Impact of Insulin Initiation and Time to Insulin Initiation on within-Person Change of Healthcare Utilization in Medicaid Enrollees with type 2 Diabetes

Background: Insulin use, time to insulin initiation, and subsequent healthcare utilization in low-income adults with type 2 diabetes (T2DM) are poorly understood. Methods: Our objectives were to examine whether 1) insulin initiation and 2) every 1-year decrease in the time from rst-line glucose-lowering agents (GLAs) to insulin initiation reduced healthcare utilization within 12 months after insulin initiation in Pennsylvania (PA) Medicaid enrollees with T2DM. We included a cohort of 12,648 PA Medicaid enrollees (age 47.3±10.3 years) with rst-line non-insulin GLAs between 2008 and 2016. Insulin users (N=3,625) were matched with non-insulin users (N=9,023) on dates of 1 st GLA prescription and propensity scores calculated based on baseline characteristics to account for potential confounders. Generalized estimating equations models estimated within-person changes in inpatient stays and emergency room (ER) visits 12 months after vs. 12 months before insulin initiation. We performed sensitivity analyses in young (18-45 years) and middle-aged enrollees (>45 to 64 years). Results: The average time from rst-line GLAs to insulin initiation was 2.0±1.7 years. Neither insulin initiation (rate ratio [RR]:1.0 [1.0, 1.1]) nor time to insulin initiation (RR: 1.0 [1.0, 1.0]) was associated with within-person change in ER visits. In young T2DM adults, insulin users had a greater subsequent increase in inpatient stays after insulin initiation vs. non-insulin users over the same time period (RR: 1.3 [1.1, 1.5]). Conclusions: In T2DM Medicaid enrollees, no reduction in healthcare utilization was observed after insulin initiation, even in early stages of pharmacotherapy. Studies investigating primary contributors to the increased inpatient use in young insulin users with T2DM are needed.


Background
Medicaid, as the primary health insurance program for low-income people in the US, covered about 8% of all individuals with diagnosed diabetes in 2009 (approximately 1.8 million adults) [1,2]. In 2012, nearly 13% of total healthcare expenditures for diabetes was provided by Medicaid (~$28 billion) [3]. For each Medicaid enrollee with T2DM, the age-adjusted annual medical cost was $14,170 in 2012, higher than $9,560 in those with commercial insurance [1], with the annual medical cost increasing to $17,830 for those with a diagnosis code of uncontrolled diabetes (HbA1c > 9%) possibly due to a higher prevalence of diabetes complications [1]. Of Medicaid enrollees with T2DM, only 64% received annual HbA1c test and even a lower portion (58%) had controlled diabetes in 2012 [1]. Improving glycemic control with insulin use may therefore be important in reducing medical costs and healthcare utilization [4,5].
Insulin initiation in the early progression of diabetes shows bene t for maintaining pancreatic beta-cell function [6] and reducing the risk of long-term microvascular complications by 24% in individuals with T2DM [7]. In 2012, 29% of Medicaid enrollees with diagnosed T2DM used insulin and the total amount of insulin reimbursement substantially rose by 462% from 2006 to 2014 [1,8]. Some studies conducted in managed care organizations showed a subsequent reduction in inpatient service after insulin initiation [9,10]. However, few studies have compared this reduction with those using non-insulin GLAs and addressed the effect of time to insulin initiation on the subsequent change of healthcare utilization in Medicaid enrollees, who have low-income, and a higher prevalence of disability and behavioral health and medical comorbidities. Our objectives were to examine whether (1) insulin initiation decreased the number of inpatient stays and ER visits within 12 months after insulin initiation; and (2) if a shorter length of time to insulin initiation from rst-line GLAs was associated with reduction in inpatient stays and ER visits within 12 months after vs. before insulin initiation in young and middle-aged Medicaid enrollees with T2DM.

Data source and study population
We obtained a Medicaid administrative claims database from the Pennsylvania Department of Human Services (PADHS) to identify the study population. The database includes demographic characteristics, pharmacy claims, and ICD9/10-CM diagnosis codes, encounters with procedure information from inpatient and outpatient settings for all Medicaid enrollees in Pennsylvania (PA) from 2007 to 2016. The study sample for the cohort was limited to 168,594 Medicaid enrollees who had at least one prescription ll for non-insulin GLAs (i.e. sulfonylurea, thiazolidinedione, inhibitors of dipeptidyl peptidase 4, sodiumglucose co-transporter-2 inhibitor, glucagon-like peptide 1 receptor agonists) from January 1, 2008 through December 31, 2016 [2] (Fig. 1). From that group, we selected 136,998 Medicaid enrollees aged 18-64 years old who were not dually eligible for Medicare on the index date (i.e. the date of the rst prescription ll for non-insulin GLAs) since pharmacy claims from Medicare were not available. To identify an incident cohort of new GLA users and allow for adequate measurement of healthcare utilization before insulin initiation, we further limited the study cohort to 45,560 enrollees with ≥ 365 days of continuous Medicaid enrollment preceding the index date and without any prescription lls for GLAs in 2007. To exclude those with type 1 or gestational diabetes, we removed enrollees if a) they had no claims with T2DM diagnostic codes (ICD-9 250 or ICD-10 E11) in any position within 6 months before or after the index date (N = 12,686); or b) were women with a birth or a terminated pregnancy within 6 months before or after the index date (N = 536). To allow for adequate measurement of healthcare utilization in the follow-up, we excluded enrollees without ≥ 890 days (2.5 years) of continuous enrollment after the index date (the date of rst-line GLAs). Additionally, insulin users without ≥ 365 days of continuous enrollment after the date of insulin initiation were excluded. A cohort of 15,356 Medicaid enrollees with T2DM was used for matching analyses (Fig. 1).

Matching
The cohort of 15,356 Medicaid enrollees were followed from the index date through the rst occurrence of any censoring event. Censoring events included the end of enrollment in Pennsylvania Medicaid, gaining dual eligibility for Medicare, or all-cause mortality. Insulin users (N = 3,715) were de ned as enrollees with ≥ 1 prescription lls for insulin during the follow-up period. The date of insulin initiation was the date of the rst prescription ll for any type of insulin after the index date. Those without insulin prescriptions in the follow-up period, de ned as non-insulin users, served as the comparison group (N = 11,641).
We used a hierarchical 1:3 matching strategy to assign a time point for non-insulin users to measure within-person change in healthcare utilization as well as to account for potential confounders of the association between insulin initiation and healthcare utilization (Fig. 2). Non-insulin users were matched with insulin users on a) the index date (the date of rst line GLAs) ± 5 days, b) continuous enrollment within 12 months before and after the date of insulin initiation, and c) propensity score (PS) with a deviation of 0.05 on the PS scale. The PS represented the probability of initiating insulin, which was estimated using a multivariable logistic regression model adjusted for covariates observed within 12 months before or on the date of rst line GLAs. Covariates in the propensity score model included age, sex, race (White, Black, other), calendar year of the index date, primary enrollment in fee-for-service (as opposed to managed care), Medicaid eligibility categories (disabled or chronically ill vs. others such as Temporary Assistance for Needy Families) [11]. Our cohort effectively excluded the expansion group under the Affordable Care Act (ACA) which was implemented in 2015 in PA due to the continuous enrollment criteria we imposed. We also included the Area Deprivation Index, an area-based measure of education, income, and occupation status, using the 9-digit ZIP code of residence [11]. Additional covariates in the propensity score model measured within 12 months before rst-line GLAs included indicators for the use of certain medications (anti-hypertensive agents, anticoagulant agents, lipid lowering agents, nitrates, and loop diuretics), indicators for the presence of diabetes-related comorbidities and complications (e.g. hypertension, obesity, depression, psychosis, congestive heart failure, nephropathy, neuropathy, retinopathy, cardio/cerebrovascular complications, and metabolic complications) [12], as well as a modi ed Elixhauser comorbidity index excluding the health conditions mentioned above that were included as separate indicators [13] (de nitions in Additional le 1).
A total of 3,625 insulin users were matched with 9,023 non-insulin users based on the date of rst-line GLAs, continuous enrollment and propensity scores. A total of 2,708 enrollees (90 insulin users and 2,618 non-insulin users) were excluded after matching since the distance of the PS within matched sets was either greater than 0.05 or greater than the rst three smallest distances of the PS among all possible pairings [14]. After matching, no signi cant differences were noted in sex, psychoses, congestive heart failure and cardiovascular complications between insulin users and their matched non-insulin users indicating that the matching process attenuated these differences between groups, though other differences in race, Fee-For-Service status, prevalence of microvascular complications and depression, use of lipid-lowering agents, and healthcare utilization remained signi cant (Additional le 2).

Primary Independent Variables
The primary exposures were the binary indicator of insulin initiation (insulin users/non-insulin users) and the continuous variable for the time from rst-line GLAs to insulin initiation, calculated by subtracting the index date from the date of the rst prescription ll for insulin. For non-insulin users, the time to insulin initiation was calculated by subtracting the index date from the date of insulin initiation for respectively matched insulin users.

Outcomes
The primary outcomes were within-person change in the number of all-cause inpatient stays and ER visits after insulin initiation or the matched date. For insulin users, the numbers of inpatient stays and ER visits were assessed within 12 months before and 12 months after the date of the rst prescription ll for insulin. For non-insulin users, the outcome was assessed 12 months before and 12 months after the date of insulin initiation for the matched insulin user. The number of inpatient stays was counted based on non-adjacent inpatient claims [15]. Adjacent inpatient claims with the same discharge and admission date were considered as hospital transfers, which were de ned as a single hospital stay for the purpose of generating the count [15]. ER visits were determined by the outpatient claims with revenue center codes of 0450-0452, 0456, 0459, 0981 or procedure codes of 99281-99285, G0380, G0381-G0385 [16]. ER visits which occurred in the period between the admission and discharge dates of any inpatient stays were excluded.

Statistical analysis
Descriptive statistics were conducted for enrollees' demographics, comorbidities, complications, and medication use measured on the date of rst-line GLAs or within 12 months before rst-line GLAs. Univariate comparisons between matched insulin users and non-insulin users were conducted using McNemar's tests for categorical variables, paired t-test for continuous variables with normal distribution, and Wilcoxon Rank sum tests for non-normally distributed continuous variables. The χ 2 tests and t tests were used to compare categorical and continuous variables in unmatched cohorts. Given that the number of inpatient stays and ER visits before and after insulin initiation were discrete and over-dispersed, generalized estimating equations (GEE) models with a negative-binomial distribution were used. The effect of time to insulin initiation after rst-line GLAs on within-person change was examined in insulin users only. In order to attenuate the confounding effect between measured covariates and the withinperson change, interaction terms between time and covariates included in the propensity score were added to the model stepwise and removed at p > 0.1 except for age, sex, and race. Collinearity was tested before adding covariates into the model. We also performed strati ed GEE models for subgroups of enrollees aged 18 to ≤ 45 years old (young adults) and from > 45 to 64 years old (middle-aged adults) since we observed the interaction term between time and age signi cantly attenuated the relative risk ratio of within-person change in inpatient stays toward the null in the primary analysis. A sensitivity analysis was performed in the subgroup of those with 2 or more prescription lls for insulin vs. noninsulin users to exclude those who initiated insulin by chance. Analyses were performed using SAS, version 9.4 (SAS Institute, Inc., Cary, NC).

Results
Baseline characteristics of insulin vs. non-insulin users in the matched cohort (N = 12,648) are shown in Table 1. After matching, insulin users (N = 3,625) and non-insulin users (N = 9,023) were different in several baseline characteristics. In the matched cohort, insulin users were 2 years younger (mean age: 45.6 ± 10.1 vs. 47.6 ± 10.6 years, P < .001) and more likely to be Black (32.3% vs. 28.9%, P < .001) than non-insulin users. Insulin users were more likely to have depression, psychosis, congestive heart failure, nephropathy and metabolic complications than non-insulin users. Insulin users were less likely to use lipid lowering agents compared to the matched non-insulin users. Insulin users had a higher number of inpatient stays (0.5 vs. 0.3; P < .001) and ER visits per person (1.9 vs. 1.4; P < .001) during the year before rst line GLAs than matched non-insulin users. Stratifying by age, baseline characteristics were similar between insulin and non-insulin users in young adults (N = 4,745); whereas in middle-aged adults (N = 7,903), insulin users were more likely to have depression, congestive heart failure, metabolic complications, and cardiovascular complications than non-insulin users (Additional le 3). However, the rate of inpatient stays did not change after insulin initiation (RR: 1.  Table 3) or in the subgroups of young and middle-aged enrollees.

Emergency Room (er) Visits
In the matched cohort, the crude numbers of all-cause ER visits within 12 months before and after insulin initiation were 2.2 (standard deviation [SD] 4.0) and 2.2 (SD 4.1) per person for insulin users, respectively.
Non-insulin users had 1.4 (SD 3.1) and 1.3 (SD 3.0) all-cause ER visits per person in 12 months before and after the matched date of insulin initiation. In the adjusted GEE model, insulin users had a higher rate of ER visit than non-insulin users (RR: 1.5 [1.4, 1.6]), Table 2). The number of ER visits decreased within 12 months after the matched date by 5% (RR: 1.0 [0.9, 1.0]). No signi cant difference was found in withinperson change of ER visits between insulin and non-insulin users either in the matched cohort (RR: 1.0 [1.0, 1.1]) or in the subgroups of young and middle-aged enrollees. Among insulin users, no association was found between the time to insulin initiation and the within-person change of ER visits (RR: 1.0 [1.0, 1.0], Table 3). The results in the subgroup of those with 2 or more re lls for insulin vs. non-insulin users were consistent with the primary analysis.

Discussion
In our claims-based study among young and middle-aged Medicaid enrollees in Pennsylvania, insulin users had more inpatient stays and ER visits than matched non-insulin users after controlling for several baseline demographics, comorbidities, and complications. However, compared with non-insulin users, insulin users had a similar pattern of change in inpatient stays or ER visits within 12 months after insulin initiation vs. before. We found that insulin users in the age group of 18-45 years had a greater increase in the number of inpatient stays after insulin initiation compared to non-insulin users in the same age group, which has not been addressed in previous studies. Among insulin users, the length of time to insulin initiation after GLAs was not associated with the change in healthcare utilization. To our knowledge, this study is the rst to investigate the association between insulin initiation and the subsequent changes in inpatient and ER services in both young and middle-aged Medicaid enrollees who have low-incomes, high levels of disability, and comorbid mental health and physical health conditions [17][18][19].
We found a higher level of inpatient and ER services utilization in insulin users than non-insulin users regardless of the time from rst-line GLAs to insulin initiation, possibly due to several unmeasurable confounders related to diabetes severity. Prior observational studies showed that insulin users had higher glycemic levels and more existing complications, which may yield substantial healthcare utilization and costs [20], than non-insulin users [17][18][19]. In a study of managed care patients with T2DM on insulin, those who initiated insulin after 1 GLA had more complications than delayed insulin initiation after ≥ 3 GLAs [9]. These studies may re ect that insulin is initiated reactively to slow down the progression of complications rather than proactively to prevent the occurrence of complications in clinical practice [21].
We employed a matching strategy with propensity scores and covariate adjustments to reduce bias by balancing characteristics before rst-line GLAs between insulin and non-insulin users. Nevertheless, our ability to test this hypothesis is limited by the fact that we are unable to adjust for HbA1c levels, lipid pro les, BMI, duration of diabetes, smoking status and other disease markers related to diabetes severity at baseline. One study reported insulin users in health maintenance organizations (HMO) had more ER visits than those using sulfonylureas cross-sectionally; however, the difference attenuated to the null after adjusting for HbA1c levels, diabetes duration, and baseline diabetes severity measured by self-reported symptoms, disease manifestations, and diagnosed complications [22]. This nding further supports that the higher level of inpatient and ER services utilization in insulin users in our study may be the result of insulin users having a higher level of comorbidity, potentially unmeasured in administrative data.
Our nding regarding a lack of association between insulin initiation and within-person reductions in ER visits is consistent with prior work. One study in individuals with T2DM from an HMO showed the number of ER visits did not change within 12 months after insulin initiation in insulin users adjusting for withinperson changes in non-insulin users (adjusted pre-post ER visits change per year: 0.1[0.0-0.2]) [22].
However, this study was conducted in a commercial HMO, with fewer racial/ethnic minorities and higher income distribution compared to Medicaid enrollees [22]. Medicaid enrollees were more likely to use ER services compared with privately insured populations, which is partially explained by their poor access to primary, specialty and other outpatient health care [23]. Some projects have been implemented in Medicaid since 2012, which included providing supportive housing for homeless adults and implementing case management in frequent ER users, to re-direct non-urgent ER cares to the most appropriate settings [24]. An HMO study found that insulin initiation was related to an increase in outpatient visits probably due to additional glucose monitoring, insulin dose adjustments or foot and eye examinations (adjusted pre-post outpatient visits change per year: 3.1[1.9-4.0]) [22]. Whether those projects to control non-urgent use of ER or an increased use of outpatient visits after insulin initiation was related to the reduction in ER visits in our study is unclear and represents a direction of further study.
Our study showed insulin users aged ≤ 45 years had an increase in inpatient stays during 12 months after insulin initiation, which has not been addressed in previous studies among adults with T2DM. Previous studies showed that after insulin initiation, the proportion of those with inpatient visits was reduced by 10% in privately insured individuals with a mean age of 55 [9], and the average number of inpatient visits in 9 months decreased by 0.16 among middle-aged adults [10]. However, neither of these studies included non-insulin users as comparators and examined the association between insulin use and inpatient services in adults aged ≤ 45 years. Young adults with T2DM may be characterized by an accelerated loss of β-cell function, a higher prevalence of obesity, and a higher risk of microalbuminuria compared to middle-aged adults with T2DM [25][26][27][28]. The evidence has previously indicated a more aggressive disease progression in young adults vs. middle-aged adults, which may yield more inpatient stays. In young adults with T2DM, insulin is usually initiated for those with marked hyperglycemia (A1C ≥ 8.5%) and any symptoms of diabetes (e.g. polyuria, polydipsia, nocturia, and/or weight loss) [29]. While we were limited to diagnoses captured in claims data, it may be that young insulin users in our cohort had more severe diabetes disease progression than non-insulin users.
Our study is the rst to observe the effect of young vs. middle age on the association between insulin initiation and healthcare utilization in low-income Medicaid enrollees. We were able to address the withinperson change of healthcare utilization adjusting for underlying changes in non-insulin users, which is rarely examined in previous studies. Additionally, our Medicaid database obtained directly from the Commonwealth captures drug and healthcare utilization from bene ciaries in managed care plans, which is not always available in Medicaid databases from the Centers for Medicare and Medicaid Services (CMS). The comprehensive reporting system for encounter data in the Pennsylvania Department of Human Service (PADHS) provides a reliable and valid measure of utilization from managed care.
A key limitation of this study is that some potential confounders were not available in our claims data, especially clinical laboratory data (e.g. HbA1c, lipid pro les, and weight), health-related behavior (e.g. smoking), and physicians' referral behavior. In addition, healthcare utilization was evaluated for 1 year before and after insulin initiation, so an extended time frame may be needed to observe long-term change of healthcare utilization after insulin initiation. Moreover, using the real-world data, we were not able to determine if insulin initiation slowed down the increasing rate of healthcare utilization in the insulin users compared to those in whom insulin was not initiated. Finally, our study population is representative of a typical Medicaid population in Pennsylvania. Given variation in Medicaid policy across the states our results may not generalize to other states.

Conclusions
In T2DM Medicaid enrollees, neither insulin initiation nor the time to insulin initiation from rst line therapy was found to be related to within-person change in inpatient stays or ER visits accounting for several baseline characteristics. However, in T2DM Medicaid enrollees with insulin therapy aged ≤ 45 years, inpatient use increased after insulin initiation at a higher rate than non-insulin users, possibly due to a more aggressive disease progression. Further studies of investigating the contributors to the increased inpatient use after insulin initiation are needed targeting young insulin users with T2DM in Medicaid.

Declarations
Ethics approval and consent to participate The University of Pittsburgh IRB reviewed this study (IRB#: PRO12100227) and ruled that since it did not involve any interaction with human subjects, only review of existing data, it did not require formal ethics approval and could be designated as "exempt". Per the University of Pittsburgh IRB the study met all the necessary criteria for an exemption under section: 45 CFR 46.101(b) (4) existing data, documents, or records Additionally, the Pennsylvania Department of Human Services (PADHS) granted formal permission to access their databases through separate data use agreements with the University of Pittsburgh.
Consent for publication: Not applicable.
Availability of data and materials: The data that support the ndings of this study are available from PADHS but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available.
Competing interests: The authors declare that they have no competing interests.  Figure 1 Flow chart of the eligible patients with type 2 diabetes (T2DM) GLA: glucose-lowering agents.