br We included all female patients
We included all female patients with a primary diagnosis of breast cancer admitted and discharged from US hospitals between January 1st, 2007 and December 31st, 2014 represented in the NIS. Males and pa-tients younger than 18 years were excluded (n = 170,644) (Supplemental Fig. 1). Inpatient stays were identified using Interna-tional Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnosis codes  (Supplemental Table 1). Previous studies have employed the same ICD-9-CM codes used in selecting this MPP+Iodide sub-group [25–28].
Our primary outcome was opioid abuse status. Opioid abuse was defined using ICD-9 codes for dependent opioid use (304.0) and non-dependent opioid use (305.5). The primary predictor of interest was long-term NSAID use. Inpatient mortality and LOS were secondary outcomes. We extracted records on demographics (patient and hospital level) including race/ethnicity (white, black, Hispanic, other), primary payer (Medicare, Medicaid, private insurance, self-pay, other), income quartile of patient's zip code, and hospital region (northeast, midwest/ Cancer Treatment and Research Communications 21 (2019) 100156
northcentral, south, west). Other covariates selected were patient baseline characteristics (nonopioid substance abuse, non-cancer related pain, past surgical history, psychosocial stress, anxiety, history or long-term use of steroids, chemotherapy or immunotherapy, Charlson/Deyo comorbidity index).
The demographic variables and secondary outcomes were available and defined in the NIS. Opioid abuse, long-term NSAID use, as well as baseline characteristics, were identified and defined using ICD-9-CM diagnosis codes (Supplemental Table 1).
Appropriate discharge weights provided by the NIS for estimate adjustments were used in our analysis to ensure accuracy and to pro-vide nationally representative adjustments. We adopted HCUP re-commendations in our analysis to account for the complex NIS design.
We conducted descriptive analyses for our study variables. We computed means and standard deviations with 95% CI for age, LOS and Charlson/Deyo comorbidity score by long-term NSAID use status.
For our categorical variables, using Rao Scott design adjusted chi-square statistics, we stratified our study characteristics by long-term NSAID status. Study variables across the levels of long-term NSAID use were reported as percentages with associated p-values. We assessed the eﬀect of long-term NSAID use on our outcomes by computing crude and adjusted odds and mean ratios from univariable and multivariable re-gression analyses . We constructed three multivariable regression models representing three outcomes: opioid abuse, inpatient mortality, and length of stay. Opioid abuse and inpatient mortality were analyzed using PROC SURVEY LOGISTIC under a generalized logit function. We employed PROC GENMOD with log link function in the regression analysis of LOS. A negative binomial distribution was specified to ac-count for overdispersion in the count variable, LOS.
All analyses were performed using SAS v9.4 [Statistical Analysis System, Version 9.4, SAS Institute Inc, Cary, NC] with statistical sig-nificance set at 95% CI and p-value at <0.05, two-tailed.
Distribution of study characteristics by long-term NSAID use
The mean age of patients of the sample was 62.1 years
proportion of patients with long-term NSAID use was 4.6% (n = 7838). Among patients with long-term NSAID use, 2% had cancer-related pain; about 30% and 0.2% used and abused opioids, respectively. Slightly more than two-thirds of portal system with long-term NSAID use were white; nearly 70% were on Medicare, and 12% had a past surgical history. Using median household income quartiles for the patients’ ZIP code as a proxy for socioeconomic status, we observed a similar distribution of income across the four categories. Few patients had symptoms of de-pression and anxiety. Table 1 shows the distribution of study char-acteristics by long-term NSAID use.
Association between long-term NSAID use and opioid abuse
Table 2A shows the results of the multivariable logistic regression model comparing long term NSAID use and opioid abuse. We found 46% lower odds of opioid abuse among patients with a history of long-term use of NSAIDs compared to patients without a history of long-term use (aOR 0.53; 95% CI [0.32–0.88]). The association between long-term NSAID use and opioid abuse was significant with a 3-fold in-creased odds of opioid abuse among breast cancer patients with cancer-related pain compared to those without cancer-related pain (aOR 2.96; 95% CI [2.34–3.74]).