Currently Viewing:
The American Journal of Managed Care October 2014
Quality of Care at Retail Clinics for 3 Common Conditions
William H. Shrank, MD, MSHS; Alexis A. Krumme, MS; Angela Y. Tong, MS; Claire M. Spettell, PhD; Olga S. Matlin, PhD; Andrew Sussman, MD; Troyen A. Brennan, MD, JD; and Niteesh K. Choudhry, MD, PhD
A Comprehensive Hospital-Based Intervention to Reduce Readmissions for Chronically Ill Patients: A Randomized Controlled Trial
Ariel Linden, DrPH; and Susan W. Butterworth, PhD
Physician Compensation Strategies and Quality of Care for Medicare Beneficiaries
Bruce E. Landon, MD, MBA; A. James O'Malley, PhD; M. Richard McKellar, BA; James D. Reschovsky, PhD; and Jack Hadley, PhD
Increasing Access to Specialty Care: Patient Discharges From a Gastroenterology Clinic
Delphine S. Tuot, MDCM, MAS; Justin L. Sewell, MD, MPH; Lukejohn Day, MD; Kiren Leeds, BA; and Alice Hm Chen, MD, MPH
Currently Reading
Increasing Preventive Health Services via Tailored Health Communications
Kathleen T. Durant, PhD; Jack Newsom, ScD; Elizabeth Rubin, MPA; Jan Berger, MD, MJ; and Glenn Pomerantz, MD
Evaluation of Collaborative Therapy Review to Improve Care of Heart Failure Patients
Harleen Singh, PharmD; Jessina C. McGregor, PhD; Sarah J. Nigro, PharmD; Amy Higginson, BS; and Greg C. Larsen, MD
Extending the 5Cs: The Health Plan Tobacco Cessation Index
Victor Olaolu Kolade, MD
Caregiver Presence and Patient Completion of a Transitional Care Intervention
Gary Epstein-Lubow, MD; Rosa R. Baier, MPH; Kristen Butterfield, MPH; Rebekah Gardner, MD; Elizabeth Babalola, BA; Eric A. Coleman, MD, MPH; and Stefan Gravenstein, MD, MPH
Ninety-Day Readmission Risks, Rates, and Costs After Common Vascular Surgeries
Eleftherios S. Xenos, MD, PhD; Jessica A. Lyden, BSc; Ryan L. Korosec, MBA, CPA; and Daniel L. Davenport, PhD
Using Electronic Health Record Clinical Decision Support Is Associated With Improved Quality of Care
Rebecca G. Mishuris, MD, MS; Jeffrey A. Linder, MD, MPH; David W. Bates, MD, MSc; and Asaf Bitton, MD, MPH
The Impact of Pay-for-Performance on Quality of Care for Minority Patients
Arnold M. Epstein, MD, MA; Ashish K. Jha, MD, MPH; and E. John Orav, PhD
Healthcare Utilization and Diabetes Management Programs: Indiana 2006-2010
Tilicia L. Mayo-Gamble, MA, MPH; and Hsien-Chang Lin, PhD
Predictors of High-Risk Prescribing Among Elderly Medicare Advantage Beneficiaries
Alicia L. Cooper, MPH, PhD; David D. Dore, PharmD, PhD; Lewis E. Kazis, ScD; Vincent Mor, PhD; and Amal N. Trivedi, MD, MPH

Increasing Preventive Health Services via Tailored Health Communications

Kathleen T. Durant, PhD; Jack Newsom, ScD; Elizabeth Rubin, MPA; Jan Berger, MD, MJ; and Glenn Pomerantz, MD
The authors describe and apply a methodology for defining tailored health communications in order to increase the number of completed colorectal cancer screenings.
To develop a methodology that stratifies members by likelihood of completing a colorectal cancer screening (CRCS). Such information can guide the communication development and the allocation of resources for tailored communication outreaches.

Study Design
Prospective study of an insured commercial population that includes randomized assignments to a control group or a treatment group for communication interventions.

A total of 46,697 members overdue (nonadherent based on available administrative data) for a CRCS were assigned to 1 of 5 graded segments as part of an interactive voice response (IVR) call. Another 400 members per outreach segment were randomly assigned to a control group and received no communication. Primary outcome: receipt of CRCS (assessed Healthcare Effectiveness Data and Information Set [HEDIS] summary data, 3 months after IVR) for each graded segment within the control and the intervention groups. Secondary outcomes: association between the CRCS rates and the identified segments; communication cost per completed screening.

Primary outcome: 1971 members completed the screening. Screening rates for the 5 graded segments were 2.4%, 3.6%, 5.0%, 7.2%, and 8.0%. Secondary outcome: c2 proportions test a significant association between the segment groups and the CRCS completion rates. The association between segment and screening rates follow the expected predicted range in both the intervention and holdout groups. Communication cost per screening: $14.84.

Segmenting members by variables that are associated with CRCS completion identified graded segments that completed the CRCSs at different rates. Tailored interventions can be developed to promote a health service for each segment. The segmentation approach proves to be beneficial.

Am J Manag Care. 2014;20(10):828-835
  • Readily available quantitative data derived from claims, demographic attributes, and past Healthcare Effectiveness Data and Information Set activity can be used to segment a population based on their likelihood to fulfill a particular health service.
  • The segmentation scheme can guide the development of a predictive outreach that supports the personalization of communications in order to improve the quality of healthcare services.
  • This communication model can be used to improve the allocation of communications-oriented resources.
In recent years, the focus on healthcare quality has increased, but insurers face simultaneous pressures to control costs. Efforts to improve quality need to be directed in a measurable and cost-effective manner. This paper describes a method for developing health communication outreaches that addresses healthcare quality in a resource-limited scenario.

The US Preventive Services Task Force (USPSTF) reviews the evidence of the effectiveness of clinical preventive measures and develops recommendations for clinical preventive services.1 These clinical recommendations are incorporated into the definition of the Healthcare Effectiveness Data and Information Set (HEDIS) measures. Implementing these recommendations has been shown to prevent and detect disease early, when it is most treatable and curable.1

The USPSTF’s current recommendation for colorectal cancer screening (CRCS) is for men and women of average risk to begin screening at age 50 years and continue through age 75 years. The screenings can be implemented as an annual high-sensitivity fecal occult blood test (FOBT); a colonoscopy every 10 years; or a flexible sigmoidoscopy every 5 years along with a high-sensitivity FOBT every 3 years.1 Early detection of high-risk lesions, polyps, or CRC itself through screenings is associated with decreased incidence of, and mortality from, CRC.

Despite the proven benefits of CRCS, only 58.6% of the targeted US population completes a CRCS.2 This is well below the national target of 70.5%,2 and may explain why it is the third-most common cancer in the United States, as well as the second-leading cause of cancer mortality.3 Total direct cost associated with treating CRC is $17 billion annually.3,4

Historically, physicians communicated about preventive care, such as CRCS, to the patient. More recently, as health plans increasingly take on risk associated with quality and outcomes, plans have augmented this physician-led activity with their own communications, since health communications have been shown to be beneficial insofar as they increase the audience’s knowledge and support of the topic, refuting misconceptions, and ultimately leading to the desired behavior.5

Typically, however, insurers’ communication outreaches are homogeneous. In fact, heterogeneity will increase the success of these outreaches. It is important to: design a tailored outreach that will resonate with the recipient (member); choose a channel the member is likely to use; deliver the message when the member is most receptive to it; and allot ample time for completion before the end of the year. This paper describes a methodology that allows development and measurement of such a communication outreach. The importance of this methodology is not only in the CRCS completion rates, but also in the proper allocation of funds for plans with scarce resources.


The study was reviewed by the Horizon Blue Cross Blue Shield of New Jersey (BCBSNJ) privacy official and business owner, both whom deemed the study to fall under “healthcare operations,” not “research,” meaning it did not require approval by the Internal Review Board.

The Figure displays the process flow for the study as well as the different populations (segmented population vs not segmented population) and the treatment groups (intervention group vs control group).

Decision Tree Creation

We gathered health, healthcare utilization, and demographic variables from 2010 and 2011 claims data for 97,821 members of an insurance plan from Horizon BCBSNJ eligible for CRCS. A majority of the decision tree variables were identified as relevant population characteristics of people completing CRCS in prior research.6-37 As suggested by prior research, we included a member’s past health behavior to predict fulfillment of a healthcare service in the future.38-41

We used a modification of IBM Statistical Package for the Social Sciences' (SPSS) χ2 automatic interaction detection growing algorithm (CHAID SPSS Predictive Analytic Software (PASW) statistics version 18)42 to build decision trees that identify the most relevant variables in CRCS completion. Decision trees separate a collection of heterogeneous records into smaller groups of homogeneous records using a directed knowledge discovery. During each step of the tree-growing process, the CHAID algorithm chooses the independent variable that has the strongest interaction with the dependent variable for the population specific to a tree leaf node.42 We considered other machine-learning techniques, but chose to use decision trees since their nonlinear nature allows tree induction to exploit structures that identify subpopulations discerned by the tree variables.42

Given the distinct variables associated with gender, we chose to develop separate models for men (45,678) and women (52,143). We used 2 different validation techniques: 10-fold cross-validation and a validation set consisting of 33% of the members; the validation methods verify the chosen variables, as well as the final results from the model.

Trees to Segments

To segment members, we split the distribution of the predicted class probability of the outcome variable (CRCS completion = 1). The predicted probability is determined for each leaf node during the model creation step and is defined as the proportion of cases in each leaf node that complete the screening. We use the class probability values to define the segments, since previous studies have shown that decision trees are known for estimating accurate probability-based rankings.43 The segment cut points within the probability distribution maximize the difference in probability values among the segments, as well as create segments that are large enough to be actionable. The goal is to make the members of a segment as homogeneous as possible while making the segments heterogeneous.

Outreach Definition

The outreach was defined as 1 interactive voice response (IVR) call with varying messaging. Depending on the number of nonadherent members and the segments’ health descriptions, an outreach segment may contain 1 or more model segments. Table 1 shows the percentage of male and female populations assigned to each outreach segment.

Power Analysis

A power analysis was performed before the launch of the intervention to determine the minimal size needed for each segment, given an estimated effect size of 2% increase for each graded segment and an α level = .05. Given that the sizes of the segmented groups were determined by the insurer’s nonadherent population and not via a recruitment method, it was determined at the launch of the communication that the comparison of the completion rate of segment 4 and segment 5 was underpowered given the segment sizes and the estimated effect size.

Outcomes Analysis

We measured the predictive nature of the methodology by measuring the CRCS completion rates and calculating the 95% CI for each of the 5 segments for the intent-to-treat population and the 2 groups (holdout group and treatment group). The study's segmented population is defined as the segmented population that were targeted to receive either the general intervention, the tailored intervention or are part of the control group. We also compare the results of the control group to the reached intervention group. The control group is defined as the segmented population that was targeted to not receive the IVR call. The reached intervention group are members that actually received the intervention (the IVR content). These members were reached by telephone and may have received either the targeted content or the general content. To measure the stability of the segments, we compared each segment’s completion rates for the 2 different treatment groups. Given the different proportions of the graded segments within the holdout and intervention groups, we did not compare the completion rates of the 2 groups; we hypothesized that the holdout group would have a higher completion rate since it has a higher proportion of members from the higher-ranked segments. A demographic comparison of the 2 groups’ segments can be found in the eAppendix tables.


1971 members completed the screening with 1873 screening completions from members assigned an outreach segment (1773 from the intervention group and 100 from the control group) and 98 screening completions from members not assigned a segment due to a lack of claims data for the calendar years 2010 and 2011. Segments predicted to have a higher CRCS completion rate did follow the predicted range in both groups. The range of segment completion rates for the study's segmented population, from lowest to highest probability, were 2.42%, 3.64%, 4.83%, 7.06%, and 8.08%.

Decision Tree Variables

For females, other cancer screenings such as breast cancer screening, as well as a personal or family history of colon cancer, are the most predictive variables and make up the topmost levels of the decision tree. These variables have been previously identified as associated with CRCS.7,10,13,18,26,28,29,37 Another variable in the tree, but not previously identified, is total claims spending. Other relevant variables included having an obstetrics/gynecological visit, completing a cervical cancer screening,12,13,26 completing a physical exam,17,18,26 the number of doctor visits,14,17-19,22,30,35 and a radiology exam. A member’s age and being coded as having unexplained abdominal pain or musculoskeletal chronic pain are also included within the female decision tree.

For men, who, as specified by the old HEDIS specification, do not have other cancer-related HEDIS measures, the most predictive variable is claims spending; this variable has not been identified as associated with CRCS in previous studies. Other variables such as a personal or family history of CRC and the number of doctor visits for the prior 2 calendar years (2010, 2011) are variables that make up the topmost levels of the decision tree. These variables have been identified as associated with CRCS in previous studies.7,10,17,19,20,29,30,37 Having a physical exam17,18,26 and a variable indicating a household’s net worth7,10,17,30,37 are other variables within the male decision tree. Household net worth is typically correlated to income and socioeconomic class, variables identified as associated with CRCS completions in previous studies.10,17,30,37 Other variables within the decision tree are the number of emergency department (ED) visits and unexplained abdominal pain.

Predictability of the CHAID Trees

Table 2 lists the area under the receiver operating characteristics (ROC) curve (AUC) for the decision trees. We list the percentage of the majority class as an indication of the predictive power of a model. AUC in the range presented are consistent with those associated with other models used to predict healthcare behavior, notably models designed to identify members at risk of hospital readmission.44

Segment Definition

Copyright AJMC 2006-2020 Clinical Care Targeted Communications Group, LLC. All Rights Reserved.
Welcome the the new and improved, the premier managed market network. Tell us about yourself so that we can serve you better.
Sign Up