This study showed that this "Hypertension Programme" was more effective than "Usual care" at a relatively small incremental cost. The base case result of ICER 1,124 Int$/LYG is highly cost-effective in our local context. Moreover, in 43% of 100,000 iterations performed in the probabilistic sensitivity analysis, "Hypertension Programme" was dominant (more effective and less costly). Overall, in 95% of cases, the programme was cost-effective.
This is the first study to include all the following aspects: original (short term) effectiveness data based on a primary source; hard outcome measurements; a long-term analysis; and a probabilistic sensitivity analysis. A literature review of four previous studies showed a combination of some methodological limitations in all of them: short-term analyses[21, 24]; intermediate outcome measures[21, 23]; a model based entirely on secondary sources; or a biased sensitivity analysis.
In our model, the major determinant of uncertainty was the discount rate used. In general, benefits of hypertension treatments are seen several years after their start. As a result, the bigger the discount rate used, the lower the final benefit obtained. This is a common problem when considering cost-effectiveness of prevention programmes. Even though different discount rates produced different outputs, they would not significantly alter decision-making (see additional file 2: graphic S1). Nevertheless, a minimum 10 year-time horizon is needed.
The probabilistic sensitivity analysis evaluated uncertainty from all variables related to costs and outcomes used in the model. For example, even though we had exact costs for drug consumption (based on individual patients' drug purchase), we also included a variable for percentage of drug coverage by the payer. This variable tried to capture different economic burdens according to the percentage of coverage provided.
Regarding variables on transition probabilities for events and outcomes, we checked consistency of local and international data before fitting the model. We worked with different sources of systolic blood pressure levels[48, 49] to try to detect possible differences in risks that could change outcomes in the model. Subtle differences among different data sources did not affect original cardiovascular risk probabilities.
Mortality data from cardiovascular events were taken from the same populations used to fit other probabilities in the model[39–42]. Given the lack of data regarding 1 year-mortality of untreated cardiovascular events, we decided to adjust these probabilities using national mortality tables (adjusted for age and cause of death) and observational studies [46, 47, 50] and to explore the range in the sensitivity analysis.
Our study's results are not directly comparable to previously published works[21, 23, 24] because they did not evaluate hard outcomes and/or have a long-term perspective. On the other hand, the German study used a model that could allow broad comparisons. In general it can be said that they had findings similar to ours. This helps to corroborate results from both studies.
Of note, basal hypertension control in the usual care group from the study used to fit the model was high -60.4%- and mean basal blood pressure was 135/75 mm Hg. In other settings, were basal control of hypertension is lower or the mean basal blood pressure is higher, a greater difference in effectiveness would be expected. For example, compared to a general elderly population in Argentina, the incremental effectiveness of "Hypertension Programme" would have been 1,22 LYG.
Even though the incremental effectiveness was relatively low for each patient, the model evaluated the effect of both types of hypertension treatment in all hypertensive patients in our population. Considering the impact of the programme in the 30,000 hypertensive patients in our setting, a total of 5,400 life years could be gained.
The model did not consider specific adverse events related to hypertension treatment for two reasons: 1) In previous studies, it was found that first-line anti-hypertensive drugs do not have more side effects than placebo; and 2) to avoid double counting, because eventual costs and consequences of adverse events in hypertension treatment would be captured by the methodology used.
This study had some limitations. First, the effectiveness study used to compare treatment strategies was not a randomized controlled trial. It was impossible to perform one in our setting because of organizational restrictions (i.e. that could not prevent contamination of interventions between study groups). Nevertheless, "Hypertension Programme" and "Usual Care" groups had similar basal hypertension control in the originally published study, as mentioned above. It did not have major methodological flaws, and its results were consistent with other studies[15–18]. Second, the study's perspective was not societal. Third, outcomes did not capture quality of life. Time-frame and budgetary restrictions prevented us from considering resource use from a societal perspective or from assessing utility measures during the original study. In addition, after the success of this demonstration study, all patients were treated according to the Hypertension Programme, precluding us from assessing any actual difference between groups. Given that the aim of the study was to inform decision makers from a third-party payer, the perspective adopted is all right. Nevertheless, a societal perspective might have given useful information and allowed analysis from other sectors. The lack of consideration of health state values (e.g. through QUALYs, etc.) is an important limitation. It is not possible to predict a possible influence of this fact. Effectiveness could have been lower (for example stroke survivors would have contributed with less than one QUALY per each year survived), but the bigger proportion of patients without cardiovascular events in the "Hypertension Programme" group could have summed more QUALYs overall. Thus, it would be interesting to address this important issue with a specific study designed "ad hoc" (to assess this effect). Fourth, only the effect of (different options of) hypertension treatment was evaluated. We chose this approach because our original experience only considered hypertension treatment. Nevertheless, a more integral approach, considering also treatment of other cardiovascular risk factors could have been adopted. This would have probably increased the effectiveness seen. Fifth, the study was based on urban populations from middle income and high income countries. Results should not be extrapolated to rural or low income populations. Finally, uncertainties inherent to the model were not explored. Because of the time-horizon chosen, it would have been impossible to avoid the use of a model, although different assumptions could have been made.
Notwithstanding, the study had several strengths. First, data on costs and effectiveness -intermediate outcomes- used to fit the model were local and at patient-level. Costs were evaluated in detail, and its real distribution was fitted in the model. Second, resources used were informed in appropriate physical units and valued in International Dollars to favour comparisons in other settings. Third, consistency of local and international data on events and mortality was checked before fitting the model. The slight differences observed did not modify model results. Fourth, a hard-outcome measure -mortality- was used. Fifth, the model was built to capture costs and outcomes of people with and without hospital attention during acute cardiovascular events as in a "real-life" scenario. Different assumptions can be made in different settings according to local access to health services and/or the rate of asymptomatic events. Finally, a probabilistic sensitivity analysis was performed with all variables included in the model. Results were robust under a wide range of assumptions.