of subjective safety (n = 98) — standardised β
	(1) MoJ + Urban%	(2) MoJ + log10(density)	(3) Realdania + Urban%	(4) Realdania + log10(density)
* p < 0.05, p < 0.01, * p < 0.001
Standardised β coefficients (all variables z-scored); SE in parentheses. * p<0.05, p<0.01, * p<0.001. n = 98.
MENAPTA %: 1st+2nd-gen population share from Middle East, North Africa, Pakistan, Turkey, Afghanistan, rest of Africa. Excl. South Africa, Israel, EU overseas territories; incl. stateless. Source: DST FOLK1C, 2022Q2.
Crime rate: DST STRAF11 criminal code offences / 1,000 inhabitants / year, 2021-2024.
Urban %: DST BY2 (2024), residents in settlements ≥ 200. Density: FOLK1C pop / commune land area (EPSG:25832).
MENAPTA %	−0.656***	−0.712***	−0.478***	−0.519***
	(0.062)	(0.064)	(0.064)	(0.070)
Crime rate (criminal code /1000/yr)	−0.270***	−0.379***	−0.251***	−0.333***
	(0.071)	(0.077)	(0.072)	(0.084)
Urban % (residents in settlements ≥ 200)	−0.060		−0.286***
	(0.071)		(0.073)
log10(population density, /km²)		0.128		−0.122
		(0.083)		(0.091)
adj. R²	0.774	0.778	0.764	0.731

Key findings:

For MoJ safety, urban% and density are both non-significant once MENAPTA and crime are in the model (β = −0.06 and +0.13, both NS). MENAPTA dominates (β = −0.66).
For Realdania safety, urban% is a significant independent predictor (β = −0.29, p < 0.001) while density is not. Consistent with Realdania’s broader “whole lokalområde” framing picking up ambient urbanity beyond crime + MENAPTA.
Urban % and log10(density) correlate r = 0.84 but urban % wins the horse-race in three of four models — plain density is diluted by empty land.

Path model: MENAPTA / urban → crime → safety

Standardised β coefficients (bias-corrected bootstrap CIs, 2000 replicates). Source code: R/path_model_safety.R and R/plot_path_model.R.

extract_effects <- function(fit, ylabel) {
  ps <- parameterestimates(fit)
  get <- function(op, lhs, rhs = NULL, lab = NULL) {
    r <- if (!is.null(lab)) ps |> filter(label == lab)
         else ps |> filter(op == !!op, lhs == !!lhs, rhs == !!rhs)
    slice(r, 1)
  }
  tibble(
    outcome = ylabel,
    direct_MENAPTA  = get("~", if(ylabel=="MoJ") "moj_safety" else "rd_safety", "menapta")$est,
    direct_urban    = get("~", if(ylabel=="MoJ") "moj_safety" else "rd_safety", "urban")$est,
    indirect_MENAPTA= get(lab = "ind_men")$est,
    indirect_urban  = get(lab = "ind_urban")$est,
    total_MENAPTA   = get(lab = "tot_men")$est,
    total_urban     = get(lab = "tot_urban")$est,
    pct_direct_MENAPTA = 100 * direct_MENAPTA / total_MENAPTA,
    pct_direct_urban   = 100 * direct_urban   / total_urban
  )
}
bind_rows(
  extract_effects(path_fits$moj, "MoJ"),
  extract_effects(path_fits$rd,  "Realdania")
) |> mutate(across(where(is.numeric), round, 3))

MENAPTA operates almost entirely through a direct (non-crime-mediated) channel: ~91% direct in MoJ, ~88% in Realdania. The direct channel likely reflects some combination of (a) crime under-reporting or antisocial behaviour not reaching STRAF11 (nuisance, low-level disorder), and (b) a composition effect — individuals in the included origin groups reporting lower safety irrespective of where they live (this is well- documented in US interpersonal trust research).

Urban% behaves differently by measure: fully mediated by crime in the MoJ model (direct effect NS, ~71% mediated), but retains a large direct effect in the Realdania model (~67% direct) — the “whole-lokalområde” framing picks up atmosphere-of-urbanity signal that crime alone does not capture.

Postnummer-level cross-check

Realdania published safety for the top-10 and bottom-10 postnumre (Table 4.4) — 20 of 603 qualifying postnumre. We joined these with postnummer-level demographic data from two DST snapshots (2014-01-01 and 2016-10-01).

pn_moj <- pn14 |> select(postnr, menapta_proxy_pct, ikke_dansk_pct) |>
  mutate(postnr = as.character(postnr))

j14 <- tryghed_pn |> mutate(postnr = as.character(postnr)) |>
  inner_join(pn_moj, by = "postnr")

tibble(
  predictor = c("Ikke-dansk % (2016)", "Ikke-dansk % (2014)",
                "Asien + Afrika % (2014, MENAPTA-proxy)"),
  n = c(19, 20, 20),
  pearson_r = c(
    cor(pn16 |> mutate(postnr=as.character(postnr)) |>
        inner_join(tryghed_pn |> mutate(postnr=as.character(postnr)), by="postnr") |>
        pull(tryg_lokalomraade),
        pn16 |> mutate(postnr=as.character(postnr)) |>
        inner_join(tryghed_pn |> mutate(postnr=as.character(postnr)), by="postnr") |>
        pull(ikke_dansk_pct)),
    cor(j14$tryg_lokalomraade, j14$ikke_dansk_pct),
    cor(j14$tryg_lokalomraade, j14$menapta_proxy_pct)
  ),
  spearman_rho = c(
    cor(pn16 |> mutate(postnr=as.character(postnr)) |>
        inner_join(tryghed_pn |> mutate(postnr=as.character(postnr)), by="postnr") |>
        pull(tryg_lokalomraade),
        pn16 |> mutate(postnr=as.character(postnr)) |>
        inner_join(tryghed_pn |> mutate(postnr=as.character(postnr)), by="postnr") |>
        pull(ikke_dansk_pct), method = "spearman"),
    cor(j14$tryg_lokalomraade, j14$ikke_dansk_pct, method = "spearman"),
    cor(j14$tryg_lokalomraade, j14$menapta_proxy_pct, method = "spearman")
  )
) |> mutate(across(where(is.numeric), round, 3))

Same sign and similar magnitude at the postnummer level as at the commune level. MENAPTA-proxy (Asien + Afrika in the DST regional grouping) gives the tightest fit, mirroring the commune finding.

Neighbour-postnummer analysis (R/analyse_postnr_neighbors.R) shows that for Skovlunde (postnr 2740) the low-safety reading is consistent with a neighbour-spillover story (adjacent Albertslund 2620 has MENAPTA-proxy 23%); for Frederikssund (3600), neighbours are whiter than the focal postnummer, so the spillover hypothesis fails there and a within-postnummer “concentrated block” explanation (Ådalsparken) is the residual candidate.

Question wordings (English translations)

MoJ (Tryghedsundersøgelsen): “Now a question about fear of crime. On a scale from 1-7 […], how safe or unsafe do you feel? By ‘neighbourhood’ we mean the area immediately surrounding your residence.” Safety % = share answering 1–3.
Realdania (Vores Livskvalitet, Appendix 4A): “Are there places in your local area where you avoid going after nightfall because you feel unsafe?” Safety % = share answering “No, I feel safe in the whole local area.”

The two constructs differ in reference radius (immediate vs whole local area) and response format (7-point scale vs forced-choice behavioural avoidance). This cleanly explains the structural offset between the two measures (MoJ mean 89%, Realdania mean 75%) and why Realdania’s spread is 2.4× MoJ’s.

Caveats and limitations

Temporal mismatch: MoJ is 2021-2024 mean; Realdania is 2024/25 single wave; MENAPTA is 2022Q2; crime is 2021-2024; urbanization is 2024-01-01. We aligned on the midpoint of the 2021-2024 window wherever possible. Commune-level MENAPTA ranks are highly stable year-on-year in DST, so the 2022Q2 snapshot is representative of the full window.
Bornholm: 2019 commune figure uses the Bornholms politikreds value (the politikreds contains only the Bornholm commune); 2021-2023 values come from the report body text rather than the maps (maps don’t include Bornholm).
Causal interpretation: the path model imposes one plausible causal ordering; reverse causation (tryghed → crime reporting behaviour) and omitted-variable bias cannot be ruled out with cross-sectional data.
MENAPTA grouping: a convenience aggregate; the constituent origin groups differ substantially in their tryghed, crime, and socioeconomic profiles within Denmark (e.g. South African vs Somali migrants). Using subcategories would likely improve the model fit but would raise the small-count stability issues we avoided by aggregating.
Realdania’s top/bund-10 postnummer table contains a printing anomaly (ranks 5–7 are in the wrong numerical order). Transcribed verbatim.

Migration and feelings of safety in Denmark

Claude (Opus 4.7, 1M context) — primary author

Emil O. W. Kirkegaard — project direction and data sourcing

18 April, 2026