Insufficient knowledge and misconceptions regarding hospice remain significant barriers to hospice utilization. Hence, understanding how patients and caregivers learn about hospice is necessary to address disparities in hospice access and utilization. Given its recent rise as a health information source, the Internet is likely a principal method patients and caregivers use to learn about hospice. In this study, I used Google Trends to examine trends in search volumes for hospice- and palliative care-related search terms on Google and relationships between search volumes, hospice utilization, and mean per-patient hospice length of stay across U.S. regions and nationally. Overall search volumes for “hospice” and “palliative care” have increased from 2014 to 2018, while those for “hospice care”, “home hospice”, and “palliative” have waned. In all regions and the U.S., search volumes for all terms were positively associated with hospice utilization. Regarding mean per-patient hospice length of stay, the U.S., but not all regions, exhibited reciprocal positive associations between search volumes for certain hospice-related terms and negative associations for palliative care-related terms. These results suggest that, to some extent, searching for hospice-related information online may encourage patients and caregivers to consider hospice and engage earlier in the hospice utilization process. These findings may influence the extent to which disparities in online health information-seeking behaviors regarding hospice are considered relevant to addressing disparities in hospice utilization.
Hospice care in the U.S. is growing, with a nearly sixfold increase in enrollees from 1994 to 2014 (Finnigan-Fox et al., 2017). Timely utilization of hospice is associated with effective pain management at the end of life, preservation of patient dignity, improved quality of life for patients, and decreased rates of depression and complicated bereavement for caregivers (Finnigan-Fox et al., 2017; Wright et al., 2008). However, an estimated 40% of eligible patients do not receive hospice care (Cagle et al., 2016), and racial minority populations (Hispanics and African Americans particularly) are underrepresented in the hospice patient distribution (Haas et al., 2007). Among key factors contributing to reluctance toward hospice utilization are insufficient knowledge about hospice services and misconceptions about hospice philosophy. These misconceptions include equating hospice with “giving up”, assuming hospice is only for the final weeks or days of life, assuming hospice disrupts ongoing patient-physician relationships, failing to realize that hospice cannot provide concurrent curative care, among others (Cagle et al., 2016; Mrig & Spencer, 2018; Shalev et al., 2019; Vig et al., 2010). Hence, understanding the strategies that patients and caregivers use to learn about hospice is a necessary step toward promoting timely hospice utilization.
Recently, individuals are increasingly utilizing the Internet as a resource for obtaining health and healthcare information, given the easily accessible and ubiquitous nature of websites and a constantly updated variety of published content. Among respondents in one national survey, 81% of adults reported Internet use, and 72% of those respondents reported seeking health information online. Although Internet use generally declines with age, those age 65 or older represent the fastest growing population of Internet users (Fischer et al., 2014). While the Internet commonly functions as a secondary or tertiary source of health information behind healthcare professionals and family or friends for older adults (Fischer et al., 2014; Magsamen-Conrad et al., 2019), it is plausible that patients and caregivers routinely utilize online resources when seeking information about hospice. The breadth of hospice-related information available online corroborates this notion (Corn et al., 2011; Finnigan-Fox et al., 2017).
The current landscape for the integration of web-based resources in patient care warrants exploration of the relationship between Internet search volume for hospice- and palliative care-related terms and hospice utilization. Understanding disparities in access to or willingness to pursue online information about hospice could reveal new insights about and help address disparities in hospice utilization.
Objective and hypotheses
Here I explore trends in Google search volumes for hospice- and palliative care-related search terms across U.S. states, relationships between said search volumes and the normalized number of hospice users across regions and the aggregate U.S., and relationships between said search volumes and mean per-patient hospice length of stay (LOS) across the same geographical distribution. Although hospice and palliative care differ in that palliative care involves symptom management concurrent with curative treatment while hospice is provided in the last six months of life post-treatment termination (Shalev et al., 2019), I exploratorily included search terms for both in this study. Given the aforementioned connection between health information-seeking behaviors and healthcare utilization, I hypothesized that search volumes for hospice- and palliative care-related search terms would be positively and reciprocally associated with hospice utilization and mean LOS and that these associations would be more pronounced in areas of greater hospice utilization.
Google search terms and volumes
To obtain Google search volumes for hospice- and palliative care-related search terms, I accessed Google Trends via the gtrendsR package (v1.4.7) in R (v4.0.2). Google Trends provides search volumes, based on the number of searches for a given term normalized to the total number of searches in a geographic area during a specified time frame. These indices are then relativized to the highest measured volume across the time frame and presented on a 0-100 scale (Bail et al., 2018; Dreher et al., 2018; Effenberger et al., 2020). Search terms included “hospice”, “hospice care”, “home hospice”, “palliative care”, and “palliative”, which were among the most popular terms related to hospice and palliative care across U.S. states based on Google Trends data and in previous studies, such as Finnigan-Fox et al. (2017). For each year (defined as from January 1 to December 31) individually from 2014 to 2018, I made separate calls to the Google Trends database using the gtrends() function. In each call, I queried for all five search terms together, specified “web” as the Google product to query, and limited search results to the geographic U.S. and to the default U.S. English setting. I collected search volumes across each state and District of Columbia for each search term. When search volume levels were unavailable for a minority of states (i.e., Alaska, North Dakota, and Wyoming in the case of “home hospice” only), they were considered negligible and treated as 0.
As measures of hospice utilization, I collected data regarding the number of Medicare hospice beneficiaries and mean per-patient hospice LOS (the total number of covered days on hospice divided by the number of hospice beneficiaries) for each state and District of Columbia from 2014 to 2018. For 2014 to 2016, these data were available in the Centers for Medicare & Medicaid Services (CMS) Hospice Utilization and Payment Public Use Files for each year, archived under the Hospice Providers subsection of the Legacy Medicare Provider Utilization and Payment dataset series. For 2017 and 2018, these data were filed as Medicare Hospice Utilization by State datasets under the Medicare Utilization for Part A archive. Since normalization of hospice user count based on Medicare beneficiary count was necessary, I obtained the total number of Original Medicare (i.e., non-Medicare Advantage) enrollees for each state and District of Columbia from 2014 to 2018 via the Kaiser Family Foundation (KFF) Total Number of Medicare Beneficiaries dataset.
Race and socioeconomic status (SES)
Race and SES have been implicated as social determinants of hospice utilization. Specifically, racial and ethnic minorities and individuals with lower SES exhibit a lower likelihood of hospice utilization (Haas et al., 2007). To account for variations in hospice utilization and search volumes for hospice- and palliative care-related terms due to race and SES, I collected data on racial composition and the median household income for each state and District of Columbia. For the former, I obtained the proportion of Caucasian individuals for each location and year from 2014 to 2018 via the KFF Population Distribution by Race/Ethnicity dataset. For the latter, I collected median household incomes for the same duration based on 2018 U.S. dollars from the U.S. Bureau of Census Current Population Survey Annual Social and Economic (ASEC) Supplement for 2019.
Internet penetration rates
Additionally, Internet access is associated with greater health literacy (Bailey et al., 2015), while disparities or disruptions in Internet access are associated with decreased online health information-seeking behaviors (McCloud et al., 2016). Hence, to account for the likelihood of higher search volumes in areas of higher Internet penetration rates (Bail et al., 2018), I collected data regarding the proportion of households with Internet access in each state and District of Columbia from the National Center for Education Services (NCES) Digest of Education Statistics subsection of their Annual Reports. These data span 2015 to 2018 (data from 2014 were unavailable). Households were characterized as having Internet access if residents could access the Internet at home via paid subscription to a mobile device-based cellular data plan; broadband Internet service installed via cable, fiber optic, or digital subscriber line; satellite Internet service; dial-up Internet service; or other service.
I associated each state and District of Columbia to their corresponding U.S. Census Bureau-designated region (Northeast, Midwest, South, or West). For the generation of map diagrams, I retrieved Federal Information Processing Standards (FIPS) codes for each state and District of Columbia via the usmap package (v0.5.1) in R.
Both Medicare hospice user and total Medicare beneficiary data were skewed right. To generate more normal distributions, I log-transformed these data. Other variables were normal and did not necessitate transformation.
For the analysis of each of the following relationships, I generated separate ordinary least-squares multiple linear regression models for each region and combinatorially for each of the five search terms. Regression coefficients in all models were standardized by z-score using the lm.beta package (v1.5.1) in R. First, I generated a set of models predicting the log-transformed number of Medicare hospice users as a function of the search volume for a given term. Covariates included the log-transformed number of Medicare beneficiaries, the proportion of Caucasian individuals in state populations, and state median household incomes. Second, I queried the inverse relationship, predicting search volume for a given term based on the log-transformed number of Medicare hospice users. Covariates for this model included the log-transformed number of Medicare beneficiaries, the proportion of Caucasian individuals in state populations, state median household incomes, and state household Internet access rates. Third, I generated a set of models predicting the mean per-patient hospice LOS as a function of the search volume for a given term. Covariates included the log-transformed number of Medicare hospice users, the log-transformed number of Medicare beneficiaries, the proportion of Caucasian individuals in state populations, and state median household incomes. Fourth, I queried the inverse relationship, predicting search volume for a given term based on the mean per-patient hospice LOS. Covariates in this model included the log-transformed number of Medicare hospice users, the log-transformed number of Medicare beneficiaries, the proportion of Caucasian individuals in state populations, state median household incomes, and state household Internet access rates.
While log-transformation of certain data obfuscates determination of effect sizes, these models still provide insight into the general positive or negative directionality of queried relationships.
Trends in search volume for hospice- and palliative care-related terms
Aggregately in the U.S., “hospice” was the most commonly searched term in each year except 2014, when it marginally trailed “palliative”. Only “hospice” and “palliative care” exhibited net increases in search volume from 2014 to 2018, with percent increases of 10.5% and 66.5% respectively. Meanwhile, “hospice care” and “palliative” experienced significant declines in search volume with percent decreases of 17.1% and 16.0% respectively; “home hospice” remained comparatively stable with a percent decrease of 5.5% (Table 1).
Fig. 1 highlights trends in search volume for hospice-related terms by state. Notably, search volume for “hospice” increased sharply in the Midwest and South from 2014 to 2015 before stabilizing in mostly the 80-100% range. States in the West exhibited the lowest volume; gradual increases have occurred, albeit not to the extent of other regions. States in the Northeast stably maintained high volume from 2014 to 2018. Contrastingly, trends for “hospice care” and “home hospice” were more variable, consistent with their lower search volume compared to “hospice”. For “hospice care”, all regions exhibited declines in mean search volume. These declines were most precipitous between 2017 and 2018, as the Northeast, South, and West experienced large (12.5-19.2%) decreases. For “home hospice”, the Northeast exhibited the highest concentration of moderate or high search volumes; however, in all regions, volumes were lower and more fluctuant than those for “hospice” or “hospice care”.
Fig. 2 highlights trends in search volume for palliative care-related terms by state. To a greater extent than for “hospice”, mean search volume for “palliative care” showed significant growth from 2014 to 2015 in all regions, ranging from a 52.1% increase in the Northeast to a 78.9% increase in the West. These volumes stabilized in subsequent years, with high-volume states concentrated particularly in the Northeast and Midwest. For “palliative”, significant declines in search volume occurred in all regions, with especially precipitous declines in the Midwest and Northeast between 2014 and 2015, before stabilizing in mostly the 41-60% or 61-80% range.
Predicting hospice utilization from search volumes
Fig. 3 summarizes linear regression models predicting the effect of search volumes for hospice- and palliative care-related terms on the log-transformed count of Medicare hospice users in each region and aggregately for the U.S. In all models, all variables had a statistically significant association (p < 0.01) with hospice user count, and the number of Medicare beneficiaries was expectedly the most significant predictor. In all U.S. models, each search term had a positive association with hospice user count. The strength of association was greatest for “hospice”, followed by “home hospice”, and lowest for “palliative”. In most regional models, hospice-related search terms exhibited a positive association with hospice user count (only “home hospice” was negatively associated in the models for the Northeast and South regions). For all three hospice-related terms, these associations were strongest in the Northeast and West, which marginally exhibited lower hospice user counts (normalized for Medicare beneficiary counts) compared to other regions. Regarding palliative care-related terms, “palliative care” had a positive association in all regions except the West, and “palliative” had negative associations, albeit of minimal strength, in all regions except the Midwest. These data suggest that despite the effects of Medicare population, race, and income, search volumes for hospice-related terms—more so than palliative care-related terms generally—correlate with hospice utilization to some extent.
Interestingly, in all regions except the Midwest, Caucasian proportion and median household income had slightly negative associations (p < 0.01) with hospice utilization, standing in contrast to existing literature (Haas et al., 2007; Wang et al., 2015). Refinement of models to include proportions of non-white races and other metrics of SES, including mean educational attainment levels, unemployment rates, and poverty rates, is warranted.
Predicting search volumes from hospice utilization
To examine the inverse effect of hospice utilization on search volumes for hospice- and palliative care-related terms, I generated the linear regression models summarized in Fig. 4. Log-transformed hospice user count did not have a statistically significant association with search volume for any term.
Interestingly, in all models, mean household income had a statistically significant association (p < 0.01)—negatively in models for all hospice-related terms (except in the West and aggregate U.S. for “home hospice”) and positively for both palliative care-related terms. Although this requires more substantiation, this may suggest that financial flexibility or insurance coverage for continued curative treatments may delay hospice-related information seeking. For the former, associations were strongest in the Midwest; for the latter, associations were strongest in the South. Caucasian proportion also exhibited a positive, statistically significant association (p < 0.01) with search volume for all terms except “hospice care” aggregately for the U.S. Regionally, only the Northeast consistently maintained this trend (p < 0.01). Notably, Internet penetration rates were not a significant predictor of volume for any term.
Predicting mean per-patient hospice LOS from search volumes
Fig. 5 summarizes linear regression models predicting the effect of search volumes on mean per-patient hospice LOS in each region and aggregately for the U.S. For “hospice” and “hospice care”, the U.S. exhibited a statistically significant positive association (p < 0.01); the West was the only region that showed a significant association, which was negative, for “hospice” (p < 0.05). For “home hospice”, the West and U.S. exhibited positive associations, while the Northeast exhibited a negative association (p < 0.01). Interestingly, for “palliative care” and “palliative”, the Midwest and U.S. had negative associations (p < 0.05).
Further, in all models, both Caucasian proportion and median household income had significant negative associations (p < 0.01 in most cases) with mean LOS in all regions, including the aggregate U.S., except the Northeast. This may further corroborate delays in hospice utilization as a result of financial flexibility.
Predicting search volumes from mean per-patient hospice LOS
To examine the inverse effect of mean per-patient LOS on search volumes, I generated the linear regression models summarized in Fig. 6. Among hospice-related terms, mean LOS had positive, statistically significant associations with volume for only “hospice” and “home hospice” for the West and U.S. (p < 0.01 in all cases, except p < 0.05 for “hospice” in the West), but not for other locations. Among palliative care-related terms, mean LOS had a negative association (p < 0.05) with volume for “palliative” in the Midwest and U.S.
In all models for “hospice” and “hospice care” and in most models for “home hospice” (except in the West and U.S.), median household income had a negative association with search volume (p < 0.01). This association was ubiquitously positive for palliative care-related terms (p < 0.01). For all terms in aggregate U.S. models, Caucasian proportion was positively associated with search volume (p < 0.01). Regionally, this relationship was positive and significant (p < 0.05 minimum) for all terms except “hospice care” in primarily the Midwest. Further, Internet penetration rates were insignificant in all models.
In this study, I used Google Trends to explore the relationship between search volumes for hospice- and palliative care-related terms and hospice utilization Regionally and aggregately in the U.S., search volumes for “hospice” and “palliative care” increased rapidly from 2014 to 2018, while those for “hospice care”, “home hospice”, and “palliative” declined. Importantly, in the aggregate U.S., search volumes for each search term were positively associated with hospice utilization; however, there were variations regionally. The inverse relationship was insignificant. This may suggest that searching for hospice-related information online may be a precursor to hospice utilization. Although differences in normalized hospice user counts per region were marginal, this relationship appeared slightly more salient in areas where per capita hospice utilization was less. Although further corroboration is needed, this may suggest that patients or caregivers may turn to the Internet to obtain information about hospice when other information sources or support structures that conduce to hospice utilization are less readily available. Regarding the relationship between mean per-patient hospice LOS and search volumes, national, but mostly no regional, reciprocal associations were found. That search volumes for hospice-related terms exhibited some positive associations may suggest that searching for hospice-related information online may allow patients and caregivers to be more adequately informed about hospice and consequently more likely to engage in upstream conversations about hospice or consider it earlier. Albeit unsubstantiated, that some negative associations were found between search volume for palliative care-related search terms and mean LOS may suggest that searching for palliative care-related information online may encourage patients or caregivers to select palliative care over hospice, based on need. These relationships are further summarized in Fig. 7.
Nonetheless, this study has several limitations. First, Google Trends does not provide absolute search volumes. This limits the interpretability of search volume patterns longitudinally, as changes for states treated as the point of reference in a given time range may be more dramatic than they really are. Albeit less accessible, one useful alternative that warrants consideration in future investigations is Google AdWords, which reports average monthly search volumes for specified terms. Albeit reported as monthly averages in the form of 10-point ranges, these volumes represent absolute levels. Additional benefits of Google AdWords include the availability search volumes for low-frequency search terms, which Google Trends often does not, and the availability of search volumes at city and county levels. Lastly, Google AdWords provides search volume results from not only Google, but also other leading Internet search engines (Bail et al., 2018).
A second limitation pertains to algorithmic confounding: search algorithms are not static, and changes therein may influence the reliability of data derived from said algorithms. (Lazer et al., 2014). Of note, Google Trends introduced an improvement to their data collection system on January 1, 2016. It is possible that this update disproportionately impacted search volume levels for lower-frequency search terms, such as “home hospice”. Separately, an inability to capture user intention when performing searches may skew a minority of search result data. Because the Google search algorithm provides autocompleted search suggestions when users begin typing search queries (Bail et al., 2018), it remains possible that some queries may be misleading (e.g., inadvertently searching “hospice care” when “hospice” was intended, searching “palliative care” when “palliative” was intended, or conflating “home hospice” and “home health”). Although this possibility cannot be readily accounted for, the search terms included in this study were general, reducing the possibility of variability in specific search suggestions based on user demographic characteristics or search history.
A third limitation pertains to an inability to rigorously assess temporality in the relationships between search volume and hospice utilization. Despite aforementioned results, it remains unclear whether searches for hospice- or palliative care-related terms precede or follow the onset of hospice utilization (i.e., it is possible that prospective or pre-contemplative hospice patients or caregivers search for said terms to learn more about hospice prior to onboarding or that patients or caregivers already enrolled in hospice search for said terms). One method to address this limitation in future investigations involves employing Google AdWords to query low-frequency search terms that may indicate temporality, such as “how to enroll in hospice” or “starting hospice”. Further, the applicability of online behaviors to offline actions remains unverified (Bail, 2014; Golder & Macy, 2014), especially considering that user characteristics are unknown. Searching for hospice- or palliative care-related information does not guarantee access to reliable sources of information (Corn et al., 2011; Finnigan-Fox et al., 2017), limiting the generalizability of discussed relationships. Searching for information alone does not account for the impacts of Internet skill and health literacy on health information-seeking behaviors (Jacobs et al., 2017; McCloud et al., 2016); hence, integration of these data into models is warranted in future investigations.
Despite these limitations, these results suggest that, to some degree, variability in Internet search volumes for hospice- and palliative care-related terms can elucidate disparities in hospice utilization. These digital trace data can be helpful in identifying areas where health information-seeking behaviors regarding hospice lag and potentially conduce to clinical interventions to increase patient and caregiver agency in pursuing hospice-related information online in addition to their Internet skill and hospice-related health literacy. Proximately, clinicians should be encouraged to connect patients and caregivers with reliable online sources of information about hospice.
|Region||Year||Number of Medicare hospice users||Number of Medicare beneficiaries||Mean per-patient hospice LOS||% Caucasian||Median household income||% of households with internet access||“Hospice” search volume||“Hospice care” search volume||“Home hospice” search volume||“Palliative care” search volume||“Palliative” search volume|
Table 1: Mean values for all variables categorized by year and by region or aggregately for the U.S.