What is new?
Key findings- •
We compared the risk of cardiovascular (CV) events in patients starting blood pressure (BP)–lowering therapy with various therapeutic regimens using the health care utilization (HCU) database of the Italian region of Lombardy.
- •
We found that patients starting on extemporaneous combination or on monotherapy had a CV risk significantly higher than patients initiating on fixed-dose combination.
- •
Confounding bias is of particular concern in interpreting such findings because combinations therapy might be more likely prescribed to patients with worse clinical characteristics. By omitting sources of selective prescribing, biased estimates of the drug–outcome association are systematically generated.
- •
We propose of addressing the possible extent of unmeasured confounders by means of an approach based on an easy-to-apply Monte Carlo sensitivity analysis (MCSA) that accounts for external adjustment with data informative of the prescribing behavior of primary care physicians.
- •
We found that the excess of CV risk among patients starting on extemporaneous combination, with respect to those on fixed-dose combination, was likely explained by differences in clinical profile of patients receiving these treatments.
- •
Conversely, we found that starting therapy with a combination of drugs significantly reduced the CV risk with respect to initiating with a single antihypertensive agent, even after external adjustment.
What this adds to what was known?- •
Although the absence of information on potential confounders is a common criticism of observational studies based on HCU data, our analysis shows that medical records database can be used to quantitatively assess confounding bias.
- •
The main suggestion from this study is that starting therapy with a combination of BP-lowering agents likely reduces the CV risk with respect to treatment with one drug alone, even in the setting of primary CV prevention.
What is the implication and what should change now?- •
There are many situations where unmeasured clinical features may act as confounders of the drug–outcome association, in particular when HCU databases represent the main data source. However, whenever data informative of the physicians' prescribing behavior are available from external data sources, the estimates validity may be easily improved by means of MCSAs. The availability of the provided SAS macro could lead to a widespread diffusion of this easily applied approach in the framework of pharmacoepidemiologic research.
Existing health databases are widespread sources of data for pharmacoepidemiologic investigations [1]. Databases collecting health information can be classified into two broad categories: those that collect information for administrative purposes, such as filling claims for payment (i.e., health care utilization [HCU] databases), and those that serve as the patient's medical record (MR) and are therefore a primary mean by which physicians track health information on their patients (i.e., MR databases). A major advantage of HCU data is that they reflect real-world clinical practice for large and unselected populations [2]. Nevertheless, studies based on HCU data have been criticized for the incompleteness of information on potential confounders such as markers of clinical disease severity, lifestyle habits, socioeconomic status, among others. In contrast, although MR data are richer of clinical and lifestyle information, they often suffer of the fact that any given practitioner provides only a piece of the care a patient receives, and specialist and hospital cares are unlikely to be recorded in a common MR database [1].
These considerations suggest that, where feasible, multiple sources should be considered. Overall, the rationale of the approach used in this article is that although the strength of drug–outcome association estimated by HCU database may be biased by unmeasured confounders, external adjustment can be attempted if additional information is available from MR database covering a similar population as that of the considered HCU database.
As a motivating example, this article compares the risk of cardiovascular (CV) outcomes among patients submitted to different blood pressure (BP)–lowering drug regimens, nominally monotherapy, fixed-dose, or extemporaneous combination of two agents.
International guidelines have recommended initiating antihypertensive therapy with a single agent in patients with mild hypertension [3], [4]. However, based in part on data from the Antihypertensive and Lipid-Lowering Treatment to Prevent Heart Attack Trial, almost 75% of patients with hypertension will require antihypertensive therapy with more than one agent to achieve recommended BP goals [5]. Furthermore, based on analyses of large numbers of published clinical trials, it has been estimated that the BP-lowering capacities of low doses of two agents used in combination are additive [6]. Consequently, initiating therapy with more than one agent may be more efficient and efficacious than sequential monotherapy. A potential additional advantage of combination therapy is the reduction of drug adverse effects by allowing lower doses of the agents used in combination. Based on a meta-analysis of a large number of trials, it has been estimated that the extra BP reduction from combining drugs is about five times greater than doubling the dose of one drug [7]. Combinations between two drugs in a single tablet (fixed-dose combination) are regarded as to offer some advantages over vis-à-vis use of liberal combinations of two or more drugs (extemporaneous combinations) because simplification of treatment is associated with an improvement of patients' compliance [8], [9], [10], [11]. However, limited evidence has still available on the impact that use of fixed-dose combinations on the rate of CV events with respect to extemporaneous combinations or monotherapy in the clinical practice.
Confounding bias is of particular concern in investigating such a problem. In fact, combinations of BP-lowering drugs might be more likely prescribed to patients with poor prognosis, such as those with certain clinical characteristics (e.g., severe hypertension and poor clinical profile). By omitting sources of selective prescribing (e.g., because these data are not available from the HCU database used for investigating this issue), biased estimates of the drug–outcome association are systematically generated. This problem can be investigated by sensitivity analysis [12]. The basic concept of sensitivity analyses is to make informed assumptions about potential confounders and quantify their effects on the observed drug–outcome association. If additional data sources can be identified (e.g., from MR database), these assumptions can be substituted by empirical estimates and then they may be used for external adjustment of the drug–outcome association [13].
We conducted a large observational study using the HCU database of the Italian region of Lombardy aimed at assessing the risk of CV nonfatal events in patients starting BP-lowering therapy on fixed-dose combination in comparison to those on extemporaneous combination or monotherapy. For the analysis presented here, we used data obtained from 700 Italian primary care physicians providing information to the Health Search Cegedim Strategic Database [HSD]. We assessed the impact of confounding bias by clinical factors unmeasured in HCU data, including severity of hypertension and chronic disease score (CDS). Although our example is from pharmacoepidemiology, the techniques we discuss can be applied to any study in which one has some external data source that may improve validity of main study estimates [14]. For this reason, we developed and supplied an SAS code useful for any application of the Monte Carlo sensitivity analysis (MCSA).
