As the costs of type 2 diabetes mellitus (T2DM) care and related clinical trials continue to rise, economically viable methods are being sought to effectively predict the relative utility of various treatment options. The high price of clinical trials has led to the development of alternative methods to collect and consolidate data. Comparative effectiveness research (CER) synthesizes existing evidence to address knowledge gaps and drive patient-focused clinical decisions and outcomes. CER methods compare the health outcomes and costs associated with interventions to determine the option with the maximum patient benefit at optimal cost. In addition to traditional CER approaches such as systematic reviews, meta-analyses, and retrospective claims analyses, Markov modeling and Bayesian analysis can be applied to predict patient outcomes in scenarios where clinical trials are not feasible. Additionally, cost-benefit, cost-effectiveness, and cost-utility analyses comprise “cost-effectiveness analyses.” Cost-benefit analysis looks solely at monetary value, while cost-effectiveness and cost-utility analyses include gains in health and quality of life, providing a ratio of cost to benefit. This paper will discuss a range of approaches to CER including Markov modeling, mixed treatment comparisons, the Archimedes model, and Bayesian statistics, and provide guidance in interpreting data from these studies in a managed care context, with a particular focus on evaluating treatments for T2DM. It will also provide guidance on common indices of comorbidity used in health economics research. Data from these models can be used to reduce treatment costs and improve the overall quality of population-level health.
(Am J Manag Care. 2011;17:S377-S383)
Healthcare delivery is extremely expensive, and costs continue to rise as demand for care increases and newer treatments are developed.1 By 2020, US healthcare expenses are expected to increase to nearly 20% of the total gross domestic product.2 Healthcare delivery for type 2 diabetes mellitus (T2DM) is particularly costly; currently, 1 out of 5 US healthcare dollars is used to treat patients with diabetes.3 Likewise, the costs of clinical trial programs required for drug development have continued to rise sharply.4 This has led to an increased focus on alternate approaches to effectively predict the relative utility of various treatment options.
Considered broadly, these new approaches can be labeled as comparative effectiveness research (CER). The US Federal Coordinating Council describes CER as “The conduct and synthesis of systematic research comparing different interventions and strategies to prevent, diagnose, treat and monitor health conditions.”5 CER enables stakeholders to work toward a common goal of improving access to effective treatments, while keeping costs in mind. CER can identify ways to lower costs through improved outcomes, such as fewer emergency department visits and hospitalizations, and can be used to stratify treatment options based on efficacy, cost, and other factors.6
CER can also provide data related to drug effectiveness, making it highly relevant to managed care decision making. Until recently, clinical trials applying active comparators were considered to be the most reliable type of CER. Such trials, however, are not only expensive to conduct, but also only predict outcomes over a short period of time.
Emerging CER approaches have started to address these issues by consolidating data from sources such as existing placebo-controlled clinical trials or patient data sets using computer-assisted analytic techniques.6,7 Some of these alternate approaches to evidence acquisition have achieved a high level of evidentiary acceptance, and many are conducted using fairly straightforward research methodologies; for example, meta-analyses and retrospective claims analyses.6,7 Other, more sophisticated approaches, such as Markov modeling or Bayesian analysis of mixed treatment comparisons (MTCs), are not as well understood by healthcare professionals.7
This paper will summarize our current understanding of T2DM health economics research, discuss the role and benefits of CER in managed care decision making for T2DM, and provide an overview of Markov and Bayesian modeling techniques, as well as guidance on how to interpret data obtained from such studies in a managed care context.
Health Economic Analyses of Antidiabetic Medications
Economic evaluations of T2DM treatment and care compare the health outcomes and costs of different interventions to determine which option, when utilized, achieves maximum patient benefit at optimal cost.1
There are 3 primary analytic methods for evaluating the economic value of an intervention that fall under the umbrella term “cost-effectiveness analysis.” These are cost-benefit, cost-effectiveness, and cost-utility analyses. Cost-benefit analysis typically measures and reports results in the form of the overall monetary value, whether costs or savings, associated with use of the treatment. Cost-effectiveness and cost-utility analyses also project monetary outcomes, but go further to measure gains in health and quality of life using a ratio of overall cost (in dollars) divided by health effect (eg, number of heart attacks prevented), or assessed as quality-adjusted life-years (QALYs).1
In a systematic review, Klonoff et al evaluated 17 cost-benefit analyses performed to evaluate common diabetes interventions; the authors used this prior research to develop a scale estimating the economic impact of various diabetes interventions.8 The Table summarizes the key findings of this systematic review, including the ratings for each type of intervention. This provides a stark demonstration of the lack of cost-effectiveness of many interventions, as well as the high utility of some underappreciated concerns in diabetes, such as preconception care. However, it is important to keep in mind that a lack of identified cost-effectiveness may simply indicate a need to modify an intervention, rather than discontinue its use.
Using Modeling to Estimate the Value of T2DM Screening Strategies
The Archimedes model is a tool that assesses the potential long-term health and economic impact of new treatments on the healthcare system from the perspective of changes in medical guidelines, processes, and practice patterns. The model has been validated by more than 50 clinical trials.9,10
A recent analysis by Kahn et al used the Archimedes model to simulate and compare 8 T2DM screening strategies, compared with a control group receiving no screening. Screening strategies included routine screening once patients reached a certain age, or at hypertension diagnosis, as well as repeated screening at various time intervals. All simulated patients were followed for 50 years, or until death, with outcomes recorded annually.9
The model found that the T2DM screening strategy associated with the highest benefit, in terms of QALYs gained, was to initiate T2DM screening for all patients at 30 years of age, with repeat screening every 6 months. Investigators found that the use of this strategy resulted in earlier diagnosis (mean, 7.8 years) compared with the control group. Additionally, the simulation showed the effects of each screening strategy on myocardial infarction, stroke, microvascular outcomes, and death over a 50-year period.9
The ability to compare the outcomes of various screening strategies, plus the benefits of early T2DM diagnosis, against the costs of frequent screening provides useful information for informed managed care decision making.
Comparative Effectiveness Research on Treatments for T2DM
One of the main purposes of CER is to synthesize existing evidence in order to address knowledge gaps and drive patient-focused clinical decisions and outcomes.11 Thus, CER can identify the most beneficial treatment based on patient characteristics and inform real-world practice.6
Two recent large, systematic T2DM reviews conducted for the Agency for Healthcare Research and Quality (AHRQ) summarize existing evidence for the use of oral agents and insulin.12,13 The AHRQ review of oral medications included outcomes from 140 randomized controlled trials (RCTs) and 26 observational trials; the results support the use of metformin as first-line monotherapy.12 In contrast, comparisons of dual drug combinations (metformin thiazolidinedione, sulfonylurea, meglitinide, dipeptidyl peptidase-4 inhibitors, glucagon-like peptide-1 agonists, or basal or premixed insulin) identified no greater benefit with any specific treatment combination. Similarly, the insulin CER review included results from 50 studies and found, for example, that premixed insulin analogs were more effective in lowering fasting glucose compared with long-acting insulin analogs and noninsulin antidiabetic agents.13 As with the oral agents review, however, this paper identified research gaps. Specifically, certain comparator designs were not available (eg, basal-bolus vs premixed insulin), and a need was observed for more real-world effectiveness data on insulin use. In some cases, these limitations precluded the ability to draw definitive conclusions.
Observational studies have also been instrumental in CER when evidence gaps from RCT findings exist, despite their methodological ability to establish a reliable relationship between cause and effect.14-16 Observational studies use real-world data to assist in decision making. Sources of “realworld data” include administrative claims, patient registries, large simple trials, resource use information collected as part of clinical trials, supplements to clinical registration studies, health surveys, and electronic medical records.16,17 However, due to the lack of standardized instruments available to appraise quality of evidence, real-world data are not yet widely used for decision-making purposes.17 This highlights the need for innovative, sophisticated techniques such as Markov modeling and Bayesian MTCs to overcome these barriers. A study by Zhou et al illustrates how, in the absence of high-quality real-world data, an analytical computer model can have significant predictive power. Zhou and colleagues built a Markov-like model of diabetes progression, based on the baseline characteristics of patients from a populationbased study in southern Wisconsin. When the model was followed over time, predicted values were consistent with actual observed outcomes such as mortality, quality of life, and cost.18
Common Indices of Comorbidity Used in Health Economics Research
A number of comorbidity indices have been developed and validated for use in health economics research. These assessments evaluate patient risk and enhance study designs by accounting for patient comorbidity assessment, mortality risk, and comedication indicators. Some key indices used in CER evaluations include:
Charlson Comorbidity Index (CCI)
The CCI predicts the 10-year mortality risk of patients with certain illnesses, based on medical record review. Patients are given scores of 0 to 5, with 0 representing the absence of comorbidities and 5 indicating a moribund state. Conditions that contribute to the CCI score are weighted by illness severity. This measure can be used to predict mortality and healthcare utilization costs.19,20
Index of Coexistent Disease (ICED)
The ICED assesses the burden of coexistent diseases by combining 2 scales that reflect: (1) burden of coexistent disease, based on disease severity; and (2) overall physical impairment, based on the level of disability caused by illness. The burden of coexistent diseases can affect multiple aspects of care, such as recovery from surgery, and affect patient outcomes and resource use.21
Chronic Disease Index (CDI)
The CDI uses data from a medication database to estimate a patient’s chronic disease burden. Although this model can miss chronic diseases that are not treated with medication, it can be used to risk-adjust a large patient population by predicting outcomes such as morbidity and mortality, costs of medical care, and healthcare utilization.22
The RxRisk model is an algorithm that uses prescription medication data to predict the cost of chronic diseases, making it useful for medical risk analyses. The use of specific prescription medications classifies a patient into a specific risk category, which, combined with the patient’s age and sex, provides an estimate of future healthcare costs.23
Interpreting Complex Analytical Approaches
As statistical approaches to data analysis become more sophisticated, they also become more complex and, for those unfamiliar with these techniques, more difficult to interpret. This prompts questions regarding how to use the data generated by this type of research, so that evidence obtained can be translated into coverage decisions in a managed care setting.24 To facilitate this process, discussions of Markov models and MTCs are presented below.
Markov modeling is used to characterize a random process that evolves over time; it is therefore especially useful for chronic disease state modeling. In a Markov model, the disease is broken into separate “states;” each state is mutually exclusive (ie, a patient cannot be in more than 1 state at a time) and represents an important clinical and/or economic event in the process of the disease being modeled. For example, in a diabetes model, states might include “on treatment—healthy,” “on treatment with neuropathy,” or “on treatment with hypoglycemia.” Transition probabilities are assigned to the movement between these states over a set period of time (known as a Markov cycle). Estimates of health outcomes and/or resource use are assigned to the states and transitions. When the model is run for a large number of cycles, it becomes possible to estimate the long-term health outcomes or costs that may be associated with a disease or a particular intervention.25
For illustration purposes, Figure 1 shows a simple example of a Markov model, with only 3 states. In this figure, there is asymptomatic disease, in which the patient has the disease, but is not experiencing any effects and has a mortality risk similar to an otherwise healthy individual. Next is progressive disease, in which the patient is experiencing symptoms and has an increased risk of mortality due to the disease. The final state is death. Death, in a Markov model, is an example of an “absorbing state,” one that is impossible to leave. In this sample model, patients can progress from asymptomatic to progressive disease. It may also be possible for a patient to move backward (ie, from progressive disease to asymptomatic disease) if their other condition improves. Also note the circular arrows in Figure 1, which indicate the possibility for a patient to remain in a particular state over more than 1 cycle.25
An important limitation of Markov modeling is that it is “memoryless,” meaning that the probability of moving from one state to another is not dependent on the history of states previously experienced. A technique used to overcome the memoryless properties of Markov models is the use of time-dependent transition probabilities. In a Markov chain, transition probabilities are assumed to be constant without regard to time.25
For example, a study used data from the National Health and Nutrition Examination Survey to create 2 cohorts (diabetes-prevalent and diabetes-incident) to put through a Markov model. This model further drew data from the UK Prospective Diabetes Study to model diabetes progression. Analysts were then able to estimate population changes in body mass index, diabetes incidence, disease progression and complications, and related costs over a 25-year period.26
Another study used data from large-scale clinical trials, epidemiologic studies, and national surveillance studies to create 2 comparator cohorts: 1 group with diabetes receiving injected insulin and another on oral antidiabetic medication. In this study, the model was run twice, once as a baseline and once with medication adherence parameters included. Following this, the outputs of life expectancy, quality-adjusted life expectancy, complications, costs, and incremental cost-effectiveness ratios were compared between interventions, both with and without adherence issues taken into account.27
Mixed Treatment Comparisons
In contrast to systematic reviews, wherein 1 pharmacological intervention is consistently compared against the same competitor using traditional meta-analysis techniques, MTCs, also known as network meta-analysis, compare the safety, efficacy, and costs of 2 or more interventions by linking them in a network. This method obviates the need to conduct a direct comparison study, and instead borrows evidence from other trials in which each respective intervention was evaluated against a common comparator (eg, placebo). The information obtained from these analyses can prove useful for medical decision making when direct comparison data are not available.28 For example, Gross et al combined the results of 18 trials in an MTC analysis to study the benefit of adding a third agent to metformin and sulfonylurea therapy (Figure 2).29
Understanding why MTCs have not been widely used until now requires a paradigm shift from traditional frequentist “objective probability” statistics. A unique feature of MTCs is that they use Bayesian statistics. In the Bayesian approach, probability describes the uncertainty of events, and takes into account not only the uncertainty intrinsic to an event (as does the frequentist approach), but also uncertainty due to incomplete knowledge or understanding of an event.30 This “subjective probability” allows researchers to concurrently incorporate a range of prior evidence (eg, outcomes related to the efficacy of a treatment, obtained from previous studies) into the statistical analysis of current data. This, in turn, produces posterior informed estimates of the parameters.31 This analytic approach is based on Bayes’ Theorem, shown in Figure 3.
Figure 4 shows a visual representation of how prior beliefs and current data combine in Bayesian analysis to comprise the posterior distribution. In this figure, the blue line represents the prior distribution for the parameters, the red line represents the current data, and the dotted line shows the posterior distribution for the parameters. The most common criticism of the Bayesian method is the subjectivity involved in the interpretation of probability used to develop the prior distribution. Even though subjectivity is unavoidable with Bayesian statistics, as the prior distribution is based partially on a “personal degree of belief,”30 the analysis can lead to scientifically defensible results if the choice of the prior distribution is based on sound evidence and reasoning.
In the absence of true comparative outcomes data, CER analyses provide the best available tools for managed care professionals to assess the efficacy of treatment options for chronic conditions such as T2DM. Newer methods being applied to CER as replacements for or supplements to frequentist statistical approaches can help to overcome some of the limitations that previously prevented the adequate analysis of cost, safety, and efficacy outcomes outside the context of a controlled clinical trial. In coming years, CER modeling is expected to open new avenues of research to clarify best practices in the treatment of T2DM. In turn, these data can be used to reduce treatment costs and improve the overall quality of population-level health.6
Author affiliations: Center on Drug and Public Policy (FTS), Department of Pharmaceutical Health Services Research, University of Maryland School of Pharmacy, Baltimore, MD (FTS, VVC).
Funding source: This activity is supported by an educational grant from Novo Nordisk, Inc.
Author disclosures: Dr Shaya and Mr Chirikov have no relevant financial relationships to disclose.
Authorship information: Concept and design (FTS); analysis and interpretation of data (FTS, VVC); drafting of the manuscript (FTS, VVC); critical revision of the manuscript for important intellectual content (FTS, VVC); and administrative, technical, or logistic support (VVC).
Address correspondence to: Fadia T. Shaya, PhD, MPH, University of Maryland School of Pharmacy, 220 Arch St, 12th Floor, Room 01-204, Baltimore, MD 21201. E-mail: firstname.lastname@example.org.