Statistical Methods in Crime Research

Criminological research asks two broad kinds of question, and statistics serves each differently. The first is descriptive: how much crime is there, of what kind, where, and how has it changed.

The second is explanatory: does a given factor, such as unemployment, policing intensity or a new law, raise or lower the likelihood of offending or victimisation. Descriptive statistics, the subject of the next section, answer the first kind of question. Tests of association, regression and time-series methods answer the second.

A statistical claim in criminology is only as strong as the research design that produced the data feeding it. A regression model run on cross-sectional survey data can show that two variables are associated, but association alone does not establish which variable came first, or whether a third, unmeasured factor produced both.

Sound criminological statistics therefore pairs the choice of test to the structure of the data (categorical, continuous, or a count; cross-sectional or longitudinal) with an honest account of what the design can and cannot support.

Statistics in this field also has a policy audience. A recognised crime survey, a court's sentencing register, or a police force's recorded-crime file feeds into decisions about resourcing and law reform, so an analyst has an obligation to state uncertainty (a confidence interval, a sample size, a margin of error) rather than presenting a single number as if it were exact.

The sections that follow work through the specific tools, from a simple rate to a synthesis across many studies, and the errors that recur when those obligations are skipped.

The level at which data is collected shapes which tool is available: a per-offender dataset supports individual-level modelling, while an aggregate dataset, a monthly count per city, supports only area-level or time-series analysis, and treating the two as interchangeable is the ecological fallacy examined in section 6.

Every method below also assumes the sample was drawn with enough care that its patterns are not simply an artefact of who happened to be counted, an issue covered in Sampling and Fieldwork with Hard-to-Reach Populations.

The most basic descriptive tool is the crime rate: the count of a given offence divided by the population at risk over the same period, conventionally expressed per 100,000 people.

A raw count is close to meaningless for comparison. A city of ten million people will record more burglaries in absolute terms than a town of fifty thousand even if the town is, per resident, far more burglary-prone. Standardising by population lets researchers compare Delhi with London, or this year with a decade ago, on a common footing.

Agencies that compile official statistics publish rates for exactly this reason. India's National Crime Records Bureau issues its annual Crime in India report with state and city rates per 100,000 population. In the United States, the Federal Bureau of Investigation's Uniform Crime Reporting programme, alongside the newer National Incident-Based Reporting System, does the same at the national, state and agency level.

England and Wales draw on two parallel sources, police-recorded crime and the Crime Survey for England and Wales, precisely because each captures a different slice of the true offending picture, as discussed further in the linked topic below.

Beyond a simple rate, researchers use ratios (such as the ratio of male to female arrestees for a given offence, or a clearance ratio comparing offences solved to offences recorded) and trend indices that rebase a rate to 100 in a chosen base year so that later years can be read as a percentage change.

A trend index is useful precisely because it strips out the scale of the original numbers and shows direction and pace of change, which is what a policymaker usually wants to know first.

All of these descriptive measures inherit the limitations of the underlying count. A rise in a recorded rate can reflect more offending, but it can equally reflect a change in reporting behaviour, a change in police recording practice, or a change in the legal definition of the offence.

This is the dark figure of crime, the gap between offences that occur and offences that enter an official count, and it is the first thing a careful reader checks before treating a movement in a published rate as a movement in real offending.

A related descriptive measure is the age-standardised or population-adjusted rate, which is used when the group being compared has an unusual age structure, since offending is not spread evenly across age groups.

Comparing raw offending rates between a young population and an older one without this adjustment can make one look more crime-prone than a fair, like-for-like comparison would show, which is why demographers and criminologists borrow standardisation techniques originally developed for comparing mortality rates across populations with different age profiles.

Once two variables are categorical, such as offence type and area type, the natural first step is a cross-tabulation, and Pearson's chi-square test checks whether the categories are associated more than chance would predict. Building that table and judging how large a sample needs to be before a claim is trustworthy are inference fundamentals shared across every quantitative forensic discipline, not something specific to crime counts.

For the mechanics, worked construction and the common misreadings of a confidence interval, see Confidence Intervals and What a Sample Can Support, and for how a sample must be drawn before any of these tests mean anything, see Sampling Strategies and Representative Data.

What is specific to crime research is picking the regression family once the design moves beyond a single two-way table. A continuous outcome, such as a sentence length in months, calls for ordinary linear regression.

A binary outcome, such as whether a released prisoner is reconvicted within two years, calls for logistic regression, whose coefficients are read as odds ratios rather than as a raw change in the outcome. Neither choice is a matter of preference: it follows mechanically from what the outcome variable is.

A crime count, such as the number of burglaries recorded in a police beat in a month, fits neither family well, and needs its own regression tools. That is the subject of the next section.

Many crime outcomes arrive as counts rather than as a continuous score or a yes/no result: burglaries per ward per month, assaults per precinct per year, or drug seizures per port per quarter.

Poisson regression is the starting model for this kind of outcome. It assumes the counts follow a Poisson distribution, whose defining property is that the variance of the counts equals their mean, and it estimates how a one-unit change in a predictor multiplies the expected count rather than adding to it, so coefficients are read as rate ratios once exponentiated.

Crime counts are rarely collected over identical windows: a ward of 40,000 residents will produce a larger raw burglary count than a ward of 4,000 even at the same underlying rate. Poisson and negative binomial models handle this with an offset, a term (commonly the log of the population) added to the model with its coefficient fixed at 1, which converts the model's predictions into a rate per unit of exposure without pre-dividing the outcome.

The Poisson assumption that variance equals the mean is frequently wrong for crime data, because some wards or months have far more variable counts than a Poisson model expects, a pattern called overdispersion. A simple diagnostic is to compare the model's residual deviance to its residual degrees of freedom: a ratio well above 1 signals overdispersion.

When it is present, criminologists switch to negative binomial regression, which adds a dispersion parameter that absorbs the extra variance and produces wider, more honest standard errors than a Poisson model forced onto data it does not fit. A further extension, the zero-inflated model, is used when an outcome has more zero-count periods than either family predicts on its own, such as a rural precinct that records no armed robbery in most months.

Choosing between these count models is a domain judgement about how crime data is actually generated, unlike the earlier choice between linear and logistic regression, which follows straight from whether the outcome is continuous or binary. See Common Distributions in Forensic Science for how the Poisson distribution itself is derived and used elsewhere in forensic work.

A time series is a sequence of measurements of the same variable, such as a monthly robbery count for one city, taken at regular intervals over time. Time-series analysis studies how such a sequence moves: whether it trends upward or downward, whether it has a recurring seasonal pattern (many property crimes rise around the winter holiday period in several countries), and how much of this month's value can be predicted from recent past values.

Models in the ARIMA family (autoregressive integrated moving average) are the standard tool for capturing these dependencies and for producing a short-term forecast of where a trend is heading if nothing changes.

Criminologists most often reach for time-series analysis to evaluate an intervention: did a new policing strategy, a sentencing reform, or a change in a specific law actually shift the trend in the outcome it targeted.

The interrupted time-series design compares the trend before an identified intervention date against the trend after it, looking for a change in level (an immediate jump or drop right at the intervention point) and a change in slope (the trend afterward rising or falling at a different pace than before). Both changes can be tested statistically against the pre-intervention pattern extrapolated forward.

The design's strength is that it uses the same city, and often the same recording practices, as its own comparison, which removes many of the confounds that plague a cross-sectional comparison between two different places.

Its weakness is that any other event coinciding with the intervention date, such as an unrelated economic shock or a separate policy introduced in the same period, will be mistaken for an effect of the studied intervention unless the researcher can rule it out, ideally by comparing against a similar area that did not receive the intervention (a comparison series) alongside the interrupted series for the treated area.

Seasonality is a further trap specific to time series. A raw month-to-month comparison in property crime can look like a real short-term surge when it is simply the recurring winter pattern reappearing on schedule.

Analysts either compare each month against the same month a year earlier, or apply a seasonal-adjustment procedure before looking for a genuine departure from the expected seasonal pattern, so that a routine seasonal peak is not mistaken for evidence that an intervention has failed or that a new crime wave has begun.

A further consideration is how much data precedes and follows the intervention date. A series with only a handful of pre-intervention months gives too few points to establish what the trend was really doing before the change. Criminologists generally prefer several years of data on each side of the intervention date, which lets the model separate a genuine step change from ordinary month-to-month noise.

A single study, however well designed, is drawn from one sample, one place and one period, and its estimate of an effect carries sampling error. Meta-analysis addresses this by combining the results of many separate studies of the same question into a single, more precise estimate, typically weighting each study's contribution by its sample size or the precision of its own estimate, a synthesis logic related to the pooling of evidence described in Mixed Methods Research in Crime and Justice.

The term itself was introduced by the statistician Gene Glass in 1976 to describe this statistical synthesis of findings across studies, and the technique has since become a standard way to summarise, for example, whether cognitive-behavioural programmes reduce reoffending across many separate evaluation studies rather than relying on any one of them alone.

Meta-analysis brings its own hazard, publication bias: studies that find a significant effect are more likely to be published than studies that find none, so a synthesis built only from published work can overstate a true effect.

Meta-analysts check for this with a funnel plot, which should show a roughly symmetric spread of effect sizes around the pooled estimate if publication bias is absent, and an asymmetric funnel is a warning sign that null results have gone missing from the literature.

Two further errors recur across the whole toolkit covered in this topic, and both are worth naming precisely. The first is the ecological fallacy, identified by the sociologist William S. Robinson in 1950: a correlation calculated from group-level data, such as a strong positive correlation between the percentage of foreign-born residents and the aggregate crime rate across cities, does not establish that foreign-born individuals within those cities are themselves more likely to offend.

The relationship observed across aggregated units can reverse, weaken, or vanish entirely at the individual level, so a claim about individuals always needs individual-level data to support it, not an inference borrowed from a group-level table.

The second is spurious correlation: two variables can move together closely over time or across places purely by coincidence, or because a third, unmeasured variable drives both, without either causing the other.

A criminologist should always ask what a proposed third factor might be, whether the data allow it to be measured and controlled for in a regression, and whether the timing of the two series is consistent with the causal story being told, before reporting a correlation as evidence of a causal relationship between them.

Both share a common remedy: matching the level of the data to the level of the claim, and treating a correlation as the start of an investigation rather than its conclusion.