Journal Logo

Research Article: Observational Study

A study on the impact of poor medication adherence on health status and medical expense for diabetes mellitus patients in Taiwan

A longitudinal panel data analysis

Lin, Chin-Shien PhDa; Khan, Haider PhDb; Chang, Ruei-Yuan PhDc,d; Liao, Wei-Chih MDa,e; Chen, Yi-Hsin MD, PhDf,g,h,∗; Siao, Sih-Yin MBAa; Hsieh, Teng-Fu MD, PhDh,i

Editor(s): Wane., Daryle

Author Information
doi: 10.1097/MD.0000000000020800
  • Open


1 Introduction

Diabetes mellitus is an incurable disease that is increasing in both incidence and prevalence.[1] In 2017, the International Diabetes Federation Diabetes Atlas showed that an estimated 451 million people were suffering from the disease worldwide.[2] Diabetes mellitus can result in various macro-vascular and micro-vascular complications that substantially impact both patient health and the medical system in general. However, adequate blood sugar control is known to reduce the risk of cardiovascular and microvascular complications for diabetes patients.[3] In this regard, adequate treatment is not only important for individual health status among diabetic mellitus patients but also has substantial implications for resource utilization throughout the medical system.

Reports indicated that patient noncompliance for diabetes treatment is around 50% in developed countries and maybe even higher in developing countries.[4] Among the key treatments for diabetes mellitus, medical compliance is a main area of concern; indeed, it has been discussed in the literature from many different perspectives.[5]

This study explored the impacts of demographic characteristics on patient medication nonadherence, medication nonadherence on health status, and medication nonadherence on different medical expenses. In this regard, previous studies have produced unconvincing evidence due to several issues, such as the difficulty of data collection, insufficient sample size, and poor sample representativeness. As such, this study examined daily medical visit data from Taiwan's National Health Insurance Research Database (NHIRD) to explore the impact of the abovementioned relationships.

2 Methods

2.1 Ethical statement

This study was approved by the Institutional Review Board of the Buddhist Taichung Tzu Chi General Hospital, Taiwan (REC103-43). However, written consent was not obtained from each patient because all study data consisted of secondary files. In this regard, all identification numbers and personal information were deleted before the data were released from the NHIRD.

2.2 Dataset

The Taiwan National Health Insurance (NHI) program is a mandatory government-run enrollment service that characterizes the Taiwanese universal single-payer health insurance system. It was initiated in 1995 and contracts with over 97% of the medical facilities in Taiwan to provide health care services. Nearly 99% of the 23 million residents of Taiwan receive medical service in contracted facilities through the NHI program.[6] Further, patients who receive NHI medical services can obtain medications and pharmacy services at any NHI-contracted pharmacies. All data related to these medical and pharmacy services are then collected and recorded in the NHIRD by the National Health Research Institutes to provide a general record of medical care.

The NHIRD contains inpatient, outpatient, prescription information, and disease diagnosis files that are coded according to the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM). Naturally, this has been a very important data source for research in medicine and health-related fields. This study extracted all daily medical visit data for diabetic mellitus patients between January 1, 2000 and December 31, 2004.

2.3 Design and participants

Inclusion criteria for patients were as follows: 18 years of age or older, diagnosed with diabetes mellitus (ICD-9-CM Codes 250.xx), complications with diabetes mellitus (ICD-9-CM Codes 271.4, 357.2, 362.01–362.02, 366.41, 648.00–648.05, 750.0–750.1, 791.5, V18.0, V77.1), and receiving insulin, biguanides, sulfonamides, urea derivatives, a-glucosidase inhibitors, thiazolidinediones, dipeptidyl peptidase 4, or a combination of oral antihyperglycemic agents. Patients who had diabetes but who were not taking anti-hyperglycemic agents were excluded from the study. For patients who made multiple physician visits during the study period, we combined their information in accordance with the date of the relevant fee, application type, hospital identification code, and case sequence number. This was done to combine all records for each calendar year into 1 single annual record for each patient.

The final derived data files were obtained by merging different files, such as those for inpatient expenditures based on admissions files, ambulatory care expenditures based on visit files, details of inpatient order files, details of ambulatory care order files, registry for beneficiary files, registry for contracted medical facilities files, and registry for drug prescriptions files. The overall sample thereby spanned a 5-year period and included patients from a variety of residential areas. After deleting missing data and summarizing all records for each calendar year into 1 annual record for each patient, this resulted in 14,602 total patients with diabetes mellitus who were taking anti-hyperglycemic medications.

We also used the database to collect and/or calculate demographic and other information, including gender, age, socio-economic status, living area, level of visited hospitals, and total annual medical visit expenses for each patient. Gender was noted as a binary (ie, 1 for male, 0 otherwise), while age was calculated for each patient based on the number of months between their respective birthday and the day of the medical visit. Socio-economic status was determined based on the amount of insurance expenses paid yearly; if expenses were greater than or equal to NT$ 120,000, then the patient was categorized as having high socio-economic status. However, such expenses amounting to less than NT$ 120,000 were used to indicate low socio-economic status. Living area was determined based on the 4 total residential areas of Taiwan: the north, center, south, and east. Next, the levels of the visited hospitals were determined based on 4 classes: Medical Centers, Regional Hospitals, District Hospitals, and General Practice Clinics. The numerical values of all variables used in this study's research model were calculated based on raw data to avoid possible biases incurred during questionnaire collection (eg, answers from memory and those disguised through social norms).

2.4 Medication adherence and follow-up

The World Health Organization defines medication adherence as the “the degree to which the person's behavior corresponds with the agreed recommendations from a health care provider.”[7] This most often refers to medication or drug compliance but can also apply to medical device usage, self-care, self-directed exercises, and/or therapy sessions. The literature is characterized by 2 approaches for measuring medication adherence, including the medication possession ratio (MPR) and proportions of days covered (PDC). MPR measures the percentage of time a given patient has access to medication; it is the sum of the days’ supply for all fillings of a given drug during a particular time period divided by the number of days in that time period. PDC is the proportion of days in the measurement period “covered” by prescription claims for the same medication or another in its therapeutic category. The MPR is a popular measurement for adherence within the context of the health care industry.[8] However, some researchers favor the PDC because it eliminates the problem of overlapping prescriptions during the specifically investigated period.

This study examined medication adherence with a focus on medication compliance, which is represented as the MPR. First, MPR is more broadly used than PDC in studies on medication adherence. Second, chronically ill patients may avoid medical facilities if their illnesses are not severe and/or if they have residual drugs; this is especially true among the elderly. As such, we preferred MPR over PDC as a proxy for medication adherence in the stipulated situations. In this study, MPR was calculated using the following formula: MPR = total number of days’ supply obtained between the first and last fills (excluding the last fill) divided by the number of days between the first and last fills.[9] High medication compliance was defined as patients with MPRs greater than or equal to 80%, while patients with MPRs less than 80% were defined as having low medication compliance. In order to more accurately describe compliance willingness, we did not count days when patients were staying in hospitals.

2.5 Main outcome measurements and covariate assessment

2.5.1 Health status and the Charlson comorbidity index (CCI)

Diabetes mellitus is a lifelong disease that may result in complications leading to substantial damage to the patient's quality of life.[10] A given diabetic mellitus patient usually experiences 3 to 4 complications simultaneously.[11] As such, a common way to measure diabetic mellitus severity is through comorbidities listed in the literature. Comorbidity refers to the presence of 1 or more additional diseases/disorders that co-occur with a primary disease/disorder.

Charlson, Pompei, Ales, and MacKenzie[12] developed the CCI to measure the severity of health status for patients with multiple diseases; here, higher indexes indicate lower survival probability.[13] Different versions of the CCI have been developed to adapt clinical comorbidity indices to specific diseases.[14–16] One of this study's aims was to elaborate the effects of medicine compliance on health among diabetes patients. In this regard, it is imperative to find a dependent variable that can be measured to reflect the health status of diabetes mellitus patients. However, there is no such commonly used measurement in the current literature. Because diabetes mellitus is a systemic disease that may cause macro- and micro-vascular disorders in multiple organs (eg, nephropathy, retinopathy, and neuropathy, which are major diabetes complications), it is also associated with other diseases or syndromes, including stroke, heart disease, and electrolyte disorders resulting from hyperosmolar hyperglycemic status. Thus, diabetes mellitus complications and comorbidities should be included when measuring the health status of a given diabetes mellitus patient. As the CCI produces a weighted sum of some especially important complications and comorbidities, it has been used to predict a variety of outcomes, including the mortality of type 2 diabetes mellitus nephropathy[17] and subsequent hospitalizations among type 2 diabetes mellitus patients.[18] Further, a study in northern Denmark revealed a positive predictive value of 82.0% (95% confidence interval; 68.6%, 91.4%) for CCI when used among diabetes patients with diabetic complications.[19] Taken together, the evidence shows that the CCI is an appropriate proxy for measuring health status among diabetes mellitus patients in regard to the complications of comorbidities.

This study used Deyo version of the CCI due to the availability of relevant variables in the database.[14] The examined comorbidities included myocardial infarction (ICM-9-CM Code 410–410.9, 412), congestive heart failure (ICM-9-CM Code 428–428.9), peripheral vascular disease (ICM-9-CM Code 443.9, 441, 441.9, 785.4, V43.4), cerebrovascular disease (ICM-9-CM Code 430–438), dementia (ICM-9-CM Code 290–290.9), chronic pulmonary disease (ICM-9-CM Code 490–496, 500–505, 506.4), rheumatologic disease (ICM-9-CM Code 710.0, 710.1, 710.4, 714.0–714.2, 714.81, 725), peptic ulcer disease (ICM-9-CM Code 531–534.9), mild liver disease (ICM-9-CM Code 571.2, 571.5, 571.6, 571.4–571.49), diabetes (ICM-9-CM Code 250–250.3, 250.7), diabetes with chronic complications (ICM-9-CM Code 250.4–250.6), hemiplegia or paraplegia (ICM-9-CM Code 344.1, 342–342.9), renal disease (ICM-9-CM Code 582–582.9, 583–583.7, 585, 586, 588–588.9), any malignancy (including leukemia and lymphoma) (ICM-9-CM Code 140–172.9, 174–195.8, 200–208.9), moderate or severe liver disease (ICM-9-CM Code 572.2–572.8, 456.0–456.21), metastatic solid tumor (ICM-9-CM Code 196–199.1), and acquired immunodeficiency disease (AIDS) (ICM-9-CM Code 042–044.9).[14]

2.5.2 Medical expenses

Medical expenses include those incurred from clinics, emergency rooms, ordinary wards, and intensive care units. These are correlated with patient medication adherence. For example, patients exhibiting proper medication adherence paid higher fees at clinics but less for hospitalizations. On the other hand, patients exhibiting poor medication adherence spent more money for services in emergency rooms, ordinary wards, and intensive care units due to comorbidities and developed complications. This study collected medical expense data per person per year, including clinical expenditures, emergency treatment expenses, hospitalization costs, and expenses incurred at intensive care units. This was done to explore how patients were affected by medication nonadherence. Moreover, in this study, we considered the severity of illness for diabetic patients and calculated their expenditures for diabetes in emergency rooms, ordinary wards, and intensive care units. We then analyzed the relationships among these costs and the degree of morbidity among diabetic patients.

2.5.3 Statistical analyses

The SAS 9.4 software was used for all data analyses. A total of 3 models were constructed to investigate the relationships among variables. In this context, the generalized estimating equations (GEE) method is commonly applied to evaluate the associations among repeated observations in panel data. The model is used to replace basic regression because repeated measurements are correlated, thereby violating the assumptions of independence found in traditional regression models.[20] Based on the characteristics of the dataset, GEE models were thus constructed to explore the relationships among the variables examined in this study.

3 Results

In addition to descriptions of patient characteristics, the impacts of these demographic variables on nonadherence, and the impacts of nonadherence on both health status and medical expenses are outlined in the following passages.

3.1 Patient characteristics

A total of 14,602 diabetes mellitus patients were derived from the dataset. Patient records were then combined to result in annual records for each patient from 2000 to 2004 (5 years). The basic statistics are listed in Table 1. As the table shows, 48.79% of patients were male and 51.21% were female, while the average age was 61.74 years with a standard deviation equal to 11.13 years. For residential area, 43.14%, 17.76%, 36.02%, and 3.07% lived in the north, center, south, and east, respectively. Next, 19.2% were of high socio-economic status, while 80.8% were of low socio-economic status. Finally, 36.1%, 20.35%, 22.67%, and 20.87% visited medical centers, regional hospitals, district hospitals, and general practice clinics, respectively.

Table 1
Table 1:
Baseline characteristics.

Medication adherence was defined as high (y = 1) when MPR was greater than or equal to 80% and as low (y = 0) when MPR was less than 80%; this method followed the threshold values established by McGovern, Tippu, Hinton, Munro, Whyte, and de Lusignan.[21] The average MPR was 0.637 with a standard deviation equal to 0.274. The proportions for high and low medication adherence were 75.27% and 24.73%, respectively.

3.2 Empirical results for impacts of patients characteristics on medication adherence

Each patient had 1 record for each year (5 years total). A GEE model was then constructed with medication adherence as the dependent variable, while demographic and medical-behavior variables were set as independent variables. Further, medical central and north Taiwan served as references for the medical institution level and residential area variables. The empirical results are listed in Table 2.

Table 2
Table 2:
The GEE results for medication adherence on demographic variables.

As Table 2 shows, medication adherence was lower among male patients and was positively impacted by age. Patients who visited different levels of hospitals revealed different levels of medication adherence. Compared with patients living in north Taiwan, those living in south and east Taiwan exhibited lower medication adherence. And there is no significant difference between those living in north and central Taiwan.

3.3 Empirical results for impacts of medication nonadherence on health status

The annual means and standard deviations of CCI for all patients are listed in Table 3. Here, an increasing trend was evident. Further, results of the ANOVA on CCI among these years were statistically significant; a post-hoc comparison also showed that differences between any consecutive years were statistically significant. The average CCI for all patients in the sample is 3.07 with a standard deviation equal to 2.29.

Table 3
Table 3:
The average CCI for all the diabetic mellitus patients from 2000 to 2004.

A GEE model was then constructed with CCI as the dependent variable, while demographic and medical-behavior variables were set as independent variables (Table 4). Except for socio-economic status, all independent variables statistically and significantly impacted CCI. Notably, MPR had a negative impact on CCI.

Table 4
Table 4:
GEE results for CCI on demographic variables and medical behavior.

3.4 Empirical results for impacts of medication nonadherence on medical expenses

Data on clinical expenditures, hospitalization costs, emergency treatment expenses, and expenses incurred in intensive care units were collected for each patient. Table 5 shows that each item increased over the 5-year period of study. Notably, hospitalization costs, emergency treatment expenses, and expenses incurred in intensive care units doubled or more during that time. In other words, medical expenses increased annually in every category of medical expenses for each patient.

Table 5
Table 5:
Different medical expenses for each year from 2000–2004.

GEE models were also constructed for each expense item to determine how the demographic and medical-behavior variables affected expenses (Table 6). As shown, MPR effects differed based on the type of medical cost. That is, MPR had a positive impact on clinical expenses but a negative impact on admissions, emergency room, and intensive care unit expenses.

Table 6
Table 6:
The GEE results for different medical expenses on demographic variables and medical behavior.

4 Discussion

As shown in Table 1, 36% of diabetes patients visited medical centers for treatment. The proportions for the other 3 types of medical institutions were 20.35%, 22.67%, and 20.87% for regional hospitals, district hospitals, and local clinics respectively, thus indicating the availability and popularity of particular medical resources in Taiwan. As shown in Table 2, medication adherence was significantly higher among patients who visited regional hospitals, district hospitals, and clinics when compared to patients who visited medical centers. Further study is needed to determine the reasons for the low medication adherence found among patients who visit Taiwanese medical centers.

This study used the CCI index as a proxy for health status among diabetic mellitus patients; here, higher index values indicated worse health status. As shown in Table 3, the index increased at a rate of 0.8 each year for all patients. As such, it is a sufficient indicator of how effectively the medical system controls diabetic mellitus. When compared to the CCI index values found among patients who visited medical centers, those found among patients who visited regional hospitals, district hospitals, and general practice clinics were 0.237, 0.471, and 0.03, respectively. This can be further studied to investigate what causes these differences.

As shown in Table 4, patients with high medication adherence had lower CCI index values than those with low medication adherence. The empirical results of the GEE model shown in Table 6 also indicate that male patients generated more medical expenses in admissions and at intensive care units than female patients. Further, age was found to have a significantly positive influence on all medical expenses except those incurred at emergency rooms. Next, patients of high socio-economic status spent less on all medical expenses except those incurred at clinics. Finally, high medical compliance was associated with increased clinical expenditures but lower emergency room expenditures in both ordinary wards and intensive care units. In other words, patients exhibiting high medical compliance made more frequent medical visits than those exhibiting low medical compliance; it is thus reasonable for these patients to have greater clinical expenditures. However, patients exhibiting high medication adherence possibly had better overall health status, which may be associated with lower amounts of hospitalization, emergency treatment, and intensive care unit expenses. Numerically, the results showed that the total expenses for patients exhibiting high medication adherence were NT 4304 dollars less than those of patients exhibiting low medication adherence (1388.16–5089.7–275.47–326.857 = 4304).

5 Conclusions

This study explored the relationships among patient demographics, medical behaviors, CCI index values, and medical expenses through an existing database. Because of the characteristics of the examined dataset, the utilization of a longitudinal approach, the amount of examined data, appropriate objectivity, sufficient representativeness, the use of a panel data analysis tool, and the employment of GEE models, this study revealed the impact of demographics and medical behaviors on patient outcomes in a more rigorous way than previous studies. As such, important policy implications can thereby be derived.

This study's results clearly demonstrate that several important factors affect medicine nonadherence. First, medication adherence was higher among female patients and the elderly, thus implying that medication adherence should specifically be further promoted among male and younger patients. Second, patients of regional hospitals, district hospitals, and general practice clinics exhibited higher medication adherence than those of medical centers, thus suggesting that medical centers should take immediate action to facilitate patient medical compliance. Third, based on the CCI results, diabetes patients who exhibited medication adherence had better control than those who exhibited medication nonadherence. Finally, patients who exhibited high medication adherence spent less each year on total medical expenses. Based on these last 2 results, diabetic patients with proper medicine compliance can maintain CCI while saving money on clinical costs.

This study had 4 main limitations. First, the NHIRD did not include several potential confounding factors, including emotional support, lifestyle, family support and conflict, and living environment. Second, MPR is measured by applying prescription refill patterns, which could overestimate real drug consumption rates.[22] Third, this study did not record other medications that patients were taking at the same time, nor did it consider clinical illnesses that may have impeded medication adherence. Finally, this study was only an investigation of the relationships among the same variables in 1 country; its approach should be extended to investigate these relationships in other countries through similar respective databases.

Author contributions

Conceptualization: Chin-Shien Lin, Haider Khan

Data curation: Yi-Hsin Chen, Wei-Chih Liao,

Formal analysis: Yi-Hsin Chen, Sih-Yin Siao

Methodology: Yi-Hsin Chen, Wei-Chih Liao

Project administration: Chin-Shien Lin, Haider Khan

Resources: Yi-Hsin Chen, Wei-Chih Liao

Supervision: Chin-Shien Lin, Haider Khan

Validation: Yi-Hsin Chen, Sih-Yin Siao, Teng-Fu Hsieh

Visualization: Yi-Hsin Chen, Wei-Chih Liao

Writing – original draft: Sih-Yin Siao

Writing – review & editing: Wei-Chih Liao, Chin-Shien Lin


[1]. Sharma M, Nazareth I, Petersen I. Trends in incidence, prevalence and prescribing in type 2 diabetes mellitus between 2000 and 2013 in primary care: a retrospective cohort study. BMJ Open 2016;6:e010210.
[2]. Cho N, Shaw J, Karuranga S, et al. IDF Diabetes Atlas: Global estimates of diabetes prevalence for 2017 and projections for 2045. Diabetes Res Clin Pract 2018;138:271–81.
[3]. Frias J, Virdi N, Raja P, et al. Effectiveness of digital medicines to improve clinical outcomes in patients with uncontrolled hypertension and type 2 diabetes: prospective, open-label, cluster-randomized pilot clinical trial. J Med Internet Res 2017;19:e246.
[4]. Khan AR, Lateef ZNA-A, Al Aithan MA, et al. Factors contributing to non-compliance among diabetics attending primary health centers in the Al Hasa district of Saudi Arabia. J Family Community Med 2012;19:26–32.
[5]. Capoccia K, Odegard PS, Letassy N. Medication adherence with diabetes medication: a systematic review of the literature. Diabetes Educ 2016;42:34–71.
[6]. Chen Y-H, Hsieh T-F, Lee C-C, et al. Estrogen therapy and ischemic stroke in women with diabetes aged over 55 years: a nation-wide prospective population-based study in Taiwan. PLoS One 2015;10:e0144910.
[7]. Jimmy B, Jose J. Patient medication adherence: measures in daily practice. Oman Med J 2011;26:155–9.
[8]. Sperber CM, Samarasinghe SR, Lomax GP. An upper and lower bound of the medication possession ratio. Patient Prefer Adherence 2017;11:1469–78.
[9]. Kabore L, Muntner P, Chamot E, et al. Self-report measures in the assessment of antiretroviral medication adherence: comparison with medication possession ratio and HIV viral load. J Int Assoc Provid AIDS Care 2015;14:156–62.
[10]. Narayan KV, Gregg EW, Fagot-Campagna A, et al. Diabetes—a common, growing, serious, costly, and potentially preventable public health problem. Diabetes Res Clin Pract 2000;50:S77–84.
[11]. Waheed S, Jamal M, Amin F. Polypharmacy and medication compliance in patients with type 2 diabetes. Int J Pharm Sci Res 2017;8:2298–301.
[12]. Charlson ME, Pompei P, Ales KL, et al. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J Chronic Dis 1987;40:373–83.
[13]. Roffman CE, Buchanan J, Allison GT. Charlson comorbidities index. J Physiother 2016;62:171.
[14]. Deyo RA, Cherkin DC, Ciol MA. Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. J Clin Epidemiol 1992;45:613–9.
[15]. Romano PS, Roost LL, Jollis JG. Presentation adapting a clinical comorbidity index for use with ICD-9-CM administrative data: differing perspectives. J Clin Epidemiol 1993;46:1075–9.
[16]. D’Hoore W, Sicotte C, Tilquin C. Risk adjustment in outcome assessment: the Charlson comorbidity index. Methods Inf Med 1993;32:382–7.
[17]. Huang Y-Q, Gou R, Diao Y-S, et al. Charlson comorbidity index helps predict the risk of mortality for patients with type 2 diabetic nephropathy. J Zhejiang Univ Sci B 2014;15:58–66.
[18]. Lau DT, Nau DP. Oral antihyperglycemic medication nonadherence and subsequent hospitalization among individuals with type 2 diabetes. Diabetes Care 2004;27:2149–53.
[19]. Thygesen SK, Christiansen CF, Christensen S, et al. The predictive value of ICD-10 diagnostic coding used to assess Charlson comorbidity index conditions in the population-based Danish National Registry of Patients. BMC Med Res Methodol 2011;11:83
[20]. Hubbard AE, Ahern J, Fleischer NL, et al. To GEE or not to GEE: comparing population average and mixed models for estimating the associations between neighborhood risk factors and health. Epidemiology 2010;21:467–74.
[21]. McGovern A, Tippu Z, Hinton W, et al. Systematic review of adherence rates by medication class in type 2 diabetes: a study protocol. BMJ open 2016;6:e010469.
[22]. Chang P-Y, Chien L-N, Lin Y-F, et al. Nonadherence of oral antihyperglycemic medication will increase risk of end-stage renal disease. Medicine 2015;94:e2051.

Charlson comorbidity index; diabetes mellitus; generalized estimating equation model; health care costs; medication adherence

Copyright © 2020 the Author(s). Published by Wolters Kluwer Health, Inc.