SCREENING FOR A CHRONIC DISEASE: A MULTIPLE STAGE DURATION MODEL WITH PARTIAL OBSERVABILITY

• Author: Frank Sloan,Arseniy P. Yashkin,Gabriel Picone,Thomas A. Mroz
• Date: 01 August 2016
• DOI: http://doi.org/10.1111/iere.12180
• Published date: 01 August 2016

INTERNATIONAL ECONOMIC REVIEW

Vol. 57, No. 3, August 2016

SCREENING FOR A CHRONIC DISEASE: A MULTIPLE STAGE DURATION

MODEL WITH PARTIAL OBSERVABILITY∗

BYTHOMAS A. MROZ,GABRIEL PICONE,FRANK SLOAN,AND ARSENIY P. YASHKIN1

Georgia State University and the Federal Reserve Bank of Atlanta,U.S.A.; University of South

Florida, U.S.A.; Duke University, U.S.A.; Duke University, U.S.A.

We estimate a dynamic multistage duration model to investigate how early detection of diabetes can delay the

onset of lower extremity complications and death. We allow for partial observability of the disease stage, unmeasured

heterogeneity, and endogenous timing of diabetes screening. Timely diagnosis appears important. We evaluate the

effectiveness of two potential policies to reduce the monetary costs of frequent screening in terms of lost longevity.

Compared to the status quo, the more restrictive policy yields an implicit value for an additional year of life of about

$50,000, whereas the less restrictive policy implies a value of about $120,000.

1. INTRODUCTION

According to the U.S. Centers for Disease Control (CDC), 75% of health-care expenditures

and 70% of all deaths in the United States are attributable to chronic diseases, including

heart conditions, cancer, stroke, and diabetes (CDC, 2009). Earlier detection of these chronic

diseases can yield substantial savings and better health outcomes. To achieve these goals,

the Affordable Care Act (ACA) of 2010 subsidizes not only primary preventive measures,

such as improvements in diet, but also secondary preventive measures. For example, since

2014 all insurance plans must cover many screening tests without any copayment. However,

the empirical evidence that increased screening will save resources or even improve health

outcomes is mixed (Cutler, 2008). Knowing earlier that an individual has a chronic disease does

not necessarily imply that screening can delay disease progression or increase longevity.

The “gold standard” for evaluating the benefits versus costs of alternative screening policies

is the randomized controlled trial (RCT). RCTs provide a simple approach to solve the problem

of unobserved heterogeneity, but their usefulness for policy evaluations can be limited in many

important situations. When the outcomes monitored are relatively rare, they can be quite

expensive. They are difficult to conduct for long follow-up periods. In the context of dynamic

decision making, the treatment protocols specified in RCTs may not yield results generalizable

to community settings. This is especially the case when there are many different outcomes

occurring over long time horizons and numerous intermediate outcomes that might require

additional interventions. These are all key issues when one studies diabetes. Moreover, the

sample size of an RCT would have to be very large and the follow-up period very lengthy for it

to have sufficient statistical power and time to measure many relevant relationships.

∗Manuscript received March 2014; revised August 2014.

1We thank seminar participants at the University of Florida, University of Alabama Birmingham, Elon University,

Twentieth European Workshop on Econometrics and Health Economics, the 2011 Annual Health Econometrics

Workshop, two anonymous referees, and the editor of this journal for valuable comments. We also thank the National

Institute of Aging (Grants R01-AG017473, and R01-HD047213) for their generous support and the use of the services

provided by Research Computing at the University of South Florida. The views expressed here are the authors’ and

not necessarily those of the Federal Reserve Bank of Atlanta or the Federal Reserve System.

Please address correspondence to: Gabriel Picone, Department of Economics, University of S. Florida, 4202 E.

Fowler Ave., CMC 207, Tampa, FL 33620. E-mail: gpicone@usf.edu.

915

C

(2016) by the Economics Department of the University of Pennsylvania and the Osaka University Institute of Social

and Economic Research Association

916 MROZ ET AL.

A potentially important alternative is to use observational data like we do in this study.

Large, longitudinal, administrative data sets are becoming increasingly available to researchers,

but their effective use requires one to address directly how unobserved heterogeneity impacts

both observed treatment choices and outcomes over time. Econometric solutions for dealing

with such issues in analyzing observational data have been developed, including for dynamic

problems like those examined here, but there is still much progress to be made. In this study, we

incorporate many of these advances and add in a key component that is relevant for studying

diabetes and many other diseases.

Diabetes mellitus is a complex chronic disease. It can affect eyesight, kidneys, cardiovascular

systems, and nervous systems affecting the lower extremities, i.e., legs and feet. The incidence

of diabetes is increasing in the United States and elsewhere, reflecting increased obesity of the

population in part. Prevalence is high among the elderly (Sloan et al., 2008).

The disease progresses through several stages, each with increasingly debilitating conse-

quences. Once an individual reaches a more advanced diabetes stage it is impossible to undo

the physiological damage; diabetes complications can eventually lead to death. The early stages

of the disease are typically nonsymptomatic, but without interventions, irreversible physiolog-

ical damage will continue to accrue. Like many other chronic illnesses, diabetes can be much

more costly to treat if detected later. Regular screening for diabetes potentially can help the pa-

tient and her physician recognize when it is appropriate to undertake behavioral modifications

and start medications and other therapies to slow the disease’s progression.

Screening is costly, and the optimal screening regimen depends on a comparison of the

marginal benefit and the marginal cost of screening. The marginal benefit from more frequent

screening depends on the probability that screening will reveal useful information and the

value of this information in slowing the disease’s progression. This is the primary focus of this

article. This study uses a dynamic multistage discrete duration model to investigate the effec-

tiveness of early detection of diabetes mellitus through screening in delaying the progression of

complications and death.

An evaluation of screening for diabetes encounters at least four econometric issues. First,

ascertainment of diagnosis in particular and care more generally is endogenous. Second is

the importance of partial observability of the disease state; the person could have the disease

for a long time without being diagnosed. Third, since individuals and diseases differ in aspects

unobservable to the researcher, there is likely unobserved heterogeneity. Fourth, the probability

of adverse outcomes increases with duration and progression of the disease.

Our estimation strategy deals with each of these four econometric problems. We address

endogeneity and unmeasured heterogeneity issues by using discrete factor models (Heckman

and Singer, 1984; Mroz, 1999). This approach has been used previously by Picone et al. (2003),

Glewwe and Jacoby (2004), Bhattacharya (2005), Mroz and Savage (2006), and Liu et al. (2010),

among others. We simultaneously account for multiple disease stages, partial observability of

disease progression, endogeneity of the timing of diagnoses, and health outcomes. We control

for partial observability by modeling empirically all potential exact times of disease (or stage)

onset and integrating over all these potential onset times. The bounds for these integrations

come from the last time period a person was known to not have the disease and the first time

the individual was known to have the disease.

These time periods during which we are uncertain about precisely when the individual pro-

gressed to the next disease stage constitute a key feature of this analysis. Not only is this an

econometric issue to be addressed, it is a real, substantive issue for analyzing disease progres-

sion and treatments. Many individuals will not recognize that they have progressed to diabetes

or to more advanced stages if they do not see a health-care professional who can diagnose

their condition. If the period of time during which the disease is present but unobserved and

untreated is long, then the individual may progress much more rapidly to more severe disease

stages, possibly resulting in amputation or death.

We find that earlier diagnosis of diabetes, and presumably the treatments that follow di-

agnosis, delays the onset of lower extremity complications (LECs) including amputation. For