Table 1 Description of the challenges and potential biases faced in this study

(1) Non-randomised design

Patients were not randomised as this was a scheduled implementation of an evidence-based intervention [6]. In some practices almost all patients on the practice hypertension register were offered the intervention, but other practices adapted the implementation strategies to concentrate initially on various sub-groups e.g. patients with poorly controlled hypertension; those of working age who were more likely to find surgery visits difficult; low risk patients; or those assessed as being more able to manage the system. As a result, the non-participating patients (comparator group) were systematically different at baseline from those in the telemonitoring group.

(2) White coat effect

Surgery readings were likely to be affected by ‘white coat effect’ [9], whereby BP tends to be higher in clinical settings compared to home settings. This causes confounding bias when seeking to determine the difference between readings taken by the telemonitoring group at home and readings taken in the comparator group in the surgery.

(3) High variability in the frequency of readings

Surgery readings were often recorded much less frequently than home readings, raising the possibility of a type of ascertainment bias whereby raising or lowering of BP is much more easily identified for those in the telemonitoring group. This is related to the problem highlighted by Goldstein (2020) whereby data may be collected at different rates or not at all, and therefore any missing data may be informative of underlying health status [1]. Patients in the Scale-up BP study were on different protocols for how frequently they should measure their BP and (probably for clinical reasons) these varied over time. For example: less frequent protocols would be required when BP is stable than when adjusting treatment to improve control. Adherence may also have varied over time.

(4) Contamination of readings

We observed that home readings are sometimes transcribed by general practitioners into practice systems which are thus indistinguishable from surgery readings, [6] making comparison of apparent surgery readings prone to error. Although this was likely to have occurred more frequently in the telemonitoring group, it could also have occurred in the comparator group (e.g. if comparator patients had access to home reading systems). In any case, this would have had the effect of diluting the intervention effect.

(5) Regression to the mean

Reductions in BP over time were likely to be affected by regression-to-the mean in both groups, but particularly in the telemonitoring group as patients were prospectively selected for telemonitoring (perhaps based on their level of BP control).

(6) Measurement error

Individual BP readings will be measured with measurement error such that they may deviate from their true value. This is expected to increase the overall variability of BP measurements.

(7) End digit preference

We had discovered that a limited level of end digit bias was present in home telemonitored readings [10], however we were unsure as to how large this effect would be in surgery based readings.

(8) Withdrawal bias

A small proportion of patients (7%) dropped out of the telemonitoring arm, and there were missing data across both groups. We found that telemonitoring patients with higher starting systolic BP were more likely to drop out [6]. The reason for this is unknown.