We examined pain reported by female patients diagnosed with early stage (I-III) hormone receptor-positive breast cancer at the West Cancer Center and Research Institute (WCCRI in Memphis, Tennessee. WCCRI is a comprehensive oncology center that collects patient-reported outcomes at each clinic visit using the Patient Care Monitor (PCM). The PCM is an electronic, tablet-based patient engagement platform used to collect information on patient-reported treatment side effects, physical and emotional symptoms, and functional status at the point of care. Patient responses are included in their electronic health record. Summary reports of patient PCM responses highlight significant changes as well as elevated symptom severity that may require further evaluation and treatment during the clinical visit (33). WCCRI has nine clinics and over 70 physicians providing a network of fully integrated cancer care that serves the tri-state area of west Tennessee, north Mississippi, and east Arkansas. WCCRI provides oncologic treatment for over 70% of all patients in the region and serves a diverse patient population.

We included female patients diagnosed with early stage (I-III) hormone receptor-positive breast cancer between 2007 and 2015 at the WCCRI. We merged the breast cancer patient data from the WCCRI and the 2010 census data including patients’ residential addresses at the time of their first diagnosis (N=3,835). According to the 2010 census, the overall population density in Tennessee was 162.9 residents per square mile and the Shelby County population density was 1,215.5 people per square mile (34). To maintain relative consistency of the setting, we limited the physical geography to Shelby County where the majority of the WCCRI patients reside. The neighborhood boundary can operate distinctively and an account of the impact of physical geography should be tailored accordingly. Therefore, we used 34 Zip codes within Shelby County, Tennessee, (N=2,584) to match the neighborhood-level data with the WCCRI data. The other inclusion criteria were availability of pain perception data and address information of the patients at the time of diagnosis. We also excluded one Zip code area that contained only two patients. Thus, we reduced the data set to patients with no missing data related to the main variables. Our final sample included 1,191 patients who reported their pain intensity at least once within 365 days since their initial diagnosis as well as patients from residential areas for which the data regarding neighborhood poverty prevalence were available (33 Zip code areas). This study has been reviewed and approved by the University of Tennessee Health Science Center Institutional Review Board, and waiver of patient consent was granted for the retrospective study.

Descriptive statistics are shown in Table 1 . We conducted group difference tests for each individual-level covariate and the significance levels of the differences were tested with a Wilcoxon test (35) for two-class variables and with a Kruskal-Wallis test (36, 37) for the variables with more than two classes ( Table 1 ). Outcome variable is the self-reported pain and patients were asked to rate their pain-related experiences as numerical rating scale scores (from 0 to 10), with higher scores indicating severe perceived pain. This self-reported pain can be regarded as a comprehensive estimate of the patient’s complex pain perception associated with physical distress related to multiple symptoms. The patients were followed up during 6 years on average and reported their pain intensity at their initial diagnosis and subsequent clinical visits.

To overcome the high noise and complexity in the outcome variable, we computed the average pain score for each patient during the observation period (from the initial diagnosis to 365 days after the initial diagnosis). Immediate pain (38) (within 3 months) (26, 39) and persistent pain (2) (during 3 years) have been important topics of discussion in many studies involving cancer patients, especially breast cancer patients (8). We focused on mean pain perception for a year after the initial diagnosis. The average number of pain reports is 5.8 (median = 4.0). Patients suffering from pain visit the clinic center more often compared to the patients with no pain shown by the significant association between positive pain perception and the number of pain reports (t-test, P < 0.001). Yet, the number of visits did not significantly differ by pain intensity within the patients who reported positive pain.

Social determinants were measured both at the individual level and at the neighborhood level. Neighborhood-level social conditions have been suggested as key factors to study how social inequalities are engraved into health inequalities (16, 40–44). We measured the neighborhood social determinants using poverty level and blight prevalence from the US census data. Poverty [the percentage of individuals below the Federal Poverty Level (FPL) within a Zip code area (45)] is an indicator of socio-economic well-being of the neighborhood. Blight (the proportion of abandoned and vacant housing units within the neighborhood) is an indicator of the quality of built environment, which is considered a stressor that negatively affects the quality of life (46, 47) and various health outcomes (24, 46, 48). Due to their high correlation, blight and poverty were analyzed separately in different models. Individual-level social determinants were measured using employment status, marital status, and insurance type (Commercial Insurance vs Medicaid) provided by the WCCRI data. For the analysis of treatment types, we included chemotherapy, which has been suggested as a possible source of chronic pain among breast cancer survivors (1). Demographic variables [age (49, 50) and race (9)] and socio-economic class (51) are known to be associated with pain-related experiences and cancer stages, which were incorporated into the analysis.

We estimated the associations between social determinants and pain with the final models adjusted for multiple individual-level covariates including biomarkers (cancer stage, chemotherapy) and demographic attributes (age and race). To examine the association between social determinants and pain, we conducted group difference tests to examine the significance of differences among different categories of independent variables. We also investigated the impact of these variables on pain intensity in an adjusted as well as unadjusted regression framework. In the unadjusted models, we analyzed pairwise regression of pain intensity against each of the covariates, revealing partial relationships. The adjusted model included all available regressors selected by model selection.

A large proportion of patients reported no perceived pain, resulting in a considerable number of zero counts in the response variable. Altogether, 452 patients (39.5%) reported pain. Considering a standard Poisson model, there were significantly more number of zero counts than expected (P<0.001). To deal with the excess zero counts, we applied the zero-inflated Poisson (ZIP) model, which fit the zero count component using a logit model and the count component using a Poisson model (52–55). The ZIP model introduces additional probability weight for zero counts to generalized linear models (GLMs). Thus, data are modeled using a two-component mixture (Poisson model for the count component and a point mass at zero), which resulted in a significant model improvement by fitting the ZIP model when compared with a standard Poisson GLM [Vuong test (56, 57) (P<0.001)]. We additionally tested and confirmed that the assumptions of the ZIP model were valid based on comprehensive model diagnostic tests including linearity, link function, response variable distribution, outliers, and multicollinearity. The results from the ZIP model were reported separately for each component. The outputs from the Poisson count data component were reported using the estimated incidence rates (exponential transformation of the estimates) and the outputs from the zero-inflation component were reported using the estimated odds ratios.

While selecting variables for the adjusted model, we started from a fully specified model including all interaction terms among the social determinant variables and other variables, and performed a backward variable selection procedure. We tested only the main covariates, namely poverty, blight, and individual-level social context variables for model selection with the other covariates fixed in the model. All statistical analyses were performed using R (version 4.0.0, the R Foundation, Vienna, Austria) and the pscl package was used to fit the zero-inflated model. All of the methods were performed in accordance with the Declaration of Helsinki and the relevant guidelines.

Table 2 has summarized the association between the absence of pain and the covariates after exponentiating the estimates, which can be interpreted as odds ratios (ORs) relative to the reference classes. Cancer stage showed the most substantial association with the probability of pain occurrence (P<0.001). With progression in the cancer stage, patients were more likely to suffer from pain. Moreover, patients who did not undergo chemotherapy had a lower probability of reporting any pain compared to patients treated with chemotherapy (P<0.001).

We found significant association between no pain and three social determinant- poverty, blight, and insurance types. In the unadjusted model, the ORs of the estimated probabilities of pain occurrence and no pain were 1.015 for every 1-unit increase in poverty (95% confidence interval [CI], 1.003–1.028) and 1.033 for every 1-unit increase in the blight level (95% CI, 1.008–1.057). These results suggest that the odds of perceiving pain in the highest-poverty area and in the highest-blight area were expected to increase by about 140% and 600%, respectively, when compared with the lowest-poverty area and the lowest-blight area. The estimated odds of the presence of pain in the Medicaid group were 1.78 times higher (Unadjusted OR: 1.397; 95% CI, 1.017–1.916 and Adjusted OR, 1.779; 95% CI, 1.164–2.717; respectively). The estimated probabilities of zero pain in the commercial insurance group and in the Medicaid group at fixed levels of other covariates (mean values for poverty and blight, reference values for the others) were 0.39 (95% CI, 0.25–0.55) and 0.26 (95% CI, 0.11–0.51), respectively, demonstrating a considerable inequality.

Table 3 demonstrates the association between pain intensity and the covariates through count components of the ZIP regression analysis. Areas with higher poverty and blight were significantly associated with greater pain intensity in both unadjusted and adjusted models. The incidence rate (IR) of neighborhood poverty in the unadjusted model was 1.01 (95% CI, 1.007–1.013), which means 1% of increase in the neighborhood poverty, the expected pain intensity increased by 1%. Pain perception was expected to increase by 81.6% in the highest-poverty area compared to the lowest-poverty area considering the range of the observed poverty variable (range: 2–61%) in our cohort (in the unadjusted model IR, 1.016; 95% CI, 1.01–1.023). Pain intensity in the highest-blight area was 2.58 times the pain intensity in the lowest-blight area. Even after adjustment for other covariates, the results were consistent with IR of 1.006 (95% CI, 1.002–1.009) for poverty and 1.007 (95% CI, 1–1.015) for blight. These findings showed that 43.2% and 43.7% increases in pain intensity were expected in the highest-poverty area and in the highest-blight area, respectively, when compared with the areas with the lowest poverty and blight.

Among the individual social context variables, employment and marital status were significantly associated with lower pain intensity. In the unadjusted models, unemployed patients showed 35.5% greater pain perception than the employed group [IR, 1.35; 95% CI, 1.152–1.583, in the adjusted model IR 1.335 (95% CI:1.128–1.58)]. Single patients showed 32% greater pain than the married group (IR, 1.32; 95% CI, 1.177–1.48). The adjusted model exhibited similar results with a slightly lower IR of 1.204 (95% CI, 1.065–1.36).