TRA-2
CONCURRENT PRESENTATIONS OF TOP RATED ABSTRACTS II: MODELING AND POPULATION HEALTH
* Finalists for the Lee B. Lusted Student Prize
Purpose: The human genotype CYP2B6 can predict plasma drug levels for the HIV drug efavirenz. We assessed the potential cost-effectiveness of CYP2B6 genetic testing to inform efavirenz dose reduction for initial antiretroviral therapy (ART) in HIV disease.
Method: We used the Cost-Effectiveness of Preventing AIDS Complications (CEPAC) microsimulation model to project quality-adjusted life expectancy and lifetime costs (2012 US dollars) for initiating efavirenz-based ART with or without CYP2B6 genetic testing. Genotyping of three CYP2B6 polymorphisms identifies patients eligible for reduced dose (200 mg/day in slow metabolizers and 400 mg/day in intermediate metabolizers) rather than standard dose therapy (600 mg/day). We assumed that after genetic testing, 47% of patients would be eligible to reduce from 600 to 400 mg and 13% eligible to reduce from 600 to 200 mg, based on published data. Cost of 600 mg efavirenz was $7,300/year, and for generic efavirenz it was $2,400/year. In sensitivity analyses we varied the probability of toxicity with and without dose reduction, treatment efficacy with dose reduction, population characteristics, and availability of less expensive generic efavirenz. We also considered a universal dose reduction strategy (i.e. no genetic testing) to 400 mg/day for all patients. Costs and quality-adjusted life years (QALYs) were discounted at 3% annually.
Result: Initiating efavirenz-based ART at CD4 counts <500/mm3 without genetic testing had a per person discounted life expectancy of 13.43 QALYs and discounted lifetime cost of $419,600. Genetic testing followed by dose reduction decreased lifetime cost by $16,200 with no change in QALYs, assuming equal efficacy after dose reduction. With no generic efavirenz available, genetic testing remained cost-effective (standard dose versus genetic testing ICER > $100,000/QALY) after dose reduction. With generic efavirenz available, standard dose was preferred if efficacy decreased 9% or more after dose reduction (Figure). In a scenario including a universal dose reduction strategy, universal dose reduction was preferred unless generic efavirenz was available and efficacy decreased 7% or more after dose reduction.
Conclusion: Genotyping of CYP2B6 may be effective to inform efavirenz dose reduction strategies and lower the cost of HIV therapy. A universal dose reduction strategy could further lower this cost with minimal life expectancy changes. Results depend on the efficacy of reduced dose regimens if generic efavirenz is available.
Purpose:
The decision to perform coronary artery bypass grafting (CABG) or percutaneous coronary intervention (PCI) is complex and depends on many factors, including age, disease morphology, and patient comorbidities. Previously, we developed a prediction model which used patients' age, sex, year of treatment, and present-on-admission diagnosis codes to generate predicted probabilities of in-hospital mortality with each treatment. The model recommended whichever treatment minimized the predicted probability of in-hospital mortality. Management was deemed discordant with the model when the actual treatment differed from the model-recommended treatment. In this study, we evaluated the association between hospital discordance rates (HDRs) and risk-adjusted in-hospital mortality.
Method:
We analyzed 550,958 discharges with either PCI or CABG as a primary procedure, using the 2009-11 AHRQ State Inpatient Databases for six states (AZ, CA, FL, MD, MI, NJ). Hospitals were included if they had >300 total such discharges; those not performing CABG were excluded. Hospital observed-to-expected (O/E) in-hospital mortality ratios were modeled using mixed effects logistic regression with random intercepts. Expected mortality was defined as the predicted probability of mortality under the recommended treatment, under the hypothesis that increasing HDR is associated with increasing O/E ratios. As a negative control, we redefined expected mortality as the predicted probability of mortality under the actual treatment administered, under the hypothesis that, using this definition, there is no association between HDR and O/E ratio.
Result:
283 hospitals met inclusion criteria. HDRs ranged from 5.2% to 61.9%, with a mean (SD) of 29.7% (8.0%). Using recommended treatment to define expected mortality rates, we found a significant association between HDR and hospital O/E ratio (see Figure): each increase of HDR of 10% associated with a factor [95% CI] of 1.12 [1.05, 1.19] change in hospital O/E ratio. In the negative control analysis, there was no such association: the change in O/E ratio associated with a 10% increase in HDR was a factor of 1.02 [0.96, 1.08].
Conclusion:
We found heterogeneity in decision making among hospitals – at least in comparison with model recommendations – which was significantly associated with risk-adjusted mortality. Large-scale decision models, if accurate and calibrated, may help improve outcomes by empirically expressing risk under competing treatments, for patients' specific sets of comorbidities.
Purpose: The beta distribution serves as a prior distribution for a binomial probability. In the case of a bivariate binomial distribution with two probabilities the corresponding prior can be a bivariate beta distribution on the unit square. This distribution can serve as a proper prior for correlated binomial responses. For example, in a Bayesian setting it can be used to model the sensitivities (specificities) of two index tests, based on the cross-classification of test results in a group of people with (without) disease.
Method: We seek a prior distribution for the bivariate binomial distribution that (i) has beta marginal distributions, (ii) has support on the unit square, and (iii) allows positive and negative correlations throughout the whole range (-1, 1). Existing bivariate beta distribution families such as the Farlie–Gumbel–Morgenstern and Sarmanov families only allow for narrower correlation ranges.
We use an additive construction scheme: Let U11, U10, U01 have a Dirichlet distribution. We define the additive version X= U11+U10 and Y=U01+U01; then X, Y have a bivariate beta distribution. We derive the joint density of (X, Y), show that it satisfies (i)-(iii), and derive central moments. We present algorithms for fitting the joint density to data and for Bayesian inference.
Result: The proposed distribution is defined by 4 positive parameters. Shown are example densities when the parameters are (1, 1, 1, 1); (4, 2, 4, 1); and (2, 2, 2, 0.5), respectively:
We demonstrate use by modeling the sensitivities (specificities) of two ultrasonographic markers for detecting trisomy 21 in liveborn infants. Extension to k dimensions follows the same construction and the k-dimensional joint distribution is defined by 2k parameters.
Conclusion: We provide a bivariate beta distribution using an additive construction scheme that allows correlations in the full range (-1, 1). This is an alternative to the Farlie–Gumbel–Morgenstern, Plackett, Mardia, and Sarmanov bivariate beta distribution families, which can have a much more restrictive correlation range.
Purpose: Current guidelines on colorectal cancer (CRC) screening do not consider gender nor do they differ by age. Recent evidence, though, has shown that polyps form earlier and progress faster in men than women. Economic models for tailoring the guidelines need efficient and accurate quantification of model parameters, especially those governing the incidence and growth of adenomas for different age-gender groups. However, this model calibration is challenging given the large numbers of model parameters and outcomes. The purpose of our study is to generate insights into calibration of discrete-event simulations for CRC screening cost-effectiveness analyses.
Methods: We developed a progressive fitting procedure to calibrate a discrete-event simulation model, which extended the 2007 Vanderbilt-NCSU CRC (VNCS) model by replacing each probability distribution governing adenoma state dwelling with a set of age-gender-specific distributions and synthesizing the probability distribution parameters into three growth rate parameters (incidence of non-advanced adenoma; progression to advanced-adenoma; and progression to CRC) for each age-gender group. Note that when VNCS was built, there were no age-gender-specific data on transitions from non-advanced to advanced adenoma.
To estimate the growth rate parameters, the developed procedure started with younger populations. It ran the simulation with smaller sample size and fewer replications for computational efficiency and progressively increased the size and replications to ensure statistical significance of the fitted results. In the calibration, we compared the average simulated non-advanced and advanced adenoma incidences with recent evidence from a 7-year observational study published recently. For each set of parameters pertaining to a gender and a 5-year age group, we created 35 hypothetical cohorts, each of which corresponds to particular birth and screening years.
Results: Our calibration procedure produced results comparable to recent evidence. We captured the features in the simulation that a male cohort had higher risks of developing both non-advanced and advanced adenomas than a female cohort of the same age; the risks for both genders were significantly different between age groups.
Conclusions: We proposed an efficient and accurate calibration procedure for economic studies of CRC screening strategy tailoring. The procedure took into account the disease natural history and exploited recent developments in model-based parameter estimation. We demonstrated its performance by updating the VNCS model. More broadly, the developed procedure can be applied to other simulation-based economic studies on individualized medicine.
Methods: Heuristically, a constant HR for leaving the collapsed state can be derived by equating the mean dwelling time (MDT) in the collapsed state to the sum of the MDT in the k states and then inverting the sum of MDT. In method 2, we observe that the distribution of the sum of the dwelling times of a continuous-time Markov model with k states and constant HRs is mathematically equivalent to that of a 1-state model with a time-dependent HR derived from the k HRs (a generalized-Erlang distribution). We analytically derived the HR of the generalized-Erlang distribution to compute the HR of the collapsed state. We applied both methods to collapse the following states in our CHC model: mild CHC from Metavir-stages F0 and F1 and moderate CHC from stages F2 and F3. The HRs of progressing from mild to moderate CHC and from moderate CHC to cirrhosis were derived from the Metavir-stage-specific HRs. To compare the accuracy of each method, we calculated the absolute and relative differences in MDT (metric-1) and the relative differences in the distributions of health states across all cycles (metric-2) between the models with the collapsed and separated states.
Results: The analytical and the heuristic methods yielded absolute (relative) differences in MDT of 0.006 (0.03%) and 1.21 (6.54%) years, respectively. In terms of metric-2 which considers the timing of events, the analytical method generated a significantly lower relative error than the heuristic method, 0.04% vs. 49.75%. The heuristic method became less accurate in reproducing the correct distributions of states at all cycles since the method used a constant HR in the collapsed state.
Conclusions: Time-dependent HR of the generalized-Erlang should be used for collapsing states to accurately reproduce the MDT and the cycle-specific distribution of states.
Method: We analyzed Medical Expenditure Panel Survey (MEPS) data, a nationally representative health survey, for the calendar years 2008, 2010, and 2011. We identified 2892 cancer survivors with cancer remission and 48141 individuals who had no history of cancer. The Physical Component Summary (PCS), the Mental Component Summary (MCS), and the SF-6D scores were used to assess HRQOL. We examined the effect of time since cancer diagnosis on these HRQOL measurements, controlling for a range of measurable characteristics in MEPS data and accounting for the MEPS’s complex survey design.
Result: For cancer survivors with remission status, the average scores of PCS, MCS, and SF-6D were 42.77 (SE = 0.322), 49.49 (SE = 0.280), and 0.74 (SE = 0.003), respectively. Compared with people without cancer, cancer survivors within 2 years of cancer diagnosis had lower adjusted scores of PCS (-4.2; p<.001), MCS (-2.0; p=.049), and SF-6D (-0.03; p<.001). Cancer survivors 2 to 5 years after cancer diagnosis had a lower PCS score (-1.3, p=.036); no significant difference in PCS score was found between the no-cancer group and cancer survivors who were surveyed 5 or more years after cancer diagnosis. Results also showed no significant difference in the MCS and SF-6D scores between the no-cancer group and cancer survivors with cancer diagnosed after 2 years. We found that adjusted mean PCS and SF-6D scores were similar among survivors of melanoma and breast, prostate, colorectal, hematological, and cervical cancers, yet were significantly lower for survivors of short-survival and other cancers. The patterns were similar for adjusted mean MCS score, except that survivors of prostate and cervical cancers had better MCS scores.
Conclusion: For cancer survivors with remission status, the effect of time since cancer diagnosis on these HRQOL measurements depends on the measurement instrument and the types of cancer. The SF-6D utility score was decreased in the first 2 years, and then returned to a “normal” level. Our findings provide useful information of calculating the weight of quality adjusted life year in cost-effectiveness analyses.