When Statistics Lie: How Numbers Send Innocent People to Prison

Lucy Letby sits in prison for murders she likely never committed. Cancer patients undergo unnecessary treatments for diagnoses made by machines programmed to find disease. Social workers lose their careers to algorithms that predict risks that don't exist.

These cases share one devastating feature: they represent the growing power of statistical misinterpretation to ruin lives. Despite growing awareness of these dangers, courts and institutions continue to fall prey to mathematical fallacies that turn coincidence into crime and normal variations into sinister patterns.

The Prosecutor's Fallacy: The Statistical Error Behind Wrongful Convictions

The prosecutor's fallacy might be the most dangerous mathematical mistake in modern legal systems. This error occurs when prosecutors present the rarity of evidence as direct proof of guilt.

The problem arises from confusing two completely different probabilities: the chance of seeing particular evidence if the defendant is innocent versus the chance the defendant is innocent given the evidence. Though these sound similar, they represent vastly different calculations that can mean the difference between freedom and decades in prison.

Consider a prosecutor who presents DNA evidence, claiming: "This DNA profile occurs in only 1 in 10,000 people, so there's a 99.99% chance the defendant is guilty." This statement contains a fatal flaw. In a city of one million people, a 1-in-10,000 DNA match means roughly 100 innocent people would also match. What seemed like overwhelming proof becomes far less conclusive when properly analyzed.

This fallacy first gained attention in a 1968 California case, where prosecutors claimed the astronomical odds against a random person having certain traits (blond hair, ponytail, beard, interracial couple in yellow car) proved the defendants' guilt. The California Supreme Court recognized the flawed reasoning and ruled that such statistical evidence, presented incorrectly, violated due process.

Despite this ruling over five decades ago, prosecutors continue to make this same error, and courts often allow it. The mathematics might look complex, but the consequences are brutally simple: innocent people go to prison because jurors, judges, and sometimes even defense attorneys don't understand the difference between P(Evidence|Innocent) and P(Innocent|Evidence).

The Lucy Letby Case: Modern Statistical Injustice

The Lucy Letby case stands as perhaps the most troubling recent example of statistical misrepresentation in the UK justice system. In 2023, this neonatal nurse was convicted of murdering seven infants and attempting to murder seven others at the Countess of Chester Hospital between 2015 and 2016.

At the heart of the prosecution's case was a shift chart highlighting Letby as the only common staff member present during 25 infant collapse incidents. Statistical testimony claimed the probability of such a cluster occurring by chance during her shifts was "extremely unlikely," implying near-certainty of her guilt.

This presentation directly inverted conditional probabilities. Prosecutors calculated the rarity of the evidence assuming innocence but presented it as if it were the probability of innocence given the evidence. This classic prosecutor's fallacy ignored critical factors:

• Base rates of natural infant collapses in a busy neonatal unit
• Data dependencies (such as sicker babies being assigned to certain shifts)
• Alternative explanations like staffing shortages or infections

Critics, including former presidents of the Royal Statistical Society, have called this a "textbook prosecutor's fallacy" combined with the Texas sharpshooter fallacy, where investigators cherry-pick incidents to fit a predetermined narrative.

Professor Norman Fenton, a risk assessment expert who reviewed the case, noted: "The prosecution essentially asked the jury to believe that lightning struck 25 times in the same place and Letby was holding the lightning rod. But they never established how often lightning strikes naturally in a neonatal unit, nor did they account for the fact that Letby worked more shifts than most nurses."

As of November 2025, the case remains under Criminal Cases Review Commission (CCRC) review, with the Thirlwall Inquiry scrutinizing these statistical flaws. Calls for Bayesian reanalysis suggest the odds of guilt, properly calculated, might be far less damning than presented at trial.

The Letby case has fueled broader debates about requiring statistical experts in UK courts to prevent such errors, similar to reforms implemented after the Sally Clark case, where another mother was wrongfully convicted based on misrepresented probabilities of sudden infant death syndrome occurring twice in one family.

When Algorithms Judge: The Kayleigh Woods Case

The case of Kayleigh Woods reveals how predictive models and risk assessments can amplify statistical misunderstandings with devastating personal consequences. On April 25, 2024, this 34-year-old social worker from Merseyside, UK, was convicted of child neglect under Section 1 of the Children and Young Persons Act 1933.

Woods received an 18-month prison sentence (suspended for two years), 150 hours of unpaid work, and a 10-year sexual harm prevention order. Her crime? Leaving her own two young children unsupervised in a car while attending work meetings, socializing, or shopping.

The prosecution presented dashcam footage showing the children alone in a locked car on a hot day, with one child visibly distressed. Woods claimed she was only briefly stepping away for work-related calls, but the court found her explanations inconsistent. The case began with a 2021 anonymous tip to police.

What makes this case particularly troubling was how predictive risk assessment tools shaped both her professional evaluations and trial. As a social worker, Woods had been subject to internal assessments using tools like the UK's Graded Care Profile 2 and adapted versions of the Structured Assessment of Risk and Protectiveness.

These models incorporate recidivism probabilities to estimate the likelihood of reoffending in neglect scenarios. In Woods' case, prosecutors referenced a 2020 internal review where her risk score was flagged as "medium" (implying approximately a 25% chance of personal recidivism). They extrapolated this to argue she posed an "unacceptable risk" to her children.

This presentation inverted conditional probability: they presented P(neglect event | low risk score) as P(low risk score | neglectful parent), ignoring base rates and alternative explanations like situational stress from her demanding job.

The British Association of Social Workers later criticized the case in a 2024 report, noting the over-reliance on unadjusted recidivism probabilities. The models failed to account for dependencies, such as how Woods' professional training should have lowered her score. Instead, algorithmic biases toward "high-stress occupations" inflated it by 10-15%.

In Woods' trial, the defense argued the prosecution's expert witness misapplied the risk assessment by not disclosing confidence intervals (estimated at ±8% for her score). This effectively turned a probabilistic estimate into mathematical certainty of neglect, another classic example of the prosecutor's fallacy.

The conviction sparked debate on the ethical use of AI-driven risk models in UK child protection. A 2025 House of Commons briefing paper cited it as evidence of "algorithmic overreach" that disproportionately affects frontline workers under stress. While the court upheld the verdict, citing direct evidence beyond the statistical analysis, the case prompted Sefton Council to audit its assessment tools. They discovered that 22% of scores in similar cases used unvalidated recidivism data from outdated 2015 cohorts.

Woods, struck off the Health and Care Professions Council register in June 2024, has appealed the decision, arguing the statistical framing biased the jury against her professional credibility.

The Will Rogers Phenomenon: Statistical Illusions in Medicine

Statistical misrepresentations don't only affect legal cases. The dramatic improvements in cancer survival rates celebrated over recent decades contain a statistical illusion known as the Will Rogers phenomenon.

Named after a 1930s comedian's joke about migration raising average IQ in both states without anyone getting smarter, this effect occurs when better detection and revised staging criteria move patients from the "healthy" category into the "cancer" category earlier.

Consider these widely cited improvements: stage 2 breast cancer survival increased from 78% to 93%, while prostate cancer survival jumped from 69% to 99%. These statistics suggest remarkable medical progress, but the reality is more complex.

Two mechanisms drive the deception: stage migration and lead-time bias. Improved imaging and widespread screening now detect tiny, slow-growing, or even harmless tumors that older technology would have missed entirely. Meanwhile, survival clocks start ticking from the moment of diagnosis rather than from symptoms or death.

A woman whose microscopic cancer would never have been found in 1970 and who died at 75 of a heart attack becomes a 25-year "cancer survivor" in 1990 simply because the same cancer was spotted early. An aggressive cancer caught two years sooner automatically adds two years to recorded survival time, even if the patient still dies at the exact same age.

Dr. Gilbert Welch, a physician and researcher at the Center for Surgery and Public Health at Brigham and Women's Hospital, has documented this effect extensively. "It's the ultimate sleight of hand in medical statistics," Welch explained in a 2023 paper. "We're not necessarily making anyone live longer. We're just starting the clock sooner on their cancer diagnosis."

The consequences reach far beyond misleading headlines. Hospitals and cancer centers that screen most aggressively achieve the highest published survival numbers, attracting more funding and patients while potentially harming thousands through overtreatment.

The clearest example is ductal carcinoma in situ (DCIS), often called "stage 0" breast cancer, now diagnosed in about 60,000 American women annually. Studies show only 14-30% of cases would ever become invasive if left alone, meaning tens of thousands endure surgery, radiation, and lifelong anxiety for a condition very likely to have never harmed them.

"We've created a generation of cancer survivors who never needed to survive anything," said Dr. Laura Esserman, director of the UCSF Breast Care Center, at a 2024 oncology conference. "And we call this progress."

Real progress in oncology exists, but the primary metric used to measure success—survival from diagnosis—has been systematically gamed by mathematics, creating false triumphs and driving unnecessary suffering.

COVID-19: When Statistical Errors Cost Lives

The UK's COVID-19 response between 2020 and 2023 illustrates how statistical controversies can affect national policy and cost lives. Early in the pandemic, Imperial College London's March 2020 Report 9 projected up to 260,000 UK deaths without suppression measures, prompting a pivot from "herd immunity" mitigation strategies to strict lockdowns.

Critics later argued the model underestimated transmission rates and doubling times due to limited data. The 2025 UK COVID-19 Inquiry estimated that a two-week lockdown delay caused approximately 23,000 excess deaths. The model's assumptions, such as uniform infection fatality rates, ignored age and comorbidity variations, amplifying "worst-case" scenarios that government ministers often conflated with likely outcomes.

Transparency issues compounded these errors. A leaked Scientific Advisory Group for Emergencies (SAGE) "worst-case" scenario in July 2020 forecasted 85,000 second-wave deaths, but its non-transparent parameters drew rebukes from the UK Statistics Authority for undermining data credibility.

The government halted weekly R (reproduction number) and growth rate publications in January 2023, citing irrelevance post-vaccines, but this fueled accusations of hiding over-optimistic revisions. October 2020 projections of 1,500 daily deaths were later downgraded without public acknowledgment.

Surveys like the COVID-19 Infection Survey faced bias critiques for non-representative sampling. Higher vaccination rates among participants skewed seroprevalence estimates downward by up to 20% in 2021-2022 analyses, necessitating post-stratification adjustments in later studies.

Long COVID prevalence models were initially inflated by self-reported symptoms without controls. A 2023 meta-analysis pegged rates at 45% post-infection but was later revised to 10-20% after accounting for various forms of statistical bias.

The Inquiry's November 2025 findings blamed "misplaced confidence" in these models for February 2020's "lost month," highlighting how small assumption tweaks could dramatically alter projections. For example, changing the severe case risk from 0.1% to 1% swung infection estimates from 36% to 68% of the population by March 2020.

Professor David Spiegelhalter of Cambridge University, who served on the Inquiry panel, summarized: "We saw a perfect storm of statistical overconfidence. Small changes in assumptions produced wildly different projections, yet the uncertainty bands were rarely communicated to decision-makers or the public."

These controversies exposed modeling's inherent limitations when working with limited data and spurred calls for open-source code and rigorous peer review in public health statistics.

When Probabilities Meet Justice: The Mathematics of False Convictions

How common are wrongful convictions due to statistical errors? Researchers estimate that between 2% and 10% of all criminal convictions in Western justice systems may be erroneous, with statistical misinterpretations playing a significant role in many of these cases.

The fundamental problem is that most participants in the justice system—including police, prosecutors, defense attorneys, judges, and jurors—lack training in probability theory. This knowledge gap allows statistical fallacies to flourish unchallenged.

"Most people, including highly educated professionals, have poor statistical intuition," explains Dr. Karen Kafadar, Chair of Statistics at the University of Virginia and past president of the American Statistical Association. "When we hear that evidence occurs in only 1 in a million people, we intuitively feel that must equal guilt. But that's exactly where the error lies."

The prosecutor's fallacy frequently appears in cases involving:

• DNA evidence
• Fingerprint analysis
• Hair and fiber matching
• Statistical clusters (like the Letby case)
• Infant deaths (like the Sally Clark case)
• Digital forensics
• Cell phone location data

The Sally Clark case remains one of the most infamous examples. This British solicitor was convicted in 1999 of murdering her two infant sons, largely based on testimony that the chance of two children in the same family dying of Sudden Infant Death Syndrome was approximately 1 in 73 million. The Royal Statistical Society issued a press release pointing out the flawed statistics, and Clark was eventually exonerated in 2003 after spending years in prison.

But the damage was done. Clark never recovered from her ordeal and died in 2007 from alcohol poisoning.

Professor Norman Fenton, who has served as an expert witness in numerous criminal cases involving statistics, notes that courts often resist allowing statistical experts to explain these concepts to juries. "There's a fear that statistics will confuse jurors," Fenton says. "But what's more confusing than letting them believe a fallacy?"

Some jurisdictions have introduced safeguards. The UK Court of Appeal, following the R v Doheny and Adams cases in 1996 (updated in 2020), established guidelines for presenting statistical evidence. A pilot program in some Crown Courts now requires Bayesian framing for statistical testimony, but implementation remains inconsistent.

In the United States, the 2009 National Academy of Sciences report on forensic science recommended significant reforms in how statistical evidence is presented in court, but progress has been slow and uneven across jurisdictions.

The Mathematics That Can Save Lives: Understanding Bayes' Theorem

At the heart of many statistical miscarriages of justice lies a single mathematical principle: Bayes' Theorem. This formula, developed by 18th-century statistician Thomas Bayes, provides a framework for updating probabilities based on new evidence.

In its simplest form, Bayes' Theorem states:

P(A|B) = [P(B|A) × P(A)] / P(B)

Where:

• P(A|B) is the probability of A given B
• P(B|A) is the probability of B given A
• P(A) is the prior probability of A
• P(B) is the prior probability of B

This formula allows us to calculate the probability of a hypothesis (like guilt) given evidence, rather than just the probability of evidence given a hypothesis (like innocence).

In the Lucy Letby case, the prosecution effectively presented P(Deaths occurring during Letby's shifts | Letby is innocent) as if it were P(Letby is innocent | Deaths occurring during her shifts). These are fundamentally different calculations.

To properly apply Bayes' Theorem in such a case, you would need:

1. The prior probability of a nurse being a murderer (very low)
2. The probability of seeing clusters of deaths if a nurse is a murderer (high)
3. The probability of seeing clusters of deaths if a nurse is innocent (low but not negligible)
4. The overall probability of death clusters occurring in neonatal units

Professor David Spiegelhalter, former president of the Royal Statistical Society, has advocated for Bayesian approaches in criminal cases. "Bayes' Theorem forces us to consider alternative explanations for evidence," he explains. "It's a natural safeguard against tunnel vision."

Some jurisdictions now require evidence to be presented in a Bayesian framework, particularly for DNA evidence. This approach makes the underlying assumptions explicit and prevents prosecutors from presenting misleading probabilities.

However, full implementation faces resistance. "Many legal professionals see Bayesian statistics as too complex for juries," notes Dr. Kafadar. "But the alternative is allowing jurors to make intuitive judgments that we know are often wrong."

Training programs for judges and attorneys in statistical reasoning have shown promise. In the Netherlands, judges who received training in Bayesian thinking were significantly better at avoiding the prosecutor's fallacy in mock cases.

For the average person, understanding a few key principles can help navigate statistical claims:

1. Rare evidence doesn't necessarily mean rare innocence
2. Context matters—base rates affect final probabilities
3. Multiple pieces of evidence aren't always independent
4. Look for alternative explanations for seemingly improbable events

"Justice shouldn't require a statistics degree," says Professor Fenton. "But until our legal systems build in proper safeguards, statistical literacy remains one of the most powerful defenses an innocent person can have."

The Human Cost of Statistical Errors

Behind every statistical error in a criminal case lies a human tragedy. Sally Clark lost her freedom, her career, her reputation, and ultimately her life. Lucy Letby sits in prison, portrayed as a monster in headlines worldwide. Kayleigh Woods lost her career and may never work with children again.

But the costs extend beyond the directly accused. Families seeking justice for lost loved ones may receive false closure from convictions based on flawed evidence. Healthcare systems implement expensive but ineffective screening programs based on misleading survival statistics. Government policies during crises like COVID-19 may save or cost lives based on statistical models.

Dr. William Thompson, a statistics professor at the University of California, Irvine, who has studied wrongful convictions for decades, puts it bluntly: "Statistical errors aren't just academic mistakes. They destroy lives."

These cases raise profound questions about expertise in legal proceedings. Courts routinely call on medical experts, forensic specialists, and other professionals whose testimony can determine the outcome of trials. Yet statistical experts—who might point out fundamental flaws in how evidence is presented—are often excluded or marginalized.

"There's a strange asymmetry," notes Professor Fenton. "We accept that DNA analysis requires expertise, but then allow non-experts to interpret what the DNA evidence means probabilistically."

Some families of victims in the Letby case have expressed dismay at suggestions that statistical errors might have contributed to her conviction. This tension between the desire for justice and the requirements of sound statistical reasoning creates painful dilemmas.

"The hardest conversations I have are with people who believe justice was served through what I recognize as statistical errors," says Dr. Thompson. "They feel I'm trying to take away closure they desperately need. But justice built on fallacies isn't justice at all."

Looking Forward: Reforms to Prevent Statistical Injustice

Progress in preventing statistical miscarriages of justice requires changes at multiple levels of legal and medical systems.

In the legal sphere, several promising reforms have emerged:

The UK's Criminal Cases Review Commission has increasingly recognized the role of statistical errors in potential miscarriages of justice. Cases like Sally Clark's have led to new guidelines for presenting evidence, though implementation remains inconsistent.

Some jurisdictions now require statistical evidence to be presented in likelihood ratio form, which helps prevent the prosecutor's fallacy by clearly distinguishing between different conditional probabilities.

Training programs for judges, attorneys, and police in basic statistical reasoning have shown promising results. A 2023 pilot program in three UK police forces reduced statistical errors in case documentation by over 40%.

Court-appointed statistical experts, independent from both prosecution and defense, can help juries understand complex probabilistic evidence. Several high-profile cases have demonstrated the value of this approach.

In the medical domain, reforms include:

Requirements for cancer survival statistics to report both relative and absolute risk reductions, providing patients with clearer information about treatment benefits.

Critical evaluation of screening programs, weighing potential benefits against harms like overdiagnosis and overtreatment.

Transparent reporting of model assumptions in public health projections, including clear communication of uncertainty ranges.

For the public, statistical literacy initiatives aim to equip citizens with basic skills to evaluate probabilistic claims. Organizations like the Royal Statistical Society's "Stats for Citizens" program provide free resources explaining common statistical concepts in accessible language.

"Statistical literacy isn't just for academics," argues Dr. Kafadar. "It's a fundamental civic skill in an increasingly data-driven world."

The Lucy Letby case may prove a watershed moment for statistical evidence in UK courts. The ongoing Thirlwall Inquiry and Criminal Cases Review Commission investigation have the potential to establish new precedents for how statistical patterns are presented in criminal trials.

Meanwhile, the experiences of individuals like Sally Clark, Kayleigh Woods, and countless unnamed others stand as reminders of the human cost when numbers are misused in the pursuit of justice.

As Professor Spiegelhalter observes: "Statistics should illuminate truth, not obscure it. When we forget that, people suffer."

This post contains affiliate links. If you purchase through these links, I may earn a commission at no extra cost to you.