Milton C. Weinstein, PhD

 To assess the cost-effectiveness of laparoscopic adjustable gastric banding (LAGB) and laparoscopic Roux-en-Y gastric bypass (LRYGB) as treatment for morbid obesity.

A Markov model was developed to simulate weight loss, health consequences, and costs for surgical treatment of morbid obesity. The model was used to estimate incremental cost-effectiveness ratios (ICERs) in terms of cost per quality-adjusted life-year (QALY) gained.

Estimates of procedure effectiveness were derived from published results of a head-to-head randomized controlled trial. Other model parameters, including complication rates, costs of treatment, adverse events and obesity, mortality rates, and utilities, were estimated from published literature and publicly available databases. Costs (2006 US dollars) and QALYs were discounted by 3% per annum.

Under conservative assumptions, both LAGB and LRYGB improved health outcomes, at a higher cost, compared with no treatment. ICERs for both LAGB and LRYGB versus no treatment were below $25,000 per QALY gained. ICERs were lower for individuals with higher initial body mass index and higher for older individuals. ICERs for men were generally higher than those of women. Sensitivity analyses showed these results to be robust to reasonable variation in model parameters and overall parameter uncertainty. Base-case ICERs for LRYGB versus LAGB were below $25,000 per QALY gained, but were highly sensitive to model assumptions.

Both LAGB and LRYGB provide significant weight loss and are cost-effective compared with no treatment at conventionally accepted thresholds for medical interventions.

	The majority of bariatric procedures in the United States are laparoscopic Roux-en-Y gastric bypass (LRYGB) and laparoscopic adjustable gastric banding (LAGB).

		Both LAGB and LRYGB lead to better healthcare outcomes and higher costs compared with no treatment. In the main analysis, LAGB results in less weight loss than LRYGB, but is associated with lower costs.

		Both LAGB and LRYGB are cost-effective alternatives to no treatment, with incremental cost-effectiveness ratios below $25,000 per quality-adjusted life-year gained; results are robust.

		Estimates of the cost-effectiveness of LAGB versus LRYGB are sensitive to assumptions about maintenance of weight loss in the medium to long term; long-term comparative studies of the 2 procedures are needed.

	Obesity, or excessive body fat, is a major risk factor for a host of detrimental medical conditions including type 2 diabetes, hypertension, dyslipidemia, osteoarthritis, and obstructive sleep apnea.1 Approximately 5.7% of American adults (>12 million people) are morbidly obese,2 with a body mass index (BMI) of >40 kg/m

	There are 3 established approaches to treating morbid obesity: lifestyle modification, pharmacologic therapy, and surgery. The first 2 approaches have been shown to have a limited impact on long-term weight control.

 The third option, bariatric surgery, has been shown to be an effective long-term treatment for morbid obesity, with mean excess weight loss exceeding 60%.

 However, the risks of complications with bariatric surgery are not trivial.

	The majority of bariatric surgeries currently performed in the United States are laparoscopic Roux-en-Y gastric bypass (LRYGB) procedures. An alternative to gastric bypass is laparoscopic adjustable gastric banding (LAGB), in which a silicone elastomer ring is placed around the upper part of the stomach, creating a new small stomach pouch so the food storage area in the stomach is reduced. Excess weight loss with LAGB, although significant, may be lower than that seen with LRYGB.

 However, LAGB, unlike LRYGB, does not involve stapling or partition of the stomach or intestine; thus, postoperative complication rates are lower, hospitalizations are shorter, and there is less risk of micronutrient deficiency compared with LRYGB.

 Moreover, the procedure is fully reversible. The objective of this study is to assess the cost-effectiveness of LAGB versus that of LRYGB, or no treatment, for patients with morbid obesity.

	We constructed and estimated a Markov model of treatment and outcomes for obese adults. The model focused on decreases in BMI and treatment-related adverse events as predictors of lifetime cost, health-related quality of life, and survival. Patients received initial treatment with LAGB or LRYGB, or received no treatment. After the initial treatment period, patients entered the Markov model in 1 of the 5 health states corresponding to their initial BMI status: not obese (BMI <30 kg/m

). The sixth health state in the model was death. Health states were associated with specific treatment costs, levels of “other” health expenditures, health-related quality of life, and mortality risks. Patients transition between Markov states by changing BMI status (losing or gaining weight) or dying. For the first 10 years of the model, surgical patients remained at risk of transient treatment-related adverse events that may be mild, moderate, or severe.

	We estimated model probabilities, costs, and utilities from a variety of published and publicly available sources, supplemented by input from a clinical expert in minimally invasive (laparoscopic) bariatric surgery. 

 reports model parameters and data sources.

The population for this analysis was limited to adults age 18 to 74 years who satisfied the clinical eligibility criteria for bariatric surgery: BMI >40 kg/m

Probabilities of treatment-related weight loss were estimated from a prospective study by Angrisani et al, in which morbidly obese patients were randomized to LAGB and LRYGB and followed for 5 years postsurgery.

 In base-case analyses, it was assumed that BMI remained constant after the 5-year follow-up period. Patients receiving no treatment were assumed to maintain a constant BMI for the duration of the model. The study by Angrisani et al was judged to be the strongest available based on its randomized head-to-head design and minimal attrition over the 5-year study period. The study limitations include a small initial sample (N = 51) and a European rather than US setting. For this reason, we also estimated the model using an alternative source for 5-year efficacy data: a systematic review of 36 English-language studies of weight loss outcomes following LAGB or LRYGB by O’Brien et al.

 Limitations of the O’Brien et al study include a preponderance of single-armed rather than comparative studies, and attrition rates that may have been underreported. Compared with the Angrisani et al study, O’Brien et al reported greater effectiveness for LAGB and shorter durability of weight loss for LRYGB (with weight regain beginning in year 3), as shown in 

	The calculation of transition probabilities between BMI categories also is described in 

Probabilities of Treatment-Related Adverse Events.

 The estimated incidences of adverse events were derived from a comparative study of the complications of LAGB and LRYGB by Parikh et al.

, defined as organ repairs such as resection of the bowel; 

, defined as operations not requiring organ repair (eg, laparoscopic repositioning of a gastric band); and 

, defined as drug therapy only (eg, antibiotic therapy for an infection). Parikh et al reported separate estimates for the probabilities of major reoperations and all adverse events for the first 30 days (early) and after 30 days (late). We estimated the early and late probabilities of moderate reoperations and medical treatments based on the relative proportions of total early and total late complications from that study. The average period of follow-up was approximately 12 months; we assumed that the yearly probability of surgical adverse events remained constant for the first 4 years postsurgery, was halved for years 5 to 10, and was zero after 10 years. Consistent with the literature, patients receiving no treatment were assumed to have no serious adverse events.

Patients undergoing LAGB or LRYGB have a 1-time probability of postoperative death from surgery in the first 30 days.

 BMI-specific mortality rates were estimated based on age- and sex-specific all-cause mortality rates derived from US life tables,

 to which we applied BMI-specific relative risk ratios from an analysis of the National Health and Nutrition Examination Survey by Flegal et al.

 Because Flegal et al reported a single aggregated mortality risk ratio for BMI >35 kg/m

, we estimated separate mortality risk ratios for the morbidly obese I, morbidly obese II, and super obese groups using linear extrapolation based on the trends at lower levels of BMI and the relative prevalence of morbid obesity categories reported by Sturm: morbidly obese I (67%), morbidly obese II (29%), and super obese (4%).

The physician costs for the LAGB and LRYGB procedures were estimated using 

codes: 43770, 90772, 90801, 97802, and 99214 for LAGB; and 43644, 90801, 97802, and 99214 for LRYGB.

 Associated inpatient costs were estimated using data from the Healthcare Cost and Utilization Project (HCUP) Database with 

International Classification of Diseases, Ninth Revision 

codes: 44.95 for LAGB, and 44.38 for LRYGB.

 The hospital charges were converted to resource costs using hospital cost-to-charge ratios published in the HCUP Database. The default schedule of follow-up visits was based on treatment guidelines of the American Society for Metabolic and Bariatric Surgery, supplemented by input from our clinical expert as: 9 visits in year 1, 2 visits in year 2, and 1 annual visit thereafter for LAGB; and 6 visits in year 1, 2 visits in year 2, and 1 annual visit thereafter for LRYGB.

	In consultation with our clinical expert, the costs of LAGB- and LRYGB-related adverse events for each level of severity were estimated using the Physician Fee Guide (physician costs),

 Within each severity category, a weighted average cost was derived, based on the relative prevalence of such complications as reported by Jan et al,

 supplemented by additional input from the clinical expert. The cost of a postoperative fatality was assumed to be equal to that of an early major reoperation. A list of the complications that were used to generate the weighted average costs is available from the authors on request.

	"Other" medical expenditures specific to each BMI level were calculated based on a study of the Medical Expenditure Panel Survey by Arterburn et al.

 Because this study reported aggregated expenditures for the BMI >40-kg/m

 group as a whole, we estimated the other medical costs for the morbidly obese II and super obese groups using linear extrapolation as described above.

 Where necessary, costs were updated to 2006 US dollars using the Medical Care Services component of the Consumer Price Index.

Estimates of BMI-specific utilities were derived from a regression analysis of EQ-5D data from the 2000 Medical Expenditure Panel Survey by Jia and Lubetkin,

 rescaled to replicate the US mean population EQ-5D score of 0.868 as reported by Sullivan and Ghushchyan.

 Jia and Lubetkin reported mean utilities for individuals with BMI >35 kg/m

 as a single group. We used linear extrapolation to estimate separate utilities for the morbidly obese I, morbidly obese II, and super obese groups, as described above.

	No estimates of bariatric-surgery—specific utility decrements were identified in the literature. Estimated utility decrements for the LAGB and LRYGB procedures therefore were derived from utilities reported for laparoscopic surgery for hernia repair in the United Kingdom,

 The utility decrement was applied at the time of surgery for 4 weeks for LAGB and 6 weeks for LRYGB, to reflect the slightly longer recovery time associated with LRYGB compared with LAGB. The utility decrement for a major reoperation was estimated as the decrement associated with a hysterectomy

 and was applied for 6 weeks. The decrement for a moderate reoperation was estimated as the decrement associated with laparoscopic hernia repair

 and was applied for 4 weeks. The decrement for medical treatment was estimated as the mean of the decrements associated with venous thrombosis, gastroesophageal reflux disease, ulcer, and gastritis as reported by Sullivan and Ghushchyan

	Cumulative costs associated with morbid obesity and its treatment, survival in life-years, and quality-adjusted life-years (QALY s) were estimated for each obesity treatment strategy. Cumulative costs include costs of initial treatment, follow-up, and treatment-related adverse events minus cost offsets generated by decreases in other health expenditures as patients lose weight. The treatments were ranked in order of increasing costs and the incremental cost-effectiveness ratios (ICERs) were calculated by comparing the more costly with the less costly, dividing the incremental cost by the incremental benefit. The base-case analyses were stratified by sex and initial BMI category, assuming an initial age of 40 years. Results also were aggregated to the population level using a patient distribution consistent with the patient sample in the LAGB arm of the Angrisani trial (80% female; initial BMI: 23% morbidly obese I, 66% morbidly obese II, and 11% super obese).9 Cost-effectiveness analyses were performed from the perspective of a third-party payer over a lifetime horizon. All costs, life-years, and QALY s were discounted at an annual rate of 3%. Base-case analyses were performed using treatment efficacy parameters derived from the Angrisani et al study.

Alternate Scenarios and Sensitivity Analyses

	We assessed the effect of uncertainty in comparative weight loss estimates by reestimating the model with an alternative source for efficacy estimates.10 In a second alternative analysis, patients were assumed to regain half of the cumulative BMI lost after 5 years over the following 5-year period. Multiple 1-way sensitivity analyses then were conducted to test the robustness of model results to changes in key parameters, particularly efficacy, rates of adverse events, and costs. Additional 2-way sensitivity analyses are discussed in 

	A second-order probabilistic Monte Carlo sensitivity analysis was conducted to assess the sensitivity of study findings to overall parameter uncertainty, particularly with regard to treatment efficacy, rates of complications, costs, and utilities. We constructed a probability distribution for the estimates of treatment efficacy by bootstrapping the simulated LAGB and LRYGB cohorts—repeatedly sampling with replacement from each cohort of patients. Uncertainty in other key model parameters was characterized by probability distributions with means equal to the default estimates. Probabilities of adverse events were assumed to follow a beta (β) distribution, with distributional parameters reflecting the underlying trial samples. Estimates of treatment costs were assumed to follow a gamma (g) distribution with standard errors assessed, subjectively, as 10% of the mean value. The parameters and distributions of the variables included in the probabilistic analysis are summarized in 

	Base-case results from the analysis using efficacy estimates derived from the article by Angrisani et al9 are reported in 

. Results of analyses are presented weighted over the patient population, and stratified by sex and initial BMI. Neither LAGB nor LRYGB is cost saving compared with no treatment; however, both LAGB and LRYGB are more beneficial than no treatment, providing an additional 2.04 (LAGB) and 2.9 (LRYGB) discounted lifetime QALY s for each treated patient. The ICERs for LAGB versus no treatment and for LRYGB versus no treatment are similar and relatively low (<$10,000 per QALY gained). Although LAGB costs an average of $5324 less than LRYGB in the aggregate population, it also is less effective, providing 0.42 fewer life-years saved and 0.86 fewer QALY s; the ICERs for LRYGB versus LAGB across patient groups lie below $15,000 per QALY gained. As the stratified results show, bariatric surgery is most cost-effective for females and for patients who have higher initial BMI. In addition, bariatric surgery is more cost-effective for younger patients (not shown).

Results for Alternate Scenarios and Sensitivity Analyses

	For comparison, results of the analysis using the O'Brien et al data

 also are presented in Table 2. Under these alternative efficacy assumptions, both procedures remain cost-effective compared with no treatment, with ICERs less than $20,000 per QALY gained; however ICERs for LRYGB versus LAGB are significantly greater than those estimated using efficacy estimates from Angrisani et al.

reports the results of a series of 1-way sensitivity analyses for the default analysis based on efficacy estimates from Angrisani et al.

 We report the results for analyses for which the range in the relevant ICERs exceeds $1000. In all instances, the magnitude of the variation in the aggregate ICER is less than $10,000 per QALY gained, and the upper limits of our estimates of incremental cost-effectiveness for LAGB and LRYGB versus no treatment lie below $25,000 per QALY . The model was not sensitive to small changes in efficacy over the initial 10-year period. All ICERs for LRYGB versus LAGB fell below $15,000 per QALY .

	However, in an analysis in which patients receiving LRYGB were assumed to partially regain weight over years 5 to 10 (as suggested by the O'Brien et al

 analysis), LAGB dominated LRYGB when using estimates of efficacy based on either Angrisani et al

 or O’Brien et al (not shown). When weight regain was modeled for patients receiving either LAGB or LRYGB, the ICER for LRYGB versus LAGB was $19,700 using estimates derived from Angrisani et al; however, LRYGB was dominated by LAGB using estimates derived from O'Brien et al.

Articles

Laparoscopic adjustable gastric banding and gastric bypass are cost-effective treatments for morbid obesity compared with no treatment.

Cost-Effectiveness of Laparoscopic Gastric Banding and Bypass for Morbid Obesity

To investigate the extent to which preferred drug lists and tiered formularies reflect evidence of value, as measured in published cost-utility analyses (CUAs). 

Using 1998-2001 data from a large registry of cost effectiveness analyses, we examined the 2004 Florida Medicaid preferred drug list and the 2004 Harvard Pilgrim Pharmacy Program 3-tier formulary, and compared cost-utility ratios (standardized to 2002 US dollars) of drugs with preferred and nonpreferred status. 

Few drugs on the formularies had any cost-utility data available. Of those that did, median cost-utility ratios were somewhat higher (less favorable) for Florida's preferred drugs compared with the nonpreferred drugs ($25 465 vs $13 085; 

= .09). Ratios did not differ for drugs on tiers 1 and 2 of the Harvard Pilgrim formulary, although they were higher for tier 3 and for excluded drugs ($18 309, $18 846, $52 119, and $22 580, respectively; 

= .01). Among therapies reported to be cost-saving or to have cost-utility ratios below $50 000, 77% had favored status in Florida Medicaid and 73% in Harvard Pilgrim. Among dominated drug interventions (reported to be more costly and less effective than alternatives), 95% had favored status in Florida Medicaid and 56% in Harvard Pilgrim. 

This study underscores the paucity of published cost-utility data available to formulary committees. Some discrepancies prevail between the value of drugs, as reflected in published cost-utility ratios, and the formulary placement policies of 2 large health plans. 

Formularies with tiered copayment structures or preferred drug lists have become increasingly popular, as plans seek to provide financial inducements for enrollees to select the least expensive drugs while avoiding the restrictions of entirely closed formulary systems.

Mounting evidence suggests that incentive-based formularies are associated with lower costs and smaller increases in drug utilization and expenditures compared with control groups.

But researchers also have found that they are associated with undesirable effects: patients suddenly faced with higher copayments are more likely to switch medications or to discontinue medications entirely.

Moreover, studies have found that cost sharing may be followed by reductions in the use of "essential" drugs, higher rates of serious adverse events, and increased use of emergency department visits and hospital days.

In theory, plan formulary placement decisions should be guided by information about a drug's overall value or cost effectiveness. Ideally, drugs with favorable cost effectiveness would receive preferential formulary status (ie, physicians would have incentives to prescribe these drugs rather than nonpreferred drugs). Instead, observers complain that plans often base patient contributions on drugs' acquisition costs and, in fact, do not consider value in formulary placement decisions.

However, there has been little empirical evidence on whether formulary policies actually reflect evidence of value in terms of cost effectiveness. The objective of this study was to examine whether the formulary drug-placement decisions of 2 large healthcare providers reflect the findings of published cost-utility analyses (CUAs). We acknowledge at the outset that this is an exploratory exercise whose purposes are to investigate the type and quality of cost-utility data available to formulary decision makers, and to further discussions about the need to move the healthcare system toward value-based benefit design. 

To measure a drug's value, we focused on CUA, a special type of cost-effectiveness analysis in which health effects are measured in terms of quality-adjusted life years (QALYs) gained. Leaders in the field have recommended this outcome as the standard for cost-effectiveness research.

Though not without limitations, CUAs are appealing as a measure of overall value for several reasons: they capture in a single measure gains from both prolongation and quality of life, they incorporate the value or preferences people place on different health outcomes, and they provide a common metric for comparing analyses of diverse interventions and conditions.

Data on the cost effectiveness or value of drugs were derived from the Harvard Center for Risk Analysis Cost-Effectiveness Registry, a database of 539 original CUAs published in the public health and medical literature from 1976 through 2001. Cost-utility analyses were included if they were published in a MEDLINE-referenced journal and provided an original cost-utility estimate. Review papers and methodological papers were excluded. More detail about the registry is provided elsewhere.

Briefly, 2 readers independently collected data on each article and then convened for a consensus review to resolve discrepancies. Data were collected on a wide range of items related to the methods used to estimate and report costs, health effects, preference weights, modeling assumptions, and study limitations. We also included a subjective assessment of overall study quality on a Likert-type scale from 1 (low) to 7 (high), which reflects readers' evaluation of the rigor, reasonableness, and usefulness of each analysis. Data on the cost-utility ratios associated with each intervention also were collected and standardized to 2002 US dollars.

Drug Data in the Cost-utility Analysis Registry. 

For this study we focused on CUAs in the registry pertaining to pharmaceutical interventions (378 CUAs reporting 898 cost-utility ratios). To exclude potentially outmoded analyses, we further restricted the sample to studies published from 1998 through 2001, leaving a final sample of 112 studies and 266 cost-utility ratios (the number of ratios exceeds the number of studies because some analyses contain more than 1 ratio). The complete list of references is available from the authors. 

We examined publicly available formularies of 2 healthcare providers, the Florida Medicaid preferred drug list and the Harvard Pilgrim Healthcare Pharmacy Program (HPHC) in Boston. These plans were chosen arbitrarily, based on the availability of formulary information, although we did set out to select plans representing different geographic regions and types of payers. 

The Florida Medicaid program lists prescription products on a preferred drug list selected by the pharmaceutical and therapeutics committee as "efficacious, safe, and cost effective choices when prescribing for Medicaid patients."

Our analysis used the March 22, 2004, updated list, which contained 1749 drugs. 

The HPHC categorizes medications into 1 of 3 tiers. Patients make the lowest copayment for drugs on tier 1, which consists of generic drugs approved by the US Food and Drug Administration (FDA). Tier 2 drugs are brand-name drugs granted preferred status (and a moderate copayment) by HPHC based on reviews of their relative safety, effectiveness, and cost.

Tier 3 contains nonpreferred drugs for which patients make the highest copayment. The actual copayment amount varies, depending on which benefit design is selected by the beneficiaries' employer group. Some prescription medications are excluded from coverage, such as Avage (tazarotene) for skin conditions and the weight-loss treatments Meridia (sibutramine) and Xenical (orlistat). 

We examined the relationship between the placement of drugs on formularies and evidence of value, as reflected in published cost-utility ratios. Specifically, we investigated whether the drugs on Florida Medicaid's preferred lists had lower (ie, more favorable) cost-utility ratios than those with nonpreferred status. Similarly, we investigated whether drugs on HPHC's lower tiers (reflecting lower copayments) had more favorable cost-utility ratios than those on higher tiers. 

We compared the median cost-utility ratios between preferred and nonpreferred drugs in the Florida Medicaid program, and among drugs with different tier classifications on the HPHC formulary. We used the Wilcoxon rank sum and Kruskal-Wallis nonparametric tests because of heteroskedasticity and heavily skewed ratio distributions.

We also sought to find examples of drugs with good overall cost effectiveness or value that were 

on preferred lists or favorable tiers, and conversely whether preferred lists or favorable tiers contained drugs with unfavorable cost effectiveness. These "discordant" pairs were then reviewed to explore possible reasons for discrepancies. Note that we excluded from our analysis combination therapies (drug A + drug B) for which the plan pays only partially (either drug A or B). For example, a CUA might analyze etoposide plus cisplatin versus gemcitabine in patients with metastatic non-small-cell lung cancer. In the Harvard Pilgrim Health Plan, etoposide is covered, while cisplatin is not. 

The 266 cost-utility ratios in our sample (representing 102 drugs) covered a wide range of disease categories, most commonly infectious diseases (31%), malignancies (17%), musculoskeletal and rheumatologic diseases (13%), and cardiovascular disease (13%) (

). The majority (59%) were drugs for chronic diseases (treatment course of 18 months or longer). In terms of sponsorship, 33% came from studies funded by the pharmaceutical industry and 25% from government-funded studies; in 34%, the funding source could not be determined from the article. 

Of the 266 cost-utility ratios pertaining to these 102 drugs, 150 (56%) were in the 0-$50 000 per QALY range (a range often cited as representing good value for money);

43 (16%) were more than $100 000 per QALY; and 25 (9%) each were cost saving or dominated (ie, meaning the cost of the drug in question was higher and its health effects lower than its comparator) (

Of 1749 drugs on the Florida Medicaid preferred list, 102 drugs (5.8%) had any recently published data from CUAs. Median cost-utility ratios for drugs on Florida Medicaid's preferred list ($25 465) were higher than those on the nonpreferred list ($13 085) (

). Median ratios for drugs on tiers 1, 2, and 3 of the HPHC plan were $18 309, $18 846, and $52 119, respectively, and $22 580 for excluded drugs (

= .01 by the Kruskal-Wallis test, meaning that at least 1 tier had either a higher or a lower median cost-utility ratio than the other tiers). 

show the distributions of cost-utility ratios according to preferred/nonpreferred status in the Florida Medicaid plan. The figure shows, for example, that for drugs on the preferred list, 29 cost-utility ratios (17% of the total number of cost-utility ratios for drugs on the preferred list) were more than $100 000 and an additional 20 were dominated; in contrast, for nonpreferred drugs, 2 ratios were cost saving and 32 were between 0 and $50 000. 

shows an analogous distribution for cost-utility ratios of drugs on the various HPHC tiers. In some instances, drugs reported to have good value received unfavorable tier placement (eg, for drugs on tier 3, 1 cost-utility result was cost saving and QALY improving, and 19 ratios were less than $50 000), and drugs with poor value received favorable tier placement (5 ratios for tier 1 were more than $100 000 per QALY and 7 ratios for tier 1 drugs reflected dominated interventions). 

Where cost-utility ratios were available for Florida Medicaid, 64% of preferred drugs (111 of 173 ratios) and 72% (34 of 47) of nonpreferred drugs had ratios below $50 000, whereas for HPHC, those percentages were 71% (115 of 163) and 58% (42 of 73), respectively (where 

The quality of CUAs varied slightly across formulary placement categories. For example, the quality score was 4.3 for CUAs of drugs on the Florida Medicaid preferred drug list vs 4.5 for CUAs of drugs on the nonpreferred list (

= .049). For the HPHC formulary, the quality score was 4.4 for CUAs of drugs on tiers 1 and 2 and 5.4 for CUAs of drugs on tier 3 (

provides selected examples of drugs with favorable cost-utility ratios that did not receive preferential status, as well as drugs with unfavorable cost-utility ratios that did receive preferential status. For example, interferon beta for multiple sclerosis and riluzole for amyotrophic lateral sclerosis received favorable status on both plans, despite high (less favorable) reported cost-utility ratios. In contrast, donepezil for Alzheimer's disease did not receive favorable tier placement, despite a favorable cost-utility ratio. 

To our knowledge, this is the first study to examine whether tiered formularies and preferred drug lists reflect evidence of value, as measured in published CUAs. Our findings underscore the challenges inherent in this exercise: only 6% of drugs on the Medicaid preferred drug list had corresponding data from recently published CUAs, for example. 

Where CUA data were available, they suggested that formulary decisions frequently reflect published evidence of value in that drugs with favorable formulary placement reflect products with "reasonable" cost-utility ratios. (Note that if we had used $100 000 per QALY as the measure of societal willingness to pay, the consistency between formulary placement and evidence of value would have been greater.) 

The results also suggest some room for improvement. There were numerous examples of reportedly cost-effective drugs excluded from preferred lists or favorable formulary tiers, as well as examples of drugs with unfavorable cost-effectiveness profiles that were included on preferred drug lists and favored formulary tiers. The data hint that HPHC's formulary may be more in-line with value (as measured by CUA) than that of Florida Medicaid. Pressure to cover drugs in public programs also could help explain why median cost-utility ratios in our sample were higher for drugs on Florida's preferred list than for those on HPHC's, as well as the fact that Florida Medicaid covered expensive cancer therapies that were dominated by other treatments. We emphasize, however, that more work is needed on this question and that we have not provided new evidence that anyone is hurt by a specific formulary decision. 

In general, we emphasize that ours is a descriptive and exploratory exercise and that one should exercise caution about inferring causality between the cost-utility literature and formulary placement decisions. It is difficult to obtain a clear reading from publicly available CUAs about the actual choices facing the formulary managers, or about the trust they place in the data. Researchers have documented the fact that methods used for conducting and reporting cost-utility analyses are highly variable.

 Moreover, the decisions of a pharmacy and therapeutics committee are complex and reflect a multitude of factors, including the size and makeup of the pharmacy and therapeutics committee and the incentives facing the director of the pharmacy.

A key problem for our study is that published CUAs may not include an appropriate comparator therapy. The extent to which drugs have treatment alternatives is a key factor in determining actual formulary placement. Some products are expensive, but because there is little or no alternative they are likely to receive favorable formulary placement. On the other hand, some drugs have many treatment alternatives or close substitutes, and as a result are assigned to a higher tier or receive nonpreferred status. In the HPHC formulary, for example, all of the drugs assigned to tier 2 or 3 had treatment alternatives; in contrast, of the 21 unique drugs assigned to tier 1, 10 were generics and 9 of the remaining 11 were arguably the drug of choice for at least 1 indication. 

However, published CUAs may exclude the relevant comparator. They may compare a new drug with placebo or with an inferior alternative, rather than to the next-best intervention, thus exaggerating the value of the drug in question. Another problem is that study authors may fail to present results of CUAs in all relevant population subgroups. Drugs may be cost effective for some indications and not others, but CUAs are not always clear about distinctions. 

A measure of the degree to which the comparator drug and target population differ can be seen in the case of interferon ?-1b for multiple sclerosis, where cost-utility ratios varied from $44 000 per QALY for patients with secondary progressive multiple sclerosis (where the treatment comparator was no treatment) to $1.9 million per QALY for interferon compared with "best practice" in ambulatory patients with secondary progressive multiple sclerosis. 

There are other problems with trying to match published CUAs to formulary decisions. One is that cost effectiveness studies often are not available at the time of FDA approval, and published CUAs of a drug tend to lag behind published clinical studies.

Another is that CUAs typically do not incorporate or anticipate generic drugs as comparator therapies, which can dramatically change cost-utility ratios. Some observers have noted that CUAs may better help with decisions about whether a new class of drugs are cost effective rather than which drugs within a class are appropriate.

Furthermore, even if the ratios do reflect a reasonable measure of societal value, it is possible that they do not reflect value from the perspective of health plans. Importantly, we lack data on the actual drug prices paid by the plans, which could be considerably lower than the drug prices used in the CUAs, and hence improve the drug's cost effectiveness. Plans–and pharmacy benefits managers, who frequently negotiate sizable discounts–may indeed find it cost effective to give preferential status to discounted drugs, even if published CUAs suggest that, based on average wholesale prices, the drug in question is not cost effective. 

Even if the prices used in the CUAs are accurate, strict rankings of cost-utility ratios may not reflect how plans wish to allocate resources. Results of CUAs, conducted from a societal perspective, may not be seen as relevant for health plan managers who worry about short-term financial shortfalls, total plan costs, and drug budget "silos."

It is difficult to determine the impact of the CUA "perspective" in our sample, because many studies are not clear about their perspective. In addition, there is inconsistency in the manner in which analysts use the term 

: some state a societal perspective, but only include direct healthcare costs and nominal nonhealth costs; others include productivity costs. 

Empirical evidence suggests that public preferences for allocating society's resources may not be consistent with a goal of maximizing QALYs subject to a budget constraint. Instead people want to give hope to and avoid discrimination against patients with chronic illness or disability even when treatments are not cost effective.

These inclinations may explain some of the discrepancies we found (eg, preferred formulary placement of drugs with high cost-utility ratios, such as drugs for multiple sclerosis and amyotrophic lateral sclerosis). 

There may be pressure to cover certain therapies even if they have unfavorable cost-effectiveness ratios. This may be a particular concern for Medicaid programs, which may be more sensitive to public pressure than private plans. 

Despite these concerns, we believe that our analyses provide important information and hope that they further policy discussions about the need to move the healthcare system toward value-based benefit design. Many reimbursement authorities abroad have used CUAs (and other forms of economic evaluation) to inform formulary decisions. In the United States, initiatives such as the US Preventive Services Task Force have begun to consider cost effectiveness in their deliberations. Moreover, although health plans may not adopt a societal perspective, they do compete for enrollees and it is not unreasonable to believe that CUAs approximate value in many and even most cases. For these reasons, we believe more consideration of data from CUAs would represent a positive step forward for US formularies. 

Despite limitations, CUA remains an important technique for measuring value. The number of CUAs has been growing and their quality may be improving.

 Some payers have found that CUAs can be an important part of a sequenced assessment of value that begins with the application of evidence-based medicine. In this characterization, consideration of clinically important benefits and harms, as well as the subpopulations for which products have specific advantages, forms the basis for assessments of medication value. Cost-utility analyses can be a useful extension of that assessment when they focus on the specific clinical situations for which the medication is most appropriate. 

Some formulary committees are beginning to incorporate evidence of value explicitly into their deliberations and in some cases are requesting or requiring that drug firms formally submit evidence of value to their formulary committees.

Our study suggests that these efforts should go further. 

Our data suggest a need for formulary policies to offer more targeted guidance for drugs. Formularies tend to be blunt instruments, broadly covering the use of a drug. Yet whether a drug is cost effective depends on the specific characteristics of the recipients (eg, age, clinical condition). Ideally, formulary policy will be defined accordingly. For example, guidelines for lipid-lowering drugs reflect patient-specific characteristics such as prior infarction. In practice, however, it is a challenge to incorporate heterogeneity in potential benefit among patients.

More and better data will be needed if we are to truly move toward value-based formulary decision making. Congress and other federal officials also might help by addressing gaps in the evidence base, for example, by funding more "pragmatic" clinical trials to aid decision makers,

and by funding cost-effectiveness analyses.

In the future it will be important to update these analyses and expand them to include other health plans. As more studies provide information on drug value, our next challenge is to incorporate value into formularies to increase the use of effective and cost-effective medications. 

From the Department of Health Policy and Management and Center for Risk Analysis, Harvard School of Public Health, Boston, Mass (PJN, PJL, DG, MCW); the Department of Health Policy and Administration, School of Public Health, University of North Carolina at Chapel Hill, Chapel Hill, NC (PJL); the Department of Health Systems Management, Faculty of Health Sciences, Ben-Gurion University of the Negev, Beer-Sheva, Israel (DG); Merck & Co, Inc, West Point, Pa (MB, ST, EM); and the Division of General Medicine, Department of Health Management and Policy, University of Michigan Schools of Medicine and Public Health, and Center for Practice Management and Outcomes Research, Ann Arbor Veterans Affairs Medical Center, Ann Arbor, Mich (ABR). 

This project was supported by grant R01 HS10919 from the Agency for Healthcare Quality and Research and an unrestricted grant from Merck and Co. 

Do Drug Formulary Policies Reflect Evidence of Value?

 To compare the cost effectiveness of prophylaxiswith a low-molecular-weight heparin with that of prophylaxis withunfractionated heparin for the prevention of venous thromboembolismin acutely ill medical inpatients.

 Cost-effectiveness analysis based on decision-treemodel.

  A hypothetical cohort of 10 000patients was assumed to receive either (1) prophylaxis with enoxaparin,a low-molecular-weight heparin, 40 mg daily; (2) prophylaxiswith unfractionated heparin, 5000 IU twice daily; or (3) noprophylaxis. We developed a decision-analytic model with parameterestimates derived from published clinical trials and other secondarysources. Then, for each strategy, we estimated the risks ofvenous thromboembolism, complications of prophylaxis and treatment(heparin-induced thrombocytopenia and bleeding), mortality,and costs of prophylaxis and treatment within a 30-day period.

 In a hypothetical cohort of 10 000 inpatients, expectednumbers of deaths attributable to venous thromboembolism ordrug complications related to both prophylaxis for and treatment ofVTE over a 30-day period were 37 with enoxaparin prophylaxis, 53with unfractionated heparin prophylaxis, and 81 with no prophylaxis.In 2001, corresponding expected costs of prevention, diagnosis,and management of VTE were $3 502 000 for enoxaparin,$3 772 000 for unfractionated heparin, and $3 105 000 for no prophylaxis.The incremental cost per death averted with enoxaparinprophylaxis versus no prophylaxis was $9100. Enoxaparin dominatedunfractionated heparin by being both more effective and lesscostly in the base-case analysis, as well as in sensitivity analyses inwhich equal efficacy and equal risks of bleeding were assumed.

 Thromboprophylaxis with this low-molecular-weightheparin represents a cost-effective use of healthcare resourcesin acutely ill medical inpatients and dominatesthromboprophylaxis with unfractionated heparin.

Venous thromboembolism (VTE) is a life-threateningvascular disease common among hospitalizedpatients. It is estimated that more than 170000 incident cases of VTE are diagnosed in hospitalizedpatients in the United States each year, and morethan 20 000 of these patients die of the condition beforedischarge.

  Because many cases of incident thromboembolismgo undetected and untreated, primary preventionin at-risk patients is of critical importance.

Several subgroups of hospitalized patients are at elevatedrisk for VTE, including those undergoing orthopedicor other major surgery, as well as those with acutecirculatory conditions such as myocardial infarction,atrial fibrillation, angina, or thrombotic stroke. TheSixth ACCP Guidelines for Antithrombotic Therapy forPrevention and Treatment of Thrombosis recommendthromboprophylaxis as standard practice for thesepatients.

 In recent years, the substantial risk of VTE inthe general medical population has emerged as animportant clinical concern also.

 Acutely ill medicalinpatients, such as those hospitalized for congestiveheart failure, chronic obstructive pulmonary disease, oracute infections, often have hypercoagulability andperiods of prolonged immobility that may place them atincreased risk of VTE.

 Because these patients are likelyto have one or more additional risk factors (eg, historyof VTE, advanced age, obesity, varicose veins,estrogen use),

 the potential benefits of thromboprophylaxisin this population may be substantial.

As the risk of VTE among general medical patientshas gained recognition, the introduction of low-molecular-weight heparin (LMWH) has provided an alternativeto thromboprophylaxis with unfractionated heparin(UFH), the previous standard. The efficacy of bothLMWH and UFH as thromboprophylaxis in selectedgroups of medical inpatients has been demonstrated inclinical trials,

 and there is evidence that LMWH mayhave a better safety profile than UFH with regard to bothhemorrhage

 and heparin-induced thrombocytopenia(HIT).

 Because LMWH is more costly, however,the question remains whether its use can be justified interms of reduced risk ofdrug-related adverseevents and, possibly,improved efficacy. Toaddress this question,we examined the costeffectiveness of prophylaxiswith an LMWH,specifically enoxaparin,versus prophylaxis withUFH in acutely ill generalmedical inpatients.

We used techniquesof decision analysis anddata from secondarysources to develop a model of the prevention, detection,treatment, and outcomes of VTE in generalmedical patients. A hypothetical cohort of 10 000acutely ill medical inpatients was assumed to receiveone of three methods of thromboprophylaxis:(1) enoxaparin, a representative LMWH (Lovenox;Aventis Pharmaceuticals; Bridgewater, NJ), 40 mgadministered subcutaneously once daily; (2) UFH,5000 IU administered subcutaneously twice daily; or(3) no prophylaxis. Our choice of dosing and frequencyof administration reflects current recommendationsfor thromboprophylaxis in medical patients.

 For eachstrategy, we estimated the 30-day risks of VTE,including deep vein thrombosis (DVT) and/or pulmonaryembolism (PE); complications of prophylaxisand therapy, such as HIT and bleeding; mortality; andthe costs of prophylaxis, diagnostic testing, and treatment.We used these estimates to examine the incrementalcost effectiveness of thromboprophylaxis,expressed in terms of the cost per death averted.

Treatment patterns and outcomes of care were modeledfor 30 days from the time of hospital admission.This time period was chosen to capture DVT and PEoccurring both in the hospital and during the periodimmediately following discharge, during which time patientsmay continue to be at elevated risk of VTE. Ouranalysis was conducted from the perspective of a thirdpartyhealth insurer and therefore considered only directmedical costs.

The structure of the decision-tree model is illustratedin 

. The clinical starting point forthe model is admission to an acute-care facility fortreatment of a serious medical condition. For thesepatients, the clinical decision of interest is the methodof thromboprophylaxis administered, either enoxaparin,UFH, or no prophylaxis (Figure 1). Patientsreceiving enoxaparin or UFH are assumed to be at riskfor drug-related adverse events, including hemorrhage("bleed") and HIT. Bleeds may be either minor or major,while HIT is classified as asymptomatic or symptomaticaccording to the presence or absences of overtclinical symptoms. Major bleeds and symptomaticHIT may lead to death, though for modeling conveniencewe assumed that patients who survive one of theseadverse events are not at risk of additional adverseevents or VTE.

Patients who do not experience HIT or hemorrhageare at risk of DVT (Figure 2). Because clinical diagnosisof DVT is imperfect, the tree reflects this uncertaintyby allowing patients to receive a true- orfalse-positive or negative clinical diagnosis. Patientswith a positive clinical diagnosis (whether true or false)are referred for duplex ultrasound and, if confirmed, fortreatment. Those with a negative clinical diagnosisreceive no confirmatory diagnostic procedures ortreatment. This model assumes that all patientswhose DVT diagnosis is confirmed by ultrasoundundergo anticoagulation therapy with enoxaparin;such patients are at risk for bleeds and HIT. Patients whosurvive treatment-related adverse events, such as bleedsor HIT, finish treatment and are not at risk of furtheradverse events or PE.

Patients who have DVT and do not experience anadverse event from its treatment are at risk for PE(Figure 3). Patients who experience PE and survive theperiod immediately after the acute event may be giveneither a true-positive or a false-negative clinical diagnosisof PE. Patients with apositive clinical diagnosis(whether true or false) arereferred for a confirmatoryventilation—perfusion (V/Q)scan or computed tomography(CT) scan and, if positive,for treatment of PE.Patients may die fromtreated 

 untreated PE orfrom other underlyingmedical conditions.

Model estimation involvesassigning probabilityestimates to each of themodel's chance nodes andcost estimates to each pathway.Base-case transitionprobabilities and datasources are summarized in

Risks of failure and theoccurrence of adverseevents for each prophylaxismethod were obtained from the published literature. Weused data from MEDENOX,

 a recent placebo-controlledclinical trial, in our model, evaluating 40 mgenoxaparin once daily as thromboprophylaxis in medicalpatients to estimate the risk of DVT with enoxaparinprophylaxis and the baseline risk of DVT amonguntreated patients. MEDENOX is the only publishedplacebo-controlled trial of an LMWH that enrolled generalmedical patients, administered a dose of enoxaparinconsistent with current recommendations,

 and diagnosedDVT using venography. We were unable to identifyany published studies that compared an LMWH toeither UFH or placebo and met these criteria. Therefore,to estimate the risk of prophylaxis failure among patientsreceiving UFH, we used the results of a recent meta-analysisof trials that compared LMWH to UFH in medicalpatients

 and applied the summary risk-ratio of 0.83 tothe risk of prophylaxis failure in the LMWH arm of theMEDENOX trial.

Because HIT and bleeds occurred infrequently in thepublished trials of thromboprophylaxis in medical inpatientsand because their detection does not depend onvenography, estimates of the rates of occurrence andseverity of these events were drawn from a wider rangeof clinical trials than those used for the efficacy estimates.

 Published estimates of "symptomatic" or"severe" HIT

 were assumed to include thrombocytopenialeading to thromboembolism. The other transitionprobabilities in the model are uniform across alltreatment arms, and most were estimated from the publishedliterature. The estimate of the probability of false-negativeclinical diagnosis of DVT was derived fromstudies examining the prevalence of DVT in asymptomaticpatients.

 For the base—case analysis, the riskof death due to the underlying medical condition wasassumed to be 10%, which is consistent with the rates ofdeath reported in published clinical trials focusing onacutely ill medical inpatients.

 Finally, because weassumed that one half of the patients with suspected PEwould undergo CT scans and one half V/Q scans, weused the average sensitivity and specificity of these 2diagnostic techniques in the model. Where multipledata sources were used, estimates were combinedacross studies by direct pooling.

The model's cost parameters include (1) study medicationsand ancillary resources, (2) laboratory procedures,(3) hospital inpatient days, (4) physician inpatientvisits, and (5) physician office/outpatient visits. All costestimates employed in the analysis are summarized in

Drug acquisition costs were estimated from averagewholesale prices

; for the weight-based doses, the totalamount of drug used was estimated assuming a 75-kgpatient. Cost estimates associated with the suppliesused to administer drugs were obtained from publisheddata.

 The costs of laboratory procedures for monitoringanticoagulation therapy were estimated using datafrom the Centers for Medicare & Medicaid Services

;and the costs of physician inpatient and office/outpatientvisits were estimated using the Medicare Resource-Based Relative Value Scale payment rates.

 Mean dailyinpatient costs attributable to adverse drug reactionsand VTE were estimated by dividing the unadjustedpayment for deep vein thrombophlebitis (DiagnosisRelated Group 128) by the average length of stay forthis diagnosis.

Although our primary focus was on costs incurredwithin 30 days of hospital admission, we included costsafter 30 days if they were incurred as a direct consequenceof events within the first 30 days (eg, 6 monthsof warfarin therapy following DVT). All costs weremeasured in 2001 US dollars; where necessary, costswere updated to 2001 using the medical-care componentof the Consumer Price Index.

 Because of the brieftime horizon of the model, discounting was notrequired.

For each of the 3 prophylaxis methods being evaluated,we used our model to estimate the total costs of prevention,detection, treatment, and outcomes of VTE in acohort of 10 000 acutely ill medical inpatients over a 30-day period. Cost effectiveness is expressed in terms ofthe cost per death averted. To compare treatment strategieswe first rank-ordered regimens by increasing costand eliminated those strategies that were found bothmore costly and less effective than (

 by) others.For each remaining strategy we then calculated theadditional cost and additional deaths averted versus thenext lowest cost strategy. The ratio of these 2 numbers(ie, the incremental cost per death averted) is the incrementalcost-effectiveness ratio.

Sensitivity analyses were performed to assess therobustness of key model parameters and to evaluatethe effect of alternative assumptions on model outcomes.To explore the cost effectiveness of thromboprophylaxisin patients with different comorbidity andrisk profiles, we analyzed subgroups of patients fromthe MEDENOX trial with heart failure, respiratory disease,or infectious disease.

 We varied the backgrounddeath rate and the underlying risk of DVT in patientsreceiving no prophylaxis, as well as the relative safetyand efficacy of the 2 active prophylaxis arms. Lastly,we examined the effect on our findings of changingassumptions about the cost of drug therapy, risk ofbleeding due to prophylaxis, progression from DVT toPE, and death from PE.

Base-case estimates of the occurrence of DVT, PE, andVTE-related death in a cohort of 10 000 acutely ill generalmedical inpatients are presented in 

. The riskof each of these outcomes is lowest among patientsreceiving thromboprophylaxis with enoxaparin and highestamong those receiving no prophylaxis. Compared tono thromboprophylaxis, enoxaparin prophylaxis isexpected to prevent 860 cases of DVT, 171 cases ofPE, and 44 deaths, while UFH is expected to prevent776 cases of DVT, 155 cases of PE, and 28 deaths.

The total costs of thromboprophylaxis, adverseevents (related to both prophylaxis and treatment ofVTE), and diagnosis and management of DVT andPE are presented in 

. Patients receiving nothromboprophylaxis had the lowest total costs, followedby those receiving enoxaparinand UFH.

Results of the incremental analysiscomparing a strategy of no prophylaxiswith the 2 active prophylactic strategiesare presented in 

. Whencompared with no prophylaxis, prophylaxiswith enoxaparin was the morecost-effective strategy, at a cost of$9100 per death averted. Unfractionatedheparin prophylaxis was dominatedby enoxaparin prophylaxis, whichwas both less costly and more effectivein preventing deaths.

. Results for thesubgroups of patients with heart failure,respiratory disease, or infectious disease were similarto those for all acutely ill medical patients (

).For each subgroup, thromboprophylaxis with enoxaparindominated that with UFH, and costs per deathaverted for enoxaparin versus no prophylaxis were evenmore favorable than in the base case.

 Our findingsalso were relatively insensitive to variations in theassumed competing risk of death (10% in the base-caseanalysis). Eliminating the competing risk of deathmade the results slightly more favorable forenoxaparin. Increasing the underlying risk ofdeath to 20% increased the cost per deathaverted with enoxaparin versus no prophylaxisby only about $1000 and did not changethe dominance of enoxaparin over UFH. Themodel was more sensitive to changes in theassumed risk of DVT. When we reduced therisk of DVT in all groups by one half (ie, to2.8% with enoxaparin, 3.3% with UFH, and7.1% with no prophylaxis), the cost per deathaverted for enoxaparin versus no prophylaxisincreased to $56 300, and enoxaparin continuedto dominate UFH. In contrast, when weincreased the risk of DVT by one half (ie, to8.3% with enoxaparin, 9.9% with UFH, and21.3% with no prophylaxis), both activemethods of prophylaxis were less costly andmore effective than no prophylaxis, withenoxaparin continuing to dominate UFH.When we conducted a similar analysis inwhich we simultaneously varied the risk ofbleeds with both enoxaparin and UFH from 50% to 150%of the base-case estimates, we found that enoxaparindominated UFH under both the high-risk and low-riskscenarios, with cost per death averted versus no prophylaxisranging from about $5000 assuming lower riskof bleeds to $14 000 assuming higher risk of bleeds.

Effectiveness and Side Effects of Enoxaparin vsUFH.

 To determine if our results depended on theassumption that enoxaparin had an efficacy advantageover UFH, we assumed equal efficacy (ie, 5.5% probabilityof DVT) for both methods of prophylaxis. Under thisassumption, enoxaparin remained both less costly andmore effective than UFH because of the reduced risk ofHIT and hemorrhage. Similarly, enoxaparin continuedto dominate UFH when equal rates of hemorrhage wereassumed for the 2 prophylactic strategies. When bothequal efficacy 

 equal rates of hemorrhage wereassumed, we found that UFH was no longer dominatedby enoxaparin, and the incremental cost effectiveness ofenoxaparin versus UFH was $43 256 per death averted.

Milton C. Weinstein, PhD

Articles

Cost-Effectiveness of Laparoscopic Gastric Banding and Bypass for Morbid Obesity

Do Drug Formulary Policies Reflect Evidence of Value?

Cost Effectiveness of Thromboprophylaxis With a Low-Molecular-Weight Heparin Versus Unfractionated Heparin in Acutely Ill Medical Inpatients

Barriers to Using Cost-Effectiveness Analysis in Managed Cae Decision Making