Statistical considerations

The statistical power is estimated for the intention-to-treat analysis and for the endpoints HbA1c and glucose control. We target the sample size of 1000 patients with complete endpoint data, which corresponds to 1429 recruited under 30% rate of loss to follow-up, as seen at CADA under routine care. This sample yields 80% power to detect at least 0.62 percentage points reduction in mean HbA1c in the intervention arm, based on the assumption that the SD of HbA1c in both arms is 3.5%, and 8 percentage points effect of the intervention on achieving glucose control, assuming that in the control arm, the proportion of patients achieving glucose control is 24%.

Trial endpoints are going to be analysed by means of regressions, in the intention-to-treat framework. The primary endpoint of HbA1c is going to be analysed with linear regression, adjusting for patient’s HbA1c at baseline, gender (the stratification variable) and for age and socioeconomic status, known from practice at CADA to be strongly predictive of the outcome. The main measure of the intervention effect is going to be the difference in mean HbA1c between study arm, with its uncertainty quantified by a 95% CI. Unadjusted difference of means with its 95% CI is going to be reported as well. Given the size of the trial, we do not expect significant baseline imbalances in any measured covariate; if any such imbalance occurs, we will perform a sensitivity analysis where such covariates are adjusted for in addition to the above.

The binary secondary endpoints (ie, glycaemic control and follow-up completion) will be analysed with logistic regressions, with adjustment for gender, age and socioeconomic status, yielding ORs. Unadjusted ORs and risk ratios and their 95% CIs will be reported as well.

The DASS-21 and health-related quality of life scores will be analysed with linear regression, adjusting for baseline values, gender, age and socioeconomic status. As these outcomes are scales, we will examine the appropriate linear regression diagnostics carefully to ensure that the assumptions of this method are satisfied well enough. In case of serious violations, we will fall back on ordinal or logistic regression, using clinically meaningful cut-off points for the scales.

The hypothesis of modification of the intervention effect by social capital will be addressed by introducing an interaction term between the arm indicator and the social capital score in the primary analysis. The hypothesis will be accepted if the 95% CI for the interaction coefficient excludes 0.

We anticipate minimal missing data at baseline, but expect up to 30% of trial participants to be lost to follow-up. If thanks to the additional outreach to all participants the actual attrition is less than 10%, we will perform complete case analysis only. Otherwise, we will perform inverse probability weighting using a predictive model of attrition built with baseline covariates. If no sufficiently strong baseline determinants of loss to follow-up are found, we will explore alternative approaches that are appropriate for the data as collected, and informed by the trial staff’s understanding of how and why the participants have been lost.

In order to improve participants’ adherence to the protocol, costs related to transport and HbA1c testing at baseline, 3-month and 12-month follow-up will be covered by the study as motivational incentives.