Study Sample
We examined claims and enrollment data for Medicare fee-for-service (FFS) beneficiaries, which represent approximately 88 % of all Medicare recipients, to identify a sample of older adults with a diagnosis of IBD and contraindications to anti-TNF therapy [21]. Patients ≥65 years old with at least 12 months of Parts A and B (hospital and medical visit) and 6 months of Part D (outpatient prescription) coverage during the years 2006–2009 were included. Data were included for up to 6 months prior to coverage by Medicare D. IBD diagnosis was ascertained using a case-finding algorithm (≥2 claims for appropriate International Classification of Diseases, 9th edition (ICD-9) codes [Crohn’s Disease: 555.xx] or [Ulcerative Colitis: 556.xx]) [22, 23]. The first 12 months of data are referred to as collected during the “baseline year” and began after the patient had at least one IBD claim. Data collected after the baseline period was considered from the “follow-up period”. Follow-up continued until December 31, 2009, disenrollment from Medicare Parts A, B or D or death, whichever occurred first. Patients were excluded if they did not have a confirmatory IBD claim by the end of the follow-up period. Drug contraindications were determined during the baseline year. Contraindication to anti-TNF therapy was defined as advanced CHF or malignancy. Advanced CHF was identified as ≥1 outpatient claim for a CHF diagnosis (ICD-9 398.91, 402.01, 402.11, 402.91, 404.01, 404.03, 404.11, 404.13, 404.91, 404.93, 414.8, 428.x) and at least 1 CHF hospitalization (primary inpatient discharge diagnosis code for CHF [ICD-9 codes: 398.91, 404.x1, 404.x3, 428.0–428.9]) [24]. Both solid tumor and hematologic malignancies were ascertained using the 2008 Elixhauser criteria, version 3.3 (ICD-9 codes: 140.0–172.9, 174.0–175.9, 179–195.8, 258.01–258.03, 196.0–199.1, 789.51, 200.00–202.38, 202.50–203.01, 203.8–203.81, 238.6, 273.3) [25]. Patients with nonbiologic contraindications comprising hematologic malignancies (defined above) and liver disease (ICD-9 codes for 571.0–571.9, 070.2–070.9, 572.2–572.4) [26] were excluded from study as they were not eligible to receive any steroid- sparing agent.
Outcome Drug Class Variables
The primary outcome variables were receipt and duration of systemic steroid therapy (prednisone, methylprednisolone, budesonide). Although the entire class of agents were included as an outcome variable, prednisone represented >95 % of all incident steroid use in this study. Systemic steroid use was identified from Medicare Part D [27] claims history during the time the patient had Part D coverage, by National Drug Codes (NDCs) using information on NDCs and therapeutic class in the Multum Lexicon™ Plus database (Cerner Multum Incorporated, Denver, Colorado). A patient day dataset was constructed with patients assigned to having received therapy on a given day based upon the ReComp algorithm [28]. Identifying drug administration days allowed computing period prevalence, treatment duration and incident drug use.
Incident steroid use was defined as a new claim for systemic steroids that started during the follow-up period, and steroid therapy days included all the days systemic steroids were received (regardless of gaps in treatment) during the follow-up. Prevalent systemic steroid use is included in the descriptive analysis only.
Explanatory Drug Class Variables
The use of home administered anti-TNF infusions (infliximab), anti-TNF injections (adalimumab), non-biologic immunomodulators, aminosalicylates, locally administered steroids and antidiarrheal therapies were also identified from Medicare Part D [27] claims history during the baseline period by the approach used for the systemic steroid outcome variable. Facility-administered anti-TNF infusions (infliximab) were ascertained from Part A and B claims for Healthcare Common Procedure Coding System (HCPCS) J-code 1745 [27, 29]. These therapies were assessed during the baseline period and included as predictors in the model for steroid use.
Other Explanatory Variables
Demographic information was obtained from the Medicare denominator file. This file was used to determine Medicaid coverage status. The sample of Medicare beneficiaries was merged with Census 2010 Summary File 3 (SF3) data (yielding socioeconomic characteristics on households). Patient zip-codes were used to assign urban status based on rural urban commuting area (RUCA) codes [30].
Patient medical and health care characteristics were ascertained from the Medicare data during the baseline period, and included a comorbidity index (Charlson index), IBD disease severity (endoscopies, surgeries), and health resource utilization (managing provider type, hospitalization, emergency department visit).
The primary provider type (primary care provider, gastroenterologist or other specialists) for IBD management was assigned as the provider with the greatest number of evaluation and management (E&M) IBD visits (Appendix 1) [31]. Assuming patients may receive one surveillance endoscopy annually [32, 33], we considered >1 endoscopy (identified on outpatient and inpatient claims as ICD-9 codes and on carrier claims as CPT codes, Appendix 1) [22] an indicator of disease severity. IBD surgeries were identified from inpatient claims for an appropriate ICD-9 procedure code (Appendix 1) [34]. Other health resource use (hospitalizations, emergency department visits), and comorbidity indices were determined from ICD-9 codes and HCPCS codes from inpatient claims, carrier claims and E&M visits, as appropriate [35].
Statistical Analysis
We provide descriptive information on all drug classes considered and model the use of incident steroid use. Patients who received systemic steroid therapy during the baseline period were excluded from regression analyses.
Since the deployment of any systemic steroids is important and the duration of steroid therapy is separately an important indicator of appropriate use, we employ a hurdle model in the analysis of incident steroid use [36]. The logistic portion of the model evaluated patient factors associated with being an incident steroid recipient. Furthermore, it is likely that individuals vary in their propensity to continue on systemic steroids due to unmeasured factors, thereby generating overdispersion and making the negative binomial a more appropriate choice for length of steroid treatment than the Poisson. The negative binomial portion was truncated at zero and assessed factors associated with steroid therapy days among incident steroid users.
Robust standard errors were used for statistical inference on regression coefficients. Because of the relatively small sample size, forward step-wise model building was employed for the hurdle model components with sociodemographic characteristics and IBD drug classes in the first 12 months included in the initial model and each candidate covariate of reasonable cell size (>10) considered. Candidate covariates with descriptive importance (eg. region) and/or marginal statistical significance (p < 0.1) were retained in the final models. Model fit for the logistic part of the model was evaluated by comparing deciles of observed and predicted percentage receiving systemic steroids using the Hosmer-Lemeshow goodness of fit statistics (p = ns). Fit of the expected number of observations from a negative binomial to the data is presented graphically (Figure 1, Chi square p = ns).
This study included a dynamic cohort with varying patient follow-up times allowing some participant’s greater time to receive systemic steroids and additional steroid therapy days; therefore, time-offsets were used in all models. The logistic time offset was defined as the natural log of the time from the beginning of the entire observation period until systemic steroids were initiated for steroid recipients or until the end of follow-up for non-recipients. For the count model, the time offset was the time from the beginning of the entire observation period until the end of follow-up for a given patient. Our time off-set mirrors the hurdle model weighting approach used by Senturk and colleagues in their examination of cardiovascular events in the dialysis population [37].
The results of the hurdle model were compared to those of fitting a Poisson model with the log link and robust standard errors to the total number of days on systemic steroids, which is the model frequently used in medical literature [38]. The same covariates were included in the Poisson and hurdle models and significance levels compared. We evaluated the Poisson model fit using the Schwartz Bayesian information criterion and the degree of over-dispersion using deviance/df.
Since the complementary log-log (CLL) link better corresponds to a binary analysis in which time is a consideration in the probability of event occurrence, a sensitivity analysis of the logistic component of the hurdle model was performed using the CLL link for the generalized linear regression of incident steroid utilization, and generated similar findings (not reported).
Statistical analysis was conducted using Stata version 13 (StataCorp, College Station, TX) and results are presented as odds ratios (OR), ratio of durations (count model), and 95 % Confidence Intervals (CI).
This retrospective cohort study was determined to be exempt from oversight by the University of Wisconsin Institutional Review Board.