Our study was nested within a trial and included relatively more patients from doctors with long consultations due to the time needed to obtain informed consent from patients in the waiting room, which affected the impact of trial interventions.[18] The concern is therefore twofold: whether the influence of slower consulting doctors or the trial interventions may change the estimates or alter the inferences, particularly for investigations. Controlling for duration of consultations did not, however, alter the estimates of the association between investigation and perceived medical need (odds ratios for slight and moderate or definite 2.87 and 25.23, respectively) perceived patient pressure (0.77 and 3.22 respectively), or patient preference (1.78 and 1.10, respectively). The trial intervention did not affect patients preference: not only were patients asked to compete the questionnaire before looking at the leaflets, but there was evidence that the leaflets did not alter patient preferences, with a similar number of patients wanting investigation with the leaflet (23%) or without the leaflet (25%).[18] The subset where patients were randomised is smaller (n=635) and therefore we would expect slightly different estimates. Nevertheless including randomisation groups (the two types of leaflets) in the model for the trial subset still did not materially alter the estimates or the inferences for patient preference for investigation (odds ratios for slight and moderate or definite 2.24 and 1.1, respectively) perceived patient pressure (0.91 and 4.21, respectively), or medical need (4.08 and 28.3, respectively). We found no effect of randomisation group for other outcomes either.

Influence of patient variables

We assessed the patient variables of being in paid work, age, sex, marital status, years of education, sickness disability, sickness certification, and practice deprivation. The main predictors for this dataset, after allowing for clustering by doctor, were: for investigation, being in paid work, sickness certification, and marital status; for prescribing, marital status; for referral, deprivation; and for examination, being in paid work and sickness certification. When including these variables in the models (which had the disadvantage of being smaller due to incomplete data in the post consultation questionnaire), the estimates of odds ratios changed by less than 25%, and most by less than 5%, with no change in the statistical inference—that is, minimal confounding. Taking investigation as the example, and using the same levels of none, slight, and moderate or definite, we estimated odds ratios with (and without) the potential confounders as, respectively: patient pressure 1.0, 2.15 (2.29), and 1.31 (1.37); medical need 1.0, 3.87 (3.71), and 20.0 (23.0); and doctor perceived patient pressure 1.0, 1.00 (1.03), and 4.76 (4.80).