Methodology
Scope
Suç Endeksi counts homicides and armed assaults in Turkey at province and district level, expresses them relative to population and indexes them to the national average. Because regular, comparable statistics on these incidents are not published at province and district level, the data is compiled from reports in the national and local press.
The count has three units. A record is an incident report extracted from a single news story. A case is the set of records describing the same incident, merged into one, and is the basic unit of the count. A person is a victim killed in a case. Each case is assigned to a year by the date of the incident and to a province and district by the place where it happened; the date and place of publication play no part in the count.
Incidents that are not reported in the press cannot be counted. The published figures are therefore lower than the official ones and are described as documented figures. The size of the undocumented part is modelled in the Estimate section and published separately at national level.
Indicators
| Indicator | Definition |
|---|---|
| Homicide | Number of people killed intentionally that are documented in the press. An offender who takes their own life is not counted. The homicide rate and index are based on this count. |
| Homicide cases | Homicide cases that resulted in death, as documented in the press. Attempts and incidents ending in injury are not counted. |
| Armed assault | Attacks documented in the press with a firearm or a knife or blade that resulted in death or injury. Whether the incident is reported as homicide, injury or robbery does not matter. Suicides, accidents and accidental discharges are not counted. |
The main indicator is victim-based. United Nations Sustainable Development Goal indicator 16.1.1 and the definition of intentional homicide in the International Classification of Crime for Statistical Purposes (ICCS 0101) of the Office on Drugs and Crime both count the person killed, and the homicide rate used in international comparisons rests on that count. The number of cases is given as a second indicator.
A province's documented rate is its documented number of people divided by its population and expressed per 100,000 people.
r_i = 100,000 × y_i / n_i
Here y_i is the number of people (or cases) documented in the province and n_i its population. Population figures come from the Address Based Population Registration System of the Turkish Statistical Institute, for the year of the data. Where district populations for a year are not available as a table, the geometric mean of the district's populations in the preceding and following years is used and the districts are scaled to the official population of the province for that year. If the population of a year has not yet been published, the population of the nearest year is used and the rates are recalculated once it is published.
From records to cases
The same incident gives rise to several news stories. The report of the incident, the follow-up story the next day, the reports of detention and arrest, the autopsy result and the court hearing each produce a separate record. If these records were not reduced to a single case, one closely followed file could by itself change the annual count of a province with a small population and its position in the index.
Records are linked to a case by two criteria. The first is agreement between the names of the victim and the offender read from the text of the reports. Where the names are not given, the second criterion applies, and records written in the same province, in a compatible district, a few days apart, whose shared details show that they describe the same incident, are merged. Reports on trials are not counted as incidents. Traffic and workplace accidents, fires, deaths in security operations, suicides and round-up articles are not counted as homicides.
The number of people killed in a case is read from the text of the reports and reviewed at case level. Only those killed are counted; an offender who takes their own life is not included. Deaths taken from accidents, fires and round-up articles, and incidents abroad, are removed.
The province and district of an incident are read from the text of the reports and reviewed; records in which the name of a month or a surname has been taken for a district name are corrected. Where the records of a case point to different provinces, the province of the case is determined from the text of the reports; where the text is not sufficient, the case is assigned to the province most of its records point to. This usually happens where the offender was caught or the trial was held in another province.
Data review
Before each year's data is published, it is read again against the text of the news reports. The review covers the merging of cases, the type of the incident, its province and district, the offender–victim relationship and the number of people killed. The corrections that come out of the review are held at record level and applied again to the raw data at every build; the raw data itself is not altered. Every published figure is thus produced in a way that shows which correction it results from.
Rates and indices
Each indicator has its own index; no general crime index combining the indicators is calculated. The index divides the province's adjusted rate by the national rate.
E_i = 100 × λ̂_i / λ̄
Here λ̄ is the national rate, that is the total number of incidents divided by the total population, and λ̂_i is the province's rate after the small-number adjustment. The index for Turkey is 100 by definition.
Small-number adjustment. The number of incidents in a province in a single year fluctuates at random. In a province with a small population, two or three incidents more or fewer change the rate noticeably, and that fluctuation carries no information about the level of crime in the province. The index, rank and class are therefore calculated not from the raw rate but from a rate adjusted by empirical Bayes. The method assumes that counts are Poisson distributed and that the true rates of provinces are drawn from a gamma distribution whose mean is the national rate.
y_i ~ Poisson(λ_i n_i), λ_i ~ Gamma(k, k / λ̄)
Under these two assumptions the posterior expectation of the province's rate has a closed form.
λ̂_i = (k + y_i) / (k / λ̄ + n_i)
The result is a weighted average of the province's own rate and the national rate. The weight on the province's own data follows from the ratio of its expected number of incidents to the number of prior incidents.
w_i = λ̄ n_i / (k + λ̄ n_i)
As the number of incidents grows, w_i approaches one and the province's own rate prevails; as it falls, the result is drawn towards the national average. k is the number of prior incidents and amounts to adding to the province an imaginary population living at the national rate together with the k incidents occurring among them.
Estimating the prior. k is not chosen but estimated from the data. The marginal distribution of the Poisson–gamma mixture is negative binomial, and k is found by maximum likelihood on that distribution.
L(k) = ∏_i Γ(y_i + k) / (Γ(k) y_i!) × (k / (k + μ_i))^k × (μ_i / (k + μ_i))^y_i
μ_i = λ̄ n_i, k̂ = argmax L(k)
The more of the difference between provinces comes from chance, the larger k̂ is and the stronger the adjustment. The prior is estimated separately for each indicator and each level; at district level the target it draws towards is not the national rate but the rate of the province the district belongs to. The estimate is confined to a reasonable range so that an unusual distribution in a single year cannot push the adjustment to an extreme. Counts and rates per 100,000 people are given without adjustment; the adjustment applies only to the index, the rank and the class.
Classes. On the maps, provinces and districts are divided into five classes by index value: under 70, 70–90, 90–110, 110–140, and 140 and over. The limits are fixed and remain comparable across years.
Region and population group. The values for regions and population groups on the province pages are calculated from the ratio of the total number of incidents to the total population of the provinces in the group. As the groups are large, no small-number adjustment is applied.
Estimate
Because incidents that are not reported in the press cannot be counted, the documented figure is a part of the true figure. The size of that part is estimated with a province-level model. The model combines three sources: the province counts in this database, the series of convicts received into prison for homicide by the province where the crime was committed, published by the Turkish Statistical Institute, and the official figures published for that year. The last of these consists of the year-end statements of provincial police departments and governor's offices together with the homicide rate series for Turkey published by the United Nations Office on Drugs and Crime.
Turkey's official homicide figure is published on the data portal of this series with a delay of one to two years. For a year whose official figure has been published, the model is not run; the official figure is given as the national total and the documented share is calculated against it.
The series of convicts carries no information about the level. It includes attempts and several offenders for one incident, and convictions become final years after the incident, so its absolute figures do not equal the true ones; in the model it serves only as a proxy for the relative level of provinces.
Observation model. The number documented in a province is the share of its true number of incidents that the press captures.
y_i ~ Poisson(λ_i n_i p_i)
λ_i is the true rate of the province and p_i the coverage of the press there. The prior for λ_i is formed from the series of convicts and the national average and written in gamma form. The documented count updates that prior through a Poisson likelihood; thanks to conjugacy the posterior is again a gamma distribution.
λ_i | y_i ~ Gamma(k_i + y_i, k_i / m_i + n_i p_i)
m_i the rate expected by the prior, k_i the strength of the prior
Coverage. Coverage is written as a function of population, because incidents are more likely to be reported in provinces with large populations.
log p_i = log c + β (log n_i − log n̄) + u_i, u_i ~ N(−σ_u² / 2, σ_u²)
The slope β is obtained in two independent ways: by comparing the coverage of the provinces whose official figures have been published, and from the way the ratio of the documented count to the series of convicts varies with population. The constant c, which sets the level of coverage, is found by equating the total corrected for provincial coverage to the national total.
Σ_i y_i / p_i = N
National total. N is the inverse-variance weighted mean of two estimates. The first divides the official figures of the provinces that published them by their shares in the model; the second is the average of the last years of the national rate series.
N̂ = (N̂₁ / s₁² + N̂₂ / s₂²) / (1 / s₁² + 1 / s₂²)
The undocumented part. The true figure is the documented count plus the expected number of those not documented. The thinning property of the Poisson distribution gives the conditional distribution of the undocumented part directly.
N_i = y_i + Poisson(λ_i n_i (1 − p_i))
By construction the estimate cannot fall below the documented count, because the documented incidents did happen. The ranges come from repeated simulations in which all sources of uncertainty are sampled together, and they contain the estimate with 90 per cent probability. The sampled sources are the national total, coverage, the true rate of the province and the fluctuation of the count.
No estimate is published at province level. The variability of documented provincial rates separates into three components: the true differences between provinces, differences in coverage, and small-number noise.
Var(log r_i) ≈ σ_λ² + σ_u² + E[1 / y_i]
The third term dominates. In most provinces the annual number of incidents is not large enough to separate the true difference from the noise, and the estimate is therefore published at national level and not at province level. The documented and estimated figures for the year are on the homicides in Turkey page, and how to read the rates and the index is explained on how to read crime data.
Change over time
Whether homicide or armed assault is rising or falling in a place is assessed by comparing the rates of the two most recent years. Because the comparison uses rates, changes in population are taken into account.
R = (y₂ / n₂) / (y₁ / n₁)
Here y₁ and y₂ are the documented counts of the two years and n₁ and n₂ the populations of those years. The change is written as R − 1.
Uncertainty. The count observed in a year fluctuates at random even when conditions stay the same. Given the total of the two years, the count of the second year is binomially distributed; the 95 per cent interval of the rate ratio is obtained from the exact interval of this binomial proportion.
y₂ | y₁ + y₂ ~ Binom(y₁ + y₂, π), π = R n₂ / (R n₂ + n₁)
This applies directly to the counts of homicide cases and armed assaults. The number of people killed, however, is clustered in cases. Because cases with more than one victim make the count fluctuate more than a Poisson distribution allows, the variance of the number of people killed is calculated as the sum of the squared numbers killed in each case. The variances of the two years are added and divided by the sum of the counts to obtain a common dispersion factor. Each year's count is divided by this factor to give an effective count, and the interval is built on the effective counts. Because the factor is common to both years, the ratio of the effective counts equals the ratio of the counts.
V_t = Σ_j m_j², φ = (V₁ + V₂) / (y₁ + y₂), ỹ_t = y_t / φ
m_j is the number of people killed in case j
Shrinkage at province level. Annual counts are small in provinces. When 81 provinces are compared together, the change in some of them looks large by chance alone. The change in a province is therefore drawn towards the change across Turkey, in the same way as the small-number adjustment of the index.
θ̂_i = w_i d_i + (1 − w_i) μ, w_i = τ² / (τ² + s_i²)
d_i is the log change in the province, s_i² the random variance of that change, μ the log change across Turkey and τ² the spread of the true changes of provinces around this common change. τ is estimated from the part of the spread of provincial changes that exceeds what chance would produce. The estimate is not fixed at a single value. τ is given a uniform prior, its posterior distribution is calculated from the provincial changes, and the change in a province and its interval are obtained over all values of this distribution. The uncertainty in τ is thus carried into the interval of each province. The uncertainty of the change across Turkey is added to a province's interval in proportion to its shrinkage. In provinces with large counts w_i approaches one and the province's own change prevails; in provinces with small counts the change approaches that of Turkey. No shrinkage is applied to regions or to Turkey.
Classification. If the 95 per cent interval of the change does not include zero, the change is classified as a clear increase or a clear decrease; otherwise as no clear change.
Annual trend. When there are three or more data years, the average annual trend across all years is given in addition to the change between the two most recent years. The trend is the weighted least-squares slope of the logarithm of the rate on the year.
log(y_t / n_t) = a + b t, annual change = e^b − 1
Each year is weighted in inverse proportion to the random variance of its count; for the number of people killed this variance is calculated with V. If the years scatter around the line more widely than chance can explain, the standard error of the slope is enlarged in the same proportion (overdispersion); a narrower scatter does not reduce it. Provincial slopes are drawn towards the slope for Turkey, as with the two-year change, and classified by the same rule.
Totals and official figures. For Turkey, the total number of people killed is also compared as the official figure or the coverage-corrected estimate. The uncertainty of the estimate comes from the model's simulations, that of the official figure from annual Poisson fluctuation. Where a province's official figures have been published for both years, the change in the official figures is given with the same method.
Consistency between years. Whether the index of a province stays the same from one year to the next is shown by two measures: the rank correlation (Spearman) of the provincial indices of the two years, and the transition table of the map classes of provinces and districts between the two years. When counts are small these measures come out low even if nothing has really changed. Each measure is therefore compared with the value expected if the true rate of no unit had changed between the two years.
The expected value is found by simulation. The true rates of the units are drawn from a gamma distribution that explains the counts of both years together; its shape parameter is estimated from the joint likelihood of the two years. The count for each year is generated from these rates with a Poisson distribution. For the number of people killed, the number of cases is Poisson and the number killed per case is drawn from the observed distribution of the two years. From the generated counts the index and class are recomputed with the calculation of the published index: the prior is estimated from the data of each year and the empirical Bayes adjustment is applied. The simulation is repeated 1,000 times for provinces and 500 times for districts; the expected value and its 95 per cent interval are taken from this distribution. If the observed value lies inside the interval, the data cannot separate a year-specific shift from chance.
For districts the reference rate is the rate of the province, and only the aggregate transition table is published; district-level classes belong to the registered layer. The proportional distribution of incidents without a recorded district over the districts, and differences in the district matching rate between years, are not part of the simulation; shifts in district classes can therefore come out larger than the simulation shows.
12-month rolling total. The course within the year is shown by the total of the previous 12 months at the end of each month. The first value is the total of the earlier year and the last value that of the latest year; the values in between show in which months the change accumulated.
S_t = Σ_{k=0}^{11} y_{t−k}
A change in the documented count carries changes in press coverage as well as changes in incidents. If the news sources monitored, or the attention the press gives to a province, change from one year to the next, the documented count can rise or fall without any change in the number of incidents. Where an official figure exists, the change is read together with it.
Subgroups
Two subgroups are reported as separate sections on the homicide page. The subgroups count only homicide cases that resulted in death; they do not enter the index or the ranking, and no rates are calculated for them. The definitions follow the homicide statistics classification of the Office on Drugs and Crime and the statistical framework for measuring femicide published in 2022 by the same office together with UN Women.
| Subgroup | Definition |
|---|---|
| Intimate partner and family | Cases in which the offender was the victim's spouse, former spouse, partner or a family member. |
| Femicide | Women and girls killed by a spouse, former spouse, partner or family member. The sex of the offender is not a criterion. |
Femicides are the part of intimate partner and family homicides in which the victim is a woman or a girl. The framework counts women killed by an acquaintance or a stranger as femicides only where a gender-related motive is established; news reports are not sufficient to establish such a motive reliably, so those cases do not enter the subgroup. As the offender–victim relationship is not reported in every story, the subgroup counts are minimum figures. The femicide map shows the number of cases per province, with the area of each circle proportional to the number of cases.
The official counterpart of the subgroup is the set of femicides that the police and the gendarmerie classify under Law No. 6284 on the protection of the family and the prevention of violence against women. In the 2016–2018 data published from these records by the Police Academy, the offender was the victim's spouse or partner in 63.5 per cent of cases and a relative in 32 per cent. The official definition therefore largely coincides with that of the subgroup, and the subgroup is compared with this series. The comparison shows two differences. Femicides are reported in the press more often than homicides in general, so the coverage of the subgroup is above the national coverage of homicide. On the other hand, cases in which the offender was a relative make up a smaller share of the press data than of the official records, so family femicides are less complete in the subgroup.
The sex of the offender is given as a breakdown for femicides, with the cases where it cannot be determined from the text shown separately. No indicator resting on the sex or age of the offender is published. Official figures on children in the justice process are published by the Turkish Statistical Institute in its statistics on children received into or brought to security units.
District level
The district index is calculated by drawing the district's rate towards the rate of the province it belongs to. The form of the adjustment is the same as at province level, and the prior is estimated separately for the district level. Districts with small populations are hatched on the map. Counts and rates are not published at district level; index classes and ranks within the province are available to registered users.
Other incident types
Intentional injury, robbery, theft, sexual assault, drug offences, fraud and smuggling are not published on the site as counts, rates, indices or rankings. For these types, incidents reported in the press are a small share of official records and are unevenly distributed across provinces, so presenting them as rates would imply that they are representative.
Incidents of these types are indexed in the institutional API as incident records at province, district and neighbourhood level, without rates. The legal definitions of these types and the ways to access the data are on the other offences page. Suicides and accidents appear in no layer.
Limitations
Homicides that are not reported in the press are not counted, and coverage rises in provinces with a strong local press, where incidents are more likely to be reported. The offender–victim relationship, the method and the characteristics of the victim are not reported in every story, and attacks for which the weapon is not specified cannot be counted as armed assaults. Coverage of armed assaults is higher in large provinces than in small ones.
The figures for each year are produced with the same definitions. Because press coverage and the news sources monitored can change from year to year, the difference between two years does not show only the change in incidents. The way the text of news reports is read, and how sensitively reports of the same incident are merged into one case, may also differ slightly from year to year; this adds uncertainty to the change between years on top of the difference in coverage. Monthly series are based on the date of the incident and fluctuate from month to month. Because the number of articles downloaded for February 2023 was about a third of that for February 2024 (2,088 against 6,621; the 6 February earthquakes may have filled the news flow), the incident count for that month is incomplete and the dip in the monthly series does not show that incidents declined. District classes are based on a single year of data and may change from year to year in districts with small populations. The estimation model rests on the assumption that the series of convicts reflects the relative level of provinces; where conviction rates differ between provinces, that assumption carries a bias into the estimate. The national rate series fluctuates between years, and before the latest year is published the estimate is carried forward from the most recent years of the series.
Sources
Population, and neighbourhood and village names: TurkStat Address Based Population Registration System. Province and district boundaries: geoBoundaries, OpenStreetMap contributors (ODbL). Legal texts: mevzuat.gov.tr. Sources of the estimate: TurkStat prison statistics (convicts received into prison for homicide by the province where the crime was committed), the homicide rate series of the United Nations Office on Drugs and Crime, and statements by provincial police departments and governor's offices. Femicide comparison: Taştan, C. and Küçüker Yıldız, A. (2019), Dünyada ve Türkiye'de kadın cinayetleri: 2016-2017-2018 verileri ve analizler [Femicides in the world and in Turkey: data and analyses for 2016–2018], Police Academy.