Study design and data source

Detailed methods are provided in S1 Text. Briefly, we conducted a historical cohort study based on nationwide data from the French National Health Insurance Information System (SNIIRAM), which contains individualized, anonymous, and comprehensive data on health spending reimbursements and demographic data[6]. Information on 100% reimbursed severe and costly chronic diseases including diabetes and hypertension, is available in the SNIIRAM[7]. These data can be linked to the French Hospital Discharge database (PMSI)[8].This database provides detailed medical information on all admissions in public and private hospitals, including discharge diagnosis ICD-10 codes and medical procedures performed during the stay. The authors were not involved in the initial data collection.

Ethics Statement

In France, only confidentiality approval from the Commission Nationale de l'Informatique et des Libertés (CNIL) [French data protection authority] is required for non-interventional observational studies; ethical approval is not mandatory (Law No. 2004–800 on bioethics, Aug. 6, 2004). CNIL approval for our data source had been obtained previously (CNIL AT/CPZ/SVT/JB/DP/CR05222O (Jun. 14, 2005); DP/CR071761 (Aug. 28, 2007)). Informed consent was not required, as patients’ data were anonymized and de-identified prior to analysis. The national health administrative database can be accessed upon authorization delivered by the CNIL, 8, rue Vivienne, CS 30223, 75083 Paris cedex 02 (http://www.cnil.fr/les-themes/sante/fiche-pratique/article/communication-des-donnees-de-sante/). A description of the process, in English, is available at http://www.institut-des-donnees-de-sante.fr/abstract/ and the e-mail contact is gipids@gip-ids.fr.

Study population

Eligible patients were registered in the French national health insurance general scheme, who initiated fibrate or statin monotherapy (2 consecutive reimbursements) prescribed by a general practitioner between 1^st July 2008 and 31^st December 2009, with T2D and hypertension (defined by at least 3 reimbursements of oral antidiabetic/antihypertensive drugs over one year during the 30 months prior to study entry) at the date of first reimbursement of a fibrate or statin, corresponding to the date of study inclusion (Fig 1). This 30-month period was the longest period available in our database. Patients had to be over the age of 50 (men) and 60 (women), these gender-specific thresholds being considered in the recommendations and guidelines provided by The French National Health Authority (HAS, Haute Autorité de Santé) as defining a cardiovascular risk factor. Because our goal was the study of the compliance of physicians regarding these guidelines, and not to challenge these thresholds, we did not discuss them.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. A. Study scheme. B. Flow chart of patients with diabetes and high cardiovascular risk in primary care initiating fibrate or statin monotherapy.

https://doi.org/10.1371/journal.pone.0137733.g001

Patients were excluded if they had received any lipid-lowering treatment during the 30 months prior to study entry. The lipid-lowering drug had to be prescribed in the context of primary prevention; patients with a history of cardiovascular event during the 30 months prior to study entry (list of diseases and codes in S1 Text) were therefore excluded.

Follow-up and exposure

Follow-up began at initiation of monotherapy: date of the first reimbursement of a fibrate or a statin between 1^st July 2008 and 31^st December 2009. Follow-up ended at the time of an event or at 31^st December 2010. Statistical analysis was based on an intention-to-treat approach.

Outcomes

The main outcome was a composite of all-cause mortality, ischemic stroke, myocardial infarction, and above-the-ankle amputation, identified by a hospital discharge diagnosis or medical procedures. In case of occurrence of the outcome, patients were censored. Then, only the first occurrence, if any, was taken into account in the survival analysis.

Potential confounders

Demographic and medical factors that could have an impact on the incidence of events were obtained from SNIIRAM and PMSI and determined at baseline: age, gender, history of diabetes, deprivation index[9], comedications (at least 3 reimbursements for specific drugs during the year before baseline), common chronic comorbidities (Alzheimer’s disease, Parkinson’s disease, depression and cancer—defined by several reimbursements for specific drugs and/or SNIIRAM data and/or specific hospital discharge diagnoses during the year before baseline) and hospitalization for any cause (during the year before baseline).

Statistical analysis

Descriptive analyses compared confounding factors between patients with fibrate or statin monotherapy using Pearson’s chi-square and Wilcoxon tests for categorical and continuous variables, respectively, and Mantel-Haenszel method for age- and gender-adjusted rates. The adjusted HR for each factor was estimated by a Cox proportional hazards model. Confounding factors (i.e. with a significant HR (p<0.05), and with a potential clinical impact on the incidence of events) were used in a final Cox proportional hazards model to estimate the adjusted HR for events of patients on fibrate or statin monotherapy.

For each variable used in a Cox proportional hazards model, the proportionality assumption was assessed graphically by survival curves (Kaplan-Meier method). If the proportionality assumption was not met, indicating a changing effect of the variable on the incidence of events over time, the variable of interest (X) was divided into several time intervals. For example:

with and P_k is a time interval such as {P₁,…P_k,…,P_n } = entire follow-up where t is time in days; n is a natural number greater than or equal to 2.

The best cut-off time corresponds to the maximum likelihood of a univariate Cox model with different splitting of the variable of interest using a bootstrap method. In more detail, the best cut-off time of the fibrate/statin variable was reached when the Cox model provided the best prediction of events in the cohort. This best prediction was achieved when the likelihood of the Cox model was maximized. All following cut-off times were tested: one or two (and more if necessary) time intervals at each month of follow-up. After obtaining the best cut-off time, the proportionality assumption was assessed graphically on each time interval. In order to improve the robustness of the calculation of the best cut-off time, which depends on the time of the events, 500 Cox models were recalculated for each cut-off time with a random resampling population, obtained from the initial cohort, with the same sample size, and the same proportion of subjects on fibrate/statin (bootstrap method). Finally, the best cut-off time must satisfy the proportionality assumption at each time interval and maximize the median of 500 likelihoods obtained by bootstrap.

For sensitivity analysis, a propensity-score method was applied to address the issue of residual confounding. The propensity score was defined as the conditional probability of using a fibrate given the covariates (same as in the Cox proportional hazards model described above)[10]. This approach used inverse probability weighting (IPW) estimators and has been developed based on the work of Hernan et al[11–13]. Briefly, this IPW method consists of obtaining a pseudo-population in which the probability of receiving one treatment rather than the other no longer depends on prognostic factors, as in a “clinical trial population”. To construct this pseudo-population, a weight is attributed to each subject; this weight corresponds to the inverse probability of receiving this treatment. In the present study, this weight corresponds to the inverse of an odds-ratio calculated with a logistic regression model in which the dependent variable is fibrate/statin and the prognostic factors are the independent variables used in the Cox model described above. Finally, these inverse probability weights, a 1-dimensional function of the propensity score, were used in a Cox model as a covariate with the dependent fibrate/statin variable in order to obtain the HR.

Study population

From 7/1/2008 to 12/31/2009, 1,027,314 patients were identified as receiving a prescription for either fibrate or statin monotherapy, with no previous reimbursement for any lipid-lowering drug for at least 30 months (new users). A total of 31,652 of these patients satisfied all inclusion criteria (Fig 1), corresponding to 27,594 (87.2%) statins new users and 4,058 (12.8%) fibrate new users.

Patients exposed to fibrates were slightly, but significantly, younger than patients exposed to statins, with a shorter history of diabetes, and they received fewer antidiabetic drugs, including insulin. These patients were also more frequently women and presented a higher socioeconomic status (Table 1). They were less exposed to cardioprotective drugs (antiplatelet and antihypertensive drugs), but received more antidepressants (Table 1). The proportion of subjects who had received a reimbursement for serum creatine kinase assay during the year before first use of the index lipid-lowering drug was 9.8% (95% confidence interval: 9.5–10.2) and 8.7% (7.8–9.6) for fibrate and statin users, respectively. A combination of fibrates and statins in the year following first use of fibrates was exceptional (18/4,058, 0.4%, fibrate users had received at least two simultaneous prescriptions for a fibrate and a statin during the first year).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Patient characteristics.

https://doi.org/10.1371/journal.pone.0137733.t001

Association with clinical outcomes

Mean follow-up was 616±166 days. At least one event was observed in 167 of the 4,058 fibrate users and 1,026 of the 27,594 statin users. Corresponding age- and gender-adjusted annual event rates were significantly different (p-value < .0001): 2.42% for fibrates (2.05%-2.79%) and 2.21% for statins (2.07%-2.35%). Significant predictors of the combined outcome in age- and gender-adjusted Cox models are indicated in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Predictors of the combined outcome in age- and gender-adjusted Cox models.

https://doi.org/10.1371/journal.pone.0137733.t002

Kaplan-Meier event-free survival curves according to the initial prescription showed that the proportionality assumption required for a Cox model calculation was not met, as the effect associated with fibrate use changed with time (Fig 2). The fibrate/statin variable was therefore divided into several time intervals. The best cut-off time was obtained by the method described above: testing each cut-off with 2 or 3 time intervals at each month of follow-up, assessing the proportionality assumption on each time interval, maximizing the likelihood obtained by bootstrap. This best cut-off corresponded to two time intervals, before and after 540 days (18 months) (S1 Fig). Lastly, the final Cox models included the fibrate-vs-statin variable with an HR before 540 days (with the statin users group as reference, HR_<540) and an HR after 540 days (HR_>540). Other variables associated with the outcomes were included with a constant HR (Table 3). In all three Cox models, 1) univariate, 2) including gender and age, and 3) including gender, age and all predictors identified in the univariate models), HR_<540 was not significantly different from 1, while HR_>540 was significantly different from 1, suggesting that the use of fibrates was associated with a poorer prognosis after this time cut-off (Table 3). Fully adjusted HR_<540 and HR_>540 were 0.95 (0.78–1.16) and 1.73 (1.28–2.32), respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Survival curves for death and cardiovascular events in patients treated with statin or fibrate monotherapy.

https://doi.org/10.1371/journal.pone.0137733.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Final Cox models with a cut-off time at 540 days.

https://doi.org/10.1371/journal.pone.0137733.t003

Sensitivity analyses

A sensitivity analysis was conducted by censoring follow-up data after a six-month period of no reimbursement of the fibrate or statin initiated by the patient at the time of inclusion, when such a period was observed (lack of persistence of exposure to the index drug). This censoring rule did not modify the findings, as the fully adjusted model yielded 1.01 (0.82–1.24) for HR_<540 and 2.00 (1.43–2.80) for HR_>540.

A treatment switch (mostly from a fibrate to a statin) or another change of lipid-lowering therapy was observed in some patients during follow-up. This information was not captured by the intention-to-treat approach. We therefore conducted a sensitivity analysis with censoring of follow-up data six months after a treatment switch. HR_<540 and HR_>540 (fully adjusted model) were then 0.89 (0.72–1.09) and 1.89 (1.38–2.58), respectively. Our findings therefore appear to be robust, and not dependent on the type of analysis: intention-to-treat in the primary analysis, or by using these two censoring rules.

As amputations are not typical outcomes used in most trials on lipid-lowering drugs, amputations were excluded from the composite cardiovascular outcome (14 of the 1,193 events in the primary analysis) and similar results were obtained.

Finally, the propensity score-based analysis taking into account the same factors as the multivariate analysis (see above) and the same cut-off of 540 days gave consistent results: HR_<540 and HR_>540 were 0.95 (0.78–1.15) and 1.78 (1.33–2.38), respectively.

Limitations

First, the present study is a retrospective observational analysis and not all possible confounders can be accounted for. This study was based on several independently collected, French nationwide medical administrative databases, resulting in a decreased risk of selection biases. These databases are considered to be comprehensive, with annual quality controls of coding. However, during acute care or long-term hospitalization, as with long-stay institutions, current medications are included in hospital costs and individual medications cannot be identified. We may therefore have unintentionally neglected older populations in whom lipid-lowering drug therapy was initiated in long-stay hospitals or institutions, but this bias is not expected to have a major impact on the generalizability of our conclusions. Due to legal restrictions concerning access to the databases, pre-inclusion use of lipid-lowering drugs could not be excluded for periods of more than 30 months. There is therefore a possibility that patients who failed to respond to previous statin therapy could have been included. A small proportion of the patients included were treated with clopidogrel; this treatment can be considered to be a marker of an undocumented cardiovascular event, but it is not expected to change the guidelines concerning the use of a statin rather than a fibrate. Sensitivity analyses including a specific approach based on propensity scores was conducted and reached similar conclusions, although we admit that the data included in the propensity score were the same as those entered in the Cox analysis, and that residual confounding remains a potential limitation. Statistical analysis was conducted according to an intention-to-treat approach, which means that a patient who initiated treatment with a fibrate and who was then switched to a statin, remained in the fibrate group for the analysis. However, the intention-to-treat approach cannot be expected to have induced the association between the fibrate covariate and clinical outcome, as this approach is conservative and would attenuate rather than accentuate any associations. Moreover, the same results were obtained in a sensitivity analysis in which data for switchers were censored after a 6-month follow-up. The main limitation concerns from the lack of behavioral data, like smoking habits, and of laboratory test results in the databases. These variables may participate to residual confounding. However, the guidelines emphasize the benefit associated with the use of statins in high-risk patients, regardless of baseline LDL-cholesterol, and the study population was undoubtedly at high risk. Follow-up did not exceed 30 months, a limitation related to recent French regulations concerning confidentiality of health insurance databases. However, this relatively short follow-up was sufficient to demonstrate an increased risk associated with the use of fibrates. The same limitation applies to the observation period prior to the study. Although unlikely, we cannot exclude the possibility that some patients may have been exposed to a fibrate or a statin before this 30-month period. According to the study design, only patients with an initial prescription for fibrates or statins prescribed by a general practitioner were included. This theoretically limits the generalizability of our conclusions. However, most patients with type 2 diabetes are managed in the primary care setting in France, especially in terms of cardiovascular primary prevention. The intention-to-treat approach could lead to misclassification of certain patients if they stopped their initial drug, with or without switching from a fibrate to a statin, for example. However, alternative types of statistical approaches were conducted with other censoring rules and the results remained substantially the same. We did not conduct analysis comparing specific molecules inside either the class of statins, or the class of fibrates, for statistical power issues. Finally, one limitation of this study is its generalizability. Data on prescription of fibrates in patients with diabetes and high cardiovascular risk are actually extremely scarce. However, they suggest a marked heterogeneity in worldwide prescribing habits, as a recent Swedish study by Eliasson et al. showed that fibrate monotherapy was used in only about 1.5% of diabetic patients (but not specifically with high cardiovascular risk)[34]. A very marked heterogeneity was also observed in the general population in USA and Canada in the study by Jackevicius et al [27]. Together with Jackevicius et al., we can hypothesize that this heterogeneity suggests that “other factors beside clinical trial evidence are influencing fibrate prescription patterns”; these factors may include commercial promotion and medical education. However, it is likely that the adverse consequences of the use of fibrates (vs statins) demonstrated in the present study also apply elsewhere, as, in Eliasson’s study, LDL cholesterol levels were higher in fibrate users than in statin users, illustrating the missed opportunity to decrease this major cardiovascular risk factor.

Clinical perspective

We believe that this work will contribute to a re-appraisal of how guidelines are accepted and implemented in primary care. In the present study, based on nationwide data, we investigated the issue of prescription of fibrates when they are not indicated. Inappropriate use of fibrates was frequent among subjects for whom guidelines are available concerning the use of lipid-lowering therapy, namely in diabetic patients with high cardiovascular risk. Moreover, this denial or ignorance of evidence-based medicine was associated with a poorer prognosis. We believe that this finding is of major interest in the current context, in which prevention of macrovascular complications of diabetes appears to have come to a halt: for example, very intensive glycemic or blood pressure control is not as effective as previously believed. Correct use of the available drugs with proven efficacy is therefore of tremendous importance.