Printed educational messages fail to increase use of thiazides as first-line medication for hypertension in primary care: a cluster randomized controlled trial [ISRCTN72772651]

Two types of PEMs addressed the identified evidence-practice gap: a two-page article, indistinguishable from the rest of the newsletter in size and style (the “insert”) and two versions of a short, directive, evidence-based PEM on a postcard-sized card stapled to the front page of informed (the “outsert”). The insert and one of the versions of the outsert (the atheoretical outsert) were developed without any explicit theory of action using input from a diverse group of physicians who identified barriers to evidence-based practice and from a communications expert. The second outsert (the theory of planned behaviour-based outsert) was developed by a group consisting of three health psychologists and two implementation researchers with experience using theories of behaviour. Full details about the development process for this outsert are described in the study protocol [13]. The addition of the “theory of planned behaviour-based” outsert allowed us to test the hypothesis that a message inspired by a psychological theory, specifically the theory of planned behaviour [14] (TPB), would be more effective in changing clinical behaviour toward more evidence-based practice than a message designed without an explicit theoretical basis. The study design is shown in Table 1. Table 2 shows the two versions of the outserts, and the insert is included in Additional file 1. The interventions were included with the July 2005 edition of informed.

Ontario has a single-payer public health insurance system, in which necessary medical and hospital care and prescription drugs are covered for all Ontario residents aged 65 years and older. The following administrative data sources from this insurance plan were used [15, 16]:

Records from these datasets were linked, using unique, encoded patient and physician identifiers and analysed at the Institute for Clinical Evaluative Sciences (ICES; www.​ices.​on.​ca), to determine which individuals were newly treated for hypertension during the study period, and which of the target physicians prescribed their antihypertensive medication.

All Ontario physicians with an active general/family practice in Ontario in 2003/2004 were eligible for inclusion. “Active” practice was defined as having a total billing volume for the year of at least $50,000 and writing prescriptions for at least 100 different patients (aged 65 years and older), with at least one prescription in at least 10 of the 12 months.

Physician identifiers were linked to the College of Physicians and Surgeons of Ontario (CPSO) number at ICES. CPSO numbers were then hand-linked to the publicly available CPSO database (www.​cpso.​on.​ca) to obtain practice addresses.

Trials of interventions aimed at changing clinical practice must be randomized at the level at which they are directed—in this case, the family physician. In a group practice, doctors may share information. To prevent contamination, we randomized at the level of the practice. Physicians were placed into practices on the basis of a shared address.

The Ontario Drug Benefit database was used to identify Ontario residents who filled one or more prescriptions for an antihypertensive medication between July 11, 2005, and July 24, 2006. To ensure that these individuals were started on an antihypertensive agent for treatment of hypertension and not another indication, the cohort was linked with the CIHI DAD, the RPDB and the ODB Database. We excluded those patients with OHIP claims within 3 years or CIHI claims within 4 years (or any antihypertensive medications prescribed within 1 year) for any of the following conditions: myocardial infarction or angina, heart failure, arrhythmias, renal disease (including nephropathy), liver disease (including oesophageal varices), stroke or transient ischemic attack, hyperthyroidism or migraines [17]. Consistent with guideline recommendations for treatment of hypertension [18], we focused on thiazide prescribing as the first-line treatment in individuals with uncomplicated hypertension. In order to ensure that the cohort included only individuals newly treated for hypertension, individuals who had filled a prescription for any antihypertensive medication in the year prior to the intervention were also excluded [17]. Only the first prescription for an antihypertensive filled during the observation year was included; all prescriptions filled on the same day were taken into consideration when determining the outcome. While data were available for individuals 65 years and over, all individuals included in the study were at least 66 years old at the time they filled their first prescription for an antihypertensive medication in order to ensure that information covering a 1-year look back period for prior prescriptions was available.

Using the OHIP physician claims database, we identified the physician seen during the 14 days prior to and including the date the antihypertensive prescription was filled. In some cases, more than one study physician was selected. (Patients who had not visited any of the study physicians during the 14-day period were not included in the analysis.) We included patient visits to the physician’s office, physician visits to the patient’s home or to a patient in a long-term care facility and physician phone calls to the patient.

The study was a pragmatic, factorial, cluster-randomized controlled trial with the physician practice as the unit of randomization. Practices were randomly assigned to one of six intervention groups by the study statistician (see Table 1), using computer-generated random numbers. Patient and physician participants were unaware of allocation and administrative data were collected without knowledge of the research under way. Full details of the study design can be seen in the published protocols [13, 19, 20].

The objectives of the study were to determine whether the format of the PEM affected the likelihood that the only medication initially prescribed for hypertension was a thiazide and whether a theory of planned behaviour-based approach to developing the message was any more effective than a PEM developed without use of explicit theory. A “successful” outcome was a first-line prescription for a thiazide and no other antihypertensive medication. Combinations of a thiazide plus potassium or a potassium-sparing medication were counted as a “success”; combinations of a thiazide plus another diuretic were counted as a “failure”. A prescription for any non-thiazide antihypertensive medication was also counted as a “failure”.

Even a 5 % improvement in the prevalence of care for common conditions represents a meaningful health benefit. Based on pilot data, Monte Carlo simulations, assuming an intracluster correlation coefficient of 0.092, three patients per physician, a baseline success rate of 0.36 and absolute intervention effects of 0.05, 0.075 and 0.10, demonstrated that a trial with 1250 practices per arm would provide over 97 % power to detect a 5 % increase, and over 98 % power to distinguish between the effects of the combined intervention and either alone, assuming the combined effect to be additive. This original power calculation was based on testing the effectiveness of the insert and one version of the outsert. We conducted a follow-up simulation to assess whether we could test a second version of the outsert (theory of planned behaviour-based), which suggested that the power would be greater than 80 %. The research opportunity that this modification presented was judged to outweigh the loss of power. These decisions were made during the design phase and prior to randomization, as detailed in our published protocols [13, 19, 20].

Logistic regression was used to test the hypothesis that the intervention affected the choice of first-line medication for hypertension. While the unit of randomization was the physician practice, outcomes were measured at the individual patient level. The logistic regression model was estimated using generalized estimating equations (GEE) to account for the clustering of patients within physician practices [21]. This method of analysis allows for the inclusion of patient-level and physician-level covariates (e.g. patient age, physician age), while at the same time accounting for possible correlations amongst patient outcomes within a practice.

Analysis was on an intention-to-treat basis. As shown in Table 3, models were fit to test: insert and both outserts combined, and insert and two outserts split (theory of planned behaviour-based vs. atheoretical). These were fit using unadjusted (regressions 1a and 2a) models and models adjusting for patient- and physician-level covariates (regressions 1b and 2b). The full model in 1b included interaction terms between the intervention and each covariate, to determine whether the impact of the intervention depended on patient and/or physician characteristics.

Patient and physician characteristics were compared between randomization arms using chi-square tests for categorical variables and Kruskal-Wallis tests for continuous variables.

Figure 1 shows the number of physicians and patients included in the study. Three quarters of Ontarians aged 66 years and older filled a prescription for an antihypertensive medication in the year following the mail-out, but almost all (92 %) had already filled a prescription for an antihypertensive medication during the preceding year. A further 38 % of those who filled a prescription for the first time were excluded because there was evidence that the antihypertensive might have been prescribed for a reason other than hypertension. We identified 38,102 individuals who were taking medication for hypertension for the first time and were able to link the prescriptions of 23,508 of these individuals to one of the physicians targeted by the intervention (Fig. 1). Thirteen percent of family physicians randomized to receive one of the interventions were excluded from the analysis because they were not linked to at least one patient newly treated for hypertension.

There were small, statistically significant but clinically unimportant, differences between the characteristics of the physicians in the six intervention groups (Table 4).

The intracluster correlation coefficient was 0.18 (95 % confidence interval 0.16 to 0.19). Intervention effects are shown in Table 3. Neither the unadjusted nor the adjusted results show evidence that any of the interventions, alone or in combination, were effective. The widest confidence interval reported in the table, an odds ratio between 0.91 and 1.21 for the insert + TPB-based outsert corresponds to a true absolute effect of the intervention lying between a decrease of 1.5 % and an increase of 3.5 %. Thus, not only did the intervention fail to achieve statistical significance but as well the confidence interval does not contain values of much practical importance.