1. Introduction

Cross-examination on statistical models is an important aspect of modern evidence law, scientific evidence, forensic litigation, artificial intelligence (AI) accountability and biotechnology regulation.

Statistical models are mathematical or computational systems used to analyse data, identify patterns, estimate probabilities, predict outcomes and support legal or scientific conclusions.

Examples include:

DNA probability calculations in criminal investigations.

Statistical models used to identify disease causation.

AI-based medical diagnosis and risk prediction.

Algorithmic tools used in criminal sentencing and predictive policing.

Statistical evidence in employment discrimination cases.

Insurance fraud detection and financial risk assessment.

Clinical-trial statistical models used to establish drug safety and efficacy.

Cross-examination tests whether the model is scientifically reliable, whether its underlying data are accurate, whether its assumptions are justified and whether its conclusions genuinely support the proposition advanced by the party relying on it.

The central legal question is:

Can a statistical model provide sufficiently reliable evidence to justify a judicial finding, and can the opposing party effectively challenge its methodology, assumptions and conclusions?

A statistical result is not automatically proof of a legal fact. Courts must distinguish mathematical probability from legal proof.

2. Meaning and Scope of Cross-Examination on Statistical Models

Cross-examination involves questioning an expert witness who has relied upon statistical analysis or a computational model.

Its principal purposes are:

Testing the qualifications and independence of the expert.

Examining the reliability of the dataset.

Identifying errors in data collection and processing.

Challenging the statistical methodology.

Testing the assumptions underlying the model.

Examining sampling errors and uncertainty.

Identifying bias and confounding variables.

Evaluating whether the model has been independently validated.

Examining whether alternative explanations were considered.

Preventing misleading interpretation of statistical results.

For example, suppose an expert states that a DNA profile has a random-match probability of one in ten million.

The opposing lawyer may ask:

What population database was used?

Was the population genetically representative?

Were related individuals considered?

Was the laboratory procedure independently validated?

Does the probability refer to a random match or to the probability that the accused is innocent?

What assumptions were made in calculating the probability?

These questions are important because a random-match probability is not equivalent to the probability of innocence.

3. Legal Framework in India

The principal legislation is the Bharatiya Sakshya Adhiniyam, 2023 (BSA), which came into force on 1 July 2024.

The following provisions are particularly relevant.

Section 39 is particularly important. It permits courts to consider opinions given by persons specially skilled in science, art or another relevant field.

Statistical modelling falls within this framework where specialist statistical or computational knowledge is necessary to resolve a disputed issue.

Section 143(2) is also significant because cross-examination is not confined to matters raised during examination-in-chief. This permits the opposing party to investigate the underlying statistical methodology, datasets, assumptions and limitations of the expert's opinion.

IndiaCode+1

4. Important Principles Governing Statistical Expert Evidence

A. Expert qualification

The witness must demonstrate relevant expertise.

A person who can operate statistical software is not necessarily qualified to explain advanced statistical inference, epidemiological modelling or machine-learning validation.

B. Transparency of methodology

The expert should explain:

The statistical method employed.

Why that method was selected.

The source and quality of the data.

The assumptions used.

The treatment of missing observations.

The confidence intervals and error rates.

The limitations of the model.

C. Reproducibility

The opposing party should be able to examine whether another qualified expert, using the same dataset and methodology, can reproduce the reported results.

D. Independence

An expert should disclose whether the model was developed, commissioned, funded or modified by an interested party.

E. Distinction between correlation and causation

A statistical association does not automatically establish that one event caused another.

For example, a statistical relationship between exposure to a chemical and cancer incidence does not, by itself, prove that the chemical caused a particular individual's cancer.

F. Avoidance of false precision

A model producing a result of 99.9% accuracy does not necessarily establish that its prediction in a particular case is 99.9% reliable.

The population, test conditions, prevalence, error rates and validation methodology must be examined.

5. Important Case Laws

1. State of Himachal Pradesh v. Jai Lal (1999) 7 SCC 280

Indian Supreme Court

Facts:

The case concerned allegations involving inflated records of diseased apples and financial loss to the State. A horticultural officer was presented as an expert witness to assess the productivity of apple orchards.

However, the expert had not demonstrated adequate scientific study or research supporting his assessment.

Legal issue:

Whether a witness's official position or practical experience alone is sufficient to establish the reliability of expert evidence.

Judgment:

The Supreme Court explained that expert testimony is advisory in nature. The credibility of an expert depends on the reasons supporting the conclusions and the underlying data and materials.

Importantly, the Court observed that an expert report does not automatically become evidence merely because it has been submitted. The expert must be examined in court and must face cross-examination.

Application to statistical models:

This judgment is particularly useful when challenging an expert who presents statistical conclusions without demonstrating:

Relevant statistical qualifications.

Scientific research supporting the model.

The source of the dataset.

The calculations behind the results.

The reasons supporting the conclusions.

Legal principle: An expert's conclusion cannot replace examination of the scientific foundation on which it rests.

JK High Court+1

2. Daubert v. Merrell Dow Pharmaceuticals, Inc. (1993) 509 U.S. 579

Landmark scientific evidence case

Facts:

The plaintiffs alleged that a pharmaceutical product caused birth defects. They relied upon scientific expert testimony concerning studies and statistical evidence.

Legal issue:

How should courts determine whether scientific expert testimony is sufficiently reliable to be admitted?

Judgment:

The United States Supreme Court established a flexible reliability assessment.

It identified several important considerations:

Whether the scientific method can be tested.

Whether it has undergone peer review and publication.

Its known or potential error rate.

Whether standards control its operation.

Whether it has gained acceptance in the relevant scientific community.

The Court also recognised vigorous cross-examination, contrary evidence and proper jury instructions as important methods of challenging admissible but questionable scientific evidence.

Application to statistical models:

A lawyer can rely on the Daubert framework to question:

Whether the statistical model has been independently tested.

Whether its error rate has been measured.

Whether its assumptions have scientific support.

Whether the model has undergone peer review.

Whether the methodology is accepted in the relevant scientific discipline.

Legal principle: Scientific evidence must have a reliable methodological foundation; impressive numerical results alone are insufficient.

LII / Legal Information Institute+1

Note: Daubert is an American judgment, not a binding interpretation of Indian evidence law. Its scientific reliability principles can nevertheless provide useful comparative guidance.

3. General Electric Co. v. Joiner (1997) 522 U.S. 136

Facts:

The plaintiff alleged that exposure to PCBs and related chemicals contributed to his lung cancer.

Experts relied on animal studies and epidemiological research to support their conclusions.

Legal issue:

Whether expert conclusions can be admitted when the underlying research does not sufficiently support the conclusion being advanced.

Judgment:

The Supreme Court upheld the exclusion of the challenged expert testimony.

It explained that a court need not accept an expert opinion connected to existing research only through the expert's unsupported assertion.

The judgment also confirmed the trial judge's gatekeeping responsibility concerning scientific evidence.

Application to statistical models:

Suppose an expert relies on several epidemiological studies to establish that a biotechnology product caused a disease.

Cross-examination should examine:

Whether the studies involved comparable populations.

Whether the exposure levels were comparable.

Whether the sample size was adequate.

Whether confounding factors were controlled.

Whether the expert's final conclusion goes beyond the actual research findings.

Legal principle: An expert cannot bridge gaps in scientific evidence merely by asserting a conclusion.

LII / Legal Information Institute+1

4. Kumho Tire Co. v. Carmichael (1999) 526 U.S. 137

Facts:

The dispute concerned expert testimony about the cause of a tyre failure.

Legal issue:

Whether the judicial reliability assessment applies only to scientific experts or also to technical and experience-based experts.

Judgment:

The Supreme Court held that the trial judge's gatekeeping obligation applies to scientific, technical and other specialised expert testimony.

Application to statistical models:

This is important when statistical evidence involves:

Machine-learning algorithms.

Predictive analytics.

Automated medical systems.

Financial risk models.

Forensic classification systems.

An expert cannot avoid methodological scrutiny merely by describing the model as technical expertise rather than science.

Legal principle: Specialised technical conclusions must also satisfy judicial reliability scrutiny.

LII / Legal Information Institute

5. People v. Collins 68 Cal. 2d 319 (1968)

Facts:

The prosecution used mathematical probability calculations based on several characteristics of a suspected couple to support its case.

The probabilities were multiplied to suggest an extremely small likelihood of another couple possessing the same characteristics.

Legal issue:

Whether mathematical probability evidence could be used without adequate factual and statistical foundations.

Judgment:

The California Supreme Court rejected the statistical approach.

It identified serious problems, including:

Unsupported probability estimates.

Failure to establish statistical independence.

Inadequate evidentiary foundations.

The danger of misleading the jury through impressive numerical calculations.

The Court described the danger as a form of trial by mathematics.

Application to statistical models:

This case is particularly valuable for challenging models that multiply several probabilities without establishing that the variables are genuinely independent.

For example, a predictive model might combine age, occupation, location and behavioural indicators as if each were statistically independent, although they may be strongly correlated.

Legal principle: Mathematical sophistication cannot cure an unsupported statistical foundation.

Justia Law+1

6. R v. Adams [1996] EWCA Crim 222

Facts:

The prosecution relied on DNA evidence and statistical calculations suggesting that the probability of a randomly selected unrelated person matching the DNA profile was approximately one in 200 million.

The defence challenged the methodology and the interpretation of the statistical evidence.

Legal issue:

Whether DNA statistical evidence could be relied upon to support a criminal conviction and how its reliability should be examined.

Judgment:

The Court of Appeal held that DNA evidence was not inherently incapable of supporting a conviction even without additional evidence.

However, its reliability depended upon scrutiny of the assumptions and methodology behind the calculations.

The Court also expressed serious reservations about using Bayes' theorem as a numerical framework for the jury's overall assessment of guilt.

Application to statistical models:

Cross-examination should distinguish between:

Random-match probability.

Likelihood ratios.

Prior probability.

Posterior probability.

Probability of guilt.

These concepts are not interchangeable.

Legal principle: A statistically powerful forensic result must be interpreted correctly and evaluated alongside the complete evidentiary record.

BAILII

7. R v. T [2010] EWCA Crim 2439

Facts:

The case concerned footwear impression evidence and the use of statistical methods to evaluate similarities between footwear and marks found at a crime scene.

Legal issue:

Whether available statistical data justified expressing footwear comparison evidence through mathematical likelihood ratios.

Judgment:

The Court of Appeal concluded that the statistical foundation was insufficient to justify the proposed mathematical calculations.

It warned against presenting a subjective expert assessment as if it were an objectively established numerical probability.

Application to statistical models:

This judgment is useful when an expert:

Uses an inadequate reference population.

Presents subjective classifications as precise probabilities.

Applies a formula without sufficient empirical data.

Claims numerical certainty where the scientific field does not support it.

Legal principle: Numerical precision must reflect actual scientific reliability, not merely the sophistication of the formula.

Lccsa+1

6. Major Areas of Cross-Examination on Statistical Models

The following areas are particularly important in biotechnology, pharmaceutical litigation, AI regulation and forensic science.

A. Dataset reliability

The lawyer must examine whether the dataset accurately represents the population or event under investigation.

Questions:

Who collected the data?

Was the data independently verified?

How many observations were included?

Were any observations excluded?

Were missing values replaced or estimated?

Was the dataset representative of the relevant population?

Were historical records used?

Were errors in the original data corrected?

Was the dataset affected by selection bias?

Can the opposing party independently inspect the dataset?

B. Sampling bias

Sampling bias arises when the observations used in a statistical model do not adequately represent the population about which conclusions are being drawn.

For example, a medical AI model trained predominantly on one demographic population may perform poorly when applied to another.

Cross-examination should establish whether the sample was representative and whether the expert considered the consequences of non-representative sampling.

C. Model validation

Validation determines whether a model performs reliably on data beyond the material used to develop it.

Important questions include:

Was independent validation conducted?

Was a separate testing dataset used?

Was cross-validation performed?

Was the model tested in different populations?

Were false-positive and false-negative rates calculated?

Was external validation undertaken?

Has the model's performance deteriorated over time?

D. Statistical significance

Statistical significance must not be confused with practical or legal significance.

A result may be statistically significant but still have little practical importance.

For example, a clinical study involving thousands of participants may identify a statistically significant difference in treatment outcomes that is medically insignificant.

Cross-examination should examine the p-value, confidence intervals, sample size, effect size and clinical relevance.

E. Confounding variables

A confounding variable may create or distort an apparent relationship between two variables.

For example, a study might identify an association between a particular pharmaceutical treatment and improved patient outcomes without adequately considering age, underlying disease severity or other treatments.

The lawyer should question whether the expert identified and controlled relevant confounders.

F. Algorithmic bias

Statistical models used in AI systems can reproduce discriminatory patterns contained in historical data.

Examples include:

Gender-based employment predictions.

Caste-related discrimination.

Racial bias in medical risk assessments.

Discriminatory insurance classification.

Biased criminal risk prediction.

Cross-examination should investigate the training data, variable selection, proxy variables, fairness testing and differential error rates.

G. Model explainability

An expert relying on a complex machine-learning model should explain how the relevant result was generated.

Questions should include:

What variables influenced the prediction?

Which variables had the greatest influence?

Can the result be independently reproduced?

Was an explainability method used?

Can the expert explain the result for this particular individual?

Were alternative models considered?

A model's inability to provide a complete explanation does not automatically make it inadmissible, but it can affect the assessment of reliability and evidentiary weight.

7. Practical Cross-Examination Questions for Statistical Experts

The following questions can be adapted for use in court.

8. Special Application in Biotechnology and Pharmaceutical Litigation

Cross-examination of statistical models has particular importance in biotechnology because scientific and statistical evidence frequently determines whether a biological product is considered safe, effective or responsible for injury.

Important examples

1. Clinical-trial litigation

Suppose a pharmaceutical company claims that a drug is safe because a clinical trial demonstrated a favourable statistical result.

The opposing lawyer may examine:

Whether the trial was adequately powered.

Whether the control group was appropriate.

Whether participants withdrew from the trial.

Whether adverse events were fully reported.

Whether the endpoints were changed.

Whether the statistical analysis plan was followed.

2. Vaccine injury litigation

Where a claimant alleges that vaccination caused a particular injury, statistical evidence may be used to examine whether an association exists between vaccination and the alleged condition.

The expert should be questioned about background incidence, temporal association, confounding factors, epidemiological studies and individual causation.

3. Gene therapy litigation

Where a gene therapy product allegedly causes unexpected biological consequences, statistical modelling may be used to assess adverse events, treatment outcomes and genetic risks.

Cross-examination should examine whether the model accounts for individual genetic variation and long-term uncertainty.

4. AI-based medical diagnosis

Where a hospital relies on an algorithm to diagnose patients, the expert may be questioned about training data, diagnostic accuracy, false negatives, demographic performance and model drift.

5. Genetic and forensic evidence

Statistical calculations may be used to establish DNA match probabilities or evaluate genetic relationships.

The expert must explain population assumptions, relatedness, allele frequencies and the difference between a likelihood ratio and the probability of guilt.

9. Consequences of Unreliable Statistical Evidence

Where cross-examination exposes serious weaknesses, several legal consequences may follow.

Exclusion: The court may exclude expert testimony that does not satisfy applicable admissibility requirements.

Reduced evidentiary weight: The court may admit the evidence but attach limited weight to it.

Contrary expert evidence: The opposing party may present an independent statistical expert.

Further disclosure: The court may require production of relevant data, methodology or supporting material, subject to applicable procedural and confidentiality rules.

Reassessment of findings: Findings based materially on unreliable statistical evidence may be challenged on appeal.

Adverse credibility assessment: Unsupported assumptions or misleading presentations may undermine the expert's credibility.

Importantly, an error in one statistical calculation does not automatically invalidate an entire trial. Its materiality and effect on the findings must be assessed.

10. Model Cross-Examination: A Practical Illustration

Hypothetical case: A biotechnology company introduces a statistical model claiming that its medical AI system predicts disease with 99% accuracy.

The opposing counsel could conduct the following examination.

Counsel: You have stated that the model has 99% accuracy?

Expert: Yes.

Counsel: Is that figure based on training data or independent testing data?

Expert: It is based on testing data.

Counsel: Was the testing dataset completely separate from the training dataset?

Expert: There was some overlap.

Counsel: So the model was not tested exclusively on independent data?

Expert: That is correct.

Counsel: Did you calculate the false-negative rate?

Expert: Yes.

Counsel: Does your 99% accuracy figure reveal the probability that a particular patient's diagnosis is correct?

Expert: Not by itself.

Counsel: Did you test the model across different demographic populations?

Expert: Only selected populations.

Counsel: Therefore, you cannot establish from this figure alone that the model has equal predictive reliability for every patient?

Expert: That is correct.

Legal significance: This examination demonstrates how a seemingly impressive statistical result can lose evidentiary strength when its underlying methodology is challenged.

11. Comparative Legal Position

The foreign decisions discussed above are comparative authorities. They do not automatically govern Indian courts.

12. Important Legal Doctrines

The following doctrines provide a useful conceptual framework:

Scientific Reliability Doctrine: Statistical evidence must have a scientifically defensible foundation.

Expert Accountability Doctrine: Experts must justify their methodology and conclusions.

Statistical Independence Doctrine: Probability calculations must not assume independence without adequate justification.

Reproducibility Doctrine: Scientific findings should be capable of independent verification.

Data Integrity Doctrine: Statistical conclusions depend upon the accuracy and integrity of the underlying data.

Causation Distinction Doctrine: Statistical association does not automatically establish individual causation.

Algorithmic Transparency Doctrine: Relevant model operations and limitations should be sufficiently disclosed to permit meaningful scrutiny.

Judicial Gatekeeping Doctrine: Courts must evaluate the admissibility and reliability of expert evidence.

Evidentiary Weight Doctrine: The probative value of statistical evidence must be evaluated in the context of the entire case.

Fair Trial Doctrine: Parties must have a meaningful opportunity to challenge technical and statistical evidence.

13. Conclusion

Cross-examination on statistical models is an essential safeguard against unreliable scientific evidence, misleading probability calculations and unjustified algorithmic conclusions.

The Indian position under the Bharatiya Sakshya Adhiniyam, 2023, particularly Sections 39, 40 and 143, provides the framework for examining expert opinions and challenging their foundations.

The most important judicial authorities include:

State of Himachal Pradesh v. Jai Lal (1999) 7 SCC 280.

Daubert v. Merrell Dow Pharmaceuticals (1993) 509 U.S. 579.

General Electric Co. v. Joiner (1997) 522 U.S. 136.

Kumho Tire Co. v. Carmichael (1999) 526 U.S. 137.

People v. Collins 68 Cal. 2d 319 (1968).

R v. Adams [1996] EWCA Crim 222.

R v. T [2010] EWCA Crim 2439.

The ultimate legal principle is that a statistical model must not be treated as an unquestionable authority. Its data, methodology, assumptions, limitations and interpretation must remain open to meaningful judicial scrutiny and cross-examination.

14. Suggested Research Areas for Advanced Legal Study

For further biotechnology and AI law research, this subject can be developed into:

Cross-Examination of AI-Generated Statistical Evidence.

Statistical Model Errors and Expert Witness Liability.

Cross-Examination of Predictive Genetic Models.

Statistical Evidence in Clinical Trial Litigation.

Algorithmic Bias and Evidentiary Fairness.

Statistical Causation in Vaccine Injury Claims.

Model Transparency and Disclosure Obligations.

Judicial Review of Black-Box Statistical Models.

Statistical Evidence and Constitutional Due Process.

Accountability for Automated Scientific Decision-Making.

Research note: The Indian statutory discussion reflects the BSA framework effective from 1 July 2024. The cited judgments should be distinguished according to their jurisdiction and precedential status.