Guidance for a causal comparative effectiveness analysis emulating a target trial based on big real world evidence: when to start statin treatment

There is an ongoing debate on the optimal use of statin treatment in the prevention of cardiovascular events. In 2016, the European Society of Cardiology (ESC) published new guidelines for the prevention of cardiovascular disease [1]. These guidelines recommend lipid-lowering treatment such as statins for excessive-risk individuals with overt cardiovascular disease, diabetes, chronic kidney disease, familial hypercholesterolemia, or very high levels of individual risk factors. Outside this excessive-risk group, the ESC Systematic Coronary Risk Evaluation (SCORE) [2] is recommended for total disease risk assessment [1]. This country-adapted SCORE system intends to assess a patient’s 10-year risk of undergoing a fatal cardiovascular event. The risk estimation is based on age, sex, total cholesterol (with an optional use of high-density lipoprotein cholesterol), smoking status and blood pressure, and it should be repeated every 5 years. Patients are divided into four risk categories, low (<1% SCORE), moderate (1–5% SCORE), high (5–10% SCORE) and very high (>10% SCORE). The target for low-density lipoprotein cholesterol (LDL-C) levels and the recommended treatment and treatment dose depend on the risk category and the LDL-C level [2]. Immediate statin treatment is recommended for a risk score ≥10 and an LDL-C level ≥70 mg/dl or a risk score ≥5 and <10, and an LDL-C level ≥100 mg/dl. However, for all risk scores ≥5, drug treatment could be considered [1,2].

The American College of Cardiology and the American Heart Association (ACC/AHA) released new recommendations in November 2013 [3], in which the threshold indication for statin therapy was lowered for primary prevention. Instead of LDL-C thresholds, these recommendations focus on total atherosclerotic cardiovascular disease (CVD) risk, which is defined by new Pooled Cohort Equations [3]. Statin treatment is recommended when the atherosclerotic CVD risk equals or exceeds 7.5%. Pandya estimated the cost–effectiveness of various risk thresholds in the US American setting and concluded that more lenient thresholds of 4% or higher would be optimal using a threshold of $100,000 per quality adjusted life year (QALY) gained [4].

This raises the question whether there is evidence that a policy oriented purely on the risk score and that is using a more lenient threshold would also be beneficial in the European setting. Vancheri and others could not demonstrate an association between an increased statin utilization and coronary mortality or its rate of change in European countries between 2000 and 2012 [5]. However, this study did not have access to data on adherence, indication for statin treatment and industry campaigns on doctor’s decision making, which could have confounded the results. Adherence to statin treatment for instance, is known to be negatively associated with mortality as well as with the intensity of statin regimen [6,7].

Real-world evidence (RWE) is a widely propagated source to gain information outside the artificial setting of clinical trials [8–10]. However, these data bear their own difficulties, including confounding, missing or nonreliable data, no clear treatment assignment, imperfect compliance, dynamic treatment regimes, switching and multiple-line treatments. Several approaches exist that may help conducting studies that allow drawing causal conclusions from observational studies and RWE data [11]. Visual, structural and statistical techniques exist to understand and estimate causal relations using RWE [11–42].

Directed acyclic graphs are graphical models that help in visualizing causal relations. In addition, causal-directed acyclic graphs help to understand the nature of confounding, to assess identifiability, and to guide the selection of the appropriate statistical analytic approach [18,22,32–37,39,41,42].

A prominent structural approach is the so-called ‘target trial’ approach described by Hernan, Robins, Cain and others [11,13–15,23,25] combined with statistical methods controlling for time-dependent confounding. The target trial approach suggests designing any observational data analysis as if one would design a randomized clinical trial. This approach is consistent with counterfactual theory in the analysis of causality, and additionally, it is structurally close to the randomized trial approach, and therefore, facilitates the communication to medical researchers familiar with clinical trial design. In addition to comparison of treatments, this approach allows researchers to answer the question of when to start a specific treatment (e.g., when to start statin) [14].

Causal statistical methods include the (parametric) g-formula, marginal structural models with inverse probability weighting and structural nested models with g-estimation. These are complex and validated statistical methods that can be applied when time-dependent confounding is present. These techniques are all based on the counterfactual approach and differ regarding their assumptions [12,13,16,17,19–21,24,26–32,38,40]. In the counterfactual approach, one estimates the outcome assuming the individual had received no treatment.

A combination of these visual, structural and statistical causal methods allows estimating causal relations using RWE and is particularly useful to determine risk thresholds for initiation of preventive medication such as statin use.

The overall goal of this project is to demonstrate how causal inference methods can be applied to RWE to obtain unbiased estimates. We focus on statin treatment to prevent CVD events in individuals with and without a history of a cardiovascular event. In the absence of a clinical trial evaluating the optimal start of statin treatment, we would like to guide the use of available real-world data and causal inference methods. We aim to provide a study protocol to answer the question of when to start statin treatment to optimally prevent CVD events. It is not the intention of this paper to present the results of the data analysis.

Real-world data analyses are particularly prone to confounding and selection bias. Unlike in a randomized clinical trial, the intervention – in our case the (time of starting) statin treatment – is not administered at random but based on individual clinical information. When the same factors that influence the decision whether or not treatment should be provided, also influence the outcome, in our case CVD events, controlling for confounding is essential for obtaining valid causal effect measures. To provide the basis for the selection of variables needed to be controlled for in a causal analysis, we generated a causal graph including variables that are directly or indirectly associated with the intervention statin treatment and the outcome CVD (Figure 1). The assumptions on the presence or absence of associations were based on published literature [43–47] and expert opinions. We searched the literature for CVD-risk evaluations and risk prediction models as well as for clinical treatment guidelines. We then asked the clinical experts which factors they base their decision on when starting or not starting statin treatment. We included those factors into the causal graph irrespective of their availability in the datasets. Some factors such as birth date and sex may not change after the treatment decision has been made and can be included as baseline variables. However, other variables, such as any disease-specific and treatment variables may change over time and should be included in the analyses as time-varying variables. The assumed causal direction is indicated by the arrows. Some variables are connected by links pointing in both directions. Those cases represent relationships with feedback loops over time.

The causal graph (Figure 1) shows that confounding can be eliminated by blocking the backdoor paths from statin treatment to cardiovascular events, that is, controlling for age, sex and the ESC SCORE, or when the ESC score is not available, all the variables being included in the ESC SCORE as well as diabetes and severe kidney disease. Similarly, as the variables included in the risk score, the variables severe kidney disease and diabetes may change over time, and therefore, should be included in the model as baselines variables and time-varying variables.

The ESC SCORE, and newly diagnosed severe kidney disease and newly diagnosed diabetes are time-dependent confounders [24,32]. On one side, they are influencing the probability of physicians prescribing the start of statin treatment, and on the other side, they are themselves affected by statin treatment (Figure 2). This creates a feedback loop between time-dependent treatment and time-dependent confounder, and hence, causal methods such as g-methods [32] need to be applied to draw valid (unbiased) causal conclusions.

In the absence of a randomized clinical trial comparing multiple strategies for when to start statin treatment, we based our protocol on the target trial approach, which emulates a trial using real-world observational data: when designing a target trial, one formulates the study protocol as if one was designing a randomized controlled trial. This approach assures that the timely sequence and analytic strategy is appropriate and biases such as immortal-time bias are avoided [25]. The target trial emulates (mimics) a randomized controlled trial, and therefore consists of the same components as a randomized controlled trial. These components include eligibility criteria, treatment strategies, assignment procedures, follow-up period, outcomes, causal contrasts of interest (i.e., the estimands), and analysis plan and start with a well-formulated research question [11].

For our statins study, each component of the protocol was carefully considered and defined according to the operationalized research question. As result, we describe each component of the target trial study protocol.

To assess the optimal (i.e., most effective) timing of starting statin treatment, several comparative strategies with different treatment starting points must be defined. Currently, the statin treatment strategy is based on the ESC SCORE, which assesses the individual 10-year risk of undergoing a fatal cardiovascular event based on age, sex, total cholesterol, smoking status and blood pressure, and the LDL-C level [1]. Our intention is to define the treatment strategies by describing the starting point of statin treatment entirely based on ESC SCORE thresholds.

As in a randomized clinical trial, the individuals in the compared strategies should be as similar as possible. Individuals are therefore included in each strategy at the time of study inclusion. Just the defined treatment starting point differs. Figure 3 shows that while being in the study, the risk may increase as time passes by, and the individual may cross the specific risk threshold of a certain given strategy. At that time point, treatment should start.

The operationalized research question is shown in the following population, intervention, comparison and outcome (PICO) structure:

Population: Men and women aged 40–75 years of the general Austrian population.

Interventions:

Comparator: No statin treatment.

Outcomes: Primary: CVD defined as composite end point major adverse cardiovascular event (MACE), that is, myocardial infarction (MI) or stroke, or cardiovascular death; secondary: overall survival.

This research question reflects the study population, treatment strategies and outcomes as defined below.

Following the target trial concept, only those individuals are included into the analyses who could be randomized to one of the four treatment strategies or the comparator. Hence, we plan to include the general population between the age of 40 and 75 years who are not classified yet as excessive risk, that is, no overt CVD, diabetes, chronic kidney disease, familial hypercholesterolemia, very high levels of individual risk factors or an acute cardiovascular event. Individuals are included into the study when they first show to have crossed the 10-year risk of undergoing a fatal cardiovascular event threshold of 1%. Note that inclusion in the study does not mean immediate statin treatment as treatment depends on the strategy-specific threshold.

Our exclusion criteria are age <40 years or >75 years, diagnosis of stroke ≤1 month prior of the start of the emulated trial (ICD10: I60–64), diagnosis of MI ≤1 month prior of the start of the emulated trial (ICD10: I21–25), diabetes (ICD10: E10–14), chronic kidney disease (ICD10: N18) and familial hypercholesterolemia (ICD10: Z82)

Individuals with and without a history of prior cardiovascular events are included in the study. The analyses will be conducted in subgroups: without prior cardiovascular events (primary prevention); and with prior cardiovascular events (secondary prevention).

Five treatment strategies (i.e., four statin strategies and the comparator no statin treatment) will be compared, differing in whether and when to start statin treatment as shown above.

All individuals enter the study at a similar state of risk, that is, when they first show to have crossed the 10-year risk of undergoing a fatal cardiovascular event threshold of 1%. This implies that individuals in the treatment strategies who start statin treatment at a high-risk threshold wait and will be monitored until they cross the risk threshold corresponding to the particular treatment strategy they are assigned to (Figure 3).

The statin treatment used in the observational study is defined by prescription of any available statin filed in the database. The dose of statin treatment is not considered for treatment definition. By applying the prescription of statins as indicator of statin use, it is assumed that individuals take the medication as it was prescribed.

Subjects will be assigned to a treatment arm at the beginning of the study. Rather than randomizing subjects to a single treatment arm, so called ‘replicates’ (‘copies’ or ‘clones’) of each subject in the dataset will be created. One replicate of each subject is then analytically being allocated to each of the treatment strategies, and subjects not following the assigned treatment protocol will be censored at the time of protocol violation. For example, for a subject A starting treatment with a risk score of 3%, the subject’s data are copied five-times and each copy is then allocated to a different treatment strategy at the beginning of the study. In the 1% threshold strategy, the replicate of subject A would then be censored right at the beginning of the study, as the protocol is to start treatment when crossing the 1% risk threshold and subject A does not start until the risk exceeds 3%. In contrast, in the 2% threshold strategy, the data of the replicate of the same subject A are censored at the time she/he has a 2% risk of cardiovascular event. As treatment is not started, it is violating the 2% threshold protocol. In the 3% threshold strategy, the data of the replicate of subject A are not censored, as treatment is started at a 3% risk, which is following the protocol. In the 5% threshold strategy and the no-treatment strategy, the data of the replicate of subject A are censored at the time of treatment start, as this is the time when the observation indicates assigned treatment violation. As censoring is informative, adjustment for baseline and time-dependent confounders of the censoring–outcome relation must be performed to derive valid causal treatment effects and is described in the analysis plan.

Individuals assigned to one of the four immediate treatment arms receive statin treatment after crossing the corresponding risk threshold (Figure 3). We allow for a ‘grace period’ of 3 months during which the treatment may be started after crossing the respective threshold, as individuals may not start treatment immediately due to nonmedical reasons, such as vacation, family holidays, travel and other personal reasons. Individuals not starting statin treatment within this grace period will be censored at the end of the grace period (i.e., 3 months after crossing the risk threshold). Individuals who had started treatment according to the protocol and stop it for more than 3 months are also planned to be censored 3 months after the last statin intake. The impact of the length of the grace period on the results should be tested in sensitivity analyses.

The individuals in the no-treatment arm must not receive statin treatment at any time. Therefore, individuals in this arm who start treatment at any time are being censored at the time of first statin use.

The study baseline is at the time of eligibility, which is the time point when an individual’s ESC-SCORE exceeds 1%. The study’s follow-up period starts at baseline and ends at the end of the 15-year follow-up time, occurrence of the outcome, loss of follow-up or treatment violation, whichever occurs first.

Our primary end point (outcome) is defined as any major cardiovascular event (MACE), including MI, stroke or death due to CVD during the time of follow-up. Our secondary end point is death from any cause during the time of follow-up. As frequency measures to be compared between the treatment arms, we use the 15-year incidences (risks) of the outcomes.

Although a clinical trial may be analyzed to estimate the intention-to-treat effect, this is usually often not possible when using existing real-world data. In contrast, when using the target trial approach, all subjects are replicated and each subject is assigned to each treatment arm. Subsequently the causal per protocol effect of the respective sustained treatment strategy is estimated, that is, the effect of the assigned treatment strategy if all individuals in the trial had sustained statin treatment until end of study as indicated in the protocol. As censoring due to treatment violation or censoring due to loss to follow-up maybe informative, methods must be applied that account for informative censoring (see below).

To follow the study protocol, several variables need to be available at baseline (e.g., age, sex, and prior diagnoses of CVD, diabetes, chronic kidney disease, and familial hypercholesterolemia). Variables that require repeated measures include diagnoses, laboratory values, behaviors and measurements of blood pressure. Depending on the data source, not all of these data might be available. Continuous information on laboratory values, behaviors and clinical measurements are often lacking. This information could be indirectly estimated or approximated using data on diagnoses or on filled prescriptions.

If not all data are available from one source, the necessary information may be obtained by linking several databases containing administrative reimbursement data, specialized registry data, clinical information and operative data.