Impact of a Telephonic Intervention to Improve Diabetes Control on Health Care Utilization and Cost for Adults in South Bronx, New York

Study Setting

The Bronx A1C (6) was a randomized controlled trial comparing the impact of a diabetes self-management print-only intervention to a telephonic self-management support intervention on diabetes control (www.clinicaltrials.gov/show/NCT00797888). The target population was adult individuals with suboptimally controlled diabetes in the South Bronx, NY, where the population is predominantly Latino or African American and poverty is highly prevalent (11,12).

The Bronx A1C methods and results were described in detail elsewhere (6,11). Briefly, 941 adults >18 years with diabetes and a recent A1C >7% (53 mmol/mol) residing in the South Bronx ZIP codes were recruited into the study from September 2008 to October 2010. A total of 443 participants were randomized to the telephonic diabetes intervention (Tele/Pr) arm and 498 were randomized to the print-only (PrO) control arm. All participants were mailed the same high-quality diabetes self-management print materials at baseline randomization and modest lifestyle incentives quarterly. The Tele/Pr arm participants′ telephone calls followed a curriculum developed for this study. They were delivered individually in Spanish or English by a health educator who was primarily supervised by a nurse who was a certified diabetes educator. Additionally, the Tele/Pr arm received up to four calls from health educators over 1 year if participant’s baseline A1C was 7–9%, (53–75 mmol/mol) or up to eight calls if baseline A1C was >9% (>75 mmol/mol). A total of 910 participants finished the intervention (15 participants withdrew and 16 died). Postintervention A1C results (available for 694 participants) showed that from baseline to follow-up, mean A1C decreased by 0.9% among the Tele/Pr arm compared with 0.5% among the PrO arm, which is a 0.4% difference (P = 0.01). The intervention had a significant effect when baseline A1C was >9% (>75 mmol/mol) but not when baseline A1C was between 7% (53 mmol/mol) and 9% (75 mmol/mol). In addition, the results of a cost analysis indicated that the intervention was delivered at moderate cost relative to the gains achieved (13). Total costs for the PrO and Tele/Pr arms were $109.18 and $269.79 per person, respectively. The cost specific to the telephone intervention was $187.61 per person (13).

Data Source

For this secondary analysis of Bronx A1C, we used data from the Statewide Planning and Research Cooperative System (SPARCS), a comprehensive New York State claims database, to measure hospital inpatient health care utilization and costs and emergency department (ED) use for Bronx A1C participants (14,15). We also used information from the New York City (NYC) A1C Registry (16) to confirm the participants’ NYC residency. The NYC A1C Registry is a laboratory-based surveillance system that was created in 2006 following an amendment to the NYC Health Code mandating laboratories to report A1C test results for NYC residents to the NYC Department of Health and Mental Hygiene.

Study Population

Health care utilization was assessed by matching SPARCS data to the Bronx A1C participants using unique patient identifiers. Hospital and ED use were calculated for each individual participant in the Tele/Pr and PrO arms from 2006, which is up to 2 years prior to study enrollment (i.e., preintervention utilization), to 2014, which is up to 4 years after study enrollment (i.e., follow-up utilization). Data were analyzed on an intent-to-treat basis, meaning all individuals who participated in the Bronx A1C and were NYC residents during both the preintervention and follow-up study periods were included, regardless of the effect of the intervention or the availability of A1C outcome data. A total of 816 of the 910 who completed the Bronx A1C intervention had data within the A1C Registry and/or SPARCS inpatient database during the study period. Similarly, 839 participants had data within the A1C Registry and/or SPARCS ED database during the study period and were included in the study. The median (interquartile range [IQR]) times during the 2-year preintervention period were 18.8 (11.1–22.1) and 19.6 (12.4–22.3) months for the Tele/Pr and PrO arms, respectively (P = 0.24). The median (IQR) follow-up times during the 4-year follow-up period were 33.9 (28.6–35.6) and 33.1 (28.9–35.7) months for the Tele/Pr and PrO arms, respectively (P = 0.99).

Study Measures

We evaluated hospitalizations and ED services using the ICD-9 Clinical Modification (ICD-9-CM) codes included in primary or secondary (up to 24) fields in SPARCS. Changes in inpatient health care utilization and cost were measured for overall all-cause hospitalizations and diabetes-related complications. These were defined as 1) short-term diabetes complications, including hyperglycemia (diabetes with ketoacidosis, diabetes with hyperosmolarity, and diabetes with other coma) and hypoglycemia (diabetes with other specified manifestations) (17); 2) long-term diabetes complications including microvascular complications (diabetic nephropathy, diabetic retinopathy, and diabetic neuropathy) and macrovascular conditions (diabetes with peripheral circulatory disorders, acute myocardial infarction, and stroke) (17,18); and 3) lower-extremity amputations, which included amputations of the upper leg, lower leg, ankle, foot, and toe (18). Each of these outcomes was considered to be related to diabetes if any of the listed diagnoses included diabetes (ICD-9-CM: 250) (18). Any change in ED health care utilization was measured overall, including the use of any ED services, regardless of cause. The SPARCS ED data used in this study are “treat and release,” meaning visits that do not result in a hospitalization. Inpatient costs were estimated by multiplying reported charges in SPARCS by a year- and hospital-specific cost-to-charge ratio (CCR) and diagnosis-specific adjustment factors, which were created by the Healthcare Cost and Utilization Project in order to account for between- and within-hospital variation in the markup of charges (19–22). The CCRs were created by comparing nationally reported charges to hospital account reports collected by the Centers for Medicare and Medicaid Services, and they were linked with SPARCS using hospital identifiers available from the American Hospital Association Linkage Files. Diagnosis-specific adjustment factors were based on the reported diagnosis-related group code, which groups hospitalizations with similar clinical and demographic characteristics. Finally, costs were adjusted for inflation to 2014 dollars using the Consumer Price Index for Medical Care from the Bureau of Labor Statistics. Since CCRs do not exist for ED services, we were unable to estimate the costs for these services.

Statistical Analysis

We compared characteristics of the two study arms using two-tailed t tests and Wilcoxon rank sum tests for parametric and nonparametric continuous variables, respectively, and χ² and Cochran-Mantel-Haenszel χ² tests for categorical variables. We used the negative binomial distribution to account for the highly skewed, over-dispersed nature of the data, which is a common issue in health care utilization and cost studies (23,24). In addition, as expected, health care utilization and cost outcomes were related and were highly and significantly correlated. For example, for all-cause hospitalization, correlation coefficients were as follows: number of hospital stays versus number of hospital days = 0.79 (P < 0.001); number of hospital stays versus hospital cost = 0.73 (P < 0.001); number of hospital days versus hospital cost = 0.86 (P < 0.001). Hence, the joint mixed modeling approach was appropriate because it allowed for the correlation structure of different related outcomes (25,26).

Models of health care utilization and cost with joint binary outcome distribution and logit link function (hospital/ED use: yes, no), count (number of hospital stays/ED visits, number of hospital days), and semicontinuous (cost) outcomes with negative binomial distribution and log link function were estimated using repeated-measures generalized linear mixed models to account for potential correlations between related outcomes and serial correlations between repeated events (e.g., hospitalizations) within individuals. Separate models were developed for the inpatient health care utilization and ED use. We developed models that included a main effect of group (study arm), a main effect of time (utilization period), and a two-way group-by-time interaction term (study arm × utilization period) to compare differences in changes in health care utilization and costs from the preintervention to follow-up utilization periods within and between the two study arms. Since the time of events was somewhat different within and between the participants, we used the spatial power covariance structure to account for unequally spaced observations among individuals. We constructed a sample variogram to guide this decision. To further account for the duration of follow-up time intervals of different length, log mean duration of time intervals was included in the models as an offset variable. The fit of the models was examined by using mixed model diagnostic statistics, such as the likelihood ratio test and Akaike’s Information Criterion and Bayesian Information Criterion measures. Results were presented as odds ratios (OR) for binary outcomes and rate ratios (RR) for count and cost data and their 95% CIs. We conducted the analysis for the whole sample and for samples with baseline A1C 7–9% (53 mmol/mol) and A1C >9% (>75 mmol/mol) at randomization. We also assessed the sensitivity of our choice of negative binomial distribution by fitting additional generalized linear mixed models using several alternative distributions, such as γ and lognormal distributions, especially for cost data. Models with negative binomial distribution had a better fit.

To measure the absolute effect of the intervention, the event-based number needed to treat was computed for the two study arms by calculating their cumulative event proportions (hospitalization) using previously published methods (27–29). The corresponding event-based number needed to treat was calculated, representing the number of individuals who need to be treated for a given time period to prevent one hospitalization event in that period (30). Since this statistic is based on absolute differences, it is helpful when examining economic implications of any health policy decision (29). All statistical analyses were performed by using SAS software version 9.4 (SAS Institute, Inc., Cary, NC). The protocol was approved by the NYC Department of Health and Mental Hygiene institutional review board.