Patient-reported outcomes
Why they matter
Many of the things treatment is for, pain, breathlessness, fatigue, function, are only observable to the patient. A trial that measures only physiological endpoints has not measured what the patient came for, and there is often a poor correspondence between the two.
Requirements that are frequently skipped
- The instrument is validated in this population and this condition, with the validation cited rather than asserted.
- The PRO is named as a primary or secondary outcome in the protocol, with a hypothesis, not collected and reported selectively.
- The scoring is prespecified, including which domains of a multidimensional instrument are primary.
- A minimal clinically important difference is stated in advance, so the result can be interpreted rather than only tested.
- The plan for missing responses is explicit, with the mechanism named and a sensitivity analysis over plausible departures.
Missingness is the analysis problem
Completion falls with time and with illness, and both are informative. A mixed model handles intermittently missing responses under a missing-at-random assumption; whether that assumption is plausible is a clinical judgement about why people stopped answering. Death complicates it further, since a questionnaire score after death is not missing, it is undefined, and models that impute one are answering a question about a population that does not exist.
Multiplicity
Instruments produce many domain scores, and testing them all produces findings. Nominate the primary domain in advance and treat the rest as descriptive, or adjust.
What to watch out for
Missing data are the defining problem, and they are rarely missing at random: patients who feel worst are the ones who stop completing questionnaires, so a complete-case analysis of a symptom score is biased toward the treatment looking better than it is. Instrument choice is the second problem: a scale validated in one condition may not measure the same construct in another, and a scale assembled for a particular trial has usually not been validated at all.
Clinical example
A trial reports improved quality of life at 12 months from a validated instrument, with 62 percent completion in one arm and 78 percent in the other. The differential completion is itself a finding, and any analysis of the responders alone is comparing two selected groups.
- Reporting of patient-reported outcomes in randomized trials: the CONSORT PRO extension (opens in a new tab) · JAMA · doi 10.1001/jama.2013.879
- Measurement of health status. Ascertaining the minimal clinically important difference (opens in a new tab) · Control Clin Trials · doi 10.1016/0197-2456(89)90005-6