Assignment Questions: Find any article using clustered data and describe: the unit of clustering; the hypothesized effects and the level at which the exposure is measured (is it a characteristic of the cluster or the observation within the cluster); and the statistical model used to estimate the effect. Describe whether there are any other statistical models that might be appropriate and whether they would be preferable (e.g., GEE vs mixed).
Article: Godwin M, Lamb M, Birtwhistle R, et al. A primary care pragmatic cluster randomization trial of the use of home blood pressure monitoring on blood pressure levels in hypertensive patients with above target blood pressure. Family Practice 2010; 27:135–142.
Unit of Clustering: The units of randomization for this study would be the physician practice (as eligible participating patients who received primary care from a family medicine physician were randomized to the same arm).
Intervention group: Group of patients with weekly BP home measurements, recording of home BP measurements, and review of home BP measurements with physician at office visits.
Control group: Group of patients with usual care (suggests BP monitoring at clinic visits).
Methods: Missing systolic and diastolic BP data from were imputed using a propensity score multiple imputation technique, the Markov chain Monte Carlo method. Imputation was carried out separately for different sexes and intervention groups. Using 10 imputed datasets, multilevel modeling was applied to analyze each of the BP measurements at each follow-up visit (Month 6 or 12) to compare the intervention with control taking into account the cluster (physician practice) samples of patients. Results from these 10 imputed datasets were combined and summarized in tables. With the same 10 imputed datasets, multilevel modeling was extended to include the baseline BP measurement as a covariate and combine the longitudinal data (two follow-up visits), so that the effect of using home BP device can be assessed taking into account the baseline and time variations. The primary outcome in this study was mean awake BP on ambulatory monitoring at 6- and 12-month follow-up and the secondary outcomes were mean BP on full 24-hour ambulatory blood pressure monitoring, mean sleep BP on 24-hour ambulatory blood pressure monitoring and BP on the BpTRU automated device, all at 6- and 12-month follow-up.
Results: Home BP monitoring did not improve BP compared to usual care at 12-month follow-up: mean awake systolic BP on 24-hour ambulatory blood pressure monitoring [141.1 versus 142.8 mmHg, mean difference 1.7 mmHg; 95% confidence interval (CI) –0.6 to 4.0, P = 0.314] and mean awake diastolic BP on ABPM (78.7 versus 79.4 mmHg, mean difference 0.7 mmHg; 95% CI –7.7 to 9.1, P = 0.398).
Alternative Model: One suggestion could have been to utilize a mixed-effects linear and logistic regression to adjust for clinics (physician practice) as the unit of randomization. Adherence to home measurements (or “tele-monitoring”) could have been defined as a binary variable by week, such that a week received a “1” if at least a certain amount of BP readings had been taken that week, and a “0” if less than a certain amount of BP readings had been taken that week. The binary outcomes for each week could then be summed for each participant and then divided by the total weeks enrolled in the study to obtain a continuous proportion of weeks’ adherent to the intervention for each participant. Other variables could also have been collected (e.g. age, sex, race, income, education, marital status, and employment status). One could then conduct bivariate mixed-effects models to test adherence to home measurements, followed by multivariable mixed-effects linear and logistic regression models with all the above stated covariates included in each model to predict adherence to tele-monitoring (continuous).
