POLYPHARMACY IN CANCER PATIENTS: HEALTH-RELATED QUALITY OF LIFE, EXPENDITURES, AND ADVERSE EVENTS

Polypharmacy (PP), often defined as the use of five or more medications, is highly prevalent in patients with cancer. As the quantity of medications for treating cancer and comorbid conditions in patients with cancer become more numerous and diverse, it is important to understand the various ways in which patient health and economic outcomes may be adversely affected by prescribed medications. The purpose of this dissertation was to investigate three distinct associations between PP and the lives of patients living with cancer by estimating how PP (1) affects healthrelated quality of life (HRQoL), (2) is associated with healthcare expenditures, and (3) affects health complications (HCs). Approximately 25% of cancer survivors, individuals who were diagnosed with cancer and are still alive, report a decreased quality of life related to physical problems, and 10% report a decreased quality of life related to emotional issues, compared to their noncancer counterparts (10% and 6%, respectively). Specifically, cancer survivors report more mobility issues, inferior health, higher psychological distress, and more mental health needs. There is scant published literature describing PP in contributing to these outcomes. This study was conducted to address this gap to better inform cancer survivors, care providers, and health policy decision makers. Cancer was the sixth most expensive condition to treat in the United States (US) in 2015. Most cancers are estimated to have a decreasing incidence and increasing survival rate for the foreseeable future. A decreasing incidence may cause overall cancer-related expenditures to decline over time, but the prevalence of cancer coupled with the aging of the US population will result in an increase in the number of cancer survivors. Thus, expenditures during treatment through end of life are expected to continue to increase in coming years, as cancer survivors are estimated to increase from 15.5 million in 2016, to 26.1 million by 2040. Common cancer-related ailments such as pain, emesis, depression, venous thrombosis, and seizures can require prescription medications. With additional medications arises the risk for a health complication (HC). A HC, for the purposes of this study, is defined as an adverse health problem related to a drug, including adverse drug reactions, worsening of disease symptoms, falls, or overdoses. Although many HCs are preventable, they represent approximately 125,000 hospitalizations, over 3.5 million physician office visits, and an estimated 1 million emergency department visits each year in the general population. Previously identified risk factors for HCs in people with cancer, depending on the type of cancer, include PP, advanced stage of cancer, higher comorbidity, gender (for colorectal cancer), older age, and prior ER visits or hospitalizations. The purpose of the studies in this dissertation was to advance understanding of the role of PP on health and economic outcomes among people with cancer. We examined two data sources: (1) a large national survey database for manuscripts 1 and 2, and (2) a large, commercial claims database of privately-insured individuals for manuscript 3; both of which included United States (US) populations. Manuscript 1: The intent of this manuscript was to evaluate if an association exists between PP and HRQoL in cancer survivors in the US. The analysis used selfreported answers to questions about various demographic and clinical information captured in the Medical Expenditures Panel Survey (MEPS) database for even years 2008-2014. Respondents, who stated they were told that they had cancer, answered questions from the SF-12v2 about their physical and mental health, which were converted to the HRQoL measures PCS and MCS used for this analysis. This study focused on comparing cancer survivors, defined as having ≥ 5 prescribed medication classes in the year of the interview, with those with less than 5 medication classes. Differences among types of cancer were also explored in both descriptive and regression analyses. This study hypothesized that PP would lead to lower HRQoL as compared to patients not having PP. Of 10.1 million survivors per year included in this study, 45% were defined as having PP. We used ordinary least squares (OLS) regression to estimate that PP was associated with a statistically and clinically significant decrease in PCS scores among cancer survivors by 3.75 points. However, PP was not associated with a significant decrease in MCS scores. As such, PP should be analyzed closely in cancer survivors to ensure the best possible HRQoL. Manuscript 2: Healthcare expenditures are increasing in the US, and that is especially true for patients living with cancer. The objective of this manuscript was to determine if PP was associated with increased direct health care expenditures, and if differences in expenditure exist according to cancer type or setting of care. This aim was accomplished by using the same years and source of data as Manuscript 1, while modeling expenditure as a dependent variable. We hypothesized that PP was associated with increased health expenditures in total, by type of cancer and by setting of care. We used OLS regression with log transformed expenditures to obtain estimates of association between PP and increased health expenditures controlling for various demographic, socioeconomic, and clinical variables. PP was present in 43.9% of the 10.6 million (per year) cancer survivors in the study. PP was associated with a mean annual adjusted healthcare expenditure per cancer survivor of $13,266 (SD $3,766), which was significantly higher than those without PP $8,573 (SD 5,082, pvalue <.0001). Cancer survivors with PP accounted for 70% of total healthcare expenditures, yet only comprised 43.9% of the population. Manuscript 3: This study focused on newly diagnosed patients with breast, prostate, colorectal, or lung cancer and investigated if an association exists between PP and nonfatal health complications (HCs). The data source used was Optum Clinformatics DataMart (Optum, Eden Prairie, MN, USA), years 2010-2015. The database contains de-identified claims information with medical, prescription drug, enrollment, and other data tables. PP was measured as the use of ≥ 5 prescribed medication classes in the quarter (3 months) following incident cancer diagnosis. HCs was the dependent variable in the analysis and included a range of medical conditions known to be caused or worsened by effects of medications including falls, fractures, gastrointestinal bleeding, and delirium. Descriptive and logistic regression analyses were conducted to assess any associations between PP and HCs in a multivariable framework. This study hypothesized that HCs would occur more frequently among patients with PP than those without PP. In the primary analysis using multivariable LR modeling, PP was associated with 31% increased odds (adjusted odds ratio: aOR) of having ≥ 1 HCs, controlling for age, region, type of cancer, comorbidities, radiation and chemotherapy treatments. PP was significantly associated with a higher risk of having ≥ 1 HC in each cancer type (aOR: breast 1.37, 95% CI: 1.31-1.42; prostate 1.27, CI: 1.22-1.32; colorectal 1.26, CI: 1.16-1.36; lung 1.25, CI: 1.11-1.40). Active chemotherapy was associated with significantly increased odds of ≥ 1 HC in colorectal (aOR: 1.35, CI: 1.21-1.50) and lung (aOR: 1.33, CI: 1.15-1.54) cancers, but not significantly associated with breast or prostate cancers. Newly diagnosed patients with breast, prostate, colorectal, or lung cancer were all at a higher risk of having ≥ 1 HCs if defined as having PP compared to those without PP. Active chemotherapy treatment was associated with increased risk of HCs in colorectal and lung cancer patients, but not in breast or prostate cancer patients.

vii Acknowledgements I would like to thank my major professor and committee co-major member, Dr.
Stephen Kogut, for his help, guidance, and dedication in seeing me through the dissertation process. I have gained substantial knowledge under his tutelage and will be forever grateful. His knowledge of pharmaceutical medications was invaluable, as I have no background in this area, and it was pivotal to this dissertation work. Dr. Kogut is a world-class managed care pharmacist and researcher, and I simply could not have chosen a better advisor as I navigated the world of pharmacoeconomics and health outcomes research.
I would also like to thank my co-major committee member Dr. Ami Vyas for her diligence in reviewing my work and providing helpful suggestions on topics I did not know much about prior to this dissertation. She helped guide me through the research process and always responded promptly whenever I had a question. Without her assistance, I would not have been able to complete this research. For these things, I will always be grateful.
In addition to my co-major committee members, I would also like to thank the       scores. Ordinary least squares regression was used to assess associations between PP and HRQoL controlling for various demographic, socioeconomic, and clinical factors.

Introduction
Approximately 25% of cancer survivors, individuals who were diagnosed with cancer and are still alive, report a decreased quality of life related to physical problems and 10% report a decreased quality of life related to emotional issues compared to their noncancer counterparts (10% and 6%, respectively). 1 Specifically, cancer survivors report more mobility issues, inferior health, higher psychological distress, and more mental health needs. 1 They also worry about recurrence of their malignancy, new types of neoplasms, 2 and the possible long-term damage their cancer treatment may cause. 3 These concerns are additional to normal apprehensions about aging and the occurrence of comorbidities. 4 Approximately 70% of cancer survivors have one or more comorbidities. 5 Many observational studies have reported that cancer patients have poorer survival if they have comorbidities. 6 Cancer has a systemic impact on both body and mind. 1 Treating these impacts usually leads to greater use of prescription medications. 7 to 51.5% in the non-cancer control group. 5 The study found that the median number of unique prescription medications was 6 for cancer survivors, but only 4 for noncancer controls, despite the majority (55%) of survivors having been diagnosed ≥ 5 years previously. 5 As cancer survivors receive an increased number of concomitant medications, they become at an increased risk of dangerous adverse event occurrence. 9 Concerns about PP arise from certain harmful situations, such as when unforeseen or unintended drug effects and drug-drug interactions result in health complications. 10 Short-term, long-term, and late effects of cancer treatments, 11 related, in-part, to prescribed chemotherapy regimens may also negatively impact cancer survivors. 12 Treatment effects include a wide variety of impacts to organs, tissues, body development, growth, mood, feelings, actions, thinking, learning, memory, social and psychological adjustment, and risk of second cancers. 12 Treating these late effects to alleviate discomfort can require additional medications such as analgesics for pain, 13 and corticosteroids to help breathe normally, 14 among other drugs for symptoms which may decrease health-related quality of life (HRQoL). 1 A retrospective cohort study of adults (21 years and older) with arthritis conducted using the MEPS, found that PP was associated with significantly lower physical HRQoL scores. 15 Based on this evidence and the negative impacts of cancer on HRQoL, investigating the relationship between PP and HRQoL in the cancer survivor population was warranted. The objective of this study was to evaluate this 6 association between PP and HRQoL among cancer survivors living in the US using a nationally representative survey database.

Study design and data source
We used a multi-year cross-sectional study design to analyze the MEPS, a publicly available database which contains survey questionnaire responses of deidentified non-institutionalized persons and their families (households), their medical providers, and employers in the US. 16 The MEPS includes five interviews over the course of 2 calendar years conducted via computer assisted personal interviewing (CAPI). The multiple interviews allow for (1) analyzing how people's healthcare changes over time and (2) minimizing recall bias. 17 The MEPS also permits weighting of the data to produce nationally representative estimates of the US population for various healthcare analyses (e.g. expenditures, utilization of resources, insurance plans). 16 Two major components are included in the MEPS: household and insurance. 16 We selected the longitudinal, medical conditions, and prescribed medicines files from the household component for this study and linked them through a unique identifier for each individual. 16 We first used the medical conditions file to find individuals who reported having been diagnosed with cancer by using the cancer specific clinical classification codes; which are defined using the Clinical Classification Software provided by the Agency for Healthcare Research and Quality (AHRQ) which clusters diagnoses codes into a manageable number of categories. 18 Respondents were defined as cancer survivors during the interview process if they answered affirmatively to the question "Have you ever been told by a doctor or other health professional that you had cancer or a malignancy of any kind?" Those who confirmed having, or had, cancer were asked what type of cancer and their age at diagnosis. 19 We also used clinical classification codes and the International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) codes to identify concurrent chronic conditions. Further details regarding MEPS have been described elsewhere. 16 Sample selection We combined the MEPS data for years 2008, 2010, 2012, and 2014 for our analyses. In the MEPS process of interviewing, individuals are followed for two years, therefore we selected even years to avoid including repeated observations.
Respondents with cancer other than nonmelanoma skin cancer, who were at least 18 years of age at the time of response, were included in this study. We excluded those who had missing, negative, or zero person-level sample weights. To limit the effect of multiple cancers on the estimated relationship between PP and HRQoL, individuals were excluded if they had more than one type of cancer. 19 We also excluded those who died during the calendar year due to possible inflated prescription counts during end-of-life care and the possible effect terminal cancer would have on HRQoL scores.
In one retrospective cross-sectional study of 4,252 hospice patients across 11 states in the US, 35% of whom had cancer, the mean number of prescriptions was 15. 20 Figure   1 shows a flowchart of inclusion and exclusion criteria.

Dependent variable
Health-related quality of life The MEPS provides a prescriptions file with therapeutic medication class information which are linked to the Multum Lexicon database for analysis. 25 We used these therapeutic class details to determine the maximum number of classes of prescription medications the individuals were prescribed in one of the rounds that coincided with our study years. We defined PP as using ≥ 5 therapeutic classes of medications in one of the rounds of interviews, which is consistent with other definitions in published literature. 15 19 For patients who could not remember, or otherwise did not provide a response for age at diagnosis, we used a statistical multiple imputation procedure to assign time since cancer diagnosis. 27 Multiple imputation is an iterative process which uses the distribution of the observed data to estimate the true value of the missing variable. Values produced were used in regression analysis with the results pooled through statistical software to make valid inferences about the parameters and standard errors. To fit the structure of the variable, we used a minimum value of 0 (years) and maximum value of 85 (years). We achieved a relative efficiency of 99.0% and 99.1% with 25 imputations for our PCS and MCS models, respectively. 28 Comorbidities were selected from a list of priority health physical conditions provided by the MEPS and included the following: arthritis, chronic obstructive pulmonary disease, diabetes, and heart disease/cardiovascular ailments. 26 We chose these comorbidities based on MEPS' recognition that they are more prevalent, expensive, or especially relevant to healthcare policy as well as their impact on physical functioning. 29 To assess the influence of mental health conditions in our study population, we selected mood disorders (bipolar and depression) and anxiety disorders, using the MEPS designated mental health disorders clinical classification codes to identify these conditions for each patient (Appendix B). We dichotomized these conditions as either present (1) or absent (0). Healthcare encounters were defined as total provider or outpatient visits obtained from the household files and categorized based on quartiles into the following groups: 0-4, 5-9, 10-19, and ≥ 20 visits.

Statistical analysis
We used chi-square tests to determine the statistical significance of differences Due to the complexity of the survey design used in the MEPS; stratification, clustering, and weighting were performed to control for clustering and unequal probability design. 31 Significance tests were all performed at the α = 0.05 level. All analyses were conducted using SAS version 9.4 (SAS Institute, Cary, NC).

Results
The study population consisted of an unweighted total of 3,281 adult cancer  However, their study did not look at mental health conditions or PCS/MCS as outcomes.
High pill burden has been associated with increased use of inappropriate medication, thus increasing the risk of adverse outcomes. 33 In a medical chart review of 244 cancer patients aged ≥70 years receiving chemotherapy, 39% of severe potential drug interactions involved chemotherapeutic agents. 34 Additionally, the authors found that cancer patients' risk of a potential drug interaction increases with each additional medication, up to 100% when 8 or more medications were being taken concomitantly. 34 These risk estimates are higher than those reported in noncancer populations. 34 However, not all PP can be considered inappropriate, as multiple medication use does occur commonly in cancer survivors and may be the result of appropriately treating multiple conditions. A closer look at the root causes should be undertaken to try to eliminate excessive risks of inappropriate PP, such as lack of integrated and coordinated care, and possible contraindicated drug-drug interactions, which may lead to adverse events. 35 Conversely, not addressing adverse situations requiring medications in a timely manner may lead to avoidable complications.
However, this study was not intended to address appropriateness of prescribed medications and requires further investigation in the future.

PCS
Minimum clinically significant differences using the SF-12v2 range from 2-5 points from the population mean of 50. 36,37 The difference in the adjusted analysis was 3.75 points, which met the lower bound of minimum clinically significant threshold. PP in cancer survivors has been a concern for many years and this study confirms that use of multiple medications is still highly prevalent and warrants further attention in all cancer survivors. More consideration should be paid to continuity of care for cancer survivors to ensure appropriate medication use and non-medication management for chronic conditions. The study findings support the need for future research aimed at identifying the classes of prescription medications and the clinically significant drug-drug interactions that may cause survivors to report decreased physical QoL measured by PCS scores. Therefore, healthcare providers should evaluate the benefits and harms of prescribing multiple medications for cancer survivors.

Limitations
Some limitations exist due to the nature of the data source. As the MEPS has a survey design depending on a person's ability to remember various life events, responses are subject to recall bias. Also, the MEPS does not capture information on stage or severity of cancer which affects both PP and HRQoL, and so we could not adjust for these variables in our analyses. PP was not assessed for association with responses to specific mental or physical health states from the SF12-v2 since we used the summarized scoring totals for PCS and MCS; therefore, we cannot allude to any specific physical or mental functioning that may have been impacted by PP. Despite controlling for various comorbidities, severity of those illnesses could not be captured.
We cannot make assumptions as to a causal effect that the cancer treatment, or the cancer itself, may have had on specific chronic conditions.
As we evaluated the association between PP and HRQoL among cancer survivors by therapeutic class, some information may have been lost due to multiple drugs being used within the same class. Also, because this was cross-sectional, we cannot determine if an individual's PCS or MCS scores changed over time with the addition or subtraction of medications. As PP is a proxy for measure of disease burden, it is likely that survivors were appropriately taking multiple medications to help address comorbid conditions rather than their comorbid conditions were due to taking so many medications. However, this paper's intent was not to address appropriate versus inappropriate PP, hence further research is needed to better understand PP's impact on HRQoL. Though the association between cancer and PP has been reported previously, 5 to our knowledge, no study had evaluated how PP is associated with HRQoL among adult cancer survivors in the US.

Conclusion
In this cross-sectional study of community-dwelling cancer survivors in the US, PP was associated with lower PCS scores in certain types of cancer and those with Type of cancer included the following categorizations: prostate (included testicular cancer and cancer of other male genitals), cervical (included uterine, ovarian, other female cancers), colorectal (esophageal, stomach, colon, rectum and anus, liver and intrahepatic bile duct, pancreas, and other gastrointestinal cancers). c Income level: low (<200% above poverty line), medium (200% to 400% above poverty line), high (>400% above the poverty line). d Arthritis, chronic obstructive pulmonary disease (COPD), diabetes, heart conditions, anxiety and mood disorders are binary values (No=not present, Yes=present) listed in Appendix B. e Chi-square statistics were used to assess significant differences.  Chronic physical condition is a binary value (No=not present, Yes=present) for the conditions listed in Appendix B. The model fit was measured by its adjusted R 2 value (0.35). Type of cancer included the following categorizations: prostate (included testicular cancer and cancer of other male genitals), cervical (included uterine, ovarian, other female cancers), colorectal (esophageal, stomach, colon, rectum and anus, liver and intrahepatic bile duct, pancreas, and other gastrointestinal cancers). c Arthritis is a binary value (Yes=present) and is listed in Appendix B.

IMPLICATIONS FOR CANCER SURVIVORS:
Cancer survivors should be aware that increased prescription medication use is associated with increased total healthcare expenditures.

Introduction
National health expenditures in the United States (US) increased by 3.9% from 2016 to 2017, and made up 17.9% of gross domestic product, totaling $3.5 trillion dollars according to the Centers for Medicare and Medicaid Services (CMS). 1 On average, this amounted to $10,739 per person in the US. 1 The estimated 2017 national expenditures on cancer care was $147.3 billion and is expected to increase to $157.8 billion in the Medicare population alone by 2020. 2 Cancer was the sixth most expensive condition to treat in the US in 2015. 3 Most cancers are estimated to have a decreasing incidence and increasing survival rate for the foreseeable future. 2  Cancer survivors face several major challenges including financial hardship, body image/self-esteem issues, and anxiety surrounding fears of long-term side-effects of treatment and cancer recurrence. 6 As part of some cancer survivors' treatment plans (e.g. breast cancer), they may take medications (adjuvant hormonal therapy) for the following 5 to 10 years to lower the risk of recurrence. 7 Adjuvant therapy, for example, may increase the quantity of medications the survivor is to define them as having polypharmacy (PP), most commonly defined as the use of ≥ 5 concomitant medications. 8 Because survivors may have already been taking numerous medications to treat comorbid conditions and for palliative care one additional medication may now qualify as reaching the PP threshold. 9 PP is known to be highly prevalent and is associated with higher prescription costs among cancer survivors. 10 The types of services and healthcare products cancer survivors require included in the national health expenditure estimates are hospital care, physician and clinical services, other professional services (specialists), dental services, home health care, nursing care facilities, medical equipment, prescription drugs, and various other services and products. 1 Hospital-based care comprised 33% of health spending (the largest percentage), whereas physician and clinical services made up 20%, and other health and personal care services totaled 5%, with the other groups (excluding prescription drugs) comprising the remainder. 1 Prescription drugs dispensed through retail pharmacies accounted for roughly 10% of the $3.5 trillion dollars spent on the total population for healthcare in 2017; 1 and expenditures on cancer treatments are expected to increase over time as new drugs tend to be more expensive than current standards of care. 11 With prescription drugs comprising a significant portion of cancer-related expenditures, this study was conducted to examine the association between the number of medications prescribed and healthcare expenditures among cancer survivors. The objective of this study was to expand current knowledge by examining the relationship between polypharmacy and direct healthcare expenditures.
Quantifying the relationship between polypharmacy and healthcare expenditure in cancer is a requisite first step to understand the need for further study in determining to what degree increased healthcare expenditure is attributable to medication-related adverse events, or if polypharmacy is merely a proxy for burden of illness.

Study design and data source
We used a multi-year cross-sectional study design and utilized the Medical Expenditures Panel Survey (MEPS) database, a publicly available de-identified nationally representative database of the US. 12 The MEPS is a set of surveys containing nationally representative non-institutionalized persons, households (families and individuals), their medical providers, and employers throughout the US since 1996. 12 The MEPS uses a 2-year, 5-panel overlapping survey design of interviews.
We first used the medical conditions file to find individuals who reported cancer by using the cancer specific diagnosis codes through the Agency for Healthcare Research and Quality (AHRQ) clinical classification code system (Appendix A). We then linked the medical conditions, prescribed drugs, and household data files through a unique identifier for each individual cancer survivor. 12 We also used these clinical classification codes and the ICD-9-CM codes to identify concurrent chronic conditions using AHRQ's Elixhauser comorbidity codes. 13 Further details regarding the MEPS have been described elsewhere. 12 Sample selection The analytic sample included cancer survivors who were defined as adults (≥ 18 years old) with cancer who (1)  Individuals also were excluded if they had more than one type of cancer due to the inability to determine an association between the person's total expenditures and one cancer type. Men with breast cancer were excluded because of small sample size and lack of generalizability to female breast cancer survivors. People under the age of 18, with missing age information, had an age at diagnosis greater than their reported age, or who died during the panel year were excluded. Figure 1  payments from all reported sources to hospitals (facility and separately billed physicians), physicians, other medical, home health providers, for other providers, for dental providers, for miscellaneous expenses, and for prescriptions (Appendix C). 14 We created 5 distinct categories for expenditures: hospital, office-based, emergency room, prescriptions, and other medical expenses. Hospital expenditures were the summation of the expenditures from the hospital outpatient visits and inpatient stays.
Other medical expenditures included dental visits, home health providers (agency sponsored and paid independent provider), vision, and other medical expenses. Officebased, emergency room, and prescription expenditures were standalone categories within the MEPS. These expenditure groupings, when summed, equaled that of the total direct annual healthcare expenditures per cancer survivor.

Key independent variable
Polypharmacy (PP) The MEPS include a prescriptions file with therapeutic medication class information which are linked to the Multum Lexicon database for analysis. 15 We used these therapeutic class details to determine the maximum number of distinct classes of prescription medications the individuals were on in one of the panels that coincided with our study years. A consensus definition of PP does not currently exist; however, the most common definition in the literature is 5 or more concomitant medications. 8 We chose 5 or more classes of medications as our definition for PP based on our review of the literature which included several studies which used classification classes. 16,17 Other independent variables Demographic variables included age group, sex, race/ethnicity, US geographic region (Northeast, South, Midwest, and West), and marital status (married or not married). Socioeconomic variables included income (low, middle, high based on poverty level), insurance status (privately-, publicly-, or uninsured), and level of education (did not graduate high school, graduated but did not attend college, and at least some college level education). Time since cancer diagnosis was calculated by subtracting age at diagnosis, a variable included in the MEPS, from the patient's reported age. For patients who could not remember their age at diagnosis or was otherwise missing from the dataset, 51.7% total missingness, multiple imputation was used to fill in these missing values. We used the fully conditional specification (FCS) method, with all variables in the model creating 40 imputed data sets. 18 These data sets were then combined to get mean estimates across all variables.
Clinical variables included type of cancer, Elixhauser comorbidity score, and number of total provider encounters. Cancer type was grouped in the following manner: breast, prostate and other male genital (included testicular cancer), cervical and other female genital (included uterine, ovarian, other female cancers), colon and other gastrointestinal (GI) (stomach, liver, pancreas, and other GI cancers), melanoma, leukemias/lymphomas and other/unspecified (included lung). Lung cancer was grouped into the "other/unspecified" group due to small sample size. We used the Elixhauser comorbidity score to assess physical and mental diseases and disorders due to its well-established validity. The Elixhauser comorbidity score is the summation of approximately 31 comorbid conditions, which are first dichotomized as being present or absent in the patient, which we then categorized based on its distribution using quartiles to 0, 1, 2 or ≥ 3 (Appendix D). 13 Survivors with both complicated and uncomplicated diabetes or hypertension diagnoses were assumed to have the complicated, more severe, state of disease for these analyses. Provider encounters were defined as total provider or outpatient visits obtained from the household files and categorized into 0-4, 5-9, 10-19, and ≥ 20 visits based on quartiles. The primary analysis was to estimate the association between PP and total healthcare expenditures. Potential covariates were assessed in univariate OLS models for their statistical significance. If a variable was significantly associated with both PP and healthcare expenditures (F test p-value <0.10) it was included for assessment in a multivariable ordinary least squares (OLS) model. Multivariable (OLS) regression models were used to assess the relationship between PP and healthcare expenditures while controlling for significant covariates. An iterative process was used to include individual covariates one at a time into the multivariable OLS model based on its F test p-value. If a covariate was insignificant after placement into the model it was removed, and the model was run again with the next covariate, until no more significant covariates remained for analysis.
In a secondary analysis, the relationships between PP and healthcare expenditures were modeled by setting of care overall, and by setting of care and type of cancer. Separate models were created for each of the log expenditures from the 5 settings of care as the dependent variables, controlling for all significant covariates from the primary analysis. OLS regression was used to analyze mean expenditures by PP for each setting overall. To estimate the mean expenditures for a woman with breast cancer, we first created a cohort of women with breast cancer, then we separately modeled the per-patient mean expenditures with each setting as a dependent variable. OLS regressions were used to find mean differences in expenditures by PP in both secondary analyses.
Due to the complexity of the survey design used in the MEPS; stratification, clustering, and weighting were performed. Significance tests were all performed at the α = 0.05 level. Analyses were conducted using SAS version 9.4 (SAS Institute, Cary, NC).

PP and prescription medication utilization
There were approximately 55 million (weighted) prescribed medications per year for the total cohort of cancer survivors: 72.5% (40.1 million (M)) of these prescriptions were to respondents defined as having PP (not shown). Those without PP were on 90 distinct therapeutic classes compared to 93 for those with PP. Of those therapeutic classes, 92.6% (88/95) were not unique between those without PP and those with PP. Antihyperlipidemic medications comprised the most commonly prescribed chronically used therapeutic class for both those with (7.0%; 2.8M weighted prescriptions) or without (9.2%; 1.4M) PP. Beta-adrenergic blocking agents were the second most prevalent therapeutic class in those with PP (4.5%; 1.8M) (   Associations between PP and healthcare expenditures As seen in Table 5, PP was significantly associated with higher total annual mean log expenditures (β= 0.60, SE=0.05, p-value <.0001) when controlling for all significant variables (age, insurance, cancer type, comorbidity, provider encounters, and time since cancer diagnosis). This estimate represents an 82% increase in the total annual mean log expenditures due to a one-unit increase of the average number of cancer survivors having PP, holding all other variables at their reference class.
Several covariates had significant differences from their referent group in their association with total annual mean expenditures. All types of cancer examined, except for cervical and other female genital cancers, were significantly different from melanoma in their association with log expenditures while controlling for PP, age, insurance, time since cancer diagnosis, comorbidity, and provider encounters ( Table   5). Colon and other GI cancers was the most significantly different (β= 0.57, SE 0.11, p-value <.0001) from melanoma (reference group) with a 76% increase in mean log expenditures. Survivors with ≥ 3 comorbid conditions had a significant 37% increase from those without any comorbidities (β= 0.31, SE 0.06, p-value= <.0001). Survivors with public insurance (β= -0.12, SE 0.04, p-value= 0.0023) and without any insurance (β= -0.42, SE 0.14, p-value= 0.0029) were associated with lower mean log expenditures than survivors with private insurance (12% and 34%, respectively).
Those aged 50-64 were significantly different from their referent group of 18-49 years (β= 0.19, SE 0.06, p-value= 0.0014) with an associated 20% increase in mean log expenditures. Lastly, the number of visits to a provider was progressively significant and by far the most associated with increased mean log expenditures, with ≥ 20 encounters having a 540% increase in mean log expenditures (β= 1.85, SE 0.08, pvalue <.0001) ( Table 5). Time since cancer diagnosis of 2 years or less was significantly different in mean log expenditures compared to cancer survivors of 3 to 5 years by an increase of 36% (β= 0.31, SE 0.08, p-value <.0001).
After applying the subgroup-specific smear factors to the retransformed (exponentiated) estimates of the adjusted mean expenditures, the annual expenditure for someone with PP was $13,226 (SD $3,766), which was $4,513 more than survivors without PP at $8,753 (SD $5,082), and was significant (p-value <.0001).
The log expenditure estimates, subgroup-specific smearing factors, and final adjusted values are presented in Table 6.

Discussion
In this study, we found that approximately 44 of 100 adult cancer survivors per year were defined as having PP. PP was associated with significantly higher mean annual direct healthcare expenditures in all analyses, including unadjusted, adjusted, and our log transformed multivariable OLS model. Unadjusted total mean expenditures for cancer survivors in our study were higher than the 2012 estimated expenditures reported by AHRQ for the general population by 89% ($15,369 vs. $8,125, respectively). 21 For survivors with PP, the unadjusted difference in mean expenditures was associated with an increase of 70% in spending, with annual spend equaling $21,652 compared to $13,414 for survivors without PP. In the adjusted analysis, PP was associated with a significant 82% increase in the estimated log expenditures compared to those without PP.
By comparing the various settings of care for cancer survivors, we found that spending in the hospital setting is higher compared to the other settings, for both those with and without PP, which aligns with prior research. 1 Hospitalization has been linked to increased medication use in older cancer patients. 9,22 However, hospitalbased expenditures for those both with or without PP were approximately 42% of spend by setting, higher than that in the general population (33%). 1 The largest differences for cancer survivors with versus without PP by setting were office-based (23.7% vs. 33.7%, respectively) and prescription medications (22.4% vs. 14.6%, respectively). These amounts were also higher as a proportion of spending by setting compared to the general population (20% for office-based and 10% for prescription medication). 1 We combined expenditures from both inpatient and outpatient hospital visits while other studies have categorized hospital costs based solely on inpatient hospitalizations versus ambulatory (outpatient) hospital visits and office-based visits. 23 This may be why hospital-based expenditures were so much higher than office-based visits in this study. Our analysis provides further evidence that cancer survivors have substantially greater direct healthcare expenditures than the general population.
Differences existed among the different types of cancers, regarding overall healthcare expenditures for those with PP compared to those without PP. In the adjusted analyses, where we controlled for all significant variables, total annual mean expenditures for those with colon or other GI cancers were the highest, although not statistically significant from other cancer types. In a 2016 study of the economic burden (defined as annual medical expenditures plus annual productivity losses) of colorectal, female breast, and prostate cancer survivors in the US, which also used the MEPS (years 2008-2012), colorectal cancer was associated with the highest annual expenditures and productivity losses of the three cancer types. 23 Various risk factors for PP among cancer patients include comorbid conditions, hospitalization, and unnecessary prescribing. 9 Most cancer survivors in the current study had at least 2 comorbid conditions. When examined closer by PP, 6% of those without a chronic condition were defined as having PP; while 78.2% of those with ≥ 3 conditions had PP. In the log transformed expenditure model, having ≥3 comorbid conditions was associated with a 37% increase in expenditures compared to not having any comorbid conditions. Due to the cross-sectional study design, we cannot determine causality, but there was a clear association between PP and expenditures.
Future research that focuses on the examination of individual comorbid conditions and the number of prescriptions an individual are on both pre-and post-cancer diagnosis would elucidate this relationship further, as it was not the emphasis of this research.
We identified one paper that examined healthcare expenditure differences among cancer survivors with PP, in which they estimated median prescription expenditures as $1,633 vs. $784 in noncancer controls, but did not analyze total expenditure values. 10 Knowing that prescription costs significantly differ among cancer survivors with PP, as well as noncancer counterparts with PP, is important for addressing disparities among cancer survivors with and without PP. One reason for the disparities is that spending on anticancer medications doubled from 2012-2017 to almost $50 billion, with all oncology drugs launched in 2017 having list prices above $100,000. 24 In the US, the cancer drug market is expected to grow 12-15% annually by 2020, up to $100 billion. 24 This growth is expected to be driven by new launches and increased uptake of existing branded oncologics. 24 However, one positive trend is that oncology drug prices have risen at a slower rate (4.7%-6.4%) on average than that of the general branded market (6.9%) from 2012-2017. 24 We chose to incorporate total healthcare expenditures by PP among cancer survivors to see differences at the person and societal levels. In so doing, we hope that policymakers could be informed about how influential PP is on the healthcare system in the US.
This study determined that the total annual expenditure estimates for US cancer survivors for the period of 2008-2014, adjusted to 2017 dollars was $162.6 billion. According to research which used SEER-Medicare data, the estimated costs of cancer care will equal $157.8 billion by 2020. 25 However, when taking into consideration the declining incidence for most cancers, improving survival rates, and increasing costs, the authors estimated the total cost could amount to $172.8 billion. 25 Our estimate concurs with this as it is in the upper range of these two estimates.
Increased healthcare costs can have negative effects on both the individual cancer survivor and society as a whole. 26 For cancer survivors, concerns over outcomes previously linked to PP include adverse drug events, drug-drug interactions, increased morbidity, decreased survival, frailty/disability, and poor medication adherence. 9 On the societal level, policymakers may have to address the increased expenditures related to prevention initiatives and various adverse health-related outcomes in this expanding vulnerable population. PP may cause increased healthcare expenditures because of additional therapeutic monitoring, lab tests, physician office visits, and follow-up care planning.
Currently in the US, the focus of various advocacy and governmental groups focuses on lowering the cost of prescription medications. Although this is certainly needed, for cancer survivors whom are mostly covered by private or public insurances, a closer look at hospital and office-based expenditures should also be highly scrutinized due to the largest proportions of expenditures being spent in those areas.

Limitations
As this was a cross-sectional study design, no claim of causality can be made.
Other limitations may exist due to the way the data was collected, through Mean total annual expenditures for cancer survivors with PP was significantly higher than for those without PP, with significant differences attributable to setting of care, intensity of utilization, and type of cancer. Understanding this association is the first step to addressing the underlying causes of expenditure differences among those cancer survivors with versus without PP. ) (2008, 2010, 2012, 2014)       Office-based Prescriptions Other medical Emergency room Table 5 2008, 2010, 2012, 2014 (N=10,580,285).

CONCLUSIONS:
Newly diagnosed patients with breast, prostate, colorectal, or lung cancer with PP were all at a higher risk of having ≥ 1 nonfatal HCs as compared to those without PP. Active chemotherapy treatment was associated with increased risk of HCs in colorectal and lung cancer patients, but not in breast or prostate cancer patients.

Introduction
PP is defined most commonly in the literature as the concomitant use of ≥ 5 medications, 1 and one study found that 80% of newly diagnosed elderly (≥ 65 years) cancer patients met this criterion. 2 Patients with cancer often receive many medications, 3  Some cancer patients may, or may not, be using 5 prescribed medications at the time of their diagnosis. However, during the course of treatment for cancer, they may add new medications resulting in PP. One concern which arises from PP among older patients is the increased risk associated with use of potentially inappropriate medications (PIMs) that may have a deleterious effect on the patient's health. PP has been associated with PIMs previously. 7,8,9 PIMs are concerning for cancer patients as one study found that, of newly diagnosed cancer patients who visited ambulatory oncology clinics, the odds of using PIMs increased by 18% for each additional medication in those defined as having PP (≥ 5 concomitant medications) compared to those without PP. 10 Common cancer-related ailments such as pain, emesis, depression, venous thrombosis, and seizures can also necessitate additional medications. 10 The increased use of combinations of medications also increase the risk of drug-drug interactions (DDIs) among cancer patients, even among those not currently receiving antineoplastic treatments. 1 DDIs can result in a lack of effectiveness of one or all the drugs, enhance toxicity, and diminish a treatment's intended outcome. 11 Potential underlying risk factors for DDIs in cancer patients include mucositis and malnutrition causing impaired absorption, edema resulting from changes in a drug's volume of distribution, or excretion changes from renal and/or hepatic dysfunction. 12 Other factors include a patient's age, narrow therapeutic index of the drugs involved, and physiologic make-up. 13 DDIs may lead to various negative outcomes, including new health complications among patients with cancer, 13 and falls resulting in fractures which may cause delays in cancer treatments and alter the trajectory of the disease, care planning, or prognosis. 14 event, as a possible negative outcome resulting in patient harm or injury due to use of prescribed medications, 15 including medication errors, adverse drug reactions, allergic reactions, and overdoses. 16 To the authors' knowledge, PP associated with HCs in newly diagnosed cancer patients have not been thoroughly investigated in a large administrative claims database. The primary objective of this study was to estimate and describe the frequency of HCs in newly diagnosed cancer patients, with or without polypharmacy, in a multivariable framework.

Study design and data source
We conducted a retrospective cohort study to estimate the associated risk

Sample selection
The study population included adult individuals (≥ 18 years old) with an incident diagnosis of cancer (breast, prostate, colorectal, and lung) who had continuous enrollment in medical and prescription insurance throughout a 12-month lookback period through the end of follow-up for the first year following cancer diagnosis. Female breast, prostate, colorectal, and lung cancer cases were selected for our study because they are considered the four major cancers by the American Cancer Society. 14 A patient had to have at least 2 cancer diagnosis claim codes (including in situ and metastasis), defined by the International Classification of Diseases, 9 th Edition, Clinical Modification (ICD-9-CM) classification system, in the primary or secondary diagnosis field, which were at least 30 days apart in either the outpatient or inpatient setting (Appendix E). The patient's first cancer diagnosis was their index date. Patients with claims of a personal history of cancer within one year prior to their first ICD-9-CM code matching were excluded from the algorithm. Individuals were excluded if their incident diagnosis was not between January 1, 2011 and September 30, 2014. Men with breast cancer were excluded because the focus was on the four most commonly occurring cancers in the US. If an individual did not have any pharmacy claims in the year of follow-up they were excluded. People with more than one type of cancer were excluded, except those with metastatic codes to capture advanced stage diagnoses. Patients with less than one full year of data following incident diagnosis were excluded, including those who died. Figure 1 shows the inclusion and exclusion criteria in greater detail.
The key independent variable (IV) of interest was PP, defined as a patient filling ≥ 5 distinct medication classes at an outpatient pharmacy in the first quarter (3 months) following incident cancer diagnosis, not accounting for overlap or switching, with a cumulative sum of days' supply of at least 7 days, during the 3-month exposure window after the index date. Since no clear definition of PP exists in the literature, 2 we chose our definition based on published literature which used distinct therapeutic classes. 17,18 These factors, coupled with other research which stated that no single cut-point was optimal in defining PP in cancer patients, 19 but that ≥ 5 daily medications was a reasonable threshold for predicting multiple adverse events in elderly cancer patients, informed our decision to use ≥ 5 therapeutic classes as our threshold for PP.
However, to examine medication use with more accuracy and in a shorter time period than the aforementioned study, we used a claims database study. Medication classes were categorized using the American Hospital Formulary Service (AHFS) Pharmacologic-Therapeutic classification system. 20 Vaccinations, due to one-time administrations, and vitamin (A-E), due to their tendency to be more over-the-counter, medication classes were excluded from this definition. 21

Dependent variable
The primary outcome variable of interest was nonfatal health complications (HCs), and was dichotomized to either 0 (zero) HCs or ≥ 1 HC. HCs consisted of both specifically coded adverse drug-related events (ADEs) and other health conditions that are often associated with adverse effects of medications (e.g. organ toxicity, blood dyscrasias, falls). HCs were grouped into the following clinically meaningful categories: cardiovascular (CV), central nervous system and psychiatric (CNS), gastrointestinal (GI), hematologic (HEMA), metabolic (METB), skeletal (SKEL), and miscellaneous adverse drug-related events (ADE). The categories were curated from published literature based on their relevance to patients with cancer, PP, or both. 22,23 The outcomes selected were based on current literature and have been either (1) well documented in cancer patients, 15,22,23,24,25,26,27,28,29,30,31 and/or (2) were considered more likely in people with PP. 32,33,34,35,36 The goal of choosing these outcomes was to provide a selective list of short-term events which could have been precipitated by the combination of drugs in a population with a lowered immune system, mostly elderly (≥ 65), and who may have been increasing their medications due to anticancer treatment. Clinical events related to common drug interactions in one study included deep vein thrombosis, upper digestive hemorrhage, various other forms of bleeding, and neutropenia. 15 Other studies mentioned the risk of falling in elderly due to PP, 47 or in those with cancer because of the risk to treatment delays and potential cancerrelated outcomes as a result. 30,33 Other examples of specific HCs include fractures and arrhythmias (See appendix H for full list). HCs were measured in patients with cancer by using a claims-based algorithm searching for these complications using ICD-9-CM diagnosis codes. As part of the inclusion criteria, patients had to have continuous enrollment in both medical and prescription claims for the year following their incident diagnosis, thus they were alive throughout follow-up. The follow-up period in which these HCs were measured was during the 3 quarters following the exposure period (quarter 1) in which the presence of PP was determined.

Covariates
Demographic covariates were assessed during the 12-month baseline period and included age, sex, and geographic region. Clinical variables assessed at baseline included type of cancer, insurance plan-type, and Elixhauser comorbidity score.
Radiation and chemotherapy treatments were assessed after exposure. Cancer type was grouped in the following manner: breast (female only), prostate, lung, and colorectal using the ICD-9-CM codes listed in Appendix E. We chose to use the Elixhauser comorbidity score, excluding the 3 codes related to cancer, to assess physical and mental diseases and disorders based on the variety of ailments contained within, and its well-established validity. 37 The Elixhauser comorbidity score is the summation of various comorbid conditions which are dichotomized to represent a condition's presence (1) or absence (0) (Appendix F). We categorized the scores based on the overall distribution into 3 categories 0, 1-2, ≥ 3 conditions. Patients with both complicated and uncomplicated diabetes, or hypertension, diagnoses claims were assumed to have the more complicated stage of the disease for these analyses. This method was used to prevent double counting of the disease if a patient had both claims. Anticancer infusions and injections were identified using Healthcare Common Procedure Coding System (HCPCS) coding system in the outpatient setting (J codes J8500-J9999). The HCPCS coding system classifies similar medical products into categories for efficient claims processing. 38 If the individual received either an outpatient pharmacy prescription and/or a J code for an antineoplastic agent during the year following their incident diagnosis, they were defined as receiving active chemotherapy. Radiation was defined through Current Procedural Terminology (CPT) and HCPCS G codes (Appendix G). 39,40 Statistical analysis Descriptive statistics were used to describe the proportions of cancer patients by PP for each covariate. Chi-square tests were used to determine the statistical significance between PP and categorical covariates, as well as between PP and HCs.
Also, the percentages of PP in patients with a HC were described according to the type of cancer. Lastly, to provide information on the number and percent of different medication drug classes filled by those with or without PP, the 20 most filled medication drug classes were described.
Logistic regression (LR) modeling was used to examine associations between individual covariates and HCs. Variables which had statistically significant (p-value <0.10) association with both PP and HC were used in the multivariable LR modeling process. The multivariable LR model examined the relationship between PP and HCs, controlling for the covariates which were significantly related to both PP and HC in the univariate LR models. Collinearity amongst covariates was assessed by examining the condition indices and variance decomposition proportions. 41 However, no two independent variables were collinear and thus no variables were removed at this stage.
Covariates were added to the model sequentially based on their negative 2 Log Likelihood statistic (-2 Log L). Model comparisons were assessed through the Likelihood Ratio Test (LRT) which produced comparison statistics among models based on their intercept and covariates using the -2 Log L, where a better fitting model had a lower -2 Log L value. 41 A manual stepwise elimination process was used to remove variables with p-values higher than 0.05 significance to determine which of the remaining variables were still significant in the multivariable model. Lastly, comparison between model performance were assessed by changes in Akaike Information Criteria (AIC), and goodness-of-fit was tested by changes in c-statistic (concordance index) values. 42 The measure of effect was the adjusted odds ratio (aOR) comparing the risk (odds) that a person having PP experienced a HC versus those without PP, controlling for all other significant covariates.
The objective of a secondary analysis was to examine the relationship among PP with HCs by type of cancer, controlling for significant covariates (Table 3). To understand the relationship, four models were created (one for each cancer type) by first including the following covariates: sex (only for colorectal and lung cancers), age, region, insurance, comorbidity score, radiation therapy, and chemotherapy treatment. In these analyses, a manual backward elimination process was used to remove covariates that were not significant. First, the variable with the largest p-value (> 0.05) was removed. Next, the model was reanalyzed to determine if any of the remaining covariates became or remained insignificant. If a variable was insignificant (p-value >0.05) it was removed. This process was continued until only significant variables remained in the model. All statistical analyses were conducted using SAS software, version 9.4 (SAS Institute Inc., Cary, NC, USA).

Results
The analytic cohort consisted of 35,336 adult cancer patients (Figure 1). Of these, 14,573 (41.2%) adults were defined as having PP in the first quarter following incident cancer diagnosis (  Presented in Figure 4 are the percentages of HCs by PP for each type of cancer. Across each type of cancer CV complications occurred the most, with HEMA HCs as the second most frequent. Differences between PP and no PP groups were statically significant at p-value < 0.05 for each cancer type, with the exception of GI in colorectal and lung, and skeletal in lung.

Primary analysis: association between PP and nonfatal HCs
To determine the association between PP and nonfatal HCs in the analytic cohort, a multivariable LR model was created controlling for age, region, type of cancer, comorbidity burden, radiation therapy, and chemotherapy ( Figure 4). Excluded from this analysis were sex (due to the gender-specific nature of breast and prostate cancers) and insurance type (due to its insignificance during the model building Chemotherapy and radiation treatments were both significantly associated with a slightly increased risk of having ≥ 1 HC in the final multivariable LR model (aOR 1.07, 1.03-1.10, p-value <.0001 and aOR 1.06, 1.02-1.10, p-value= 0.0012, respectively). Age ≥ 75 years old was significantly associated with an increased risk of having the outcome of interest compared to those aged 50-64 years (aOR 1.39, 1.33-1.45, p-value <.0001). The Northeast was significantly associated with an increase in risk of having ≥ 1 HCs compared to those in the Midwest (aOR 1.08, 1.02-1.14, p-value= 0.0088). Figure 4 presents additional results pertaining to comorbidity level and use of chemotherapy or radiation.
Secondary analysis: associations between PP and HCs by type of cancer Four multivariable logistic regression models were created to assess the association between PP and HCs for each type of cancer ( Table 3). As mentioned previously, sex was excluded as an explanatory variable from the analysis for breast and prostate cancers, due to those cancers being sex-specific. Across all four models PP, age, and comorbidity were significant predictors of HCs. The association between PP and ≥ 1 HC and other main findings by type of cancer are described next.
In the model for women with breast cancer, PP was associated with a 37% increase in the odds of having ≥ 1 HC in the follow-up period (aOR 1.37, 1.31-1.42, pvalue <.0001) compared to those without PP. Each age group was significantly different from those aged 50-64 years old, with the oldest having a 26% increase in risk (aOR 1.26, 1.17-1.35, p-value <.0001). The West was the only region significantly different from the Midwest and associated with a decreased risk of having ≥ 1 HC by 18% (aOR 0.82, 0.77-0.88, p-value <.0001). The number of comorbid conditions and radiation therapy were significant, but chemotherapy was not (  Table 3.

Discussion
We used a large administrative claims database to describe the association between PP and the risk of having ≥ 1 HCs among newly diagnosed patients with breast (female), prostate, lung, and colorectal cancer controlling for significant covariates (age, sex, radiation therapy, chemotherapy, comorbid conditions, and geographic region). We also estimated associations between each type of cancer and HCs controlling for those covariates. In each multivariable LR model, PP was associated with a greater than 25% increase in the risk of having ≥ 1 HC.

Polypharmacy
In our study, we found that greater than 40% (2 in 5) of adult patients with newly diagnosed breast, prostate, colorectal, and lung cancers were defined as having PP in the first quarter following diagnosis. One study, which defined PP as ≥ 5 distinct medications, reported the prevalence of PP to be 64% in cancer survivors; however, this was a cross-sectional study with a more liberal definition of polypharmacy, which summed the medications used over two years. 43 Three studies reported the overall prevalence of PP in newly diagnosed cancer patients to be 80% (patients aged ≥ 65 years in US), 3 57% (in patients aged ≥ 70 years in Australia), 44 and 35% (patients also ≥ 70 years in Denmark). 45 However, all studies varied in their setting and collection methods. In the study that reported overall PP of 35%, lung cancer had the highest percentage of patients with PP (40.9%), compared to the other types of cancer: 32.9% (breast), 29.9% (colorectal), and 32.3% (prostate). 45 These rates were slightly lower than our results; however, that study was a case-control study where the controls did not have a cancer diagnosis at the index date. Although we did not have the same study design or source population, our results showed that PP, by type of cancer, was also highest in patients with lung cancer (64.0%).
PP was associated with a significantly higher risk of having ≥ 1 HC in all analyses, including unadjusted and adjusted LR models. By grouping HCs, we found that cancer patients with PP had higher proportions of HCs for different body systems compared to those without PP. For example, complications involving the cardiovascular system were more than double (19.2%) in patients with PP compared to those without PP (8.9% p-value <.0001). A study by Barber et al found that certain hormone therapies in breast and prostate cancer patients increased cardiac arrhythmias. 24 In a review of the impact chemotherapy has on cardiac arrhythmias, Tamargo et al reported inducement of a direct cardiac effect that can also be initiated or maintained by substrates created by comorbid conditions or the chemotherapy. 25 Hematologic HCs were the second most common, with 9.9% of patients with PP having at least one compared to 5.5% in those without PP. The hematologic HCs included in this study are well-established outcomes in patients with cancer; especially venous thromboembolisms and pulmonary embolisms which are known to increase after surgery and chemotherapy treatment. 26 The results of the primary analysis showed that PP was highly significant in its association with the risk of having HCs by 31% when controlled for age, region, type of cancer, comorbidity, chemotherapy, and radiation therapy. This means that patients with PP, which comprise 40% of those newly diagnosed with the four most common types of cancer, have a 31% higher risk of health complications overall. Polypharmacy has been associated with increased use of potentially inappropriate medications, which can cause adverse health outcomes among older patients. According to a study by Lund et al, which analyzed the Surveillance, Epidemiology, and End Results (SEER)-Medicare database, among 19,318 breast, 7,283 colon, and 7,237 lung cancer patients age 66 years and older, the number of PIMs changed after initial diagnosis of cancer during follow-up (6-23 months duration). 27 The increase in PIM dispensing was directly related to chemotherapy initiation in the first six months. They reported that for women with breast cancer PIMs decreased, while those with colon or lung cancer saw an increase. In our analysis, a decreased aORs for breast cancer patients, and increased aORs for lung, compared to the reference group (colorectal cancer) may be caused by a similar PIM risk. Lund et al did not study prostate cancer, but with the watchful-waiting or active surveillance approach recommended by the National Comprehensive Cancer Network (NCCN), a lack of additional medications for treatment may also decrease the risk of PIMs and thus decrease the odds of HCs. 46 The secondary analysis of PP among cancer types revealed cancer-specific differences for PP and some of the covariates. PP had the largest estimated risk in breast cancer patients of the four main cancers, with an increased risk of 37%. One explanation for this may be the influence of the covariates, specifically that chemotherapy was not significantly associated with HCs. Lund et al found that of 19,318 newly diagnosed patients with stage I-III breast cancer, PIMs declined (40% to 34%) after diagnosis and leveled off as chemotherapy use was curtailed beginning 3 months after incident diagnosis until 23-months follow-up. 27 For women with early stage breast cancer, they often receive surgery followed by radiation then hormone therapy, but not chemotherapy. 47 According to Edwards et al, women with any number of comorbid conditions are less likely to receive chemotherapy compared to those who have none. 48 Therefore, for the women who receive chemotherapy, they may have an advanced stage of breast cancer, and the risk of complications would not be significantly different. Our results showed that women between the ages of 65 and 74 years had a lower risk of HCs compared to those 50-64 years and this lack of chemotherapy may be why. As chemotherapy is not recommended for early stages of breast cancer in adults over 70, 48 or with having a high number of comorbidities, our findings suggest that these newly diagnosed breast cancer patients were in situ or invasive, but not metastatic. Whereas, those aged ≥ 75 years had the highest number of comorbid conditions (38.5%: not shown) compared to the reference group which had the largest percentage without comorbidity (35.5%: not shown). Also, radiation therapy was associated with more HCs which is logical since side-effects linked to radiation therapy may lead to exacerbating underlying conditions. The youngest age group was associated with a higher risk for HCs, which could be explained by 58% (not shown) of those aged 18-49 having no comorbidities, indicating they may have had a more aggressive form of cancer, as 59% of those aged 18-49 received chemotherapy treatment compared to 56.0% in the reference group. This higher rate of chemotherapy may have directly led to an increase in HCs.
Patients with prostate cancer and PP had a 27% increased risk of having ≥ 1 HC when controlling for age, region, and comorbidity score. Like patients with breast cancer, chemotherapy was not significantly different between those who had ≥ 1 HC and those who had none during follow-up. One explanation would be that men with prostate cancer tend to be diagnosed in their late 60s and early 70s, and the median age in this study was 69 years. Standard of care for patients with low-risk prostate cancer thus does not usually involve chemotherapy but may include hormone therapy.
Radiation therapy was also not significantly associated with the outcome of interest.
Differences in HCs from those 50-64 years old were also significant for those 18-49, but in prostate cancer younger age was protective (35% decrease in risk) because younger people, on average, had fewer comorbidities (43.4% of 18-49 had none compared to 31.7% in 50-64, 25.1% in 65-74, and 28.7% in ≥75) . Whereas those aged 65-74 were not significantly different than the reference group, but those ≥ 75 were significantly associated with an increased risk (71%) for HCs.
Patients with colorectal cancer and PP had a 26% increase in risk of having ≥ 1 HC. Unlike breast and prostate cancer, colorectal cancer occurs in both men and women. However, in the analysis men and women did not significantly differ in risk for the outcome. As with prostate cancer, younger age  was associated with a decreased risk (39%) and older age with increased risk (73%) of HCs. Also differing from breast and prostate cancer patients, chemotherapy was associated with an increased risk of HCs (35%). One explanation for the lowered risk in younger people, despite an increased risk associated with chemotherapy, could be that younger people had the lowest number of comorbidities (34.1% had none in 18-49 years old) compared to the referent age group (25.5%). Conversely, 75% of those ≥ 75 years had at least 1 comorbid condition.
PP was associated with a 25% increase in risk for the outcome in patients with lung cancer after controlling for sex, age, comorbidity, chemotherapy, and radiation.
Men had a 22% higher risk for having ≥ 1 HC than women. Again, since men smoke more and have shorter life spans in general than women, so at the advanced age when being diagnosed with lung cancer we would expect men to have a greater risk for HCs.
Both chemotherapy and radiation were significant. We would expect this to be the case since most lung cancers are diagnosed at a late stage. 49 Although surgery may be undertaken in limited scope, treatment often relies on chemotherapy and radiation to eliminate the disease. Having 3 or more comorbid conditions compared to no conditions increased the risk by 69%. Comorbid conditions such as COPD and emphysema are known to occur in people with lung cancer at diagnosis, which would increase the risk of having HCs.
We also noted differences in the association between HC events and type of cancer in the final multivariable LR model. In breast and prostate cancer patients, results showed these cancer types were less likely to have a HC compared to colorectal or lung cancer, and may be explained, in part, by the status of chemotherapy treatment. Being on chemotherapy treatment in both breast and prostate cancers was not significantly associated with HCs in their respective models (Table 2). In one study, which measured drug-related problems (DRP) (e.g. inappropriate drug, adverse drug reaction) in elderly (mean age 71.1 years) cancer patients, 77.6% were taking ≥ 3 chronic medications concurrently with intravenous chemotherapy and reported to have an average incidence of 3 DRP. 31 Interestingly, adverse drug reactions were reported to be caused by chemotherapeutic agents 85% of the cases; whereas, potential drugdrug interactions were related to chronic use medications 92.6% of the cases. 31 Similar to this analysis, the study on drug-related problems found a statistically significant increase in the odds of having a DRP when taking ≥ 5 medications. 31 However, intensity and duration of chemotherapy were unmeasured confounders in the analysis.

Limitations
Although efforts were made to address temporality by defining PP in the 1 st quarter following incident diagnosis, no assurances can be made that the individual was actively taking the medication preceding the event, or that any combination of medications directly caused the event to occur. Also, although comorbid conditions were controlled for with a summary score, no assessment was made in baseline to assess if the HCs were incident cases, thus allowing for the HCs to be chronic in nature. Further research is warranted that would focus more closely on individual cancers and HCs resulting from concomitant use of medications.
As with any administrative database analysis, the underlying data may lead to misclassification of some individual's cancer or comorbid status. Neither severity nor stage of cancer are included within the database as standalone variables, and hence were not controlled for in the analyses. As such, determination of stage or grade of cancer was not possible. Stage or grade of tumor would be a critical confounding variable, as these would determine the course of action for these patients regarding surgery, chemotherapy, and radiation treatments.
We were unable to conduct any analyses regarding race, as we did not have this variable in the database. Incidence rates for the four most common cancers studied in this manuscript vary by race. For instance, African Americans have higher incidence rates for prostate, colorectal, and lung cancers compared to White, Hispanic, and other racial/ethnic groups. 50 Intensity of infusion chemotherapy nor strength or dosing of prescription anticancer agents were analyzed for this analysis. The definition used to classify a newly diagnosed cancer patient as PP was based on the number of distinct medication classes and a minimum days' supply during the first quarter following diagnosis. This definition inherently may lead to under-or overestimation of the number of patients with PP because most adherence rates for chronic medications would require reaching 80% adherence. Some definitions of PP have counted individual medications, including counting infusions over their day of activity, which would mean counting them more than once per month to account for administration cycles. Also, we did not account for infusions or injections which may have not been related to anticancer treatment. The focus of defining PP was for outpatient pharmacy filled medications and therefore inpatient drug usage, over-the-counter, and complementary and alternative drugs were not included as potential contributors to PP in this analysis.
Although medications were described in this analysis, no formal statistical tests were conducted to assess associations between their concomitant use and HCs. We examined common HCs associated with PP and cancer patients. The study was designed to use medication class because the mechanism of action within drug class would be the same despite different ingredients.

Conclusions
Newly diagnosed patients with breast, prostate, colorectal, or lung cancer who had PP were all at a higher risk of having ≥ 1 health complication compared to those without PP. When analyzing by type of cancer and controlling for age, sex, comorbidity, chemotherapy and radiation therapy, PP was associated with an increased risk of HCs by over 25% per cancer type. Active chemotherapy treatment was associated with increased risk of HCs in colorectal and lung cancer patients, but not in breast or prostate cancer patients. Chemotherapy was dichotomized into two groups based on absence or presence of at least 1 outpatient prescription claim using American Hospital Formulary System (AHFS) of classification coding or a Healthcare Procedure Coding System Level II (HCPCS) in the range of J8500-J9999 in the follow-up year post-index claim. c Elixhauser Comorbidity Score is the summation of a dichotomized variable for absence or presence of various health conditions found in Appendix F. In this analysis, 4 of the original 31 disease (states) coding groupings were excluded as 3 related to cancer conditions and 1 related to an outcome of interest (arrhythmias). Baseline refers to the time from the index date (first cancer diagnosis) up to 365 days prior to the index date. d Code sets for health complications (HCs) are in Appendix G. HCs were assed in the 2 nd , 3 rd , and 4 th quarter following a patient's incident cancer diagnosis. e CNS= central nervous system. 1.2 Other diabetes* 1,633 1.5 Notes: Total number of unique prescription classes filled for those without PP and those with PP were 48,116 and 107,619, respectively. *= drug class name was diabetes mellitus, but to not confuse it with biguanides (metformin) they are listed as Other diabetes.    (1). aORs in bold font indicate statistical significance where the 95% confidence interval did not cross 1.0 at alpha < 0.05. NS = not significant during backward elimination modeling. Since each type of cancer was modeled separately, aORs for variables without statistical significance are not shown. N/A = not applicable to breast and prostate cancer models due to sex-specific inclusions. Model c-statistics by type of cancer were as follows: breast 0.65; prostate 0.67; colorectal 0.68; lung 0.67. 657 Note: *= must have both ICD9CODX and CCCODEX codes. Due to inclusion/exclusion criteria for the cancer survivor population in the study, the Elixhauser coding for lymphoma, metastatic cancer, and solid tumors without metastasis are excluded as comorbid conditions. MEPS uses 3-digit ICD9CODX and CCCODEX codes. The search algorithm only counted a medical condition as present or absent, and no double-counting occurred if a patient had both the ICD9CODX and CCCODEX codes. If a survivor had both diabetes complicated and uncomplicated, preference was given to complicated. If a survivor had hypertension complicated and uncomplicated, preference was given to complicated.