Leveraging Predictive Analytics to Derive Patient Adherence Drivers

Leveraging Predictive Analytics to Derive Patient Adherence Drivers
Ewa J. Kleczyk, Ph.D., Executive Director, Commercial Effectiveness Analytics, Symphony Health Solutions and Derek Evans, Senior Vice President, Symphony Health Solutions
Abstract:
Understanding therapy adherence and its factors is an important part of managing healthcare costs and improving patients’ health outcomes. Leveraging the Cox Promotional Hazard Model and health claims data can aid in identifying the ‘risk factors’ to staying on the treatment over time, and inform optimized and efficient resource deployment strategies. This article will review the methods often utilized for adherence evaluation, as well as introduce a case study, assessing the factors driving adherence long-term.
 
Keywords: Adherence, Risk Factors, Cox Proportional Hazard Model, Health Claims Data
Introduction
Patient adherence to therapy is an essential part of improving patient health outcomes.  Non-adherence to therapy can increase healthcare costs in the long term, as it can lead to increased comorbidities and concomitant therapies, as well as decreased quality of life.  It is estimated that the economic cost of non-adherence to treatment is at 100 billion to 300 billion dollars.1 This cost includes lost wages, which result from increased disease burden, in addition to the cost of the avoidable health care.  

Understanding not just the current level of adherence within a population, but the factors, which impact adherence, are an important part of managing healthcare costs.  This article will review the currently leveraged measures of adherence, and the ‘Whys’ driving patient’s continuation on the prescribed treatment.  It will also discuss a case study, assessing adherence and its drivers.     

Practices for Measuring Adherence
Patient adherence to treatment is important to improving patient outcomes. There are a few standard measures of adherence, including Medical Possession Ratio (MPR), Proportion of Days Covered (PDC), Refill Compliance Days (RCD), Continued Measure of Medication Gaps (CMMG), Medication Refill Adherence (MRA), and Duration of Therapy (DT), to name a few.2  The most often employed approaches to measure adherence include MPR and PDC metrics. 

MPR is best used for assessing adherence for a single product by measuring the total days a product is supplied to a patient verses the total time, which elapses over the treatment period.  Generally speaking, MPR is best for measuring adherence for a single product with a daily treatment schedule, such as the case with diabetes and hypertension.3

PDC, on the other hand, allows for the assessment of adherence when concomitant therapies should be taken into consideration. It is better suited for therapeutic areas where treatment switching and multiple-medication use is common.3 Multi-therapy treatment often occurs in the oncological and immunological therapeutic areas.

According to the Pharmacy Quality Alliance (PQA), MPR can be biased, as it tends to overestimate adherence, as it double counts overlap periods when switching therapies.3  Due to these limitations PDC has become the more preferred measurement of adherence; however, it may underestimate adherence rates in situations where a patient refills their medication earlier than scheduled.3

While these measures inform whether the level of adherence to treatment is within a given range for a given population of patients, they are static measures, which do not provide insights into the factors which are contributing to adherence within a particular patient population. 

Understanding Drivers of Adherence
To enhance the adherence analysis, it is beneficial to understand the factors driving product continuation overtime. These drivers are usually referred to as ‘risk factors’ to adherence, and include patient’s cognitive ability, attitudes towards treatment, demographics and socioeconomic variables4, the number and types of comorbidities, and concomitant therapies. The ‘risk factor’ may either increase or decrease the probability that a patient will remain adherent to treatment.5

For example, the amount of information shared by a physician, as well as a patient’s ability to retain and remember the physician’s treatment recommendations, has a significant impact on the likelihood for staying on therapy long-term.4,6  On the other hand, patient adherence tends to decrease for treatments aimed for disease prevention.4 Demographic, socioeconomic, and health-related cost variables are also often cited as ‘risk factors’ in empirical research, especially as it relates to the economic variables such as out-of-pocket costs accrued by patients.7

Identifying significant ‘risk factors’ allows brand marketing teams to utilize and optimize their resources more efficiently, based on the factors’ size and probability of non-adherence, and ultimately implement a marketing strategy in a way that is both effective and timely. For example, if changing managed care organizations (MCOs) decreases adherence to treatment, the brand team can design programs proactively that may offer physician support in obtaining prior authorization approvals after a plan change in order to prevent switching or dropping therapy.8 

Adherence Research Approach Overview
There are many qualitative and quantitative ways to identify and measure the impact of ‘risk factors’ on patient adherence. For example, researchers might conduct qualitative research studies that include face-to-face interviews with patients or physicians to facilitate collection of information related to patient’s adherence and perceived barriers to staying on therapy over time. These types of studies tend to inform cultural, cognitive, brand perceptions, and physician/patient relationship based factors.9   

As qualitative research studies are often limited by the sample size of participants, patient or physician quantitative survey studies might inform additional adherence ‘risk factors,’ including impact of out-of-pocket cost or comorbid conditions on adherence.9   

Working with the primary market research data may sometimes result in a limited ability to collect in-depth information on behavior-based topics of interest or track the actual behavior of the respondents. As a result, it is often recommended to leverage secondary data to help provide additional insights not captured in the survey based approaches.10 For example, actual patient history available from health claims data can be included in analyzing adherence to medical treatment to provide more comprehensive and detailed insights into the patient treatment journey to accurately measure and estimate the impact of selected drivers on the length of staying on the therapy. More detailed information about health claims data will be provided in a later section of this article. 

The econometric methods often leveraged to estimate the impact of the ‘risk factors’ on patient adherence include techniques from simple correlation analysis to multivariate cross-sectional and time-series regression analysis (i.e. logistic regressions). These types of methodologies allow to either identify correlations between variables and discontinuation of therapy, and/or in addition, identify the causality and regression estimates of each individual factor on patient adherence.5  

In this article, the Cox Proportional Hazard Model is discussed as an approach to study adherence of medical treatments. The technique can be leveraged to identify drivers of adherence otherwise known as ‘risk factors,’ and the level of impact healthcare variables have on patient adherence to treatment.

Leveraging Cox Proportional Hazard Model for Deriving Adherence 'Risk Factors'
The Cox Proportional Hazard Model represents a group of survival models, often used in biostatistics, that relate the time that passes before some event occurs to one or more covariates that may be associated with that quantity of time.11  

The original model, introduced by Cox in 1972, was adopted in economics and business areas, and several variations and changes were proposed to the base model over the years. These changes included accounting for unobserved heterogeneity and selection of entry for non-random entry models.12

In this article, the Cox Proportional Hazard Model is introduced for estimating the impact of ‘risk factors’ on drug adherence. The benefits of the model include capturing time variance, which allows modeling of not just the overall impact of an event, but the impact that the event has over time based on its timing relative to other events. This ensures that the ‘risk factors’ probabilities and their associated impacts account for the time dynamics on adherence. Inclusion of time allows for the estimation of the probability of adherence at a selected point in time, while taking into account the impact of multiple factors.13   The concept of time varying covariates in the survival model is presented graphically in Figure 1. 

Figure 1: Explaining Time Varying Covariates


Defining the ‘Risk Factors’ to Adherence
The Cox Proportional Hazard Model is often defined via the following functional form:
  1. λ(t|z)=λ0(t)ez1β1+···+zpβp = λ0(t)ezT β,
where z is a p×1 vector of covariates (represented here as ‘risk factors’), such as treatment indicators, prognostic factors, etc., and β is a p×1 vector of regression coefficients (Cox, 1972).11 

Furthermore, λ(t|z = 0) =λ0(t). So, λ0(t) is often called the baseline hazard function, and it can be interpreted as the hazard function for the population group with z = 0. The baseline hazard function λ0(t) in model 1 can take any shape as a function of t. The only requirement is that λ0(t) > 0.11 

For any two sets of covariates (represented here by ‘risk factors’) z0 and z1,
  1. λ(t|z1) λ(t|z0) = λ0(t)ezT1 β λ0(t)ezT0 β =e (z1−z0)T β, for all t ≥ 0,
which is a constant over time. Equivalently,
  1. logλ(t|z1) λ(t|z0) = ( z1 −z0)Tβ, for all t ≥ 0.
With one-unit increase in zk while other covariate values are being held constant, the formulation is presented as:
  1. logλ(t|zk + 1) λ(t|zk) = log(λ(t|zk+1))−log(λ(t|zk)) = βk.
Therefore, βk is the increase in log hazard (i.e., log hazard-ratio) at any time with unit increase in the kth covariate zk.  Equivalently,
  1. λ(t|zk + 1) λ(t|zk) =eβk, for all t ≥ 0.

So eβk is the hazard ratio associated with one-unit increase in zk. Furthermore, since
  1. Probability [t ≤ T<t+∆ t|T ≥ t,z] ≈ λ(t|z)∆t,
than
  1. Probability [t ≤ T<t+∆ t|T ≥ t,zk + 1] / Probability [t ≤ T<t+∆ t|T ≥ t,zk] ≈ eβk, for all t ≥ 0.
k can be interpreted as the ratio of two conditional probabilities of discontinuation of prescribed therapy in the near future given that the patient is on therapy at any time t.11

Since
  1. λ(t|zk + 1)−λ(t|zk) λ(t|zk) =eβk −1,
as a result, eβk − 1 can be interpreted as the percentage change (increase or decrease) in hazard defined as discontinuation of prescribed therapy with one-unit increase in covariates (otherwise known as ‘risk factors’), zk while adjusting for other covariates. 11 

Using Health Claims for Adherence Analysis
The healthcare industry is continuously generating large amounts of data. This is driven by record keeping, compliance & regulatory requirements, new technologies, and patient care. Historically, most data were stored in hard copy form and were very static (i.e. paper files, x-ray films, scripts).  The current trend is moving toward rapid digitization of these large amounts of data, and new technologies that regularly monitor a host of variables.  As a result, volume, velocity, and variety of healthcare data is rapidly changing. This rapid change is being driven by mandatory compliance and reporting requirements with the goal of improving the quality of healthcare delivery, while reducing healthcare costs. These massive quantities of data (often referred to as ‘big data’) hold the promise of supporting a wide range of medical and healthcare functions, including among others clinical decision support, disease surveillance, and population health management.14,15

Using health claims data is often recommended for adherence modeling and finding ‘risk factors’ for staying on therapy long-term. For the purposes of this article, data for billing purposes (e.g. CPT and ICD 9/10 codes) is introduced along with medication (NDC codes) consumption in the healthcare claims bundle. While alternative data sources, such as clinical data (electronic medical or healthcare records (EMR or EHR), and medical images), lab data or genomics data and behavioral data are all readily available, the volume, variety, velocity, and veracity might not be present in alternative data sources.16,17,18 Examples of data assets available in the healthcare industry for analysis are shown in Figure 2. 

Figure 2: Health Claims Data Overview19


Health claims can help capture the patient’s interaction with the healthcare system in a comprehensive manner over an extended duration of time. They are almost universally available, and have the benefit of being comprised of structured data, which aids in processing and analysis. Claims data are often at the patient level, and provide information on drugs dispensed by pharmacies, procedures performed, including those often used in oncological and immunological areas, as well as plan and copay information, and approvals of treatment by providers. Claims data often includes hospital transactional data that allow for linking of in- and out- patient treatment to further enhance the completeness of patient care.20

The ability to track patient longitudinally can often be preserved to a high degree via collecting claims data overtime based on known patient variables common between the different datasets (i.e. name, address, date of birth). Companies collecting healthcare data often have their proprietary algorithms to preserve the longitudinal data information, and merge information over time.  On the other hand, changes in insurance providers might not always allow to capture the claims longitudinally if the data aggregation does not include all variables important for the process of assembling patient level datasets.21

Figure 3: Patient Level Health Claims Data Overview20


Figure 3 presents an example of a patient level health claims database, which brings together vast claims sources—medical, hospital, and prescription—to offer a consistent view across prescriber, payer, and patient dimensions. The database captures longitudinal information on more than 274 million patients over the last 12 years, which allows an understanding of patient medical history, insurance plan changes, diagnosis, comorbidities, selected treatments, and in- and out- patient care.  The patient level information can be easily associated with more than 1.8 million healthcare providers, as well as 13.1 million employer groups within the healthcare environment, to allow for cost benefit analysis, as well as provider segmentation and targeting.20

With specific views and tools, the patient level claims data can answer key questions and facilitate critical commercial processes within sales, marketing, and managed markets. For example, it can present a comprehensive view of a given health event, allowing for its evaluation from many different angles, and development of insight-driven strategies and programs. By enhancing claims data with information from non-retail invoices and point-of-sale data, and then applying adjustments for products abandoned at the pharmacy, the health claims data provides a detailed and nearly complete view into the brand’s journey.20

The health claims data is often aggregated to the healthcare provider level to provide a comprehensive physician level view, allowing for in-depth analysis of treatment algorithms.  The healthcare provider data can often be merged and supplemented with other data sources from other industry via using unique physician identifiers such as ME, NPI and/or DEA numbers. Most physician level data sets include at least one of the identifiers, allowing for a comprehensive analysis across disparate data sources.20 

Although there are many benefits to using health claims data as presented above, the data might not always fully capture patients’ medical history, due to the differential rates of capturing medical claims or a lack of reporting of laboratory test results. For example, hospital claims usually have low coverage rates of rendered services and might include inconsistent reporting formats, which impact the ability to track in-patient treatments and the linking of other data sources to out-patient treatment after hospitalization.21 The differentiated capture rates of therapies, especially for infused and injectable drugs, might also result in small sample sizes, especially for rare and orphaned disease therapies, which might cause difficulties in studying those therapeutic areas in-depth. 

One of the limitations of claims data is its lack of specific confirmatory information, which results in medium recall and medium precision for characterizing patients.22 For example, health claims data often does not include results of lab tests, which disallow tracking precisely disease progression and understanding of the physician decision making process in selecting treatments. This might be especially important in oncology and immunology, where laboratory test results impact the treatment pathways chosen for each patient.23

Some of the gaps in the healthcare claims data can be supplemented with the Electronic Medical Records (EMR) or laboratory claims data to provide missing information in the health claims data. For example, EMR data can often provide an in-depth and comprehensive view of patient’s history over time. The EMR data can often be merged in with the health claims data; however, due to an often limited sample size for each EMR vendor, only a limited sample of patients can have fully supplemented medical history. In addition, other data assets, such as imaging, genomics, biosensor readings, and consumer and promotional events datasets, can be merged with the health claims data to provide a more comprehensive view of patient disease progression, treatment pathways, and exposures to promotional events during the treatment decision making process.24


Case Study: Estimating the Adherence Drivers via Survival Analysis
Introduction
Consider a chronic therapeutic area that impacts about 10% of Americans, of whom nearly half of the affected individuals are aware or properly diagnosed. The often recommended therapies for diagnosed patients include a variety of treatments from branded products and generics, to over the counter medications. Generic and branded products account for more than 95% of the prescribed and recommended treatments.25 

Product A is a leading branded treatment in the therapeutic area. There is also a generic version of the product, accounting for more than 80% of the market.25 Marketing leaders would like to understand the positive and negative drivers of adherence to ensure efficient and optimized allocation of resources to drive the brand’s maximum performance in-field.  

The objectives of the case study, therefore, are:
  1. To measure patient adherence for a key branded drug, called Product A. 
  2. To understand the ‘risk factors’ impacting patients staying on the treatment over time. 
Methodology
The case study leverages patient level health claims data, as well as the Cox Proportional Hazard Model to identify ‘risk factors,’ and the associated level of impact of healthcare variables on patient adherence for Product A.

Analysis Assumptions:
  • Timeframe: 2011 - 2015
  • A stable panel of patients who used more than one treatment (including Product A), and were new to the therapeutic area at the time of study 
  • Right censoring: a proper adjustment based on month of entry into the market
  • Testing for the statistical significance at the minimum of 90% CI of various patient groups for each of the following covariates:
    • Demographics
    • Comorbidities
    • MCO Plan Changes
    • Procedures
    • Prices
    • Transaction Types (Mail Order vs. Retail)
  • Cox Proportional Hazard Model is leveraged to estimate the impact of covariates on drug adherence, defined here as the ‘risk factors’ to discontinuing Product A. 
  • The model outputs are presented as: 
    • Relative Risk Ratios that represent the strength of the association with the covariate:
      • Less than 1, implies that the covariate decreases the risk for discontinuing Product A.
      • Greater than 1, implies that the covariate increases the risk for discontinuing Product A.
  • Risk Probabilities that are provided in absolute values, and their direction is noted by the accompanying Relative Risk Ratios. The probability values represent the percentage increase or decrease in risk for therapy discontinuation.
Case Study Results
Within the case study therapeutic area, overall most patients maintained their treatments long term, including Product A, and were highly loyal to their product of choice with more than 70% of patients continuing treatment after 3 years. These results are consistent with other empirical studies, in which patients treated for chronic diseases were more likely to stay on therapy long-term.4,26  The results were also consistent between MPR and PDC measures of adherence. The patients in the stable panel had a variety of comorbid conditions, including fatigue and hormonal issues.25

Table 1: Risk Factors Driving Patient Adherence / Risk for Discontinuation of Product A at 90% CI
Risk Factors Relative Risk Ratios Risk Probabilities
Plan Changes 3.92 292%
Mail Order Delivery 0.47 53%
Mental Health 1.29 29%
Vitamin Deficiency 0.74 26%
Gender (Females) 1.15 15%
High Medication Burden 1.09 9%
Out-of-Pocket Costs 1.01 1%


The predictive analytics results, leveraging the Cox Proportional Hazard Model, suggest that the following ‘risk factors’ are important at the minimum of 90% CI in driving adherence or as defined in this study the risk of discontinuing Product A over time. As shown in Table 1, there are several ‘risk factors’ impacting adherence for Product A (or increasing the risk for discontinuing Product A): plan changes, gender (F), medication burden, out-of-pocket costs, as well as selected comorbidities, including mental health diagnosis. 

These results are consistent with previous empirical research, studying the impact of covariates on staying on therapy long-term. For example, in her article, Fullman found that medication adherence levels increased with third party co-insurance payments, while decreased with a higher cash burden accrued by patients.7 In addition, mental health and specifically, depression diagnosis, was also found as one of the strongest predictors of patient non-adherence to medical treatments.6 In this study, mental health diagnosis increases the risk for Product A discontinuation by 29%. 

As cited above, gender is also an important predictor in negatively driving adherence, and increasing the risk for discontinuing Product A by 15%. This result, although perhaps counter intuitive, has been mentioned in previous empirical research studies. There are many causes of this phenomenon, including often cited income disparities between women and men, medications’ side effects and efficacy impact, as well as the role women play in caring for their families in society. Since women are more often the primary caregivers, they tend to ensure their loved ones are cared for first before attending to their own needs, and are also more likely to consider how the prescribed medication will affect them and impact their day to day activities.27 

In this case study, the highest impact of ‘risk factor’ driving non-adherence are changes in insurance plans. As health claims data spans across multiple insurance plans, it often allows us to investigate the price elasticity of prescribed treatments, and the impact of local managed care organizations on patient treatment choices compared to national-footprint plans. Results of such studies can aid revisions in contracting with MCOs to improve coverage, and therefore patient adherence long-term. 

In this case study, insurance plan changes cause the highest rate of switching away from Product A (nearly 300% increase in risk), most often progressing to generics. The changes usually occur January through March each year, causing some patients to drop from the market altogether, if they are not able to remain on Product A as their preferred choice. Interestingly, many patients might be on the drug in the following year, due to favorable changes to insurance coverage or help with product pre-authorization. This finding was confirmed in a previous empirical study authored by Fendrick and Chernew, in which allowing health plans the flexibility to cover more services outside of the deductible, enhanced consumer choice and increased adherence.28

On the other hand, mail order deliveries of Product A decrease the risk of patients discontinuing the therapy long-term by 53%. Mail order deliveries provide a more convenient way of refilling the medication, and therefore drive adherence over time. These results are in agreement with an empirical article conducted by All Scripts Holding Company, which analyzed the association between drug delivery channels and adherence.  The study offered strong evidence that drug home delivery was associated with greater odds of being adherent with prescribed treatments for Medicare patients.29 

Other covariates of interest were also included and analyzed in this case study, but they were not found statistically significant at the 90% CI. For example, patient education and income levels were not statically significant in driving Product A’s adherence. Previous empirical studies present mixed results on the impact of socioeconomics variables on non-adherence, with some citing significant impact while others implying a limited correlation. The results usually are dependent on the homogeneity of the studied population, and the therapeutic area of interest.30,31 Furthermore, blood tests, often performed to monitor progression of the condition, did not impact staying on therapy long-term either. Other empirical studies cite, however, that frequent blood testing is a good measure of chronic disease progression monitoring and optimizing patient treatment, while increasing drug adherence.32 The insignificant results might suggest the need for tracking the test outcomes to inform changes in the treatment from disease progression, which may in turn lead to increased patient adherence. 

Recommendations
Leveraging the survival analysis model provides the opportunity for realizing additional insights into not only the adherence level for Product A, but the marketing and sales strategies, as well as programs that could help lower the probability for non-adherence.  

Understanding the key drivers and their associated impacts on non-adherence can aid in driving brand success in an efficient manner over time, while greatly impacting patient outcomes via improved compliance and adherence. For example, knowing that mail order deliveries of medications increase patient adherence may lead to extension of the programs, while ensuring patient convenience in refilling their medications.  

Research Limitations and Future Direction
The primary limitation of the study is leveraging only health claims data and not accounting for other factors related to the patient’s treatment journey, including laboratory results. Understanding in-depth the progression or sub-type of the disease might add significantly to the evaluation of drivers of adherence. 

In addition, including more information related to patient cognitive abilities in comprehending the prescribed treatment dosing and schedule recommendations, as well as understanding patients’ perceptions of the prescribed treatment, disease progression, and seriousness of the condition, may also provide an additional explanation for the adherence levels, as well as its drivers and barriers. 

As mentioned in the earlier section of the article, linking other datasets, including primary market research information, EMR, and genomics, might further define and help evaluate the drivers and ‘risk’ factors of patient adherence via presenting a more comprehensive view of patients and their experiences on the prescribed therapy.  

Furthermore, the data may allow investigating the health and economic outcomes of therapy, as well as evaluating how the current sales and marketing strategies and tactics, as well as other programs implemented to increase patient education and adherence, drive the desired outcomes. These results could inform an optimized and efficient allocation of resources to maximize desired health outcomes. 

Finally, future developments and direction in this research area may lead adherence assessment into leveraging other analytics approaches, including exploratory machine learning algorithms, in order to increase predictive accuracy of the survival models. Through the use of social media data and natural language processing, additional insights might also be gained into the more ‘personal’ effects currently not captured in health claims data.

Acknowledgment
The authors would like to thank Hanna Socaciu for her assistance in writing this article, as well as the anonymous reviewers for their valuable comments and suggestions to the publication. 

About the Authors
Ewa J. Kleczyk, PhD is an Executive Director of Commercial Effectiveness Analytics with Symphony Health Solutions, with experience in primary and secondary market research and marketing analytics for pharmaceutical clients. 

Derek Evans is a Senior Vice President leading the Commercial Effectiveness Practice at Symphony Health Solutions, with experience in primary and secondary market research, marketing analytics and strategic consulting for pharmaceutical clients.


References

1 Iuga AO, McGuire MJ. Adherence and Health Care Costs. Risk Management Healthcare Policy. 2014; 7: p. 35-44.

2 Sudeep K, Cleves MA, Helm M, Hudson TJ, West DS, Martin BC. Empirical Basis for Standardizing Adherence Measures Derived from Administrative Claims Data among Diabetic Patients. Medical Care. 2008 November; 46(11): p. 1125-1133.

3 Nau D. Proportion of Days Covered (PDC) as a Preferred Method of Measuring Medication Adherence. Article. Springfield, VA.

4 Beena J, Jimmy J. Patient Medication Adherence: Measures in Daily Practice. 2011 May; 26(3): p. 155-159.

5 Mann DM, Howard D, Ponieman H, Leventhal EA. Predictors of Adherence to Diabetes Medications: The Role of Disease and Medication Belifes. Journal of Behavioral Medicine. 2009; 32: p. 278-284.

6 Martin LR, Williams SL, Haskard KB, DiMatteo MR. The Challenge of Patient Adherence. Therapy Clinical Risk Management. 2005; 1(3): p. 189-199.

7 Fullman P. The Drivers that Impact the Patient Adherence. PM360. 2012 October 1.

8 McKethan A, Benner J, Brookhart A. Health Affair Blog. [Online].; 2012 [cited 2017 January 10. Available from: http://healthaffairs.org/blog/2012/08/28/seizing-the-opportunity-to-improve-medication-adherence/.

9 Muller S, Kohlmann T, Wilke T. Validation of the Adherence Barriers Questionnaire - an Instrument for Identifying Potential Risk Factors Associated with Medication Related Non-Adherence. BMC Health Services Research. 2015; 15: p. 153.

10 What Researchers Mean by Primary and Secondary Data. A Quarterly Publication of the Institute for Work & Health: At Work. 2015 Fall; 82: p. 2.

11 Cox DR. Regression Models and Life-Tables. 1972; 34(2): p. 187-220.

12 Huynh KP, Voia MC. Mixed Proportional Hazard Models with Continuous Finite Mixture Unobserved Heterogeneity. [Online].; 2016 [cited 2017 January [Syracuse University]. Available from: https://www.maxwell.syr.edu/uploadedFiles/cpr/events/cpr/papers2016/voia.pdf..

13 Therneau T, Crowson C, Atkinson E. The Comprehensive R Archive Network. [Online].; 2016 [cited 2017 January 10. Available from: https://cran.r-project.org/web/packages/survival/vignettes/timedep.pdf..

14 Raghupathi W. Healthcare Informatics: Improving Efficiency and Productivity. In S K, editor. Data Mining in Health Care.; 2010. p. 211-223.

15 Burghard C. Big Data and Analytics Key to Accountable Care Success; 2012.

16 Dembosky A. Data Prescription for Better Healthcare. Financial Times. 2012 December 12; 19.

17 Feldman B, Martin EM, Skotnes T. Big Data in Healthcare Hype and Hope. 2012 October.

18 Fernandes L, O’Connor M, Weaver V. Big Data, Bigger Outcomes. Journal of AHIMA. 2012;: p. 38-42.

19 Denny JC. Chapter 13: Mining Electronic Health Records in the Genomics Era. PLoS Computational Biology. 2012; 8(12): p. e1002823.

20 Symphony Health Solutions. [Online]. [cited 2017 January 10. Available from: www.symphonyhealth.com.

21 The Future of Health Insurance: A Road Map Through Change. Report; 2015. Report No.: CSG No. 1506-1546714 Northeast.

22 Big Data Analytics for Healthcare. In Tutorial Presentation at the SIAM International Conference on Data Mining; 2013; Austin, TX.

23 Emons M. Integrated Patient Data for Optimal PAtient Management: The Value of Laboratory Data in Quality Improvement. In Proceedings of the 24th Arnold O. Beckman Conference.; 2001: Clinical Chemistry. p. 1516-1520.

24 Wilson J, Bock A. The Benefit of Using Both Claims Data and Electronic Medical Record Data in Health Care Analysis. White Paper. Eden Prairie, MN:; 2012. Report No.: 12-27912.

25 Maughn K, Kleczyk E. Assessing Adherence and Usage Behavior to Identify Product Usage among Chronically Ill Patients with Idiosyncratic Treatment Plans. In PBIRG Annual General Meeting; 2016; Bocca Raton, FL.

26 Ho PM, Bryson CL, Rumsfeld JS. Medication Adherence. Its Importance in Cardiovascular Outcomes. Circulation. 2009 June; 119(23): p. 3028-35.

27 Pearson C. Medication Adherence: Are Women Worse then Men at Taking their Meds? The Huffington Post. 2013 May 5.

28 Fendrick M, Chernew ME. Precision Benefit Design - Using “Smarter” Deductibles to Better Engage Consumers and Mitigate Cost-Related Nonadherence. JAMA Internal Medicine. 2017 January 7.

29 Iyengar RN, Lefrancois AL, Frazee SG. Association of Pharmacy Dispensing Channel with Medication Adherence Among Medicare Part D Patients Having Comorbid Conditions. 2014. Poster Presentation.

30 Billings ME, Auckley D, Benca R, Foldvary-Schaefer N, Iber C, Redline S, et al. Race and Residential Socioeconomics as Predictors of CPAP Adherence. Sleep. 2011; 34(12): p. 1653-1658.

31 Falagas ME, Zarkadoulia EA, Pliatsika PA, Panos G. Socioeconomic Status (SES) as a Determinant of Adherence to Treatment in HIV Infected Patients: a Systematic Review of the Literature. Retrovirology. 2008 February 1; 5(13).

32 White T. Stanford Medicine. [Online].; 2015 [cited 2017 February 15. Available from: http://med.stanford.edu/news/all-news/2015/04/adherence-to-blood-thinner-best-with-pharmacist-management.html.