We anticipated that passive dissemination of guidelines would have little effect on the outcomes of interest [12
] and therefore was an appropriate control intervention. We decided that randomizing practices rather than physicians or patients would be the most feasible design for practical reasons, and also the most appropriate to avoid contamination from the intervention to the control group.
The methods have been described in more detail in the study protocol (Protocol S1
The study participants were all general practices in two geographically defined areas of Norway; the practices all used one or the other of two eligible electronic medical record systems.
Patients started on medication for hypertension or hypercholesterolemia during the study period and all patients already on treatment that consulted their physician during the trial were included in the analyses. The eligibility criteria for patients are described in more detail under “Outcomes.”
We developed our intervention through a process of identifying barriers to implementation of the recommendations and tailoring the intervention to address these. This is described in more detail elsewhere [19
]. Box 1
shows the various elements of our intervention.
The intervention was initiated through an educational outreach visit carried out between May and December 2002 by pharmacists recruited and trained specifically for this purpose.
During the outreach visit the main elements of the guidelines were presented, with special emphasis on cardiovascular risk estimation, choice of first-line drugs for hypertension, and treatment goals. A printed copy of the guidelines and a one-page version were given to the physicians, including a chart to aid the estimation of cardiovascular risk.
A software package was installed during the visit. This enabled us to extract data and immediately, during the visit, present the physicians with data on their performance of risk estimation, choice of antihypertensive drugs, and achievement of treatment goals (audit and feedback). The software package also included computerized reminders (“pop-ups”) on the computer screen. These were triggered at the patient's first visit following a recording of an elevated blood pressure (>140/90 mm Hg) or cholesterol level (total cholesterol > 5 mmol/l [190 mg/dl] or low-density lipoprotein [LDL] cholesterol > 3 mmol/l [115 mg/dl]).
If the patient had not been prescribed blood-pressure- or cholesterol-lowering drugs, the physician was reminded of the recommendation to carry out cardiovascular risk assessment and was given the option of starting a computer program for risk assessment. Recommendations on choice of drugs were also given, and the physician was given the choice of printing out patient information material.
If the patient was already on blood-pressure- or cholesterol-lowering drugs, the pop-up reminded the physician of recommended treatment goals and asked if the physician would like to print out patient information material.
Within 3 d after the outreach visit, a member of the research team called the clinic to confirm that they were not experiencing problems with their computers as a result of our visit.
The doctors who were invited to participate in the study were given information about the objectives of the study and the practical impact it might have on their practice. We obtained written consent from all practices. We submitted the research protocol to the Regional Committee of Research Ethics, which considered ethical approval unnecessary.
The Norwegian Data Inspectorate approved the handling of the data.
We chose three main outcome measures, all aimed at physician behavior regarding the pharmacological management of primary prevention of cardiovascular disease for the 12 mo following the outreach visit. All outcomes were recorded at the patient level. Because the eligibility criteria varied from one outcome to another, the number of patients included in the analysis varied across outcomes. Baseline data for the 12 mo preceding the intervention were also collected.
Patients with established cardiovascular disease were excluded, with the exception of the outcomes related to treatment goals for lipid-lowering therapy, since these are the same across patient groups. Because antihypertensive drugs are also prescribed for the treatment of thyrotoxicosis and migraine, we excluded data from patients with these diagnoses. We used extracted medical record data on prescribing and diagnoses to identify patients who were to be excluded. For example, patients who had been given a prescription for nitroglycerin were assumed to have coronary heart disease, and thus were excluded from most analyses.
Patients were considered to be previously untreated if they had hypertension (blood pressure > 140/90 mm Hg) or hypercholesterolemia (total cholesterol > 5 mmol/l [190 mg/dl] or LDL cholesterol > 3 mmol/l [115 mg/dl]) but no prescription for the corresponding medication had been recorded for 24 mo preceding the outreach visit.
The primary outcomes were the following: (1) the proportion of patients prescribed thiazides among patients prescribed antihypertensive drugs for the first time, (2) the proportion of patients with a cardiovascular risk assessment among all those started on antihypertensive or cholesterol-lowering treatment (excluding patients already on either type of medication), and (3) the proportion of patients with a recorded level of cholesterol (total or LDL) or blood pressure satisfying the specified treatment goals among all patients on the corresponding treatment for at least 3 mo. For cholesterol we also included patients on secondary prevention since the treatment goals are similar.
The secondary outcomes, as prespecified in the research protocol (Protocol S1
]), were the following: (1) the proportion of patients reporting that they were involved in the decision-making process before drug treatment for hypertension and/or elevated cholesterol was started, (2) the level of risk among patients started on treatment, (3) the proportion of patients with risk above 20% among those started on treatment, (4) the level of risk among patients not started on treatment for whom blood pressure and cholesterol level were recorded, (5) the proportion of prescriptions of thiazides or beta-blockers to patients who were prescribed antihypertensive drugs for the first time, (6) the proportion of prescriptions of angiotensin receptor blockers or alpha-blockers to patients who were prescribed antihypertensive drugs for the first time, (7) for patients with diabetes, the proportion of patients with a recorded level of cholesterol (total or LDL) or hypertension satisfying the specified treatment goals among all patients on the corresponding treatment (for cholesterol we also included patients on secondary prevention since the treatment goals are similar), (8) the proportion of patients reaching the specified treatment goal for blood pressure, and (9) the proportion of patients reaching the specified treatment goal for cholesterol level.
All outcomes were calculated from data extracted from the practices' medical record systems, with two exceptions: (1) the use of cardiovascular risk assessment tools by physicians before starting medication and (2) the level of patient involvement. Patients potentially eligible for inclusion in these analyses were identified from the medical record data, after which we interviewed the prescribing physician (or a colleague) per telephone. We enquired about the use of risk assessment tools for each patient they had started on treatment during the intervention period, and the physicians gave their answers based on notes from the medical records. The physicians assisted us by sending a questionnaire to the patients who had been started on medication. The questionnaire consisted of one question, asking to what extent the patients felt that they had taken part in the decision to start drug therapy. The answer was given on a five-point scale, from “none” to “fully.” Only patients reporting that they took no part at all were counted as not being involved in the decision to start treatment.
For two outcomes—(1) the proportion of patients for whom cardiovascular risk had not been estimated and (2) the level of risk among patients not started on treatment—the analysis was based on a random sample of eligible patients.
During interviews with physicians it was often possible to figure out whether a practice was in the intervention group. Investigators assessing outcomes and conducting analyses were blinded to the allocation of practices.
All analyses were by intention to treat. In response to comments from peer reviewers, we decided to deviate from our protocol and use adherence, rather than non-adherence, to recommendations as outcome measures. This had no impact on our findings.
In order to demonstrate a 25% relative reduction in non-adherence with the guidelines—with a power of 80% and a statistical significance level of 5%—in outcome measures between the control and intervention groups, we estimated that we needed a sample of approximately 140 practices in total (Cluster Randomisation Sample Size Calculator version 1.0.2, Health Services Research Unit, University of Aberdeen, Aberdeen, United Kingdom). The adjusting factor (intracluster correlation coefficient) was conservatively estimated to be 0.2, based on data from a previous study [20
]. More detail is available elsewhere [18
Block randomization was done within two geographical areas (Oslo and Tromsø) to ensure balance in the number of practices in the intervention and control groups. The size of the blocks varied randomly between two, four, and six. A colleague not directly involved in our research project generated the allocation list using software from http://www.randomization.com
. We gave her identification numbers representing each recruited practice, and she informed us whether the practice was allocated to the intervention or control group.
The generalized estimating equations (GEE) approach was used for the analysis of the binary outcomes, and the mixed effects linear regression model (a two-stage nested analysis of variance) for the continuous outcome (risk level), using the baseline log odds for compliance and the baseline mean risk level for each practice, respectively, as covariates [21
]. The analysis was performed using PROC GENMOD and PROC MIXED in SAS 9.1.3 (SAS Institute, Cary, North Carolina, United States).
The outcome from a GEE approach is an odds ratio; the adjusted relative risk was calculated from the odds ratio by using the overall proportion of compliance in the control group [22
Originally we planned to use a parametric method, which depends on an underlying distribution of the sample observations. This is not the case for the GEE approach. We therefore used the GEE approach rather than the method we had specified in the protocol. This change was made before we analyzed the data.