Study Design

A retrospective longitudinal cohort design with baseline, precancer and postcancer periods were used. Using the date of cancer diagnosis as the index date, a 24-month window before the index date was constructed. This 24-month period was split into two 12-month periods (ie, baseline and precancer period). This was necessary because identification of preexisting chronic conditions and NCHs during the precancer period involved the use of inpatient claims. To avoid circular reasoning, the first 12-month period was used as the baseline period, during which the types of chronic conditions and other independent variables were measured. The 12-month period before the diagnosis of cancer was considered the precancer period and this period was used to derive NCHs. The postcancer period consisted of 12 months after the diagnosis of cancer. To ensure a robust study design, the NCHs were measured repeatedly every 120 days during the precancer and postcancer period, yielding a total 6 repeated measures per individual. Thus, t₁, t₂, and t₃ represented the precancer period and t₄, t₅, and t₆ represented the postcancer period (Figure 1).

Data Sources

SEER-Medicare Linked Data: We used the NCI's SEER cancer registries data linked with Medicare administrative claims data. As of now, the SEER data represent 28% of the US population from 18 registries and consist of a total of 7,397,159 malignant cancer cases, and most of these cancer cases are diagnosed among the elderly population aged 65 years and older.⁸ Medicare is the US Government–mandated insurance program covering 97% of the US population aged 65 years and older.⁹ A total of 93% of men aged 65 years and older in SEER has been linked to Medicare enrollment records.^9,10 In the current study, we used data from the SEER-Medicare linked database of prostate cancer cases diagnosed between 2002 and 2009 and their linked claims between 2000 and 2010. Furthermore, we included elderly men aged 67 years and older because we required at least 2 years of health care utilization data before diagnosis of prostate cancer.

Study Cohort: The study population was based on 289,701 men diagnosed with incident prostate cancer between January 1, 2002, and December 31, 2009. Men with multiple cancers (N=25,785), those diagnosed with prostate cancer during the autopsy (N= 2,944), and those younger than 67 years who died during the study period (N=111,643) were excluded. Furthermore, the cohort was further restricted to those with continuous fee-for-service Medicare Part A and Part B enrollment during the entire study period (N=101,302). To reduce the misclassification bias, individuals with newly diagnosed conditions during the precancer and postcancer periods (N=40,544) were excluded. After excluding individuals with missing information on race/ethnicity, county, and in situ, unknown, and advanced cancer stage, the final study cohort comprised 57,589 elderly men with incident prostate cancer (Figure 2).

Key Dependent Variable

Noncancer Hospitalizations: Inpatient admissions during the precancer and postcancer periods were derived from the MEDPAR file. Hospitalizations with the primary diagnosis of prostate cancer were considered

View Full Size

Study design.

Citation: Journal of the National Comprehensive Cancer Network J Natl Compr Canc Netw 14, 2; 10.6004/jnccn.2016.0022

Independent Variables: Our study used the widely known Andersen's Behavioral Model of Health Services Use (ABM) to identify the potential independent factors associated with NCHs in our study population.¹³ The ABM and adapted versions of this model have been extensively used in health services research to study the factors associated with health care service use and health care outcomes. The ABM states that the use of health care services is determined by individual and contextual determinants, which can be categorized into 4 groups: (1) predisposing factors, (2) enabling factors, (3) external environment factors, and (4) need factors (Figure 3).

Predisposing Factors: These factors represent the unique feature of individuals who tend to engage more or less use of health care services. Age, race/ethnicity, and marital status were the predisposing characteristics. Age at the time of diagnosis (67–74 years, ≥75 years), race/ethnicity (White, African American, Hispanic, and other), and marital status (married, divorced/separated, unmarried, and others) were identified from the Patient Entitlement and Diagnosis Summary File (PEDSF).

Enabling Factors: These factors serve as conditions that enable individuals to use health care services. Income, education, access to care, and the initial cancer treatment were the enabling factors. The PEDSF file provides 2000 Census information on education level and median income. Access to care was measured with at least one visit to a primary care provider during the baseline period.¹⁴ Receipt of active treatment was identified using inpatient, outpatient, and carrier files with appropriate International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) diagnostic and procedure codes; Current Procedural Terminology (CPT) or Healthcare Common Procedure Coding System (HCPCS) codes; or revenue center code within the 6 months after the index date (eAppendix 1).^15–17 We classified the cancer treatment into 4 groups based on a hierarchy: (1) radical prostatectomy (RP), (2) radiation therapy (RT), (3) hormone therapy, and (4) none of these therapies.

External Environment Factors: These factors facilitate the use of health care services related to the structure of services in the geographic areas near the individuals. External environment characteristics comprised the individual's region (Northeast, South, North-Central, and West) and county-level health care resources, such as radiation oncology units and urology units.¹⁸ County-level health care resources were measured by the number of radiation oncology and urology units, which were derived from the Area Health Resource Files (AHRF). For the purposes of our analyses, we categorized

View Full Size

Study cohort development chart for study population of elderly Medicare beneficiaries diagnosed with prostate cancer.

Citation: Journal of the National Comprehensive Cancer Network J Natl Compr Canc Netw 14, 2; 10.6004/jnccn.2016.0022

Need Factors: These factors require the need to use health care services at the individual level. We classified the common types of chronic conditions using the organ domains.¹⁹ We selected this approach because of the synergism in treatment and self-management approaches. For example, the management of cardiometabolic conditions, such as diabetes and heart disease, have similar treatment and self-management strategies. Among men with prostate cancer, cardiometabolic diseases, respiratory diseases (eg, chronic obstructive pulmonary disease, asthma), and mental health conditions (eg, depression, other mental conditions) are commonly reported. These common chronic conditions were classified into 7 mutually exclusive categories: (1) cardiometabolic conditions only, (2) mental health conditions only, (3) respiratory conditions only, (4) cardiometabolic and mental health conditions, (5) cardiometabolic and respiratory conditions, (6) all 3 chronic conditions, and (7) none of these chronic conditions. However, because of insufficient numbers for the category “respiratory and mental health conditions,” this category was excluded from analysis.

All of the chronic conditions were identified during the baseline period using 1 inpatient claim or 2 outpatient claims for each of the chronic conditions listed in eAppendix 3. A very small percentage of men (<5%) had chronic conditions other than the cardiometabolic, mental health, and respiratory conditions; 42 other conditions were derived using the list of chronic conditions from published literature.¹ We calculated the total number of other chronic conditions during the baseline period and categorized these as “zero or one condition” and “two or more conditions.” The AJCC TNM classification system was used for prostate cancer staging. The cohort was restricted to elderly men with no-node-positive or nonmetastatic prostate cancer. Three groups of cancer stage at diagnosis were defined as T1 or less, T2, and T3 or T4.

Statistical Analyses

Because the NCHs were measured repeatedly during each of the 6 periods, the observations were no longer independent and were correlated. To account for correlated data, we used a statistical approach suggested by Liang et al,²⁰ which involves using the generalized estimating equations (GEE) with the specific correlation matrix. The general form of the model is given in the following equation:

View Expanded

where p_ij denotes the probability of NCHs for individual i, in period j. Time_ij denotes a specific period j for individual, i. CC_i denotes presence of particular types of chronic conditions at baseline period for individual, i. The model also included other independent variables, such as predisposing, enabling, need, and external environment characteristics measured for individual, i. βs are the regression coefficients to be estimated. We specified 3 types of correlation structures—exchangeable, autoregressive, and unstructured—to control for correlations between repeated observations.²¹ Because the parameter estimates of the association between NCH and types of chronic conditions remained the same under different correlation structures, only the results from GEE models with exchangeable correlation structures are reported. For ease of interpretation, the predicted probabilities of NCHs between the precancer and postcancer periods were compared. These probabilities were calculated using the parameter estimates from the GEE models, assuming an additive effect.²² Statistical analyses using GEE models were performed using STATA Statistical Software: Release 13 (STATA-13; StataCorp LP, College Station, TX). The predicted probabilities were derived using the “margins” command in STATA-13. The term “risk” is used to describe the results associated with odds ratios (OR). Although the risk ratio and OR will yield different estimates, for events with low prevalence (≤10%),²³ such as NCHs, these 2 measures will produce similar results.

Study Population

The study cohort (N=57,489) was primarily white (82.4%), within the age group of 67 to 74 years (58.3%; mean age, 74.6 years; SD, 5.1), and married (68.5%). Most men in the study cohort (80.5%) received treatment for cancer (Table 1).

NCHs During the Precancer and Postcancer Periods

Table 2 reports the rates of NCHs during each of the periods. The rates of NCHs were the highest during the postcancer period (t₄; ie, 120 days immediately after the cancer diagnosis) at 5.1% and lowest during the precancer period (t₂) at 3.2%, representing an increase of 1.9 percentage points. Among the categories of chronic conditions, the rates of NCHs during the postcancer period were significantly higher compared with the precancer period. Among elderly men with all 3 types of types of chronic conditions, the rates were 20.8% during the postcancer period (t₅) and 14.3% during the precancer period (t₃), an increase of 6.5 percentage points. Among elderly men with none of the chronic conditions, the rates were 2.6% during the postcancer period (t₄) and 1.3% during the precancer period (t₁), an increase of 1.3 percentage points. Similar increases were seen for other chronic condition categories.

Table 3 displays the ORs and adjusted ORs and 95% CI using the unadjusted (model 1) and adjusted GEE models (models 2 and 3) on NCHs with exchangeable correlation structures. Model 1 revealed that elderly men were 37% more likely to have NCHs during the postcancer period compared with the precancer period (OR, 1.37; 95% CI, 1.32, 1.41). In model 2, with adjustments for only the types of chronic conditions, elderly men were 38% more likely to have NCHs during the postcancer period (adjusted OR, 1.38; 95% CI, 1.33, 1.42) compared with the precancer period. In model 3, which controlled for all the independent variables, elderly men were 38% more likely to have NCHs during the post-cancer period compared with the precancer period.

Table 1.

Characteristics of Elderly Medicare Beneficiaries With Incident Prostate Cancer: SEER-Medicare Linked Database, 2000–2010^a

View Table

Probabilities of NCHs by Precancer and Postcancer Period

Table 4 displays the probabilities of NCHs in the precancer and postcancer periods for the types of chronic conditions. Based on the fully adjusted model (model 3), an increase in the probability of NCHs was observed for the postcancer period (P=.027) compared with the precancer period (probabilities: P=.028 vs .036; change in P=.009). Elderly men with all 3 types of chronic conditions had the greatest increase in the probabilities of NCHs during the post-cancer period (P=.141) compared with the precancer period (P=.107), whereas elderly men with none of the types of chronic conditions had the lowest change in the probability of NCHs (P=.021) during the postcancer period compared with the precancer period (P=.016).

Table 2.

Number and Percentage of Elderly Medicare Beneficiaries With Incident Prostate Cancer With Noncancer Hospitalizations During the Precancer and Postcancer Periods: SEER-Medicare Linked Database, 2002–2010^a

View Table