Study Design and Data Sources
To answer these questions, we combined nationally representative data sources, population-based cohort studies, Medicaid claims data, and published data on the effect of DPP-like LCI and ran simulations using the CDC-RTI diabetes model . Owing to great heterogeneity in demographic, epidemiological, and economic characteristics between state Medicaid programs, we present state-specific analyses for eight states (Alabama, California, Connecticut, Florida, Iowa, Illinois, New York, and Oklahoma) that capture the country’s regional and demographic heterogeneity and represent approximately 50% of the country’s adult Medicaid population. We present population-size-weighted average and/or cumulative estimates for the combined data of the eight states as main results and report additionally state-specific estimates. Details on the selection criteria for the states are presented in Online Appendix A-M1.
The study was conducted in compliance with ethical standards and in all studies from which data were used participants gave informed consent.
Characteristics and Size of the Eligible Population
We used clinical eligibility criteria close to those defined by the Medicare DPP, i.e. a BMI ≥ 25 kg/m2 and a laboratory result of either Hba1c ≥ 5.7% or a fasting plasma glucose (FPG) level ≥ 110 mg/dL . As there is no compelling evidence on the program’s feasibility and effectiveness in the disabled population, and as most dually eligible beneficiaries will be eligible for DPP-like LCI through the Medicare DPP [26, 27], we restricted our analyses to non-disability-based Medicaid beneficiaries aged 19–64 years with full benefits.
Population Size and Characteristics
We sampled participants without diabetes and insured under Medicaid or with a family income below 138% FPL from the nationally representative National Health and Nutrition Examinations Surveys (NHANES, waves 2006–2016) who matched the age, sex, and race/ethnicity characteristics of Medicaid beneficiaries without diabetes in Medicaid claims files (2008–2012) for the eight selected states. The prevalence of people at high risk of type 2 diabetes and their demographic and clinical characteristics were then taken from this merged NHANES–Medicaid claims data set. We then combined data on the total number of non-disability-based adult beneficiaries with full benefit with the estimated prevalence of people with high risk of type 2 diabetes to calculate the number of non-disability-based adult beneficiaries with full benefit that are at high risk of type 2 diabetes [28, 29]. Details of these steps are described in Online Appendix A-M2 and A-M4.
Design and Input Parameters of the Simulation Scenarios
Intervention and Comparators
We compared in-person DPP-like LCIs delivered by trained and certified clinic staff, community health workers, peers in the workplace and church and community settings, as well as virtual programs, as delivered in several studies in the Medicaid population, with a counterfactual of routine care advice for people who are identified as having increased risk for type 2 diabetes in their usual care setting [17, 30]. DPP-like LCI programs focus on healthy eating, physical activity, and coping skills and generally consist of 16 weekly core sessions over 4 months plus 8 monthly follow-up sessions. Programs have been adapted for various ethnic and racial groups [9, 31,32,33,34], and evidence from various studies has shown that the delivery of LCI versions tailored to the needs of the Medicaid population is feasible and results in clinically relevant weight loss [17, 18, 35]. Recent demonstration projects further indicated that the tools and infrastructure built by the Centers for Disease Control and Prevention (CDC) and its partners [16, 36,37,38] might be successfully used to facilitate implementation of DPP-like LCI in state Medicaid programs [16, 30].
Cost and health effects of the LCI were projected using the decision analytic CDC-RTI diabetes computer simulation model. The CDC-RTI diabetes cost-effectiveness model is a Markov model that uses annual transition probabilities to simulate cohorts through different health states including ‘pre-diabetes’ (i.e. people at high risk for type 2 diabetes), type 2 diabetes, and death. Each health state is associated with a distinct set of costs for treatment and quality of life (QoL) decrements and the model accumulates incremental costs and health benefits, measured in quality-adjusted life years (QALYs) in each intervention arm .
The disease pathways and complications that are modelled in the diabetes module include nephrology, neuropathy, retinopathy, coronary heart disease, and stroke. The respective key transition probabilities are mainly based on data from the United Kingdom Prospective Diabetes Study (UKPDS)  and the risk equations of the American College of Cardiology/American Heart Association (ACC/AHA) .
The ‘prediabetes’ module follows individuals from the time of diagnosis of ‘prediabetes’ to diagnosis of type 2 diabetes or death, whichever comes first. People with ‘prediabetes’ may already have some complications at diagnosis of ‘prediabetes’ and may also experience coronary heart disease, stroke, early stages of nephropathy and neuropathy, or death while in the ‘prediabetes’ phase. Most of the model’s disease progression parameters are based on data of  the DPP study, the UKPDS and the ACC/AHA risk equations [39,40,41].
In both disease modules, intervention effects can be modelled through changes in the annual probability of transitioning from ‘pre-diabetes’ to type 2 diabetes, as well as changes in BMI, systolic and diastolic blood pressure, and total cholesterol and high-density lipoproteins.
The model has been validated against the results of large longitudinal studies/trials  and has been used successfully for economic evaluations of various prevention and treatment strategies in clinical and non-clinical settings [21, 23, 42]. Details of the model and simulation structure are provided in Online Appendix A-M3.
Details on the data sources and methods for estimating Medicaid-specific input parameters are described in Online Appendix A-M4–A-M10. The most important model parameter is the effect of the LCI on type 2 diabetes incidence and modifiable risk factors. To obtain valid and reliable estimates on these effectiveness parameters we used systematic reviews that tested the efficacy of LCIs versus routine care in RCTs [10, 43, 44], reviews on randomized and non-randomized studies that tested interventions modelled after the DPP in more real-world settings [9, 45], observational data from the NDPP registry  as well as observational data from studies that implemented DPP-like interventions in the Medicaid population . Following this combined evidence, we assumed that the LCI induces a type 2 diabetes risk reduction of 24% in years 1 and 2, of 12% in years 3–10 and of 6% in years 11–25. Conservatively, we also assumed that the intervention induces a weight loss of 2 kg in the years 1–2 and no effect on other risk factors. We assumed that these effectiveness parameter did not differ between LCI delivery modes (for details on these assumptions see Online Appendix A-M6).
Other crucial input parameters comprise characteristics of the Medicaid population at high risk for diabetes (directly estimated from Medicaid claims and NHANES data, for details see Online Appendix A-M4), their annual background probability for developing type 2 diabetes [estimated from the National Health Interview Surveys (NHIS), the Atherosclerosis Risk in Communities (ARIC) Study, and the Coronary Artery Risk Development in Young Adult (CARDIA) Study, for details see Online Appendix A-M5], the cost for recruitment, referral and delivery of the DPP-like LCI (based on previous studies and current practice, for details see Online Appendix A-M7 and A-M8), as well as the costs (directly estimated from Medicaid Analytic eXtract files of the eight states, for details see Online Appendix A-M9) and QoL decrements [estimated from the Medical Expenditure Panel Survey (MEPS), for details see Online Appendix A-M10] associated with diabetes and its complications.
An overview of the resulting parameters is described in Table 1. For example, the annual probability of developing type 2 diabetes of a Medicaid enrollee eligible for LCI are between 4% and 8%, the combined costs of recruitment, referral and delivering of the DPP-like LCI are around US$800, annual excess costs of treating diabetes versus remaining in the pre-diabetes state are around US$1400, the QoL decrement for diabetes is −0.04 and the QoL decrements for complications lies between –0.03 (myocardial infarction) and –0.08 (stroke).
State-Specific Parameters and Assumptions
For the clinical and demographic characteristics of the population at high risk of type 2 diabetes, the annual background incidence of type 2 diabetes, and the costs of treating diabetes and its complications we could derive state-specific input parameters and used them in the state-specific model scenarios. For the effectiveness and the costs of the DPP-like LCI and the impact of diabetes and diabetes-related complications on health-related QoL we had no state-specific data and assumed that they are the same in each of the 8 states (for details see Table 1).
A healthcare system perspective was chosen because the societal perspective includes indirect costs that are not directly relevant to the Medicaid program or other payers in the healthcare system . We simulated individuals at high risk for type 2 diabetes over 5, 10, and 25 years from the start of a DPP-like LCI. Twenty-five years was chosen as maximum time horizon as this approximately coincides with the longest follow-up of current LCI studies and as every effect beyond this time horizon was considered to be quite hypothetical. Both costs, consisting of costs for referral, intervention, and treatment of diabetes and complications, and health effects, described in QALYs, a measure that combines length and QoL, were discounted at 3% annually. Costs are indexed to the year 2018. Incremental costs and QALYs were used to calculate incremental cost-effectiveness ratios (ICERs). To capture structural and stochastic uncertainties, we conducted univariate and probabilistic sensitivity analyses. In the univariate sensitivity analyses we varied crucial model parameters by ± 50%. In the probabilistic sensitivity analyses we permuted parameters simultaneously (for details, see Online Appendix A-M11). We also estimated the maximal intervention cost at which the ICERs are below US$50,000/QALY and US$100,000/QALY in the base case analysis . Analysis and reporting are based on the recommendations of the Consolidated Health Economic Evaluation Reporting Standards .
Return on Investment (ROI) from a Health Care System and Medicaid Perspective
Monetary return on investment (ROI) from a health care system perspective equals the cost outcome from the cost-effectiveness analyses. Given the specific Medicaid policy context, we conducted additional analysis in which we considered factors relevant to the ROI for state Medicaid programs. First, non-disability-based Medicaid enrollees are generally not eligible for Medicaid beyond the age of 64 years. We therefore assumed that savings that occur from preventing type 2 diabetes and its complications beyond age 64 years won’t be captured by the Medicaid system . Second, Medicaid enrollees typically move in and out of Medicaid eligibility, a phenomenon often referred to as ‘churning’. Data show that average non-disability-based Medicaid beneficiaries are enrolled 8.6 months or 72% of the fiscal year in Medicaid . In our adjusted ROI model scenario, we therefore pragmatically assumed that until Medicaid beneficiaries turn 65 only 72% of savings that occur from preventing type 2 diabetes and its complications will be captured by Medicaid (for details, see Online Appendix A-M12).
Population Health, Health Equity and Cost Impact
To estimate the expected upfront investments and the long-term cost and health impact on a population level, in a next step, we combined data on the number of expected participants with the per-participant ROI estimates. Furthermore, using the CDC-RTI model and the background type 2 incidence of race/ethnicity and income strata in the Medicaid and non-Medicaid populations, we calculated the cumulative type 2 incidence in the general US adult population with and without implementing LCI for eligible Medicaid beneficiaries at high risk of type 2 diabetes. We then calculated the absolute and relative narrowing in the difference of the cumulative diabetes incidence between white and non-Hispanic black, and Hispanics, and between people below and above 138% FPL in the general US adult population. For all those analyses, we assumed that 20% of eligible beneficiaries participated in DPP-like LCI (for details, see Online Appendix A-M13).
Analyses and simulations were run in 2018.