An Evaluation of Different Strategies for Sampling Controls in an Online Case-Crossover Study of Acute Flares in Knee Osteoarthritis

Objective: To evaluate bias and precision of exposure-outcome effect estimates from three control sampling strategies in a case -crossover study. Methods: Online case-crossover study investigating eight physical activity-related triggers for acute flares in knee osteoarthritis. Exposures were measured in hazard periods (<= 24 hours before self-declared flare onset). Control period exposure was measured in three ways: (1) four scheduled questionnaires over 13-weeks, (2) "usual" physical activity levels ascertained at baseline, (3) over three days before flare onset. Derived odds ratios, 95% confidence intervals and standard errors were compared. Results: Of 744 participants (mean age 62.1 [SD 10.2] years; 61% female), 493 reported 714 flares. Selecting controls from scheduled questionnaires, independent of hazard periods, yielded predominantly odds ratios in the expected direction (exposure "a lot" versus exposure "not at all", range: 0.57-3.22). When controls were sampled at baseline (range: 0.01-1.42) or immediately before a flare (range: 0.30-1.27) most odds ratio estimates were inverted. Standard errors of the log odds ratios were smallest when controls were sampled from scheduled questionnaires (range: 0.264-0.473) compared to controls sampled at baseline (range: 0.267-0.589) or immediately before a flare (range: 0.319-0.621). Conclusion: Our findings are sensitive to control sample selection. Under certain conditions, different patterns could be attributed to over reporting and social desirability bias, where people may want to present themselves more positively about their "usual" physical activity levels, at baseline. Exposure measurement at the time of a flare may be less precise and more susceptible to recall bias due to systematically reporting exposures differently during a flare, compared to control measurement independent of flares.