Spurred by the accuracy with which companies like Google and Netflix use large amounts of data to anticipate our interests, there is growing investment in "Big Data" applications to healthcare. For example, the FDA’s postmarket surveillance program—the Sentinel Initiative&mdash;analyzes billions of drug prescriptions for adverse events.

 However, if we’re not careful, big data healthcare could cause harm.

Data are “big” if they either: 1) represent many more subjects than a typical randomized clinical trial (RCT)—tens of thousands or more&mdash;and/or 2) they include a broad range&mdash;hundreds or more&mdash;of clinically relevant patient and provider characteristics. Such data can extend the reach of clinical research to include study of rare events, heterogeneous treatment effects, long-term outcomes, and other topics difficult or impossible to study with RCTs.

Some have suggested that big data will rapidly improve healthcare delivery. For instance, finding insufficient guidance in the medical literature, physicians at Stanford’s Packard Children’s Hospital used electronic medical record data to help make an urgent decision about using anticoagulation medication in a lupus patient.

 The strongest proponents of such big data applications believe that with enough information, causal relationships reveal themselves without an RCT.

Are they right? For clinical applications, this is a vital question. For instance, for every 5 million packages of x-ray contrast media distributed to healthcare facilities, about 6 individuals die from adverse effects.

 With big data, we learn that such deaths are highly correlated with electrical engineering doctorates awarded, precipitation in Nebraska, and per capita mozzarella cheese consumption (correlations 0.75, 0.85, and 0.74, respectively).

However, because we cannot conceive of a causal mechanism, it is obvious that these variables play no causal role in x-ray contrast media deaths. That such high correlations can be easily mined from big data is concerning nonetheless, because it is not always trivial to assess whether they are telling us something useful. For example, observational data reveal that proton pump inhibitor (PPI) use is associated with pneumonia incidence.

 This could be causal because a mechanism is plausible—gastric acid reduction could increase bacterial colonization&mdash;but perhaps the association arises because other factors drive both PPI use and pneumonia incidence.

Faced with this kind of uncertainty, there is temptation to insist that only an RCT can convince us of causation; however, the very promise of big data is its potential to see what RCTs won’t, thereby improving care in ways that RCTs cannot. Although caution is warranted, we should not dismiss big data too quickly. The way forward requires careful selection of observational research designs coupled with rigorous testing for violations of key assumptions on which causal inference relies.

A useful technique for testing a big data finding—like whether PPIs increase the risk of pneumonia&mdash;is to probe for other factors that might be driving it, called a “falsification test.” A falsification test begins with the question, “If something else is driving this observational finding, where else would it show up?” It explores the association of the suspected causal variable (eg, PPI use) with other outcomes (eg, not pneumonia) or in other populations that we have good reason to believe should not be affected by it. The key is to select other outcomes or populations that are likely to also be affected by factors that could be driving the suspected causal relationship.

 raise doubts that the PPI-pneumonia relationship is causal by showing that its use is also associated with outcomes for which no clear causal mechanism exists: chest pain, urinary tract infections, osteoarthritis, rheumatoid arthritis flares, and deep venous thrombosis.

 applied big data and falsification tests to analyze outcomes from use of common second-line medications for type 2 diabetes: sulfonylureas (SUs) and thiazolidinediones (TZDs). Using 10 years of Veterans Health Administration data merged with Medicare claims, they analyzed over 80,000 patients for mortality and hospitalization outcomes—end points for which no prior RCT was adequately powered. The study found that SUs are associated with a 50% increase in risk of mortality and a 68% increase in avoidable hospitalizations compared with TZDs. Study findings were based on variation in the relative rates with which physicians prescribe the 2 medications. Can we interpret them as causal? Put another way, are prescribing patterns random in a way akin to an RCT’s group assignment?

To probe the assumption that they are, the authors applied several falsification tests. First, if prescribing patterns were random, like an RCT’s group assignment, observable patient and provider characteristics should be balanced across groups stratified by high and low rates of physician prescribing. If factors aren’t balanced—whether in an RCT or in an observational study like this&mdash;one should suspect flawed randomization and reject a causal interpretation. Prentice et al observed such balance in demographics, diagnoses, and provider-quality variables, increasing confidence that prescribing pattern served a valid randomizing role.

Next, they also conducted falsification tests using 2 populations that were similar to the study population in every way except disease severity: a healthier population taking only metformin and a sicker population that had been prescribed metformin and then insulin. For neither of these populations was the SU/TZD prescribing pattern statistically significantly related to outcomes (see the Figure

). This could not have been the case had an omitted factor been the missing link in the association of SUs with bad outcomes. These falsification test results further increase confidence that the associations of SUs with greater mortality and avoidable hospitalization risks are causal.

Big data can be useful in healthcare; they can expand the reach of evidence-based medicine into domains not accessible with RCTs, including postmarket surveillance of drugs. But to fulfill that ambition, big data must be coupled with rigorous observational methods. Falsification tests can illuminate when a relationship is less likely to be causal, potentially sparing practitioners from making grave mistakes. Without them, we run the risk of letting our enthusiasm about big data get ahead of the science and what is best for patients.

 VA Boston Healthcare System, Boston University, and Harvard University (ABF), Boston, MA; Northeastern University (SDP), Boston, MA.

 This work supported by Health Services Research and Development Service of the US Department of Veterans Affairs (grant no. IIR 10-136), which did not in any way participate in the writing of or exert any influence over the content of this manuscript.

 The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article. The contents do not represent the views of the US Department of Veterans Affairs, Boston University, Harvard University, or Northeastern University.

 Concept and design (ABF, SDP); acquisition of data (ABF, SDP); analysis and interpretation of data (ABF, SDP); drafting of the manuscript (ABF, SDP); critical revision of the manuscript for important intellectual content (ABF, SDP); obtaining funding (ABF, SDP).

 Austin B. Frakt, PhD, VA Boston Healthcare System, 150 S Huntington Ave (152H), Boston, MA 20130. E-mail: frakt@bu.edu.

1. Findlay S. Health policy brief: the FDA’s Sentinel Initiative. 

 website. http://www.healthaffairs.org/healthpolicybriefs/brief.php?brief_id=139. Published June 4, 2015. Accessed June 8, 2015.

2. Frankovich J, Longhurst CA, Sutherland SM. Evidence-based medicine in the EMR era. 

. 2011;365(19):1758-1759. doi:10.1056/NEJMp1108726.

3. Wysowski DK, Nourjah P. Deaths attributed to x-ray contrast media on US death certificates. 

4. Prasad V, Jena AB. Prespecified falsification end points: can they validate true observational associations? 

. 2013;309(3):241-242. doi:10.1001/jama.2012.96867.

5. Pizer SD. Falsification testing of instrumental variables methods for comparative effectiveness research [published online August 21, 2015]. 

6. Prentice JC, Conlin PR, Gellad WF, Edelman D, Lee TA, Pizer SD. Capitalizing on prescribing pattern variation to compare medications for type 2 diabetes. 

. 2014;17(8):854-862. doi:10.1016/j.jval.2014.08.2674.&ensp;

Articles

COMMENTARY

Big Data analyses are observational, raising threats to causal inference. Validity checks help, but we must not let enthusiasm about Big Data obscure the science.

The Promise and Perils of Big Data in Healthcare

Recent studies have revealed significant variation in medication adherence among patients with chronic conditions. Little is known about the effect of continuity of care (COC) on changes in medication adherence. This study aims to identify medication adherence trajectories among patients with newly diagnosed diabetes, as well as to examine the association of COC and medication adherence among various adherence trajectories. 

This study utilized a longitudinal design with a 6-year follow-up, from 2002 to 2008, under a universal health insurance program in Taiwan. Subjects 18 years or older with type 2 diabetes that was newly diagnosed in 2002 were included in the study. The main outcome was medication adherence measured by medication possession ratio each year. Group-based trajectory models were used for analysis. 

Four medication adherence trajectories were identified: persistent adherence (39.9%), increasing adherence (27.5%), decreasing adherence (12.0%), and nonadherence (20.6%). Patients with high or medium COC index scores were more likely to be adherent to medications than those with low COC index scores in all of the trajectory adherence groups.

This study demonstrated the heterogeneity in patients’ medication adherence and identified 4 distinct trajectories of medication adherences among those with newly diagnosed type 2 diabetes. Improving COC may lead to better medication adherence in all of the adherence trajectory groups.

	This study employed longitudinal data to identify medication adherence trajectories among patients with newly diagnosed diabetes and to examine the relationship between continuity of care (COC) and medication adherence among various adherence trajectories. 

		This study demonstrated the heterogeneity in medication adherence and identified 4 distinct trajectories of medication adherence: persistently high, increasing, decreasing, and nonadherence. 

		Better COC between patients and their physicians was consistently associated with higher medication adherence in every trajectory adherence group. 

		More attention should be directed toward designing new strategies to improve COC.

Diabetes is a prevalent condition characterized by significant complications and serious consequences.

 Medication adherence is the most important determinant of diabetes management; its effects are greater than those of dietary control or lifestyle changes.

 However, medication adherence rates vary markedly, from 36% to 93%, among patients with diabetes.

 Nonadherence to medications is associated with worse glycemic control, which, in turn, can increase micro- or macrovascular complications and healthcare utilization.

 Therefore, improving adherence to medications is an important task for policy makers worldwide. Medication adherence is also one of the main issues considered in value-based insurance design.

Although the issues that affect medication adherence have been examined, 2 gaps remain in our knowledge about this issue. First, the majority of studies concerning medication adherence have used a static measure of medication adherence (ie, at a single point in time).

 The few studies that have used a longitudinal design either did not focus on the dynamic changes in medication adherence,

 or only focused on the changes between 2 time points.

 Furthermore, studies focusing on the changes between 2 time points could not capture the heterogeneity in medication adherence changes that occur after the second time point. Until now, only 2 studies utilized advanced dynamic measurement, capturing the short-term heterogeneity in medication adherence and examining the factors affecting the trajectory patterns of medication adherence.

Continuity of care (COC) is significantly associated with favorable health outcomes, especially among patients with chronic conditions.

 These patients usually receive care from several physicians and have complex medication regimens that often result in fragmented care. The medical home demonstration project in the United States intends to improve care coordination among various healthcare providers, and thus improve COC between patients and their physicians.

 However, the majority of previous studies have focused on the effects of COC on healthcare utilization and patient satisfaction.

 There is limited and inconclusive evidence regarding the effects of COC on medication adherence.

 Furthermore, to the best of our knowledge, the effects of COC on medication adherence among patients with various adherence trajectories have been relatively unexplored.

Group-based trajectory models may provide an alternative approach to modeling the change of outcomes and individual differences in changes of outcomes.

 Its purpose is to discover distinctive patterns of individuals with homogeneous longitudinal trajectories within the population and to explain individual-level differences at the group level.

 Using group-based trajectory analysis, this study aimed to identify medication adherence trajectories within 6 years after the initial diagnosis of diabetes, and to examine the association of COC and long-term medication adherence among patients with various adherence trajectories.

Healthcare systems in Taiwan and in many Asian and Central and Eastern European countries largely focus on specialist and hospital care. There is no referral requirement; patients are free to visit physicians at community clinics or hospital outpatient departments for each episode according to individual preference.

 Since the implementation of the compulsory National Health Insurance (NHI) program in Taiwan in 1995, empirical studies have found that the NHI has significantly improved general access to healthcare.

 The average number of annual physician visits in Taiwan is one of the highest in the world, with approximately 13 visits per person in 2010; patients are often criticized for their doctor-shopping behavior.

 Therefore, the features of having free choice of providers for every visit without referral requirement and a very high number of physician visits under the NHI program might hamper information sharing or mutual trust between patients and their physicians; this can subsequently result in the deterioration of COC and poor patient adherence to medications. The easy access to ambulatory care and comprehensive claims data under the universal health insurance system facilitates the need for investigations into the effects of COC on medication adherence among patients with diabetes.

The NHI claims data set from 1999 to 2008 was obtained from the National Health Research Institutes of Taiwan. A longitudinal study design was used for this study. We identified patients with type 2 diabetes (T2D) according to the 

International Classification of Diseases, Ninth Revision, Clinical Modification

) codes of 250.xx, excluding the type 1 diabetes codes of 250.x1 and 250.x3, in the claims data. First, we defined the index date of the first claim with the diagnosis of T2D for each patient in the year 2002. New patients with T2D were identified by their lack of T2D-related claims over the 3 years (1999-2002) before the index date of diagnosis in 2002.

Patients were then included in the analysis if they: 1) were aged at least 18 years on the index date; 2) were alive throughout the study period to ensure comprehensive follow-up observations; 3) had at least 3 physician visits in every year of the study period given the inapplicability of COC indicators in patients with very few visits (this criterion has been adopted in previous studies

); 4) had a prescription for 1 or more oral antihyperglycemic medication on the index date of the initial diagnosis to capture the appropriate timing of initial prescriptions for new patients

; 5) had at least 1 prescription for an oral antihyperglycemic medication in every year of the study period to ensure that the patients included in the study were those who required ongoing pharmacological therapy; and 6) had no insulin prescriptions during any of the years in the study period. The unit of analysis was patient-years. As a result, a total of 12,123 patients and 72,738 patient-years were included in the analysis.

COC was the main independent variable in this study. This study calculated the COC score based on outpatient services only. Considering the variation and very high number of physician visits in Taiwan, we chose the continuity of care index (COCI) as our predictor variable because the COCI is minimally sensitive to the number of physician visits by patients.

 The COCI was composed of the number of different physicians seen and the number of visits to each physician. The COCI value ranges between 0 and 1, with a higher value representing better COC. The equation for the COCI is available in the PDF. 

represents the total number of physician visits, 

 is the number of visits to the same physician (

 is the number of physicians. In a study by Bice and Boxerman (1977),

 the summation term in the numerator was the sum of the number of unreferred physicians. Because of the lack of referral arrangements in Taiwan, we used the total number of physician visits in the analysis. Also, because the COCI values have no inherent clinical meaning, we categorized them into 3 equal tertiles (low [0-0.23], intermediate [0.24-0.43], and high [0.44-1.00]) according to the distribution of scores across the entire study population for each year.

Medication adherence was defined using the medication possession ratio (MPR), which was calculated as the ratio of the total days with a supply of a prescribed medication divided by the total number of days in each year during the 6 years following the index date. The MPR calculation also included medication days based on refill visits for oral antihyperglycemic drugs.

Antihyperglycemic drugs were identified by their anatomical therapeutic chemical codes and included biguanides (A10BA), sulfonamides or urea derivatives (A10BB), combinations of oral blood glucose—lowering drugs (A10BD), alpha glucosidase inhibitors (A10BF), thiazolidinediones (A10BG), dipeptidyl peptidase 4 (DPP-4) inhibitors (A10BH), and other blood glucose-lowering drugs, excluding insulin (A10BX). When the patients were hospitalized, the days on which they took the prescribed medications were excluded from the denominator in the MPR calculation. In addition, the patients may have received multiple medications for different numbers of days during a visit; therefore, the calculation of MPR was based on the prescription that covered the longest days (usually for chronic conditions such as diabetes). In this study, patients were considered to be adherent to oral antihyperglycemic medications if the MPR was between 80% and 120% for each year.

 The trajectories of medication adherence were constructed using the adherence statuses in the 6 years after the initial diagnosis of diabetes.

Several confounding factors were controlled for in the regression models. The time-varying variables were health status measures, including: 1) hospitalization in the previous year; 2) the diabetes complication severity index (DCSI)

; 3) the chronic illness with complexity index (CICI)

; and 4) multiple medications in a prescription. The DCSI was used to adjust for the severity of diabetes complications and consisted of scores from 7 categories of complications: cardiovascular complications, nephropathy, retinopathy, peripheral vascular disease, stroke, neuropathy, and metabolic disorders. The CICI was used to adjust for comorbidity in patients with multiple chronic diseases. We excluded diabetes-related complications to avoid duplicating the comorbidity effect presented by the DCSI. In the analysis, we counted the number of comorbidities based on the CICI scores. The multiple medications per prescription were dichotomized by 3 medications or more.

The time-constant variables were the patient’s sex, age, and the location of residence (rural or urban) in the baseline year. Due to a lack of information on a patient’s residence in the NHI claim data, we took 2 steps to approximate the residence for each patient. First, we identified 3 types of beneficiaries who were enrolled into NHI via occupational unions or local government agencies—farmers and fishermen/members of occupational unions, low-income families, and local residents who were enrolled via household registry office&mdash;where the zip code of enrollment organization was considered as the proxy for patient’s residence. For the abovementioned subjects, this study only included the NHI policy holders and their spouses because they were presumably living together. For the remaining subjects, we used the zip code of the most frequently visited clinic/hospital for minor illnesses (

 codes 460-466, 480-487) as proxies. Rural-urban designations were based on the population density of the township of the patient’s residence. Those townships with the lowest 30% of population density were defined as a rural area in Taiwan.

We used a semi-parametric, group-based trajectory model to identify distinct subgroups of individuals who followed similar patterns of medication adherence over time.

 This model was performed with the SAS version 9.2 PROC TRAJ procedure (SAS Institute, Cary, North Carolina).

We employed the Bayesian information criterion (BIC) to determine the optimal number of trajectories. Because of the large sample size in this study, we anticipated that the BIC value might decrease along with the increase in the number of trajectories; nonetheless, we only considered 1 to 4 trajectories to facilitate the interpretation of the adherence patterns from a practical perspective. Then, we determined the shape of the trajectories by testing the cubic terms of time for each trajectory first, indicating the use of cubic terms of time to model the probability that a patient within each trajectory would be adherent. To ensure the principle of parsimony, if the cubic and quadratic terms were not significant in a given model, subsequent analysis was performed with constant and linear terms.

 Once the optimal number and shape of the trajectories were chosen, we estimated the group membership probability for each subject and assigned each subject to the specific trajectory group with the highest group membership probability. Because the distribution of medication adherence was binary, we fit a logit model to examine the relationship between COC and medication adherence while taking into account the time-constant and time-varying covariates.

 presents the characteristics of the study subjects with newly diagnosed T2D in 2002. The mean MPR value was 75.18% and the COCI score was 0.39; the mean age was 55.62 years and males comprised 52.82% of the sample. Regarding health status, the rate of hospitalization in the previous year was 17.17%; 60.66% of the subjects had a DCSI score of 0 and 15.61% had a score of 2 or higher.

Using the group-based trajectory model, we identified 4 distinct medication adherence trajectories based on the BIC scores (

, the persistent adherence trajectory accounted for 39.9% of the subjects, and their medication adherence was constantly high. The increasing adherence trajectory accounted for 27.5% of the subjects, who showed an increasing trend in their adherence status over time, whereas the decreasing adherence trajectory included 12% of the subjects who showed a continuous declining trend in their medication adherence. The nonadherence trajectory accounted for 20.6% of the subjects, and this group consistently had the lowest probability of medication adherence.

presents the relationship between COC and the likelihood of medication adherence among various adherence trajectories. The COCI score estimate was positively and significantly associated with medication adherence in each trajectory group. This association implies that patients with high or medium COCI scores were more likely to adhere to medication use compared with patients with low COCI scores, regardless of their trajectory groups. In addition, older patients or female patients were less likely to be in the nonadherence trajectory group than in other groups. Patients living in the rural area were more likely to be in the nonadherence trajectory group than in other groups.

The objective of this study was to construct trajectories of medication adherence among patients with newly diagnosed T2D first, and then to examine the association between COC and medication adherence among the various adherence trajectories.

Previous studies that constructed medication trajectories were restricted to short-term follow-up. Riegel et al (2012) examined the trajectories of medication nonadherence in adults with heart failure over a 6-month period and identified 2 distinct trajectories,

 and Franklin et al (2013) described patients’ adherence to the lipid-lowering medication “statin” during the 15 months following the first prescription.

The present study identified 4 distinct medication adherence trajectories (persistent, increasing, decreasing, and nonadherence) during the 6-year follow-up period after the initial diagnosis of diabetes. We noticed that only 39.9% of the subjects were always adherent to medication and that 20.6% of the subjects were consistently nonadherent during these 6 years. We also found that older patients were less likely to be in the nonadherence trajectory group. The findings were consistent with previous studies which suggested that adherence increases with age.

 In addition, this study revealed that patients living in the rural area were more likely to be in the nonadherence trajectory. On the contrary, Egede et al (2011) reported that adherence to antidiabetic medication was higher among rural versus urban veterans; however, when accounting for the effect of race/ethnicity on medication adherence, they identified that rural-dwelling Hispanic veterans had lower medication adherence compared with their urban counterparts.

 The factors associated with the medication adherence patterns deserve further investigation.

Among the 4 medication adherence trajectories, we found that medication adherence was positively associated with the degree of COC. The findings of this study are consistent with those of previous studies

; however, Robels and Anderson (2011) found that a positive association between COC and adherence to antihypertensive drugs may exist only within a limited range of illness severity.

 This positive association between COC and medication adherence may be attributable to better information sharing and better goal alignment between patients and physicians, as suggested by Donaldson (2001).

Better information sharing implies that physicians are more familiar with the patients’ prognosis and medication use over time. Better alignment between goals and treatments may increase the interpersonal relationship between patients and physicians, which, in turn, increases the patients’ trust.

 Increased communication and trust may, subsequently, improve patients’ medication management by improving adherence to medications, for example. This finding suggests that COC is an important factor that influences medication adherence for patients with T2D, regardless of the type of medication adherence trajectory.

Strategies that aim to increase the COC between patients and physicians should be encouraged; for example, the introduction of the medical home project in the United States

 to enhance coordinated care for patients might improve the COC for project enrollees.

This study has several limitations. First, this study did not include unobserved (eg, health-seeking behavior) or unavailable (eg, illness severity) characteristics in the regression models. We included several proxy indicators for illness severity in the regression model, such as the DCSI score, which might reduce the bias related to the confounders that were not incorporated into the models. Second, MPR only indicates medication possession rather than medication consumption, which may lead to an overestimation of medication adherence. Nevertheless, MPR has been commonly used in previous studies to quantify medication adherence and is a valid predictor of healthcare outcomes.

 Third, patients without medication prescription at the first time of diabetes diagnosis and patients who received insulin treatment in any of the years during the study period were excluded from the analysis, which might lessen the representativeness of the study subjects. Finally, there are certain unique aspects of Taiwan’s healthcare system, such as free choice of providers for every visit without referral requirements, which may limit the generalizability of the findings to other healthcare systems.

This study demonstrated the heterogeneity in medication adherence and identified 4 distinct trajectories of medication adherence for patients with newly diagnosed T2D. Better COC between patients and their physicians was consistently associated with higher medication adherence in the 4 trajectory adherence groups. More attention should be directed toward designing new strategies to improve COC after the diagnosis of diabetes.

 Department of Public Health, College of Medicine, Fu Jen Catholic University (CCC), New Taipei City, Taiwan; Institute of Health Policy and Management, College of Public Health, National Taiwan University (SHC), Taipei City, Taiwan.

 The study was supported by the grants from the Ministry of Science and Technology (MOST 103-2410-H-002-202-MY3; MOST 104-2410-H-030-087-MY2) and the National Health Research Institute (NHRI-EX104-10225PI) in Taiwan. The funding source had no role in the study.

 The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

POLICY

This study identified 4 distinct medication adherence trajectories and found that continuity of care was consistently associated with medication adherence in every trajectory group.

Continuity of Care and Changes in Medication Adherence Among Patients With Newly Diagnosed Diabetes

To describe adherence with United States Preventive Services Task Force (USPSTF) colorectal cancer (CRC) screening recommendations over a 10-year period in a large, continuously insured screening population at average risk for CRC.

Insured members (N = 151,638) who turned 50 years old between January 1, 2000, and December 31, 2004, and were at average risk for CRC were included in the analysis. Subjects were categorized as adherent, inadequately screened, or screening-naïve based on their level of adherence with USPSTF CRC screening guidelines. Outcomes considered were age at initial CRC screening and CRC screening tests received over the 10-year period.

Of the 151,638 subjects in the cohort, only 97,518 (64%) were adherent with current CRC screening recommendations. An additional 18,050 (12%) were considered inadequately screened and 36,070 (24%) were screening-naïve. In those subjects who received some form of CRC screening, the average age at screening initiation was 53 years—3 years past the age recommended by current guidelines. Of those subjects who were inadequately screened, nearly half (46%) received only 1 fecal occult blood or fecal immunochemical test over the 10-year period.

 In a sample of continuously insured average-risk individuals aged 50 to 54 years, CRC screening was initiated later and performed less frequently than recommended in USPSTF guidelines.

		Longitudinal adherence to current United States Preventive Services Task Force colorectal cancer (CRC) screening recommendations was evaluated in a large population of continuously insured individuals at average risk for CRC. 

		Only 64% of subjects were adherent to current CRC screening recommendations. 

		Average age at screening initiation with any test was 3 years past that recommended by guidelines.

Colorectal cancer (CRC) is the third most commonly diagnosed cancer and the second leading cause of all cancer-related deaths in the United States.

 Compared with other cancer types, CRC has the third-highest global economic impact after lung and breast cancers.

 In the United States, the annual cost for CRC was estimated at $14 billion in 2010, and this is expected to increase as the population ages.

CRC places a substantial burden on healthcare systems, primarily due to detection in later stages. Forty percent of US patients diagnosed with CRC between 2004 and 2010 had localized disease, for which the 5-year relative survival rate was 89.8%, while 20% of patients had metastatic CRC, which has a 5-year relative survival rate of only 12.9%.

 The burden of CRC is largely preventable through early detection and removal of abnormal growths before symptoms develop.

 Furthermore, CRC treatment costs are lower when patients are diagnosed in earlier stages of disease.

Evidence-based CRC screening guidelines developed by the United States Preventive Services Task Force (USPSTF) in 2008 recommend that adults at average risk for CRC begin screening at age 50 and continue until age 75 with either high-sensitivity fecal occult blood test (FOBT) or fecal immunochemical test (FIT) annually, flexible sigmoidoscopy (FS) every 5 years with a high-sensitivity FOBT every 3 years, or colonoscopy every 10 years.

 USPSTF guidelines do not recommend any of the above strategies over the others, but direct patients to use a routine CRC-screening strategy. These guidelines are aligned with recommendations by other US professional organizations, such as the Institute for Clinical Systems Improvement

Although it is widely accepted that early detection through CRC screening provides the best potential to reduce the clinical and economic burdens of CRC,

 the effectiveness of CRC screening is limited by low uptake and adherence. In fact, 1 in 3 eligible individuals are not appropriately screened for CRC.

 Despite the fact that CRC screening rates have slowly and steadily increased over the last few decades in the United States, only 65.1% of adults had up-to-date screening—well below the national target of 80% by 2018,

 and lower than for other screening-amenable cancers, such as breast (72.4%) and cervical (83%).

While several studies have assessed CRC screening behavior,

 very few US population-based studies have evaluated longitudinal adherence to CRC screening guidelines.

 Data from these longitudinal studies are less recent and may not reflect current adherence to CRC screening guidelines. In addition, some studies have used shorter time frames, which could lead to an underestimation of CRC screening adherence, especially for individuals choosing colonoscopy every 10 years as their periodic screening strategy.

The objective of this analysis was to describe the longitudinal adherence with USPSTF-recommended CRC screening over a 10-year period in an average-risk CRC screening population continuously enrolled in commercial health plans. We suspected that a longer study time frame would not reveal improved longitudinal adherence to CRC screening guidelines.

This was a retrospective analysis of patient-level data from a large, national US administrative health claims database (Clinformatics DataMart, affiliated with Optum) containing records for medical, pharmacy, and laboratory services. The database contains more than 111 million unique commercial and Medicare members, with 12.1 million members having 5 or more years of continuous enrollment in 2008-2012, and it is geographically diverse across the United States, with greatest representation in the south and midwest US Census regions. Medical (professional, facility) claims include 

International Classification of Diseases, 9th Revision, Clinical Modification

) diagnosis codes and procedure codes; Current Procedural Terminology, version 4, procedure codes; Healthcare Common Procedure Coding System procedure codes; site of service codes; provider specialty codes; and health plan and patient costs. All data were de-identified and accessed in compliance with the Health Insurance Portability and Accountability Act (HIPAA). No identifiable protected health information was extracted or accessed during the course of the study and the database used in this study has been certified as de-identified. Pursuant to HIPAA, the use of de-identified data does not require institutional review board approval or waiver of authorization.

All individuals who turned 50 years old between January 1, 2000, and December 31, 2004, and had continuous medical eligibility for at least 10 years, beginning with the year that they turned 50, were included and tracked over this 10-year period. Individuals who had claims for CRC screening prior to the year they turned 50 were excluded, including those with any claims for colonoscopy (screening or diagnostic), FS, FIT or FOBT, double-contrast barium enema (DCBE), or computed tomographic colonography (CTC) tests. Additionally, individuals with evidence of high-risk history were excluded, including those with any of the following: a diagnosis of cancer prior to the index CRC screening test (

 diagnosis codes: 153.x,154.0, 230.3, 230.4, 230.7, 211.3, 211.4, 211.9, 235.2, 239.0), a personal history of colorectal polyps (V12.72), a personal history of cancer (V10.9), a family history of gastrointestinal cancer (V16.0), Crohn’s disease (555.1, 555.9), or ulcerative colitis (556).

Included subjects were classified as adherent, inadequately screened, or screening-naïve, based on alignment with current USPSTF screening guidelines.

 Subjects categorized as adherent had at least 1 colonoscopy claim, or at least 2 FS claims, or 1 FS and 5 annual FIT/FOBT test claims (FS and FIT/FOBT could occur in the same year), or 1 FIT/FOBT test claim per year over the 10-year observation period. Subjects categorized as inadequately screened had only 1 FS claim or fewer than 10 annual FIT/FOBT test claims over the 10-year observation period. Screening-naïve subjects had no claims for a CRC screening test over the 10-year observation period. Receipt of DCBE or CTC was not considered in this study, as these tests are not recommended by current USPSTF guidelines.

The number and proportion of subjects 50 years or older who were adherent with CRC screening guidelines were assessed. Additionally, the average age at initiation of CRC screening and the types of CRC screening tests received were assessed. Within the adherent subgroup, the most common screening strategy received and the age at first colonoscopy were assessed. In inadequately screened subjects, the average number of FIT/FOBT tests received, and the utilization of FIT/FOBT and FS over the 10-year period were determined.

Subject characteristics were analyzed descriptively using 

 tests for dichotomous variables. Descriptive statistics (mean, median) were computed for all outcomes. The data analysis for this paper was generated using SAS version 9.2 (SAS Institute, Cary, North Carolina).

As depicted in the Figure, of the roughly 5.8 million enrollees who turned 50 years old between January 1, 2000, and December 31, 2004, 265,796 had continuous medical eligibility for at least 10 years, beginning with the year that they turned 50. Of these, 151,638 were eligible for inclusion in the cohort (ie, they were neither high-risk nor screened prior to age 50).

Subject characteristics for the entire cohort and by adherence category are presented in 

. In the overall cohort, most subjects were commercially insured (98.9%) and only a small number of patients (1.1%) aged less than 61 years were insured through Medicare for disability reasons. Just under half of all subjects were male (47.4%) and the majority of subjects lived in southern (38.8%) or midwestern states (30.5%). For 45% of subjects, race was unknown; for the remainder, 45% were Caucasian, 4.9% African American, 3.7% Hispanic, and 1.4% Asian. Over the 10-year observation period, 97,518 (64.3%) subjects were adherent to USPSTF CRC screening recommendations; 18,050 (11.9%) were screened, but not adherent to screening recommendations (inadequately screened); and 36,070 (23.8%) were screening-naïve. There was no significant difference between the proportion of males and females who were adherent (64.5% vs 64.2%; 

 = .22); however, a significantly greater proportion of females than males were inadequately screened (14.3% vs 9.2%; 

 <.001) and a significantly greater proportion of males than females were screening-naïve (26.3% vs 21.5%; 

Subjects could be considered adherent based on more than 1 screening strategy. Of the 97,518 adherent individuals, 97,081 (99.6%) were adherent because they received at least 1 colonoscopy; 1946 (2%) were because they received at least 2 FS tests; 614 (0.6%) were because they received at least 1 FS test and at least 5 years of FIT/FOBT tests; and 268 (0.3%) were because they received at least 1 FIT/FOBT test per year over the 10-year observation period.

Of the 18,050 inadequately screened individuals, 773 (4.3%) were inadequately screened because they received only 1 FS test and no annual FIT/FOBT tests, 371 (2.0%) because they received 1 FS test and between 1 and 4 annual FIT/FOBT tests, and 16,906 (93.7%) because they received no FS test and fewer than 10 annual FIT/FOBT tests over the 10-year study period. Nearly half (8330 of 18,050; 46%) of inadequately screened subjects received just 1 annual FIT/FOBT test (with or without a FS test) over the 10-year period. Inadequately screened subjects received a mean of 2.6 annual FIT/FOBT tests over the 10-year study period.

 shows the distribution of ages when subjects in the cohort received their first CRC screening test. The mean and median age at the time of first CRC screening test were 53 and 52 years, respectively, for both the adherent and inadequately screened subgroups. Almost half (47,085 of 97,081; 49%) of subjects who were adherent because they received at least 1 colonoscopy had their first colonoscopy after age 53 (

In this retrospective analysis of patient-level data from a large, national US administrative health claims database, we found that just under two-thirds of subjects (64.3%) were adherent to USPSTF CRC screening recommendations, primarily due to a single screening colonoscopy.

Our results are in line with current national survey data.

 Of those aged 50 to 75 years who responded to the 2012 Behavioral Risk Factor Surveillance System telephone survey, 65.1% reported being up-to-date with 1 of the CRC screening tests recommended by USPSTF.

 Up-to-date was defined as FOBT within the past year, FS within the previous 5 years, and FOBT within the previous 3 years, or colonoscopy within the previous 10 years. However, these self-reported data are likely to overestimate true adherence to CRC screening guidelines, and it may not be possible to generalize these results to the entire US population. Other national data show similar levels of adherence. National Health Information Survey data show that 58.6% of adults are up-to-date with 1 of the USPSTF-recommended CRC screening methods.

 Our database analysis aligns with results from national surveys and provides further evidence that CRC screening levels remain below the national target of 80%.

 A strength of this study is that it examined claims data from a large, insured population over a 10-year observation period rather than on less reliable self-reported data. The result that adherent subjects were primarily adherent because of screening colonoscopy is unsurprising as individuals are more likely to adhere to screening options that require fewer screenings over time.

Nearly half of inadequately screened subjects in our sample (46%) received only 1 FIT/FOBT test over the 10-year period. This result is consistent with previous studies.

 Fenton et al examined adherence with biennial (every other year) FOBT screening recommendations in over 11,000 Washington state health plan members aged 52 to 78 years who were previously screened with FOBT.

 During the subsequent 2-year observation period (2002-2003), only 44.4% adhered to repeat biennial FOBT screening.

 Gellad et al followed patient records forward 5 years to evaluate adherence to repeat annual FOBT testing through a large retrospective cohort study of 1 million patients aged 50 to 75 years who were seen at Veterans Affairs medical centers across the United States.

 Similar to our study results, very few subjects were adherent with annual FOBT testing. Only 14.1% of men and 13.7% of women completed at least 4 FOBTs over the 5-year study period. This demonstrates that good programmatic adherence to annual screening with FIT/FOBT may be difficult, based on retrospective analysis. However, a prospective population-based CRC screening study conducted by van Roon et al

 found participation rates that ranged between 63.2% and 64% for 7501 Dutch individuals sent FIT screening tests at 1-, 2-, or 3-year intervals. One reason for the higher adherence rates may be the interaction that took place with individuals through study invitations and follow-up mailings.

Even more concerning than the low adherence rates demonstrated in this study is that about a quarter of insured subjects (24%) received no screening at all. Survey studies have demonstrated that having health insurance or higher income is correlated with a greater likelihood of adhering to CRC screening guidelines.

 Claims data cannot provide insight into the reasons that these subjects did not complete a single CRC screening test over the 10-year period. Therefore, future studies should attempt to answer this question.

Studies have suggested that acceptance and completion of certain CRC screening test options may vary by race or ethnicity.

 Unfortunately, race was not known for 45% of subjects in our analysis, making it difficult to understand finer differences in adherence by race. However, similar to national survey data, we also found adherence to CRC screening was significantly lower for Hispanic subjects compared with all other ethnic groups combined (

 <.001). Bazargan et al triangulated data from a prospective survey, medical charts, and clinic logbooks for 306 underserved minority (primarily Hispanic) patients 50 years or older seen at a large outpatient primary care clinic in Los Angeles to examine correlates of physician recommendation, scheduling, and actual completion of CRC screening. Only 13.7% and 11.4% of participants had sigmoidoscopy and colonoscopy, respectively,

 per National Comprehensive Cancer Network guidelines. Further, in the population assessed by Bazargan and colleagues, only 44% and 59% of scheduled sigmoidoscopy and colonoscopy procedures were completed, respectively.

 This study, as well as others, suggests that invasive screening tests may be a barrier to screening for certain population subgroups.

Another important finding from our study is that about three-fourths (74.7%) of subjects who received a CRC screening test did not receive their first test at age 50. The average age at initial CRC screening was 53 years of age—3 years past that recommended by current guidelines.

 This is consistent with a study by Sinicrope et al that analyzed medical record data for approximately 21,000 average-risk individuals 50 years or older residing in Olmsted County, Minnesota, to explore longitudinal trends in CRC screening between 1997 and 2008.

 Individuals were categorized as current-, ever-, and never-screeners according to their level of adherence to American Cancer Society CRC screening guidelines.

 They found that although CRC screening tripled over the study period, individuals were delaying initiating CRC screening by up to 5 years.

CRC screening is known to be lower in certain population subgroups, such as people with low income or low educational attainment, recent immigrants (under 10 years), and those without health insurance, a usual source of care, or recent physician contact.

 Our long-term study found that even in a continuously commercially insured screening population, longitudinal adherence to CRC screening guidelines was less than ideal. Longitudinal adherence to CRC screening may be improved with the availability of a test that is noninvasive,

 requires less frequent testing, is widely accessible,

 and is accompanied by provider and patient support.

 Alternatively, the success of CRC screening may depend more on patient education, outreach, and monitoring than the type of test used. A recent review of population-based CRC screening programs

 found 4 factors that played a role in the success of screening programs, independent of the type of screening test used: 1) having a dedicated funding source for the program, 2) having a structured policy in place for responding to positive screening results, 3) having an outreach program to educate and increase screening rates, and 4) having a plan in place for rescreening. These 4 factors may be challenging for most community clinics to implement because of the costs involved, but may not be challenging for the larger managed care organizations.

The results of this study must be interpreted with appropriate consideration of the limitations inherent in all retrospective database analyses. Since claims data are collected primarily for payment purposes, coding errors may occur. Additionally, the presence of a screening test code on a claim is not positive confirmation that the test was actually performed. This is especially true for FOBT/FIT tests performed at home. Claims also do not provide the results of screening tests, so adherence to follow-up testing—for instance, measuring adherence to diagnostic colonoscopy in subjects with positive FOBT or FS tests&mdash;is not possible. Additionally, claims data do not provide information about whether or not programmatic/educational interventions may have been used to increase patient participation and compliance, such as a mailed invitation or reminder about completing a test.

 Furthermore, the study design does not allow for causal inferences of results to know what recommendations subjects were given by their physicians or to ascertain the reasons that subjects did not adhere to their CRC screening regimen. Lastly, the sample includes only continuously commercially insured individuals aged 50 to 54 years at average risk of CRC, and so the results may not be representative of the general population.

In a sample of continuously insured individuals aged 50 to 54 years at average risk for CRC, screening was initiated later and performed less frequently than recommended in guidelines. Only 64.3% of patients were adherent to USPSTF CRC screening recommendations and the average age at screening was 3 years later than recommended.  These results suggest a significant gap in care that should be addressed to improve early detection rates and, ultimately, reduce the clinical and economic burden of CRC.

The authors would like to thank Blake Krebs for analytical support and Michael Maschio, Virginia Rosen, and Sarah Peirce-Sandner for reviewing the manuscript.

CLINICAL

Over a 10-year time frame, longitudinal adherence to current colorectal cancer (CRC) screening guidelines was less than ideal in a large, continuously insured US population at average risk for CRC.

Longitudinal Adherence to Colorectal Cancer Screening Guidelines

Policy makers have growing interest in price transparency and in the kinds of tools available to consumers. Health plans have implemented price estimator tools that make provider pricing information available to members; however, systematic data on prevalence and characteristics of such tools are limited. The purpose of this study was to describe the characteristics of price estimator tools offered by health plans to their members and to identify potential trends, challenges, and opportunities for advancing the utility of these tools.

Between 2014 and 2015, we conducted a national Web-based survey of health plans with commercial enrollment (100 plans, 43% response rate). Descriptive analyses were conducted using survey data. 

Health plan members have access to a variety of price estimator tool capabilities for commonly used procedures. These tools take into account member characteristics, including member zip code and benefit design. Despite outreach to members, however, challenges remain with respect to member uptake of such tools. 

Our study found that health plans share price and provider performance data with their members. 

	We conducted a national survey of health plans to examine characteristics of provider price estimator tools available to the commercially insured population. Health plans offer these tools to their enrollees and take into account member characteristics such as zip code and benefit design. 

		Price estimates are available for commonly used services, including outpatient surgery and radiology. 

		Plans use various approaches, including digital and social marketing, to increase member use of tools. 

		Early evidence shows the positive impact of such tools on cost savings. Additional research is needed to increase the use of these tools and broaden their impact.

In an effort to slow the rising costs of healthcare, consumer groups, state governments, and health plans have called for greater price transparency in the healthcare market.

A national survey of health plans shows that price estimator tools available to commercial plan enrollees provide price information for a variety of services at the provider level.

February 2016

The Promise and Perils of Big Data in Healthcare

Continuity of Care and Changes in Medication Adherence Among Patients With Newly Diagnosed Diabetes

Longitudinal Adherence to Colorectal Cancer Screening Guidelines

Characterizing Health Plan Price Estimator Tools: Findings From a National Survey

The Value of Value-Based Insurance Design: Savings From Eliminating Drug Co-Payments

USPSTF Colorectal Cancer Screening Guidelines: An Extended Look at Multi-Year Interval Testing

Molecular Testing Patterns in Metastatic Non—Small Cell Lung Cancer

Inappropriate Ordering of Lumbar Spine Magnetic Resonance Imaging: Are Providers Choosing Wisely?

The Effect of Medicare Advantage Enrollment on Mammographic Screening

Case-Mix Adjustment for Diabetes Indicators: A Systematic Review