How Police Leaders Can Read Data Better

Police data literacy is not about turning leaders into statisticians. It is about making them difficult to fool with a number.

A performance figure can be calculated correctly and still answer the wrong question. Professional experience can spot that something looks odd. It can also mislead you.

So when a number feels wrong, do not choose between the dashboard and your instinct.

Test both.

Start with the number: how was it made?

Then turn the same scepticism on yourself: what evidence would show that my instinct is wrong?

That is the habit this article is arguing for. Professional instinct should start the enquiry. It should not finish it.

Treat every number as a claim

The phrase “the data says” gives a number more authority than it deserves.

Data does not speak.

People decide what to collect, what to exclude, how to classify it, what to divide it by, which period to compare it with and how to present the result.

A police dashboard sits at the end of that chain.

A burglary figure is not burglary itself. It is the number of incidents that were reported, recognised as potential crimes, recorded, classified and retained in the system under the rules operating at the time.

A missing-person figure may count reports rather than unique people.

A use-of-force total may count incidents, records or individual tactics.

A stop-and-search total counts searches, not unique people.

This does not make the data bad. It tells you what kind of question to ask.

The Office for Statistics Regulation describes the principle as intelligent transparency: users should be able to understand where a figure came from, how it was produced, the definitions used and the important limitations. Office for Statistics Regulation.

A leader should expect the same standard internally.

Use instinct to decide what to test

Police leaders build pattern recognition.

You know what a normal Friday night looks like. You know when one prolific offender is distorting demand. You know that a town-centre population at 23:00 looks nothing like the census population that lives there. You notice when a result does not fit.

That experience is useful because it tells you where to look.

It becomes dangerous when it tells you the answer before you have looked.

The better move is to turn the instinct into a proposition that can be tested.

If robbery appears to have fallen 35% and your reaction is “that cannot be right”, there are several plausible explanations:

• the data is wrong;
• the recording process changed;
• the comparison period is unusual;
• your experience is concentrated in one part of the force;
• robbery genuinely fell.

The professional response is not to choose your preferred explanation.

Ask what evidence would distinguish between them.

If robbery genuinely fell, would calls for service, ambulance data, victim reports or neighbouring areas move in the same direction?

If recording changed, does the break in the series line up with a policy, system or training change?

If your instinct is based on a particular locality or shift pattern, is the force-wide result being driven by somewhere else?

That is how experienced judgement becomes analytical value.

Rebuild the number before you debate the story

BBC Radio 4’s More or Less has spent years making a useful habit look deceptively simple: ask where the number came from, what it means, whether the scale is large or small, and which definition is being used. Open University interview with Tim Harford.

For police leaders, I would make the sequence explicit.

Before accepting an important number, ask:

What is being counted?
People, incidents, crimes, calls, searches, arrests, tactics, outcomes or records?

Where did it come from?
A live operational system, a manually maintained spreadsheet, a national return, a survey or an analyst’s extract?

Who or what is missing?
Missing ethnicity, unreported crime, incomplete records, non-responders, victims who disengaged?

What was estimated?
Weighted, imputed, projected, modelled, rounded or carried-forward values?

What is the denominator?
Residents, officers, calls, crimes, victims, hours, households or something else?

Compared with what?
Last month, last year, another force, a matched area, a target, an expected value or a control group?

What changed in the measurement process?
Recording rules, system migration, training, reporting access, staffing or classification?

These are not technical questions for an analyst to worry about later.

They decide whether the headline is capable of supporting the decision being asked of you.

What police leaders can learn from the ONS productivity rethink

The most useful part of the recent ONS productivity story is not that statisticians found an error.

It is that they challenged a familiar method when the result stopped making enough sense.

UK labour productivity is commonly expressed as economic output divided by labour input. For output per hour, that means the denominator matters enormously: if estimated hours worked are too high, measured productivity is pushed down.

The existing approach relied heavily on Labour Force Survey estimates of hours. Where a household could not be contacted at a later survey wave, a previous response could be rolled forward once. That is not an absurd rule. It is a practical response to missing data.

But as response rates fell, the weakness in the assumption became more important. The previous response could describe a normal working week even when the person was now on annual leave or otherwise working fewer hours. ONS concluded that this contributed to an upward bias in measured hours worked. ONS, component approach to measuring labour productivity.

ONS did not simply defend the old series because it was established.

It rebuilt the measure.

The new component approach uses different sources for the parts they measure best, including employer data for usual paid hours, then makes explicit adjustments for annual leave, sickness, bank holidays, overtime and other reasons actual hours differ from usual hours. ONS says the new method reduces the measured post-financial-crisis productivity slowdown by around half, while not making the productivity puzzle disappear. Post-2008 output-per-hour growth is estimated at 1.3% a year rather than 0.7% between 2009 and 2019. ONS methodology, 17 September 2026. ONS productivity blog.

That is the lesson for policing.

When a result looks wrong, do not immediately discard it. Ask what assumptions sit inside the measure. Ask whether those assumptions still hold. Compare the figure with another source. Rebuild it if necessary.

The important question is not:

“Can I trust this dataset?”

It is:

“What would have to be true for this measure to mislead me?”

That is a much more useful question in a performance meeting.

The denominator decides the question

“Per 1,000 population” sounds objective.

It is actually a choice.

Stop and search demonstrates the problem well. The Home Office uses resident population to calculate common force-area rates, but also warns that these figures do not account for transient populations such as commuters and tourists, can hide substantial variation within force areas, and need to be read alongside the concentration of searches among particular age and sex groups. Home Office, year ending March 2025.

The resident denominator is not “wrong”.

It answers a particular question.

It tells you how recorded searches compare with the population that lives in the area.

It does not recreate the population actually present in the places and at the times where searches occurred.

It does not tell you who was exposed to police activity.

And it does not explain why any disparity exists.

A useful leadership question is:

What population would need to be in the denominator to answer the question I actually care about?

That question will often reveal that the argument is ahead of the measure.

Police activity helps create police data

Some performance data measures things the police have partly produced themselves.

More proactive drug enforcement can increase recorded drug offences.

More stop and search creates more stop-and-search records.

Better crime-recording compliance can increase recorded crime.

An easier online reporting route can increase demand visible to the force.

A team instructed to focus on a specific offence can generate more detections in that category because it is looking harder.

HMICFRS has repeatedly inspected crime-data integrity because recording practice, systems, supervision and culture change the number that reaches the dashboard. That means a rising number can sometimes reflect improved visibility or recording rather than worsening underlying harm.

The reverse also follows.

A falling number can look positive while victims are becoming less visible.

Do not ask only:

“Did the number go up or down?”

Ask:

“What part of the process could have moved it?”

Comparators are questions in disguise

Most performance numbers become persuasive only when placed next to another number.

“Up 18%.”

“Below target.”

“Worst in the region.”

“Better than similar forces.”

Each phrase hides a methodological decision.

Up 18% compared with which period?

Similar forces in what respect?

Worst by counts, rates, harm, outcomes or trend?

Below a target that was derived from what evidence?

Month-on-month comparisons can be distorted by seasonality. Year-on-year comparisons can hide a recent change. Force comparisons can be distorted by geography, population, demand and recording practice. Targets can tell you whether the organisation met an ambition without telling you whether the ambition was sensible.

The question is not whether a comparator exists.

It is whether the comparator answers the question you think it answers.

Percentages can remove the very thing leaders need to see

A rise from 1 to 2 is a 100% increase.

A rise from 1,000 to 1,100 is a 10% increase.

The larger percentage represents one extra event.

The smaller percentage represents one hundred.

A percentage is often useful because it removes scale.

That is also why it can mislead.

For operational decisions, require both:

• the absolute number;
• the relative change.

If a rare event has “doubled”, ask from what baseline.

If serious violence is “up 40%”, ask how many incidents that represents.

If an outcome rate improved, ask how many additional victims received the outcome.

A trend tells you what moved. It rarely tells you why

The most common analytical mistake in performance management is moving directly from:

the number changed

to:

our intervention caused it.

A before-and-after chart cannot establish that by itself.

Crime might have fallen after an operation because the operation worked.

It might also have fallen because of seasonality, imprisonment of a prolific offender, displacement, a recording change, regression from an unusual spike or a wider regional trend.

The best question for the analyst is not:

“Can you show that the operation worked?”

It is:

“What else could explain this pattern, and can we test those explanations?”

That changes the role of analysis.

The analyst is not there to decorate the operational narrative.

They are there to try to break it.

If the explanation survives, confidence increases.

Critical does not mean cynical

Once leaders learn that every dataset has limitations, another failure mode appears:

“The data is unreliable. I will trust experience.”

That is not data literacy.

It is a different form of overconfidence.

A resident-population stop-search rate can have limitations and still identify an important disparity.

Police-recorded crime can be affected by reporting and recording and still show meaningful changes.

Survey data can contain sampling error and non-response and still be the best available evidence.

The aim is not to demand perfect data.

It is to make the conclusion no stronger than the measure.

The 90-second test

A leader should be able to challenge most performance claims quickly.

Not by doing the analyst’s job.

By exposing whether the analysis is ready to support a decision.

STOP

What exact claim are we being asked to accept?

Turn “crime is down” into something testable:

Recorded robbery in this district was 18% lower in the last 12 weeks than in the same 12 weeks last year.

TRACE

Where did the figure come from?

What was counted?

What is missing?

What denominator and comparator were used?

Did the recording or collection process change?

TEST

What else could produce the same pattern?

What would we expect to see if our preferred explanation were true?

Does another source, area, cohort or period support it?

DECIDE

What does the evidence justify?

A description?

A working hypothesis?

A deployment decision?

A claim that an intervention worked?

Those are different levels of certainty.

Make analytical challenge part of leadership culture

The leader does not need to out-statistic the analyst.

They do need to create a room where weak claims are uncomfortable.

For every important performance figure, I would expect five things.

The source is visible.

The measure is defined.

The denominator and comparator are stated.

Material limitations sit beside the number rather than in an appendix.

The analyst is expected to challenge the preferred story.

The College of Policing’s senior leadership standards explicitly connect decision-making, performance management and the use of relevant data and evidence. Its guidance on police data also says that data should be assessed for accuracy and carefully analysed before it is used to guide decisions or allocate resources. College of Policing — leadership standards. College of Policing — reporting and analysing data.

Humberside Police’s self-service Power BI work, published by the College, is useful for the same reason. The value was not merely giving supervisors dashboards. Leaders were trained to examine the data and the dashboards were built into performance meetings and governance. College of Policing — self-service Power BI dashboards.

More data does not automatically produce better decisions.

Better questions do.

A good police leader should be wary of two shortcuts:

“The data says it, so it must be true.”

and:

“My experience says it is wrong, so the data must be bad.”

Neither is good enough.

Question the number.

Test your instinct.

Then make the decision on what survives both tests.

---

Sources and further reading

• College of Policing — Leadership standards
• College of Policing — Reporting and analysing data held on PNC and LEDS
• College of Policing — Self-service Power BI dashboards
• Office for Statistics Regulation — Standards for the Public Use of Statistics, Data and Wider Analysis
• Office for National Statistics — Component approach to measuring labour productivity, UK: 1997 to 2024
• Office for National Statistics — Labour Force Survey quality and methodology information
• Home Office — Stop and search, arrests and mental health detentions, year ending March 2025
• Open University — More or Less: interview with Tim Harford