Page 38 - GHES-3-2

Global Health Economics and
Sustainability
Random price generators in health policy

terms in some runs), capturing the effects of electronic
billing and insurance plans. They controlled whether the
computerization process affected enrollees according
to their plans: Medicare and Medicaid enrollees versus
commercial insurance (study run with the Glimmix SAS
procedure) (Huttin, 2017).
However, approaches with random parameters also
tend to outperform latent class models in analyzing
heterogeneity in medical markets; Zhou & Bridges (2019)
also studied treatment heterogeneity regarding diabetes.
Therefore, using random price generators can be viewed
Figure 1. Approaches used for the development of the economic model as a new milestone on random parameters in a choice
on medical markets
Source: Extracted from Prof Huttin CC ‘s communication at Euro modeling approach for physicians.
conference, July 3–6th, ESPOO, Finland, 2022 (with permission of Prof The ML model uses a combination of two random
Huttin CC-ENDEPUSresearch, Inc).
Abbreviation: IAHPR: International academy of health preference parameters for prices and three non-random parameters
research. for demographic variables: age, gender, and obesity. The
statistical estimates for the three non-random parameters
As research is increasing in all these fields, the were estimated predictors from the previous studies using
methodological discussions around using random cumulative logistics. These variables were used in the
generators and simulated studies are growing and will baseline model for simulations of the ML model.
benefit the development of such an approach for analyzing The randomization of prices was performed using
medical markets. Some methodological issues and an Excel formula around some prices collected with
limitations discussed in the next section were identified Redbook for the main drug classes of the three categories;
in the original study by Huttin & Hausman (2021) of the this experiment helps to identify workable random
physician choice model on type 2 diabetes. price parameters to import into the database; several
experiments on the generation of such random variables
3. Original physician choice models using have been tested as well as two sequences Halton and
random drug price parameters: Promising uniform; issues of correlation between the number of
results for medical markets? draws and number of sequences should be addressed for
larger scaled studies (Mariel & Meyerhoff, 2018; Ellis et al.,
The economic model development for medical markets 2019), however for the original study, it did not create
and the statistical studies on analytical datasets extracted issues and the model could converge with draws limited
from the US National Ambulatory Medical Care Survey to 200. Price parameters with interaction terms were
used different approaches. The data were transformed to introduced to represent enrollees on public versus private
run ML models using random price parameters. insurance plans in the US. This approach may be extended
The main statistical model used in three chronic to include Medicaid patients and additional classifications
conditions (asthma, hypertension, and Type 2 diabetes) of Medicare patients with supplemental plans versus or in
were cumulative logit models using socioeconomic combination with commercial plans; however, in further
predictors and proxy for insurance profiles based on the studies, additional classifications or more complex cost-
classification from the administrative data set on insurance sharing arrangements required more extensive datasets
types and payment types. The studies were useful in and samples to comprehensively represent public and
identifying the main socioeconomic predictors and the private plans.
impact of types of insurance profiles; however, such models The ML model was run with a STATA procedure
also require additional partial odds ratio testing. called “asmixlogit” and then imported into Matlab for
Figure 2 presents the study results and illustrates the comparison between simulated data and the original
issue of odds ratio testing. data from the analytical dataset on diabetes. The different
steps of the study are described in two papers by Huttin
In such models, several types of random variables have & Hausman (first a technical note at MIT in 2018 and a
been used. For instance, in the study on Information and published paper in 2021) and Hahn, Hausman & Lustig
Communication Technologies (ICT) diffusion in primary (2020), the study was presented at a joint ISPOR/IAHPR
care, some random variables were used (as interaction conference in 2019 in Basel.

Volume 3 Issue 2 (2025) 30 https://doi.org/10.36922/ghes.3579