Main Steps in the Criminological Research Process

Every study begins with a problem, not a method. A researcher usually starts with a broad area of interest, such as domestic burglary, youth gang membership, or sentencing disparity, and has to narrow it into a question specific enough that a feasible study can answer it.

A workable research problem states the population, the variables of interest and, where relevant, the setting: not "why does crime happen" but, for example, "does completing a prison education programme reduce reconviction within three years among adult male offenders released from a given facility".

Narrowing the problem is also a feasibility exercise. A student or agency researcher has to ask whether the needed records are accessible, whether victims or offenders can realistically be reached for interview, and whether the time and budget available match the scale of the question. A problem that requires ten years of follow-up data is not wrong in principle, but it is the wrong choice for a project with a six-month deadline.

The literature review comes next, and it is not a summary exercise tacked onto the introduction; it actively shapes the study. Reading prior work on the topic shows which theories have already been tested, which measures other researchers used and how well they performed, and which findings are contested or unreplicated.

A well-known example of how existing data can itself become a full study is Emile Durkheim's 1897 study Suicide, which drew on official mortality statistics across regions and religious groups rather than collecting new data, and used the review of prior explanations to justify a sociological rather than purely psychological account.

A thorough review also surfaces gaps: a variable nobody has measured, a population nobody has sampled, or a period, such as the years after a major legal reform, where data has not yet caught up. Locating a genuine gap is what turns a topic into a contribution rather than a repetition of settled findings.

A useful review is organised thematically rather than as a list of one-line summaries: grouping studies by the mechanism they propose, the population they studied, or the method they used makes it easier to see where findings converge and where they contradict each other.

Researchers also weigh the strength of the underlying evidence as they read, noting which claims rest on a single small study and which have been replicated across several samples, since a hypothesis built on an unreplicated finding is a weaker starting point than one built on a pattern that recurs across settings and countries.

Feasibility and ethics are also checked at this stage, well before any instrument is drafted. A researcher weighs whether the population of interest, such as serving prisoners or trafficking survivors, can be approached at all without disproportionate risk to them, and whether the institution holding the relevant records, such as a police service or a court registry, is likely to grant access within the project timeline. A problem statement that looks intellectually interesting but depends on cooperation nobody has actually secured is not yet a workable research problem.

Once the literature review has clarified what is known, the researcher states what the study will test. A hypothesis is a specific, testable prediction about a relationship between variables, for example that unemployment is positively associated with property crime rates at the neighbourhood level.

Karl Popper's 1959 The Logic of Scientific Discovery argued that a hypothesis is only scientifically useful if it is falsifiable, meaning there is some possible observation that would count against it. A prediction phrased so broadly that no data could ever contradict it does not function as a working hypothesis.

Exploratory studies, by contrast, often start with a research question rather than a hypothesis, particularly when little prior work exists. Instead of predicting a direction, the researcher asks something open, such as "how do first-time offenders describe their decision to stop offending", and lets the design stay flexible enough to follow what the data show.

The research design is the overall plan for answering the question, and the choice depends heavily on whether the question is about prevalence, cause, process or lived experience. A cross-sectional survey captures a population at one point in time and suits prevalence questions.

A longitudinal design follows the same subjects over years and is the only design that can properly test claims about change or developmental pathways: Marvin Wolfgang's 1972 Philadelphia birth-cohort study, published as Delinquency in a Birth Cohort, tracked all boys born in Philadelphia in 1945 from their tenth to their eighteenth birthday and remains a reference point for cohort methodology in criminology.

Whichever design is chosen, each abstract concept in the hypothesis has to be operationalised into something measurable. "Reoffending" might be operationalised as a new conviction within a fixed follow-up window, or as a new arrest, or as self-reported offending, and each choice yields a different number from the same underlying population, so the operational definition has to be stated and defended before data collection starts.

A directional hypothesis states not just that a relationship exists but which way it runs, for example that programme completion lowers reconviction rather than merely that the two are related, and a directional prediction is held to a stricter statistical test than a non-directional one.

Choosing between an experimental, quasi-experimental, cross-sectional or case-study design at this stage is rarely about which design is objectively best; it is about which design can answer the specific question given the ethical constraints, the population's size, and the time available before the study has to report. This trade-off is a design decision, not an afterthought. Reviewers checking a published study routinely ask which of these designs was used and why.

Very few criminological studies can examine an entire population, so the design has to specify how a manageable subset will represent it. The first decision is the sampling frame: the actual list or record system from which cases will be drawn, such as a court's case-management database, a prison's release register, or a household address list for a victimisation survey.

A frame that omits part of the population, for example a court database that only covers cases that reached trial and excludes those settled by plea, will bias every conclusion drawn from it regardless of how careful the later analysis is.

Probability sampling, where every unit in the frame has a known chance of selection, supports statistical generalisation to the wider population; simple random sampling, stratified sampling by region or offence type, and cluster sampling by police station or prison wing are the common variants.

Non-probability sampling, such as purposive sampling of offenders who match a specific profile, or snowball sampling used to reach hidden populations like active drug users, trades generalisability for access, and is often the only realistic option for populations that will not appear in any official list.

Large national victimisation surveys illustrate how sampling choices shape the resulting picture of crime. The National Crime Victimization Survey in the United States and the Crime Survey for England and Wales both use large household samples designed to capture crimes that victims never reported to police, which is why their estimates of victimisation regularly exceed the police-recorded totals for the same offence categories.

Sample size decisions balance statistical power against cost and access. A study aiming to detect a small effect, such as a modest reduction in reoffending from a low-intensity intervention, needs a substantially larger sample than one looking for a large, obvious effect, and researchers typically calculate the minimum sample size before fieldwork begins rather than collecting data and hoping it is enough.

Non-response is a further threat to any sampling plan: even a well-designed probability sample can end up unrepresentative if the people who decline to participate differ systematically from those who take part, for example if victims of more serious offences are less willing to be interviewed than victims of minor theft.

Comparing basic characteristics of respondents against known population figures, where such figures exist, is a standard check for whether non-response has introduced this kind of bias. A frame drawn up years earlier can also go stale, so it is worth checking whether the underlying records still reflect current administrative boundaries before drawing the sample from them.

The instrument is whatever tool will actually gather the data: a structured survey questionnaire, a semi-structured interview schedule, a coding sheet for extracting variables from case files, or an observation protocol for time spent in a courtroom or custody setting. Each item on the instrument has to trace back to a specific variable in the research design; questions included out of general curiosity rather than analytic need waste respondent time and dilute the instrument's focus.

Before an instrument goes anywhere near the full sample, it is tested in a pilot study on a small group similar to the intended participants. A pilot checks whether questions are understood as intended, whether response categories cover the real range of answers, how long the instrument actually takes to complete, and whether any coding rules for open-ended answers need revision.

Travis Hirschi's self-report delinquency methodology, developed for his 1969 study Causes of Delinquency, helped establish self-report surveys as a way to measure offending that never reaches police records, but that method only works if the wording of each item is carefully piloted so respondents do not under-report sensitive behaviour or misread a question.

When a pilot study reveals a serious problem, such as a question that most respondents interpret in two different ways, the researcher revises the instrument and, where the problem was substantial, pilots it again before committing to full data collection. Treating a failed pilot as a normal part of the process, rather than a setback to route around, is what keeps a flawed instrument out of the field.

Ethics review typically happens alongside instrument design, since a reviewing committee needs to see the actual questions and procedures, not just a project summary. In the United States, federally funded human-subjects research is reviewed under the Common Rule by an Institutional Review Board, applying the informed-consent and risk-minimisation principles set out in the 1979 Belmont Report.

In the United Kingdom, university and funder-level research ethics committees apply frameworks such as the Economic and Social Research Council's ethics guidance, and Indian universities and the Indian Council of Social Science Research similarly require ethics clearance for studies involving human participants before fieldwork can begin.

Instrument design also has to anticipate how the answers will eventually be analysed. A question with an open-ended free-text answer generates rich detail but takes far longer to code than a fixed-category question, and building the coding scheme after the fact, rather than while the instrument is being drafted, often reveals that a key distinction was never actually asked about. A short coding manual, agreed before fieldwork starts, keeps this consistent.

Data collection in criminology splits broadly into primary collection, where the researcher gathers new data directly, and secondary collection, where the researcher extracts data from records already held by an agency. Primary methods include structured interviews, self-administered surveys, and direct observation in a courtroom, police station or correctional setting; secondary methods include coding variables out of court files, prison records or police incident logs.

Fieldwork brings practical problems that a desk-based design cannot fully anticipate: gatekeepers at a prison or police service may restrict access, victims may decline to be interviewed about a traumatic event, and offender samples are often the hardest to retain over a longitudinal follow-up because participants move, reoffend and are reincarcerated, or simply stop responding.

Building a realistic attrition allowance into the original sample size, and keeping detailed contact records for a longitudinal study, is standard practice for exactly this reason.

When the data comes from official records rather than direct contact with people, the researcher still has to document the extraction procedure precisely: which fields were coded, how missing or ambiguous entries were handled, and whether two coders cross-checked a sample of records for consistency. Inter-coder reliability checks matter here because two research assistants reading the same case file can code an ambiguous offence category differently unless the coding manual is explicit.

Informed consent, confidentiality and, where interviews touch on victimisation or offending, a clear referral pathway to support services are standard safeguards carried over from the ethics review stage into the actual data collection, not separate obligations that end once approval is granted.

Some designs combine both approaches in a single study, collecting a quantitative extract from case files alongside a smaller set of qualitative interviews with practitioners, so that the numeric pattern and the practitioner's account of why it arises can be checked against each other.

This kind of mixed-methods triangulation is particularly common in evaluations of a specific intervention, where the file data shows whether an outcome changed and the interviews explain how staff and participants experienced the programme that produced it.

The researcher's own presence can also affect the data. An interviewer's tone, an observer's visible note-taking in a courtroom, or even the fact that a survey is sponsored by the police can change how honestly a respondent answers, an effect generally described as reactivity.

Standardising how each interview or observation session is conducted, and where possible using the same trained researchers throughout, reduces this source of inconsistency across a large fieldwork exercise. A brief written protocol makes this practical rather than aspirational.

Once collection ends, raw data is cleaned, checked for errors and coded into a form the chosen method of analysis can use. Quantitative data typically moves into statistical software for descriptive summaries and inferential tests, such as comparing reconviction rates between a treatment and a comparison group; qualitative data, such as interview transcripts, is usually coded thematically to identify recurring patterns in how participants describe their experiences.

Interpretation is where the analysis is related back to the original hypothesis or research question, and it is also where a researcher has to be explicit about limitations: a small or non-random sample, missing data, or a design that can show association but not causation. Overstating what a correlational design can support, for example treating a link between two variables as proof that one causes the other, is one of the most common weaknesses reviewers flag in criminological research.

Dissemination is the final step and is not optional. Findings are written up as a report, journal article, thesis or conference paper and, for most academic outlets, go through peer review, in which other researchers assess the design, analysis and conclusions before publication. Agencies that supplied data or access, such as a police service or prison, are usually given a findings summary as part of the access agreement, which is also how research findings feed back into practice.

In India, the National Crime Records Bureau publishes an annual "Crime in India" report drawing on police-recorded data supplied by every state, and independent criminological studies regularly use this published series as a secondary data source or as a benchmark against their own primary findings, which is one concrete way dissemination in one study becomes raw material for the next.

Some journals and funders now ask researchers to register a study's hypotheses and planned analysis before data collection begins, a practice intended to stop a hypothesis being quietly rewritten after the results are already known. Whether or not a given study is formally registered, keeping the original hypothesis, the final analysis and any post hoc exploration clearly separated in the write-up is what allows a reader to judge how much confidence the findings deserve.

A results section usually reports effect sizes and confidence intervals alongside any statistical significance test, since a result can be statistically significant in a very large sample while the underlying effect is too small to matter for policy, and reporting only a p-value hides that distinction from the reader. Readers of a published study should expect both figures to be reported together. Omitting either number invites a misleading headline finding.