Keele Research Repository

Background
Prognosis research studies (e.g. those deriving prognostic models or examining potential predictors of outcome) often collect information on time-varying predictors after their intended moment of use, sometimes using a measurement method different to that which would be used. We aimed to illustrate how estimates of predictor-outcome associations and prognostic model performance obtained from such studies may differ to those at the earlier, intended moment of use.

Methods
We analysed data from two primary care cohorts of patients consulting for non-inflammatory musculoskeletal conditions: the Prognostic Research Study (PROG-RES: n = 296, aged >50 years) and the Primary care Osteoarthritis Screening Trial (POST: n = 756, >45 years). Both cohorts had collected comparable information on a potentially important time-varying predictor (current pain intensity: 0–10 numerical rating scale), other predictors (age, gender, practice) and outcome (patient-perceived non-recovery at 6 months). Using logistic regression models, we compared the direction and magnitude of predictor-outcome associations and model performance measures under two scenarios: (i) current pain intensity ascertained by the treating general practitioner in the consultation (the intended moment of use) and (ii) current pain intensity ascertained by a questionnaire mailed several days after the consultation.

Results
In both cohorts, the predictor-outcome association was substantially weaker for pain measured at the consultation (OR (95% CI): PROG-RES 1.06 (0.95, 1.18); POST 1.04 (0.96, 1.12)) than for pain measured in the questionnaire (PROG-RES 1.34 (1.20, 1.48); POST 1.26 (1.18, 1.34)). The c-statistic of the multivariable model was lower when pain was measured at the consultation (c-statistic (95% CI): PROG-RES 0.57 (0.51, 0.64); POST 0.66 (0.62, 0.70)) than when pain was measured in the questionnaire (PROG-RES 0.69 (0.63, 0.75); POST 0.72 (0.68, 0.76)), reflecting the lower OR for pain at the consultation.

Conclusions
Prognostic research studies ideally should measure time-varying predictors at their intended moment of use and using the intended measurement method. Otherwise, they may produce substantially different estimates of predictor-outcome associations and model performance. Researchers should report when, how and where predictors were measured and identify any significant departures from their intended use that may limit the applicability of findings in practice.