Adjudicating Pharmacogenomic Phenotypes

Why Phenotype Adjudication Matters

A pharmacogenomic association is only as clear as the outcome being compared. “Drug toxicity” may refer to a laboratory abnormality, a symptom documented in a note, a clinician’s diagnosis, or a serious event that prompted treatment discontinuation. These are not interchangeable outcomes. If researchers combine them without rules, or count an event that occurred before the patient took the drug, the resulting association may be difficult to interpret even when the statistical calculation is correct.

Phenotype adjudication makes the outcome definition operational: reviewers apply prespecified criteria to available clinical records, while exposure confirmation establishes that the person received the drug during the period in which the outcome could occur. The aim is not to make every record certain. It is to apply transparent rules consistently, record uncertainty, and assess whether classification errors could change the estimated association.

Specify the Phenotype Before Reviewing Records

A useful phenotype protocol states what qualifies as an event and when it must occur. For a hypothetical drug-associated liver injury outcome, the protocol might specify the laboratory thresholds, the required timing after treatment initiation, how alternative documented causes are handled, and whether an event is definite, probable, possible, or absent. The exact criteria should follow the clinical question and available measurements; a label such as “liver toxicity” alone is not a reproducible rule.

Also specify how the analysis treats borderline and incomplete records. For example, a primary binary outcome might count definite and probable cases as events, classify clearly documented non-cases as no event, and leave records with insufficient evidence indeterminate. Indeterminate records should not silently be treated as unaffected. Their number and characteristics should be reported, with a sensitivity analysis when a plausible alternative classification could affect the result.

Reviewers should use the same evidence window and case rules for every participant. When feasible, they should be blinded to genotype and to the association result being evaluated. Blinding reduces the risk that knowledge of a genotype or hypothesis influences how ambiguous evidence is interpreted. It does not correct missing or inaccurate records, and it does not guarantee that reviewers will agree.

Agreement Is Not the Same as Validity

Two reviewers can agree because the records are clear, because they share the same interpretation, or because most records are non-cases. Agreement is useful evidence about reproducibility, but it does not establish that the resulting phenotype is clinically correct. If a validation standard is available, sensitivity is the proportion of true cases classified as cases, and specificity is the proportion of true non-cases classified as non-cases.

For a simple summary of two reviewers’ agreement beyond chance, Cohen’s kappa is \( \kappa=(P_o-P_e)/(1-P_e) \), where \(P_o\) is the observed proportion of matching classifications and \(P_e\) is the proportion expected to match from the reviewers’ marginal classification proportions. Kappa depends on the prevalence of classifications and the reviewers’ marginal rates, so it should be reported with the underlying counts rather than treated as a universal quality threshold.

Confirm Drug Exposure in the Risk Window

A prescription order establishes that treatment was intended, not necessarily that treatment was dispensed, administered, or taken. The appropriate evidence depends on the setting. An inpatient study may have medication-administration records; an outpatient study may use dispensing data, patient reports, or other documentation. Each source has limitations. A dispensing record does not prove ingestion, while an order alone may include people who never started treatment.

Define the exposure window in relation to the outcome window. If the endpoint is a toxicity that can arise only after treatment starts, an event documented before the first confirmed dose should not be counted as an incident on-treatment toxicity. Conversely, a gap in records should not automatically be interpreted as proof that exposure stopped. The protocol should state how uncertain starts, interruptions, and timing are handled. The eligible denominator must match the estimand: for example, risk among confirmed drug initiators is different from risk among everyone for whom a prescription was written.

How Outcome Misclassification Changes an Association

Let \(Y\) denote the true binary toxicity outcome and \(Y^*\) the recorded or adjudicated binary outcome. Within genotype group \(g\), define sensitivity \(Se=P(Y^*=1\mid Y=1,g)\) and specificity \(Sp=P(Y^*=0\mid Y=0,g)\). If sensitivity and specificity are the same across genotype groups, classification errors are independent of genotype conditional on the true outcome, and the outcome is binary, then the observed case proportion \(q_g=P(Y^*=1\mid g)\) satisfies:

where \(p_g=P(Y=1\mid g)\) is the true risk in that group. The first term counts true cases correctly classified as cases; the second counts true non-cases incorrectly classified as cases. Rearranging gives:

Subtract the expression for genotype group 0 from the expression for group 1. When the same \(Se\) and \(Sp\) apply in both groups, the constant term cancels:

The quantity \(Se+Sp-1\) is Youden’s index. If classification is better than chance, so \(Se+Sp>1\), this model scales the observed risk difference toward zero. This result is specific to the stated assumptions and the risk-difference scale. It does not justify saying that every kind of nondifferential misclassification always attenuates every effect measure. For example, risk ratios and odds ratios need not be changed by the same factor, and genotype-dependent sensitivity or specificity can distort comparisons in other directions.

If \(Se\) and \(Sp\) are known or estimated in a suitable validation sample and \(Se+Sp>1\), the equation can also be rearranged to estimate the true risk:

This is a correction of point estimates under a measurement model, not a substitute for validation. It can yield values outside the interval from 0 to 1 when the observed data and assumed error rates are incompatible, and uncertainty in the validation estimates should be carried into uncertainty intervals for corrected associations. The validation sample must provide relevant estimates of classification performance for the population and review process being studied.

Common Pitfalls and Reporting Priorities

• Treating the prescription as exposure. The wrong reasoning is that an order proves the patient took the drug. The correction is to define the exposure evidence required and use a denominator aligned with the target population.
• Using an outcome label without operational criteria. The wrong reasoning is that a diagnosis code or note label means the same thing in every record. The correction is to prespecify measurable or documented criteria and a rule for uncertain cases.
• Equating agreement with accuracy. The wrong reasoning is that high reviewer agreement establishes a valid phenotype. The correction is to report agreement separately from validation against an independent standard, when available.
• Assuming misclassification always biases toward no association. The wrong reasoning ignores the effect measure and the error process. The correction is to state the assumptions and scale, as in the risk-difference derivation, and consider differential errors explicitly.
• Letting genotype knowledge influence chart review. The wrong reasoning is that a plausible genotype makes ambiguous symptoms more likely to be a case. The correction is to blind reviewers to genotype when feasible and apply identical criteria across groups.

A clear report describes the case definition, timing rules, exposure evidence, reviewer training and blinding, disagreement resolution, and number of indeterminate records. If validation data exist, report how they were obtained and whether classification performance could differ across genotype groups or clinical settings. These details help readers judge whether the estimated association concerns the intended toxicity phenotype among the intended exposed population, rather than an artefact of inconsistent charting or exposure assignment.