Ethical approval
This study utilized publicly available, de-identified data from the 2021 HINTS-SEER dataset. As this secondary analysis did not involve direct interaction with human participants, it was deemed exempt from institutional review board (IRB) approval at Virginia Commonwealth University in accordance with federal guidelines (45 CFR 46.104). All participants in the original HINTS-SEER survey provided written informed consent prior to participation, and data were collected under protocols approved by the National Cancer Institute.
Data source
The HINTS-SEER database was developed by the National Cancer Institute (NCI) to oversample cancer survivors and includes self-reported sociodemographic, lifestyle, and behavioral information from cancer survivors participating in the Iowa, New Mexico, and Greater San Francisco Bay Area SEER registries [41]. The HINTS-SEER dataset is unique in that it links HINTS survey responses to key data elements from the SEER program cancer registries, providing additional information on participants’ cancer diagnoses. Stratification of sampling frames was performed based on years since diagnosis and race/ethnicity, excluding nonmalignant tumors and non-melanoma skin cancers. Once consent was received, participants were selected from the sampling frame and were sent a self-administered postal questionnaire [41]. HINTS-SEER (2021), collected between January and August 2021, includes a total of 1,234 responses [41]. The overall response rate was 12.6% with very little demographic difference between those who responded and the sample frame population in each registry [41]. Of the 1,234 participants, 172 observations were removed due to missing data for at least one variable; therefore, the analytic dataset includes n = 1,062 participants.
Sociodemographic variables
Demographic characteristics
Participant age (years), gender (male or female), and race/ethnicity (Non-Hispanic White or Other) were included in the analyses.
Income
Pretax household income was self-reported by HINTS-SEER participants and included as a categorical variable with four levels: $0–$19,999, $20,000–$49,999, $50,000–$99,999, and $100,000 and up. HINTS-SEER categories were combined from $0–9,999, $10,000–$14,999, $15,000 to $19,999, $20,000 to $34,999, $35,000 to $49,999, $50,000 to $74,999, $75,000 to $99,999, $100,000 to $199,999, and $200,000 and up to facilitate meaningful analysis.
Employment status
Employment status was self-reported and operationalized as a binary variable representing Employed or Unemployed.
Educational attainment
Highest level of educational attainment was self-reported. Levels of educational attainment were as follows: Less than high school, High School, Tech Training or Some College, and College Grad or Postgrad. HINTS-SEER categories were combined from less than 8 years, 8–11 years, 12 + years, or completed high school, post-high school training other than college, some college, college graduate, and post-graduate to facilitate meaningful analysis.
Independent variables
Urban or rural residence
Rural/Urban designation was based on the patient’s county population using the 2013 Rural–Urban Continuum Codes. Urban–Rural Designations were categorized as follows: Urban (codes 1–3) or Rural (codes 4–9). The 2013 Rural–Urban Continuum Codes were combined from three metro areas and six non-metro areas to facilitate meaningful analysis [42]. Rural areas served as the reference category in all regression models.
Area-level disadvantage
Two measures were used to assess area-level disadvantage: the Social Deprivation Index (SDI) and the Areas of Persistent Poverty (APP) scale. The SDI is a composite measure using predetermined weighted factor loadings of seven demographic characteristics included in the American Community Survey (ACS) to quantify the socioeconomic variation in health outcomes [43]. The SDI was selected as the primary measure of area-level social deprivation due to its availability within the HINTS-SEER dataset structure at the county level. The present study used the 2019 SDI at the county level to determine the social deprivation of each participant’s county of residence [43]. The SDI was used as a continuous variable to capture the range of deprivation levels across counties. APP are counties that have maintained poverty rates of 20% or more for more than 30 years [44], as seen through the 1990 decennial Census, 2000 decennial Census, and 2021 small-area income and poverty estimates or the 2014–2018 five-year data series from the ACS [44]. The list of APP was last updated in 2024 using data from the 2022 Small Area Income and Poverty Estimates (SAIPE) released by the Census Bureau. The APP listing and HINTS-SEER dataset were merged using county Federal Information Processing Standards (FIPS) codes. The categories of this variable were as follows: Yes (is an APP) and No (is not an APP).
Confounding variables and variable selection
Individual-level sociodemographic variables (age, gender, race/ethnicity, income, employment status, education) were included a priori in multivariable models as potential confounders based on established literature linking these factors to lifestyle behaviors and mental health outcomes [16, 17, 19, 20].
To assess potential multicollinearity among covariates, we computed variance inflation factors (VIF). All VIF values were well below the conventional threshold of 10 and below the more conservative threshold of 5 (mean VIF = 2.06; maximum VIF = 4.32 for the highest income category), indicating that multicollinearity was not a substantive concern in our models. Although income, education, and employment are conceptually related dimensions of socioeconomic status, the empirical evidence confirms these variables are not sufficiently collinear to distort coefficient estimates, and all were retained in models given their distinct theoretical contributions as confounders in cancer survivorship research.
Dependent variables
Obesity
Participant body mass index (BMI) was calculated using self-reported height and weight. Obesity was categorized as having a BMI ≥ 30, consistent with standard CDC and WHO definitions of obesity [45].
Smoking status
Smoking status was derived from self-reported data and dichotomized to either Current Smoker or Never/Former Smoker.
Anxiety and depression
Anxiety and depression were assessed using the Patient Health Questionnaire 4 (PHQ-4). The PHQ-4 total score (0–12) was derived from summing the scores of four questions asking how often a patient has felt the following symptoms over the last two weeks: feeling nervous, anxious, or on edge; not being able to stop or control worrying; feeling down, depressed or hopeless; and having little interest or pleasure in doing things [46]. The PHQ-4 has demonstrated reliability and validity for screening depression and anxiety [47]. Total scores were dichotomized to a binary variable representing Low (0–5) and Moderate/Severe (6–12) levels of participant depression and anxiety, consistent with standard PHQ-4 scoring conventions [46].
Statistical analysis
Descriptive statistics, including the frequency and percentage for categorical variables and the mean and standard deviation for continuous variables stratified by urban versus rural residency, were calculated to describe cancer survivors’ characteristics and lifestyle behaviors. Next, bivariate logistic regression models were fit to determine the associations of each explanatory variable and demographic variable with each of the three outcome variables (Obesity, Smoking Status, and Anxiety/Depression), to assess univariate associations before multivariable modeling.
Multivariable logistic regression models were used to determine how individual-level and area-level factors are associated with each of the three outcomes while controlling for demographic characteristics. Finally, multivariable logistic regression models stratified by urban/rural classification were fit to determine the effects of individual and area-level factors in urban and rural areas separately, allowing us to examine potential differential effects by geographic context.
The HINTS-SEER uses sampling weights to ensure valid inferences from the responding sample to their respective population, correcting for nonresponse and noncoverage biases. Due to the limited number of observations in rural areas, replicate weights variance estimation was not used, as there were insufficient observations to compute jackknife standard errors. Therefore, Taylor series linearization estimation methods as provided in the HINTS-SEER documentation were used to account for the complex sample design of the HINTS-SEER dataset [41]. Stata 15.0 was used for all statistical analyses.
Model fit
To evaluate the adequacy of the fitted logistic regression models, we assessed goodness-of-fit using the Hosmer–Lemeshow test adapted for complex survey designs. Results indicated good model fit for the obesity model and the depression/anxiety model. The smoking model produced a significant goodness-of-fit statistic (p < 0.0001), likely reflecting the very low prevalence of current smoking in the analytic sample (3.9%). Because the poor fit of the smoking model—driven by sparse observations across covariate patterns—limits the reliability and interpretability of stratified estimates, smoking was excluded from the stratified multivariable models.
Effect modification
We evaluated effect modification by urbanicity using two sets of interaction analyses in the pooled survey-weighted logistic regression models, restricted to the two outcomes with adequate model fit (obesity and depression/anxiety). First, we included SDI × urban interaction terms to test whether the effect of area-level deprivation on each outcome differed by urban–rural status. Second, we included race × urban interaction terms to test whether race/ethnicity associations differed by urban–rural status. Adjusted Wald tests were used to evaluate the joint statistical significance of the interaction terms for each set of models.
Sensitivity analyses
We conducted three sets of sensitivity analyses to evaluate the robustness of findings. First, we re-fit all primary models substituting two alternative rural–urban classification measures: the Urban Influence Codes (UIC) [48] and the National Center for Health Statistics (NCHS) Urban–Rural Classification Scheme [49]. Second, we re-fit all models substituting a three-category metro-adjacency variable—distinguishing urban counties, nonmetro counties adjacent to a metropolitan area, and nonmetro counties not adjacent to a metropolitan area. Third, to relax the assumption of linearity in the log-odds for the SDI, we re-fit all models using a dichotomized SDI variable (counties classified as ‘Deprived’ or ‘Not Deprived’ based on the median county-level SDI distribution by state). Fourth, to assess whether the simultaneous inclusion of SDI and APP in primary models introduced collinearity bias, we re-estimated all primary multivariable models for obesity and depression/anxiety including SDI alone (without APP) and APP alone (without SDI) in separate model specifications.
