Search

Objectives. In 2016, the Second Panel on Cost-effectiveness in Health and Medicine updated the seminal work of the original panel from 2 decades earlier. The Second Panel had an opportunity to reflect on the evolution of cost-effectiveness analysis (CEA) and to provide guidance for the next generation of practitioners and consumers. In this article, we present key topics for future research and policy. Methods. During the course of its deliberations, the Second Panel discussed numerous topics for advancing methods and for improving the use of CEA in decision making. We identify and consider 7 areas for which the panel believes that future research would be particularly fruitful. In each of these areas, we highlight outstanding research needs. The list is not intended as an exhaustive inventory but rather a set of key items that surfaced repeatedly in the panel’s discussions. In the online Appendix , we also list and expound briefly on 8 other important topics. Results. We highlight 7 key areas: CEA and perspectives (determining, valuing, and summarizing elements for the analysis), modeling (comparative modeling and model transparency), health outcomes (valuing temporary health and path states, as well as health effects on caregivers), costing (a cost catalogue, valuing household production, and productivity effects), evidence synthesis (developing theory on learning across studies and combining data from clinical trials and observational studies), estimating and using cost-effectiveness thresholds (empirically representing 2 broad concepts: opportunity costs and public willingness to pay), and reporting and communicating CEAs (written protocols and a quality scoring system). Conclusions. Cost-effectiveness analysis remains a flourishing and evolving field with many opportunities for research. More work is needed on many fronts to understand how best to incorporate CEA into policy and practice.

Background. Simulation models designed to evaluate cancer prevention strategies make assumptions on background mortality—the competing risk of death from causes other than the cancer being studied. Researchers often use the U.S. life tables and assume homogeneous other-cause mortality rates. However, this can lead to bias because common risk factors such as smoking and obesity also predispose individuals for deaths from other causes such as cardiovascular disease. Methods. We obtained calendar year-, age-, and sex-specific other-cause mortality rates by removing deaths due to a specific cancer from U.S. all-cause life tables. Prevalence across 12 risk factor groups (3 smoking [never, past, and current smoker] and 4 body mass index [BMI] categories [2]) were estimated from national surveys (National Health and Nutrition Examination Surveys [NHANES] 1971–2004). Using NHANES linked mortality data, we estimated hazard ratios for death by BMI/smoking using a Poisson regression model. Finally, we combined these results to create 12 sets of BMI and smoking-specific other-cause life tables for U.S. adults aged 40 years and older that can be used in simulation models of lung, colorectal, or breast cancer. Results. We found substantial differences in background mortality when accounting for BMI and smoking. Ignoring the heterogeneity in background mortality in cancer simulation models can lead to underestimation of competing risk of deaths for higher-risk individuals (e.g., male, 60-year old, white obese smokers) by as high as 45%. Conclusion. Not properly accounting for competing risks of death may introduce bias when using simulation modeling to evaluate population health strategies for prevention, screening, or treatment. Further research is warranted on how these biases may affect cancer-screening strategies targeted at high-risk individuals.

Purpose. When using state-transition Markov models to simulate risk of recurrent events over time, incorporating dependence on higher numbers of prior episodes can increase model complexity, yet failing to capture this event history may bias model outcomes. This analysis assessed the tradeoffs between model bias and complexity when evaluating risks of recurrent events in Markov models. Methods. The authors developed a generic episode/relapse Markov cohort model, defining bias as the percentage change in events prevented with 2 hypothetical interventions (prevention and treatment) when incorporating 0 to 9 prior episodes in relapse risk versus a model with 10 such episodes. Magnitude and sign of bias were evaluated as a function of event and recovery risks, disease-specific mortality, and risk function. Results. Bias was positive in the base case for a prevention strategy, indicating that failing to fully incorporate dependence on event history overestimated the prevention’s predicted impact. For treatment, the bias was negative, indicating an underestimated benefit. Bias approached zero as the number of tracked prior episodes increased, and the average bias over 10 tracked episodes was greater with the exponential compared with linear functions of relapse risk and with treatment compared with prevention strategies. With linear and exponential risk functions, absolute bias reached 33% and 78%, respectively, in prevention and 52% and 85% in treatment. Conclusion. Failing to incorporate dependence on prior event history in subsequent relapse risk in Markov models can greatly affect model outcomes, overestimating the impact of prevention and treatment strategies by up to 85% and underestimating the impact in some treatment models by up to 20%. When at least 4 prior episodes are incorporated, bias does not exceed 26% in prevention or 11% in treatment.

Clinical journals increasingly illustrate uncertainty about the cost and effect of health care interventions using cost-effectiveness acceptability curves (CEACs). CEACs present the probability that each competing alternative is optimal for a range of values of the cost-effectiveness threshold. The objective of this article is to demonstrate the limitations of CEACs for presenting uncertainty in cost-effectiveness analyses. These limitations arise because the CEAC is unable to distinguish dramatically different joint distributions of incremental cost and effect. A CEAC is not sensitive to any change of the incremental joint distribution in the upper left and lower right quadrants of the cost-effectiveness plane; neither is it sensitive to radial shift of the incremental joint distribution in the upper right and lower left quadrants. As a result, CEACs are ambiguous to risk-averse policy makers, inhibit integration with risk attitude, hamper synthesis with other evidence or opinions, and are unhelpful to assess the need for more research. Moreover, CEACs may mislead policy makers and can incorrectly suggest medical importance. Both for guiding immediate decisions and for prioritizing future research, these considerable drawbacks of CEACs should make us rethink their use in communicating uncertainty. As opposed to CEACs, confidence and credible intervals do not conflate magnitude and precision of the net benefit of health care interventions. Therefore, they allow (in)formal synthesis of study results with risk attitude and other evidence or opinions. Presenting the value of information in addition to these intervals allows policy makers to evaluate the need for more empirical research.

Purpose. The authors propose a Bayesian approach for estimating competing risks for inputs to disease simulation models. This approach is suggested when modeling a disease that causes a large proportion of all-cause mortality, particularly when mortality from the disease of interest and other-cause mortality are both affected by the same risk factor. Methods. The authors demonstrate a Bayesian evidence synthesis by estimating other-cause mortality, stratified by smoking status, for use in a simulation model of lung cancer. National (US) survey data linked to death registries (National Health Interview Survey [NHIS]—Multiple Cause of Death files) were used to fit cause-specific hazard models for 3 causes of death (lung cancer, heart disease, and all other causes), controlling for age, sex, race, and smoking status. Synthesis of NHIS data with national vital statistics data on numbers and causes of deaths was performed in WinBUGS (version 1.4.1, MRC Biostatistics Unit, UK). Correction for inconsistencies between the NHIS and vital statistics data is described. A published cohort study was a source of prior information for smoking-related mortality. Results. Marginal posterior densities of annual mortality rates for lung cancer and other-cause death (further divided into heart disease and all other causes), stratified by 5-year age interval, race (white and black), gender, and smoking status (current, former, never), were estimated, specific to a time period (1987-1995). Overall, black current smokers experienced the highest mortality rates. Conclusions. Bayesian evidence synthesis is an effective method for estimation of cause-specific mortality rates, stratified by demographic factors.

Background. Evaluating cancer screening often requires modeling the underlying disease process and not observed disease, particularly in the absence of direct evidence linking screening to a survival benefit. Methods. To illustrate a potential error in modeling disease progression among healthy persons with a history of a precancerous lesion, we constructed 2 models with 4 basic health states (disease free, presence of a precancerous lesion, presence of cancer, dead), calibrated to predict the same 10-year cancer incidence. We assumed a homogeneous cohort enters each model free of disease, the probability of developing a precancerous lesion was greater for patients with a history of a prior lesion, and the screening test was perfect and riskless. In one model, we assigned a higher transition probability from a precancerous lesion to cancer in those with a history of a previously removed lesion; in the other, we assumed it was equal to those with no history. Results. Using the 1st model, life expectancy without screening was 2.4 months longer than with screening. This error did not occur using the 2nd model, in which the transition from precancerous lesions to cancer was not conditional on a history of a lesion. This modeling error's magnitude was examined under a variety of assumptions. Conclusions. We have identified an important error to avoid when modeling the underlying disease process in evaluating screening programs for cancers associated with precancerous states.

Background. Many patients with multiple sclerosis (MS) eligible for β -interferons or glatiramer acetate have chosen to forgo or discontinue treatment. The objective of this study was to evaluate risk attitude as a patient characteristic related to treatment choice for patients with MS. Methods. Sixty-two MS patients completed a survey on treatment history, risk preference, and socioeconomic and clinical variables. Multinomial logistic regression was used to assess the relationship between treatment choice and risk attitude. Risk attitude was measured using a standard gamble question on short-term health outcomes. Results. More risk-seeking patients were less likely to choose treatment compared with more risk-averse patients (P < 0.01). For patients who discontinued treatment, the explanatory variable of significance was severity of side effects (P < 0.05). Conclusions. Risk attitude is a patient characteristic related to treatment initiation in patients with MS. This could be an important factor to consider when identifying optimal treatment decisions for individual patients.

An important assumption made when constructing a Markov model is that all persons residing in a health state are identical. Failure to adjust for population heterogeneity caused by unobserved variables can therefore cause bias in model results. The authors used a simple model to evaluate the potential impact of heterogeneity bias, defined as the percentage change in the life expectancy gain with an intervention predicted by a model that does not adjust for heterogeneity (unadjusted model) compared to one that does (adjusted model). The life expectancy gains were consistently greater in the unadjusted model compared to the adjusted model (positive bias). For an annual probability of developing disease of 1%, the heterogeneity bias exceeded 50% when the relative risk of disease with the heterogeneity factor versus without the factor was greater than 15 and the prevalence of the heterogeneity factor was between 5% and 25%. When constructing decision-analytic models, analysts need to be cognizant of unobserved factors that introduce heterogeneity into the cohort. This analysis provides a general framework to determine when issues of heterogeneity may be important.

Clinical decision models often rely upon survival models predicated on disease-specific hazard functions combined with baseline hazard functions obtained from standard life tables. Two biases may arise in such a modeling process. First, life expectancy estimates may be biased even if estimates of survival probabilities are unbiased (misestimation bias). In sim ulation studies, the authors discovered that the magnitude of misestimation bias is larger as life expectancy increases, sample size decreases, and censoring percentage increases. In the context of a simple decision analysis, they found that imbalances in the sample sizes for the data used to estimate the parameters among different strategies resulted in non- optimal decisions in the long run. The second bias stems from misspecification of the survival model itself (misspecification bias). Using a simple cost-effectiveness model, the authors found that life expectancies and incremental cost-effectiveness ratios differed depending on whether an excess-mortality or a proportional-hazards model was specified. In addition, a predictable pattern was observed for these two survival models when extrapolated to other age and gender groups. Key words: Markov models; Markov-cycle tree; decision making; excess-mortality model; proportional-hazards model; life expectancy. (Med Decis Making 1995;15:158-169)