Statistical analysis

Our sample size calculation ensured adequate numbers of self-reported flares to sufficiently power the primary analysis identifying potential triggers of knee OA flare¹⁵. The target sample was 434 participants experiencing a flare. Where available, we compared sociodemographic characteristics of participants of responders and non-responders, and response rates by recruitment method. Summary descriptive statistics were derived for participant demographics and self-reported knee characteristics, as well as features of flares, symptoms, and consequences during flares. Baseline data were collected for left and right knee. For descriptive knee characteristics, the worst knee was selected and reported based on highest ‘average pain’ score on a 0–10 numerical rating scale (NRS). If scores were equal, the knee with the highest ‘worst pain in the last week’ followed by the highest ‘current pain’, followed by random selection were selected. If left, or right knee score was missing, the available knee score was included.

With each participant acting as their own control in the analysis, reported trigger exposures in the hazard periods were compared to those reported in the control periods, using conditional logistic regression with m:n matching, so each participant could potentially contribute multiple hazard and control periods³⁶. All available controls were utilised and may have occurred before and/or after the hazard period. Odds ratios (OR) with 95% confidence intervals (CI) were derived. Time trends in exposure were examined by comparing exposure frequency across the Scheduled questionnaires at weeks 1, 5, 9 and 13, and before and after the flare. A sensitivity analysis of the primary analysis was restricted to flares that had control periods occurring before and after the hazard period. To explore the exposure induction period, trigger exposure status was compared for day of completion/flare and the 3 days earlier. The proportion of flares deemed unexpected was reported. Sensitivity analyses were conducted by restricting analysis to first flare per participant (to examine carryover effects for multiple flares), flares reported within 3 days of flare onset (to reduce the potential impact of delayed recall), flares reported after baseline (i.e., excluding people who reported to be experiencing a flare at study entry, which could influence how initial questions were answered), spontaneous self-directed flares reported without prompt (i.e., not at Scheduled questionnaire time points¹⁵), and participants recruited from GP registers (for whom study eligibility was objectively confirmed).

Analysis was restricted to participants who reported Event-Driven flare questionnaires and indicated via Daily questionnaires their flare had ended. Unresolved flares were excluded from the analysis as the distribution of flare duration was more skewed (median [interquartile range] 19 [5, 87] days) than resolved flares (6 [3, 11] days). Flare characteristics, and daily course and consequences (pain, bothersomeness, medication use in last 24 h) were analysed descriptively. The Kaplan–Meier curve for time to flare resolution was derived. Effects of covariates on time to flare resolution were assessed using accelerated failure-time model, with generalised gamma survival distribution yielding unadjusted time ratios and 95%CI.

Analysis was restricted to participants who completed at least two Scheduled questionnaires: the prespecified minimum amount of engagement with study follow-up required for inclusion. Poisson regression with robust standard errors examined the effect of selected baseline participant or knee symptom characteristics on the number of reported flares over the 13-week follow-up period. Unadjusted incidence rate ratios (IRR) and 95%CI were calculated.

All analyses were conducted using STATA V.15.0 (Stata Corporation, Texas, USA).