Potential Role of Network Meta-Analysis in Value-Based Insurance Design
Published Online: August 21, 2014
James D. Chambers, PhD, MPharm, MSc; Aaron Winn, MPP; Yue Zhong, MD, PhD; Natalia Olchanski, MS; and Michael J. Cangelosi, MA, MPH
It is estimated that more than 10 million Americans 50 years or older have osteoporosis and more than 43 million have low bone mass.1 A chronic condition characterized by low bone mass and deterioration of bone microarchitecture, osteoporosis increases patients’ risk of bone fracture. Most common fractures include those of the vertebrae, hip, and wrist, and result in substantial morbidity and medical and hospital costs. It is estimated that approximately 2.05 million osteoporosis-related fractures occur annually in the United States, costing about $16.9 billion.2 Osteoporosis management focuses on reducing fracture risk, and includes lifestyle modification (eg, smoking cessation and alcohol moderation), weight-bearing exercises, and treatment with pharmaceuticals. Various classes of pharmaceuticals are indicated for osteoporosis treatment, among them bisphosphonates (including alendronate, ibandronate, and risedronate), recombinant parathyroid hormone (teriparatide), selective estrogen receptor modulator (raloxifene), and monoclonal antibodies (denosumab). osteoporosis therapies—alendronate, ibandronate, risedronate, raloxifene, teriparatide—are generally available through a health plan’s pharmacy benefit.
Historically, drug co-payments and coinsurance have been largely based on drug cost, and did not take treatment benefit into account. However, this paradigm is changing, and a move toward value-based insurance design (V-BID) is gathering momentum.3-5 Essentially, the aim of V-BID is to improve healthcare and reduce costs by encouraging high-value care— care that offers clinical benefits at a reasonable cost—and discouraging low-value care. First proposed more than a decade ago, various value-based health insurance programs are now established and subject to much discussion and evaluation in the medical literature. V-BID has been applied to many indications, including diabetes, hyperlipidemia, and asthma.6-9 In contrast to existing approaches to V-BID in which copayments are reduced to encourage use of broad categories of high-value care (eg, statins for hypercholesterolemia or hyperglycemic medications for type 2 diabetes mellitus), we sought to use comparative effectiveness evidence to prioritize treatments within a drug class, using drugs that treat osteoporosis as a case study.7,10 However, as osteoporosis treatments have not been adequately compared against one another in head-to-head studies, we could not rely on clinical trial data to inform our approach, and therefore we used network meta-analysis to synthesize the necessary comparative evidence.
Network meta-analysis is a statistical approach to synthesizing comparative effectiveness evidence and is a generalization of meta-analysis, combining head-tohead clinical trial evidence with statistically inferred indirect comparisons across treatments not studied within head-to-head clinical studies.11,12 The approach requires a connected network of clinical trials. For example, in a data set consisting of pairwise comparisons (eg, A compared with B, and A compared with C), the relative efficacy can be inferred for those comparisons not studied directly (eg, in the previous example, B compared with C through common comparator A). Thus, through a combination of direct and indirect evidence, the network meta-analysis provides the relative efficacy of the whole network. Network meta-analysis has 2 principal roles: first, to strengthen inference of relative effectiveness between a pair of treatments through the combination of direct and indirect evidence; and second, to infer relative efficacies between treatments that have not been evaluated in a head-to-head study. Furthermore, the approach allows estimation of the probability that each included treatment is most effective, an important consideration for V-BID. Network meta-analysis methods are increasingly used and are promoted by health technology assessment agencies including the National Institute for Health and Clinical Excellence (NICE) and the Canadian Agency for Drugs and Technology in Health (CADTH).13-15
The objectives of this study were to construct a V-BID for osteoporosis treatments using comparative effectiveness evidence synthesized from a network meta-analysis; and to illustrate the potential of this approach through estimation of the number of avoided osteoporosis-related fractures and of cost savings using a simulation model. These estimates were specific to a large California-based health plan. We considered the osteoporosis treatments alendronate, ibandronate, risedronate, raloxifene, and teriparatide in this research, as these drugs are commonly available through drug formularies. We excluded the intravenously administered osteoporosis treatments denosumab and zoledronic acid, as they are not typically part of a tiered drug formulary and not subject to the same co-payment structure.
We used claims obtained from a large California-based private health insurance plan. These data specified the number of patients with an osteoporosis diagnosis, and, for these patients, which one of the included treatments they received. We also used this plan data as the source of drug acquisition cost and patient co-payments.
This study consisted of 4 steps. First, we identified clinical trials of osteoporosis treatments that were included in a 2012 Agency for Healthcare Research and Quality (AHRQ) systematic review. The trials that met our inclusion criteria became part of our study. Second, using these studies, we performed a network meta-analysis to synthesize the baseline fracture risk for vertebral, hip, and nonvertebral/nonhip fracture, and the comparative effectiveness of competing treatments. Third, we simulated a V-BID by adjusting the co-payments of the existing pharmacy benefit in accordance with the synthesized comparative effectiveness data, with co-payments eliminated, ie, reduced to zero, for the most effective treatments. Fourth, we simulated the impact of co-payment adjustment in terms of fracture reduction and cost savings for the health plan. The details of each step follow.
Step 1: Study Identification and Review. We relied on studies reported in the 2012 AHRQ report Treatment To Prevent Fractures in Men and Women With Low Bone Density or Osteoporosis: Update of a 2007 Report.16 Included studies were limited to those that: (i) included adults with low bone density or with osteoporosis; (ii) examined a pharmacological intervention reported within the private health insurance plan’s claims data; (iii) reported vertebral, hip, and/or total fractures; (iv) lasted a minimum of 6 months; and (v) were randomized controlled trials. Eligibility criteria of each of the included studies are listed in the eAppendix (available at www.ajmc.com). Pairs of reviewers read each article to confirm reported counts for vertebral, hip, nonvertebral, and nonvertebral/nonhip fractures, and in some instances contacted the original authors for clarification.
Step 2: Network Meta-Analysis. Using the extracted data, we performed a Bayesian network meta-analysis to estimate 2 pieces of information—underlying fracture risk, ie, the fracture risk in untreated patients, and comparative effectiveness of the various agents both to the reference treatment (placebo) and to one another.
We conducted the network meta-analysis using a random- effect, binomial logit-linked model implemented in WinBUGS 1.4. We used a binomial distribution because of the binomial nature of fractures, and the logit-link assumes a linearity of effects on the logit scale. We determined to use a random effects model after our inspection of the deviance information criterion showed no significant difference in goodness of fit between fixed and random effects models. The networks of studies for each end point (ie, vertebral fractures, hip fractures, and nonvertebral/nonhip fractures) shared the same structure, with individual studies comparing treatments to placebo, though the individual patient counts varied across the outcome networks. Characteristics of the studies populating these networks are provided in the eAppendix. For each end point, ie, vertebral, hip, and nonvertebral/ nonhip fractures, we estimated the underlying fracture risk from the placebo arms of each study. We estimated the comparative effectiveness of the various agents to one another (relative risk and 95% credible interval) indirectly through the reference treatment, ie, placebo. We additionally examined a synthetic end point—total fractures—as this provided a summary efficacy end point with which to rank treatments.
PDF is available on the last page.