Keiki Hinami, MD, MS

METHODS

RESULTS

DISCUSSION

METHODS

RESULTS

DISCUSSION

METHODS

RESULTS

DISCUSSION

Hospitalists are a growing workforce numbering over 40,000 physicians, one‐third of whom are women.[1] Flexibility of work schedules and control over personal time have been the traditional selling points of the specialty.[2] Multiple studies of physician work life reveal growing physician dissatisfaction and a high prevalence of burnout.[3] To mitigate burnout risk, leaders in hospital medicine recognize the importance of creating a sustainable profession that offers both job and career satisfaction as well as work‐life balance and, importantly, fairness within the work environment.[4] Although success in some of these endeavors has been realized sporadically, sustaining work‐life balance and fairness in the specialty remains a work in progress, whereas evidence of high job attrition and pay inequities remain.[5, 6]

Pay inequity for women relative to men continues to be pervasive in medicine, including among early‐career physicians, researchers, and various specialists.[6, 7, 8, 9, 10, 11, 12, 13] The earnings gap seems to persist for physicians, even as federal efforts such as the Fair Pay Act of 2013 and the Paycheck Fairness Act of 2014 aim to end wage discrimination.[11, 14] Differences in specialty, part‐time status, and practice type do not mitigate the disparity.[8, 10, 15] Additional explanations have been proposed to explain the variability, including gender differences in negotiating skills, lack of opportunities to join networks of influence within organizations, and implicit or explicit bias and discrimination.[12, 16, 17, 18, 19, 20]

The earnings gap is also a consequence of what is commonly called the glass ceiling.[18, 19] Most agree that obstacles to fair advancement of women include absence of collaborative environments and role models who have successfully achieved work‐life balance.[17, 20, 21, 22] Somewhat surprisingly, women leaders in medicine seem to suffer greater income disparity than nonleaders; this income gap is prevalent among leaders in other elite professions as well.[7, 23] It is unknown whether women physicians' emphasis on work‐life balance, seen repeatedly in surveys, explains any of the pay disparity.[24] Little research to date has examined whether work‐life priorities of women in hospital medicine differ from men.

In this study, we sought to examine differences in job priorities between men and women hospitalists. In particular, we examine the relative prioritization of substantial pay to job satisfaction. We also examined gender differences in work patterns and earnings to explore potential sources of the persistent gender earnings gap.

METHODS

We analyzed data from the 20092010 Hospital Medicine Physician Worklife Survey, the design of which is detailed elsewhere.[4] Briefly, a 118‐item survey was administered by mail to a stratified sample of hospitalists from the Society of Hospital Medicine database and 3 large multisite hospitalist groups. A single survey item asked respondents to identify up to 4 out of 12 most important domains to their satisfaction with a hospitalist job. The domains were distilled from focus groups of nationally representative hospitalists as described previously,[4] and the survey item allowed up to one‐third of these domains to be identified as respondents' personal priorities. The list included: optimal variety of tasks, optimal workload, substantial pay, collegiality with other physicians, recognition by leaders, rewarding relationships with patients, satisfaction with nurses, optimal autonomy, control over personal time, fairness within organization, ample availability of resources to do job, and organizational climate of trust and belonging. We tabulated and ranked the frequency with which respondents selected each satisfaction domain by gender. Due to the nonstandard format of the survey item, we a priori decided to analyze only responses that were completed as instructed.

We also used demographic data including detailed work characteristics, clinical and nonclinical workload, total pretax earnings in 2009 as a hospitalist, and self‐identification as leader of their hospital medicine group. Respondent characteristics were tabulated and gender differences were tested using the t test, rank sum test, and the Fischer exact test as appropriate. We also listed the number of nonrespondents for each item. In estimating gender differences in earnings, we opted to use multiple‐imputation techniques to more conservatively account for greater variance inherent in the presence of missing data. Consistent with existing guidelines,[25] we demonstrated that item responses were not missing monotonically by visually inspecting patterns of nonresponse. We further demonstrated that data were missing at random by showing that response patterns of completed survey items did not predict whether or not a given variable response was missing using logistic regression models. We found no significant differences between respondents with complete and missing data. We verified that appropriate regression models for each variable on every other variable converged. We used Stata 13.1 (StataCorp, College Station, TX) to perform multiple imputations using chain equations (mi impute chain) to create 10 imputed tables for 7 normally distributed continuous variables using the ordinary least squares method, 3 non‐normally distributed variables using the predictive mean matching method, 2 nonordinal categorical variables using the multinomial logit method, and 1 binary variable using the logit method.[26, 27] Gender, pediatric specialty status, region of practice, and whether or not respondents prioritized substantial pay for job satisfaction were used as regular variables without missing data points.

Differences in earnings were assessed using a multivariate ordinary linear regression model applied to the imputed datasets fitted by forward selection of explanatory variables using P < 0.20 in bivariate analysis for inclusion and manual backward elimination of all statistically nonsignificant variables. We tested the significance of the women leader interaction term in the final parsimonious model. We used the usual significance threshold of P < 0.05 for inferences. Our analysis of publicly available anonymous data was exempt from IRB review.

RESULTS

Of the 816 survey respondents (response rate 25.6%), 40 either omitted the item soliciting work priorities or completed it incorrectly. Data from the remaining 776 respondents were used for the present analysis. Respondent characteristics are tabulated in Table 1. The characteristics of hospitalists by age, gender, specialty, practice model, and practice region were representative of US hospitalists from other surveys.[28]

Several gender differences were seen in the characteristics of hospitalists and their work (Table 1). Women compared to men hospitalists were less likely to be leaders, more likely to be pediatricians, work in university settings, and practice in Western states. Women compared to men, on average, were younger by 3 years, worked fewer full‐time equivalents (FTEs), worked a greater percentage of nights, and reported fewer billable encounters per shift. They were also more likely to be divorced or separated.

Job satisfaction priorities differed for women and men hospitalists. Table 2 lists job satisfaction domains in descending order of the frequency prioritized by men. The largest proportion of women and men prioritized optimal workload. However, although substantial pay was prioritized next most frequently by men, more women prioritized collegiality and control over personal time than substantial pay.

Key differences in individual characteristics, work patterns, and indicating substantial pay as a priority were associated with self‐reported total earnings in 2009 from respondents' work as a hospitalist. As shown in Table 3, the inclusion of detailed productivity measures such as FTE, days of monthly clinical work, and estimated number of daily billable encounters yielded a model that explained 33% of variance in earnings. After adjusting for significant covariates including pediatric specialty, practice model, geography, and amount and type of clinical work, the estimated underpayment of women compared to men was $14,581. Hospitalists who prioritized substantial pay earned $10,771 more than those who did not. The female x leader interaction term testing the hypothesis that gender disparity is greater among leaders than frontline hospitalists was not statistically significant ($16,720, P = 0.087) and excluded from the final model.

DISCUSSION

In a national stratified sample of US hospitalists, we found gender differences in job satisfaction priorities and hospitalist work characteristics. We also confirmed the persistence of a substantial gender earnings disparity. Lower earnings among women compared to men hospitalists were present in our data after controlling for age, pediatric specialty, practice model, geography, type of clinical work, and productivity measures. The gender earnings disparity noted in 1999[6] persists, although it appears to have decreased, possibly indicating progress toward equity. We showed that women hospitalists' relative tendency not to prioritize pay explains a significant portion of the residual income gap.

Hoff examined hospitalist earnings in a large national survey of hospitalists in 1999. Our cohorts differed in age and experience (both lower in the Hoff study). An estimated $24,000 ($124,266 vs. $148,132) earnings gap between women and men was greater than our estimate of $14,581 following an interval of 10 years. Although survey items differed, both studies found that women were less interested in pay than men when considering a hospitalist job. Hoff also found that work setting and attitudes about pay and lifestyle were significantly related to earnings. We extended the previous analyses of gender differences in job satisfaction priorities, work, and demographic characteristics to explain the earnings gap and understand how it may be remedied.

When considering job satisfaction, we found that more men than women prioritized substantial pay and that prioritization of substantial pay was directly related to higher earnings. Therefore, fewer women prioritizing pay partly explains women's lower earnings. Reasons for why fewer women prioritize pay were not assessed in this study but may include factors like being part of 2‐income households and competing commitments.[29] Priorities may even be influenced by empirically observed gender differences in discussions of financial matters, governed by cultural norms. Such norms may implicitly sanction employers to offer women less pay than men for the same or similar work. Women may disadvantage themselves by negotiating less or less well than men for higher starting and promotion salaries. They may be perceived more negatively than men when they do negotiate pay, leading to unintended negative consequences such as loss of social networks, decreased likability, and even loss of job offers.[29, 30, 31, 32]

More women prioritized optimal variety of tasks (6th most prevalent among women and 10th among men). Women who highly rate optimal variety of tasks as a job satisfier may choose positions in which they teach, perform research, and participate in hospital committees and quality‐improvement work, but offer lower pay. Yet hours per month doing nonclinical work was not significantly different between men and women, nor associated with earnings differences in our earnings models. Understanding whether women self‐select into hospitalist jobs with like‐minded colleagues to achieve complementary fit or end up supplementing their skills with hospitalists with different priorities may inform strategies to reduce the gender income disparity.[5] Unlike disparities between various hospital medicine groups, systematic disparities within practices risk generating low levels of organizational fairness and burnout among employees.

Not surprisingly, productivity was positively associated with earnings but did not fully account for the gender earnings gap. Our data demonstrated that women, on average, were associated with work characteristics that expectedly generate less compensation. For example, women were younger, more often part time, academic, pediatric, less often leaders, and reported fewer billable encounters compared to men. These differences account for some of the earnings gap between men and women, but these factors were controlled for in the earnings model. In addition, our analysis may have underestimated the gap by not incorporating loss of fringe benefits from part‐time status and not comprehensively counting incentive pay associated with high productivity. Other work patterns more commonly associated with women suggest an imbalance in reimbursement. More women than men work nights that are often compensated at higher rates than daytime work, yet their average pay was less, suggesting that compensation for night work may need to be adjusted to reflect its unique burdens and responsibilities.[33]

Although the gender pay gap was not more extreme among leaders compared to frontline hospitalists in our data, the trend, nonetheless, underscores an important consideration. Whereas clinical work is paid for in mostly measurable ways, pay for leadership may be influenced by intangible factors such as reputation, negotiation, and confidence that may disadvantage women relative to men.[7, 19, 23, 34, 35, 36] Efforts to overcome implicit gender bias should be most effective when we consciously couple fair promotion of women to leadership with fair compensation commensurate with their male peers.[21]

Our data are vulnerable to nonresponse bias.[1, 4, 5] Post hoc analyses demonstrated that distributions of age, gender, practice model and region of our respondents were similar to other nationally representative cohorts of hospitalists. Consequently, we believe our data can make valid estimates about a nationally representative sample of hospitalists. However, we acknowledge several additional weaknesses of self‐reported data, including recall bias and accuracy of productivity figures, which were rounded to variable significant digits by respondents. Earnings analysis using this data was intended to be exploratory, but the findings echoed analyses using more authoritative data sources.[11] Still, we made inferences conservatively by adopting multiple imputation techniques for dealing with nonresponse surveys in adherence to established reporting guidelines.[25] We also note several limitations relevant to multiple imputations. The greater prevalence of missing data for survey items soliciting earnings and the number of billable encounters suggest they were not truly missing at random as assumed. However, we showed that missingness is unrelated to the variables under study, justifying use of the technique. The wider measures of variance derived from multiple imputations make us vulnerable to not detecting associations that may exist.

The gender earnings gap found in hospital medicine echoes the gap found in multiple medical specialties, including but not limited to pediatrics, academic medicine, gastroenterology, and plastic surgery.[7, 8, 9, 11, 12, 13, 37] Hospital medicine employment models and practice patterns have important structural differences compared to previously studied populations that could mitigate factors contributing to women physicians' lower earnings. However, despite well‐defined working hours, lack of control over the number of patient encounters per day and high prevalence of hospital‐employed practice models, the gender earnings gap persists. We showed that lower prioritization for pay may reflect the self‐selection of women into lower paying jobs. Unmeasured factors, including implicit bias and differences in negotiations, social networks and mentoring opportunities[38, 39] may also contribute to pay differences between men and women hospitalists. As hospital medicine tackles gender inequities and other disparities, strategies to assess and address fair physician compensation must be on the table.

Hospitalists are a growing workforce numbering over 40,000 physicians, one‐third of whom are women.[1] Flexibility of work schedules and control over personal time have been the traditional selling points of the specialty.[2] Multiple studies of physician work life reveal growing physician dissatisfaction and a high prevalence of burnout.[3] To mitigate burnout risk, leaders in hospital medicine recognize the importance of creating a sustainable profession that offers both job and career satisfaction as well as work‐life balance and, importantly, fairness within the work environment.[4] Although success in some of these endeavors has been realized sporadically, sustaining work‐life balance and fairness in the specialty remains a work in progress, whereas evidence of high job attrition and pay inequities remain.[5, 6]

Pay inequity for women relative to men continues to be pervasive in medicine, including among early‐career physicians, researchers, and various specialists.[6, 7, 8, 9, 10, 11, 12, 13] The earnings gap seems to persist for physicians, even as federal efforts such as the Fair Pay Act of 2013 and the Paycheck Fairness Act of 2014 aim to end wage discrimination.[11, 14] Differences in specialty, part‐time status, and practice type do not mitigate the disparity.[8, 10, 15] Additional explanations have been proposed to explain the variability, including gender differences in negotiating skills, lack of opportunities to join networks of influence within organizations, and implicit or explicit bias and discrimination.[12, 16, 17, 18, 19, 20]

The earnings gap is also a consequence of what is commonly called the glass ceiling.[18, 19] Most agree that obstacles to fair advancement of women include absence of collaborative environments and role models who have successfully achieved work‐life balance.[17, 20, 21, 22] Somewhat surprisingly, women leaders in medicine seem to suffer greater income disparity than nonleaders; this income gap is prevalent among leaders in other elite professions as well.[7, 23] It is unknown whether women physicians' emphasis on work‐life balance, seen repeatedly in surveys, explains any of the pay disparity.[24] Little research to date has examined whether work‐life priorities of women in hospital medicine differ from men.

In this study, we sought to examine differences in job priorities between men and women hospitalists. In particular, we examine the relative prioritization of substantial pay to job satisfaction. We also examined gender differences in work patterns and earnings to explore potential sources of the persistent gender earnings gap.

METHODS

We analyzed data from the 20092010 Hospital Medicine Physician Worklife Survey, the design of which is detailed elsewhere.[4] Briefly, a 118‐item survey was administered by mail to a stratified sample of hospitalists from the Society of Hospital Medicine database and 3 large multisite hospitalist groups. A single survey item asked respondents to identify up to 4 out of 12 most important domains to their satisfaction with a hospitalist job. The domains were distilled from focus groups of nationally representative hospitalists as described previously,[4] and the survey item allowed up to one‐third of these domains to be identified as respondents' personal priorities. The list included: optimal variety of tasks, optimal workload, substantial pay, collegiality with other physicians, recognition by leaders, rewarding relationships with patients, satisfaction with nurses, optimal autonomy, control over personal time, fairness within organization, ample availability of resources to do job, and organizational climate of trust and belonging. We tabulated and ranked the frequency with which respondents selected each satisfaction domain by gender. Due to the nonstandard format of the survey item, we a priori decided to analyze only responses that were completed as instructed.

We also used demographic data including detailed work characteristics, clinical and nonclinical workload, total pretax earnings in 2009 as a hospitalist, and self‐identification as leader of their hospital medicine group. Respondent characteristics were tabulated and gender differences were tested using the t test, rank sum test, and the Fischer exact test as appropriate. We also listed the number of nonrespondents for each item. In estimating gender differences in earnings, we opted to use multiple‐imputation techniques to more conservatively account for greater variance inherent in the presence of missing data. Consistent with existing guidelines,[25] we demonstrated that item responses were not missing monotonically by visually inspecting patterns of nonresponse. We further demonstrated that data were missing at random by showing that response patterns of completed survey items did not predict whether or not a given variable response was missing using logistic regression models. We found no significant differences between respondents with complete and missing data. We verified that appropriate regression models for each variable on every other variable converged. We used Stata 13.1 (StataCorp, College Station, TX) to perform multiple imputations using chain equations (mi impute chain) to create 10 imputed tables for 7 normally distributed continuous variables using the ordinary least squares method, 3 non‐normally distributed variables using the predictive mean matching method, 2 nonordinal categorical variables using the multinomial logit method, and 1 binary variable using the logit method.[26, 27] Gender, pediatric specialty status, region of practice, and whether or not respondents prioritized substantial pay for job satisfaction were used as regular variables without missing data points.

Differences in earnings were assessed using a multivariate ordinary linear regression model applied to the imputed datasets fitted by forward selection of explanatory variables using P < 0.20 in bivariate analysis for inclusion and manual backward elimination of all statistically nonsignificant variables. We tested the significance of the women leader interaction term in the final parsimonious model. We used the usual significance threshold of P < 0.05 for inferences. Our analysis of publicly available anonymous data was exempt from IRB review.

RESULTS

Of the 816 survey respondents (response rate 25.6%), 40 either omitted the item soliciting work priorities or completed it incorrectly. Data from the remaining 776 respondents were used for the present analysis. Respondent characteristics are tabulated in Table 1. The characteristics of hospitalists by age, gender, specialty, practice model, and practice region were representative of US hospitalists from other surveys.[28]

Several gender differences were seen in the characteristics of hospitalists and their work (Table 1). Women compared to men hospitalists were less likely to be leaders, more likely to be pediatricians, work in university settings, and practice in Western states. Women compared to men, on average, were younger by 3 years, worked fewer full‐time equivalents (FTEs), worked a greater percentage of nights, and reported fewer billable encounters per shift. They were also more likely to be divorced or separated.

Job satisfaction priorities differed for women and men hospitalists. Table 2 lists job satisfaction domains in descending order of the frequency prioritized by men. The largest proportion of women and men prioritized optimal workload. However, although substantial pay was prioritized next most frequently by men, more women prioritized collegiality and control over personal time than substantial pay.

Key differences in individual characteristics, work patterns, and indicating substantial pay as a priority were associated with self‐reported total earnings in 2009 from respondents' work as a hospitalist. As shown in Table 3, the inclusion of detailed productivity measures such as FTE, days of monthly clinical work, and estimated number of daily billable encounters yielded a model that explained 33% of variance in earnings. After adjusting for significant covariates including pediatric specialty, practice model, geography, and amount and type of clinical work, the estimated underpayment of women compared to men was $14,581. Hospitalists who prioritized substantial pay earned $10,771 more than those who did not. The female x leader interaction term testing the hypothesis that gender disparity is greater among leaders than frontline hospitalists was not statistically significant ($16,720, P = 0.087) and excluded from the final model.

DISCUSSION

In a national stratified sample of US hospitalists, we found gender differences in job satisfaction priorities and hospitalist work characteristics. We also confirmed the persistence of a substantial gender earnings disparity. Lower earnings among women compared to men hospitalists were present in our data after controlling for age, pediatric specialty, practice model, geography, type of clinical work, and productivity measures. The gender earnings disparity noted in 1999[6] persists, although it appears to have decreased, possibly indicating progress toward equity. We showed that women hospitalists' relative tendency not to prioritize pay explains a significant portion of the residual income gap.

Hoff examined hospitalist earnings in a large national survey of hospitalists in 1999. Our cohorts differed in age and experience (both lower in the Hoff study). An estimated $24,000 ($124,266 vs. $148,132) earnings gap between women and men was greater than our estimate of $14,581 following an interval of 10 years. Although survey items differed, both studies found that women were less interested in pay than men when considering a hospitalist job. Hoff also found that work setting and attitudes about pay and lifestyle were significantly related to earnings. We extended the previous analyses of gender differences in job satisfaction priorities, work, and demographic characteristics to explain the earnings gap and understand how it may be remedied.

When considering job satisfaction, we found that more men than women prioritized substantial pay and that prioritization of substantial pay was directly related to higher earnings. Therefore, fewer women prioritizing pay partly explains women's lower earnings. Reasons for why fewer women prioritize pay were not assessed in this study but may include factors like being part of 2‐income households and competing commitments.[29] Priorities may even be influenced by empirically observed gender differences in discussions of financial matters, governed by cultural norms. Such norms may implicitly sanction employers to offer women less pay than men for the same or similar work. Women may disadvantage themselves by negotiating less or less well than men for higher starting and promotion salaries. They may be perceived more negatively than men when they do negotiate pay, leading to unintended negative consequences such as loss of social networks, decreased likability, and even loss of job offers.[29, 30, 31, 32]

More women prioritized optimal variety of tasks (6th most prevalent among women and 10th among men). Women who highly rate optimal variety of tasks as a job satisfier may choose positions in which they teach, perform research, and participate in hospital committees and quality‐improvement work, but offer lower pay. Yet hours per month doing nonclinical work was not significantly different between men and women, nor associated with earnings differences in our earnings models. Understanding whether women self‐select into hospitalist jobs with like‐minded colleagues to achieve complementary fit or end up supplementing their skills with hospitalists with different priorities may inform strategies to reduce the gender income disparity.[5] Unlike disparities between various hospital medicine groups, systematic disparities within practices risk generating low levels of organizational fairness and burnout among employees.

Not surprisingly, productivity was positively associated with earnings but did not fully account for the gender earnings gap. Our data demonstrated that women, on average, were associated with work characteristics that expectedly generate less compensation. For example, women were younger, more often part time, academic, pediatric, less often leaders, and reported fewer billable encounters compared to men. These differences account for some of the earnings gap between men and women, but these factors were controlled for in the earnings model. In addition, our analysis may have underestimated the gap by not incorporating loss of fringe benefits from part‐time status and not comprehensively counting incentive pay associated with high productivity. Other work patterns more commonly associated with women suggest an imbalance in reimbursement. More women than men work nights that are often compensated at higher rates than daytime work, yet their average pay was less, suggesting that compensation for night work may need to be adjusted to reflect its unique burdens and responsibilities.[33]

Although the gender pay gap was not more extreme among leaders compared to frontline hospitalists in our data, the trend, nonetheless, underscores an important consideration. Whereas clinical work is paid for in mostly measurable ways, pay for leadership may be influenced by intangible factors such as reputation, negotiation, and confidence that may disadvantage women relative to men.[7, 19, 23, 34, 35, 36] Efforts to overcome implicit gender bias should be most effective when we consciously couple fair promotion of women to leadership with fair compensation commensurate with their male peers.[21]

Our data are vulnerable to nonresponse bias.[1, 4, 5] Post hoc analyses demonstrated that distributions of age, gender, practice model and region of our respondents were similar to other nationally representative cohorts of hospitalists. Consequently, we believe our data can make valid estimates about a nationally representative sample of hospitalists. However, we acknowledge several additional weaknesses of self‐reported data, including recall bias and accuracy of productivity figures, which were rounded to variable significant digits by respondents. Earnings analysis using this data was intended to be exploratory, but the findings echoed analyses using more authoritative data sources.[11] Still, we made inferences conservatively by adopting multiple imputation techniques for dealing with nonresponse surveys in adherence to established reporting guidelines.[25] We also note several limitations relevant to multiple imputations. The greater prevalence of missing data for survey items soliciting earnings and the number of billable encounters suggest they were not truly missing at random as assumed. However, we showed that missingness is unrelated to the variables under study, justifying use of the technique. The wider measures of variance derived from multiple imputations make us vulnerable to not detecting associations that may exist.

The gender earnings gap found in hospital medicine echoes the gap found in multiple medical specialties, including but not limited to pediatrics, academic medicine, gastroenterology, and plastic surgery.[7, 8, 9, 11, 12, 13, 37] Hospital medicine employment models and practice patterns have important structural differences compared to previously studied populations that could mitigate factors contributing to women physicians' lower earnings. However, despite well‐defined working hours, lack of control over the number of patient encounters per day and high prevalence of hospital‐employed practice models, the gender earnings gap persists. We showed that lower prioritization for pay may reflect the self‐selection of women into lower paying jobs. Unmeasured factors, including implicit bias and differences in negotiations, social networks and mentoring opportunities[38, 39] may also contribute to pay differences between men and women hospitalists. As hospital medicine tackles gender inequities and other disparities, strategies to assess and address fair physician compensation must be on the table.

Hospitalists are a growing workforce numbering over 40,000 physicians, one‐third of whom are women.[1] Flexibility of work schedules and control over personal time have been the traditional selling points of the specialty.[2] Multiple studies of physician work life reveal growing physician dissatisfaction and a high prevalence of burnout.[3] To mitigate burnout risk, leaders in hospital medicine recognize the importance of creating a sustainable profession that offers both job and career satisfaction as well as work‐life balance and, importantly, fairness within the work environment.[4] Although success in some of these endeavors has been realized sporadically, sustaining work‐life balance and fairness in the specialty remains a work in progress, whereas evidence of high job attrition and pay inequities remain.[5, 6]

Pay inequity for women relative to men continues to be pervasive in medicine, including among early‐career physicians, researchers, and various specialists.[6, 7, 8, 9, 10, 11, 12, 13] The earnings gap seems to persist for physicians, even as federal efforts such as the Fair Pay Act of 2013 and the Paycheck Fairness Act of 2014 aim to end wage discrimination.[11, 14] Differences in specialty, part‐time status, and practice type do not mitigate the disparity.[8, 10, 15] Additional explanations have been proposed to explain the variability, including gender differences in negotiating skills, lack of opportunities to join networks of influence within organizations, and implicit or explicit bias and discrimination.[12, 16, 17, 18, 19, 20]

The earnings gap is also a consequence of what is commonly called the glass ceiling.[18, 19] Most agree that obstacles to fair advancement of women include absence of collaborative environments and role models who have successfully achieved work‐life balance.[17, 20, 21, 22] Somewhat surprisingly, women leaders in medicine seem to suffer greater income disparity than nonleaders; this income gap is prevalent among leaders in other elite professions as well.[7, 23] It is unknown whether women physicians' emphasis on work‐life balance, seen repeatedly in surveys, explains any of the pay disparity.[24] Little research to date has examined whether work‐life priorities of women in hospital medicine differ from men.

In this study, we sought to examine differences in job priorities between men and women hospitalists. In particular, we examine the relative prioritization of substantial pay to job satisfaction. We also examined gender differences in work patterns and earnings to explore potential sources of the persistent gender earnings gap.

METHODS

We analyzed data from the 20092010 Hospital Medicine Physician Worklife Survey, the design of which is detailed elsewhere.[4] Briefly, a 118‐item survey was administered by mail to a stratified sample of hospitalists from the Society of Hospital Medicine database and 3 large multisite hospitalist groups. A single survey item asked respondents to identify up to 4 out of 12 most important domains to their satisfaction with a hospitalist job. The domains were distilled from focus groups of nationally representative hospitalists as described previously,[4] and the survey item allowed up to one‐third of these domains to be identified as respondents' personal priorities. The list included: optimal variety of tasks, optimal workload, substantial pay, collegiality with other physicians, recognition by leaders, rewarding relationships with patients, satisfaction with nurses, optimal autonomy, control over personal time, fairness within organization, ample availability of resources to do job, and organizational climate of trust and belonging. We tabulated and ranked the frequency with which respondents selected each satisfaction domain by gender. Due to the nonstandard format of the survey item, we a priori decided to analyze only responses that were completed as instructed.

We also used demographic data including detailed work characteristics, clinical and nonclinical workload, total pretax earnings in 2009 as a hospitalist, and self‐identification as leader of their hospital medicine group. Respondent characteristics were tabulated and gender differences were tested using the t test, rank sum test, and the Fischer exact test as appropriate. We also listed the number of nonrespondents for each item. In estimating gender differences in earnings, we opted to use multiple‐imputation techniques to more conservatively account for greater variance inherent in the presence of missing data. Consistent with existing guidelines,[25] we demonstrated that item responses were not missing monotonically by visually inspecting patterns of nonresponse. We further demonstrated that data were missing at random by showing that response patterns of completed survey items did not predict whether or not a given variable response was missing using logistic regression models. We found no significant differences between respondents with complete and missing data. We verified that appropriate regression models for each variable on every other variable converged. We used Stata 13.1 (StataCorp, College Station, TX) to perform multiple imputations using chain equations (mi impute chain) to create 10 imputed tables for 7 normally distributed continuous variables using the ordinary least squares method, 3 non‐normally distributed variables using the predictive mean matching method, 2 nonordinal categorical variables using the multinomial logit method, and 1 binary variable using the logit method.[26, 27] Gender, pediatric specialty status, region of practice, and whether or not respondents prioritized substantial pay for job satisfaction were used as regular variables without missing data points.

Differences in earnings were assessed using a multivariate ordinary linear regression model applied to the imputed datasets fitted by forward selection of explanatory variables using P < 0.20 in bivariate analysis for inclusion and manual backward elimination of all statistically nonsignificant variables. We tested the significance of the women leader interaction term in the final parsimonious model. We used the usual significance threshold of P < 0.05 for inferences. Our analysis of publicly available anonymous data was exempt from IRB review.

RESULTS

Of the 816 survey respondents (response rate 25.6%), 40 either omitted the item soliciting work priorities or completed it incorrectly. Data from the remaining 776 respondents were used for the present analysis. Respondent characteristics are tabulated in Table 1. The characteristics of hospitalists by age, gender, specialty, practice model, and practice region were representative of US hospitalists from other surveys.[28]

Several gender differences were seen in the characteristics of hospitalists and their work (Table 1). Women compared to men hospitalists were less likely to be leaders, more likely to be pediatricians, work in university settings, and practice in Western states. Women compared to men, on average, were younger by 3 years, worked fewer full‐time equivalents (FTEs), worked a greater percentage of nights, and reported fewer billable encounters per shift. They were also more likely to be divorced or separated.

Job satisfaction priorities differed for women and men hospitalists. Table 2 lists job satisfaction domains in descending order of the frequency prioritized by men. The largest proportion of women and men prioritized optimal workload. However, although substantial pay was prioritized next most frequently by men, more women prioritized collegiality and control over personal time than substantial pay.

Key differences in individual characteristics, work patterns, and indicating substantial pay as a priority were associated with self‐reported total earnings in 2009 from respondents' work as a hospitalist. As shown in Table 3, the inclusion of detailed productivity measures such as FTE, days of monthly clinical work, and estimated number of daily billable encounters yielded a model that explained 33% of variance in earnings. After adjusting for significant covariates including pediatric specialty, practice model, geography, and amount and type of clinical work, the estimated underpayment of women compared to men was $14,581. Hospitalists who prioritized substantial pay earned $10,771 more than those who did not. The female x leader interaction term testing the hypothesis that gender disparity is greater among leaders than frontline hospitalists was not statistically significant ($16,720, P = 0.087) and excluded from the final model.

DISCUSSION

In a national stratified sample of US hospitalists, we found gender differences in job satisfaction priorities and hospitalist work characteristics. We also confirmed the persistence of a substantial gender earnings disparity. Lower earnings among women compared to men hospitalists were present in our data after controlling for age, pediatric specialty, practice model, geography, type of clinical work, and productivity measures. The gender earnings disparity noted in 1999[6] persists, although it appears to have decreased, possibly indicating progress toward equity. We showed that women hospitalists' relative tendency not to prioritize pay explains a significant portion of the residual income gap.

Hoff examined hospitalist earnings in a large national survey of hospitalists in 1999. Our cohorts differed in age and experience (both lower in the Hoff study). An estimated $24,000 ($124,266 vs. $148,132) earnings gap between women and men was greater than our estimate of $14,581 following an interval of 10 years. Although survey items differed, both studies found that women were less interested in pay than men when considering a hospitalist job. Hoff also found that work setting and attitudes about pay and lifestyle were significantly related to earnings. We extended the previous analyses of gender differences in job satisfaction priorities, work, and demographic characteristics to explain the earnings gap and understand how it may be remedied.

When considering job satisfaction, we found that more men than women prioritized substantial pay and that prioritization of substantial pay was directly related to higher earnings. Therefore, fewer women prioritizing pay partly explains women's lower earnings. Reasons for why fewer women prioritize pay were not assessed in this study but may include factors like being part of 2‐income households and competing commitments.[29] Priorities may even be influenced by empirically observed gender differences in discussions of financial matters, governed by cultural norms. Such norms may implicitly sanction employers to offer women less pay than men for the same or similar work. Women may disadvantage themselves by negotiating less or less well than men for higher starting and promotion salaries. They may be perceived more negatively than men when they do negotiate pay, leading to unintended negative consequences such as loss of social networks, decreased likability, and even loss of job offers.[29, 30, 31, 32]

More women prioritized optimal variety of tasks (6th most prevalent among women and 10th among men). Women who highly rate optimal variety of tasks as a job satisfier may choose positions in which they teach, perform research, and participate in hospital committees and quality‐improvement work, but offer lower pay. Yet hours per month doing nonclinical work was not significantly different between men and women, nor associated with earnings differences in our earnings models. Understanding whether women self‐select into hospitalist jobs with like‐minded colleagues to achieve complementary fit or end up supplementing their skills with hospitalists with different priorities may inform strategies to reduce the gender income disparity.[5] Unlike disparities between various hospital medicine groups, systematic disparities within practices risk generating low levels of organizational fairness and burnout among employees.

Not surprisingly, productivity was positively associated with earnings but did not fully account for the gender earnings gap. Our data demonstrated that women, on average, were associated with work characteristics that expectedly generate less compensation. For example, women were younger, more often part time, academic, pediatric, less often leaders, and reported fewer billable encounters compared to men. These differences account for some of the earnings gap between men and women, but these factors were controlled for in the earnings model. In addition, our analysis may have underestimated the gap by not incorporating loss of fringe benefits from part‐time status and not comprehensively counting incentive pay associated with high productivity. Other work patterns more commonly associated with women suggest an imbalance in reimbursement. More women than men work nights that are often compensated at higher rates than daytime work, yet their average pay was less, suggesting that compensation for night work may need to be adjusted to reflect its unique burdens and responsibilities.[33]

Although the gender pay gap was not more extreme among leaders compared to frontline hospitalists in our data, the trend, nonetheless, underscores an important consideration. Whereas clinical work is paid for in mostly measurable ways, pay for leadership may be influenced by intangible factors such as reputation, negotiation, and confidence that may disadvantage women relative to men.[7, 19, 23, 34, 35, 36] Efforts to overcome implicit gender bias should be most effective when we consciously couple fair promotion of women to leadership with fair compensation commensurate with their male peers.[21]

Our data are vulnerable to nonresponse bias.[1, 4, 5] Post hoc analyses demonstrated that distributions of age, gender, practice model and region of our respondents were similar to other nationally representative cohorts of hospitalists. Consequently, we believe our data can make valid estimates about a nationally representative sample of hospitalists. However, we acknowledge several additional weaknesses of self‐reported data, including recall bias and accuracy of productivity figures, which were rounded to variable significant digits by respondents. Earnings analysis using this data was intended to be exploratory, but the findings echoed analyses using more authoritative data sources.[11] Still, we made inferences conservatively by adopting multiple imputation techniques for dealing with nonresponse surveys in adherence to established reporting guidelines.[25] We also note several limitations relevant to multiple imputations. The greater prevalence of missing data for survey items soliciting earnings and the number of billable encounters suggest they were not truly missing at random as assumed. However, we showed that missingness is unrelated to the variables under study, justifying use of the technique. The wider measures of variance derived from multiple imputations make us vulnerable to not detecting associations that may exist.

The gender earnings gap found in hospital medicine echoes the gap found in multiple medical specialties, including but not limited to pediatrics, academic medicine, gastroenterology, and plastic surgery.[7, 8, 9, 11, 12, 13, 37] Hospital medicine employment models and practice patterns have important structural differences compared to previously studied populations that could mitigate factors contributing to women physicians' lower earnings. However, despite well‐defined working hours, lack of control over the number of patient encounters per day and high prevalence of hospital‐employed practice models, the gender earnings gap persists. We showed that lower prioritization for pay may reflect the self‐selection of women into lower paying jobs. Unmeasured factors, including implicit bias and differences in negotiations, social networks and mentoring opportunities[38, 39] may also contribute to pay differences between men and women hospitalists. As hospital medicine tackles gender inequities and other disparities, strategies to assess and address fair physician compensation must be on the table.

Despite widespread efforts to predict 30‐day rehospitalizations among discharged general medical patients,[1, 2, 3] not many strategies have incorporated patient‐reported outcome (PRO) measures in predictive models.[4] This despite the many longitudinal studies of the ambulatory population that demonstrate the higher likelihood of hospitalizations among those who score poorly on General Self‐Rated Health (GSRH),[5, 6, 7] baseline or declining Health‐Related Quality of Life,[8, 9, 10, 11, 12] psychological symptoms,[13, 14] and physical symptoms assessments.[15] One of the few existing studies that included PRO measures in 30‐day readmission models showed the predictive value of the 12‐item short form (SF12) Physical Component Score.[16] Others showed that persistent symptoms were associated with readmissions in patients with heart disease.[17, 18]

The paucity of efforts to connect PRO measures to utilization may be due to the limited availability of these measures in routine clinical records and the incomplete knowledge about how various PRO measures may fluctuate during episodes of acute illnesses and their treatments during hospitalizations. Health perception measures reflect both enduring features like self‐concept as well as dynamic features like a person's immediate health status.[19] As such, GSRH reflects the presence of chronic illnesses but is also responsive to acute events.[20, 21] Similarly, Health‐Related Quality of Life measures are dynamic as they decline around episodes of acute illness but are stable over a longer time window in their tendency to recover.[22] We do not know how fluctuations in measures of symptom burden, perceived health, and quality of life around the hospital‐to‐home transition may differentially inform readmission risk. Using a longitudinal cohort study, we addressed 2 questions: (1) How do PRO measures change when measured serially during the hospital‐to‐home transition? (2) How does the relative timing of each PRO measure variably inform the risk of subsequent utilization events including hospital readmissions?

METHODS

We conducted a longitudinal cohort study using data originally collected for a trial (ClinicalTrials.gov Identifier NCT01391026) of an intervention that was shown to have no associations with variables evaluated in this study. Patients were recruited from the John H. Stroger Hospital of Cook County, an urban safety‐net hospital that serves 128 municipalities in northeastern Illinois including the City of Chicago. Patients were eligible if they (1) were admitted to the general medical wards, the medical intensive care unit, or the cardiac care unit between May 2011 and February 2012; (2) had a clinic appointment in the Hospital's general medicine clinic (GMC) in the prior 12 months to facilitate follow‐up; and (3) were able to communicate independently in English or Spanish. Randomly selected patients were approached during their hospitalization and consenting subjects completed an in‐person questionnaire on the day of discharge. Subjects were contacted by telephone around 30, 90, and 180 days thereafter to complete follow‐up questionnaires; we began calling patients around 2 weeks prior to the target day anticipating noncontact on the first attempts. All telephone interviews were conducted by research assistants who had no clinical training and who did not give care‐related advice to patients based on their survey response. A few patients whose follow‐up survey window straddled the date of a scheduled clinic appointment were invited to complete the questionnaire in the GMC's waiting area using computer kiosks enabled with audio computer‐assisted self‐interview technology described elsewhere.[23] The Charlson Comorbidity Index was calculated inclusive of diagnostic codes assigned over 3 months preceding the index hospitalization.[24]

The following instruments were administered at each interview. The physical symptom severity portion of the Memorial Symptom Assessment Scale (MSAS) solicited the severity rank (none/a little bit/somewhat/quite a bit/very much) of 17 physical symptoms in the last week; the score was calculated by averaging the severity rank of the 12 most common symptom in the sample.[25, 26] The Patient Reported Outcomes Measurement Information System (PROMIS) Global Health Short Form is an instrument assessing GSRH (1 item), Social Activities (1 item), Global Physical Health (4 items), and Global Mental Health (4 items including a single‐item quality‐of‐life measure). Fatigue and pain for Global Physical Health, and emotional health for Global Mental Health were assessed over the past 7 days. Each of the 2 Global Health scores was standardized to a national mean of 50 and standard deviation of 10.[27]

The rate of survey completion at each follow‐up was calculated. Characteristics of participants were tabulated. Characteristics of patients censored prior to study completion were compared with patients with complete data. Box plots for MSAS physical symptom severity, and Global Physical and Mental Health scores were constructed to illustrate the comparisons of the mean scores between each consecutive survey period using t tests assuming unequal variance. A similar box plot of GSRH illustrated the comparison of the median score between consecutive surveys using the rank sum test. Hospital‐based utilization events were defined as either an emergency department visit or hospitalization at 1 of the 2 hospitals of the Cook County Health & Hospitals System (CCHHS). After accounting for patient data censored due to death (date reported by family) or withdrawal from study, Kaplan‐Meier curves showing time to first hospital‐based utilization event during each interval between surveys were drawn separately for above‐ and below‐median MSAS, Global Physical and Mental Health scores, and for poor or fair versus good, very good, or excellent GSRH assessment. The null hypothesis that the survivor functions were equal between the better and worse median quantiles or GSRH categories was tested using the log‐rank test at 14 and 30 days from survey completion. Hazard ratios for time to first utilization event within 14 days of each survey were calculated for the MSAS score, Global Physical and Mental Health as continuous variables, and GSRH response categories relative to poor using bivariate and multivariate Cox proportional hazard equations. Multivariate models incorporated the following 5 covariates: at least 1 utilization event during the study period prior to the survey, Charlson score, age, gender, and race/ethnicity category. Likelihood ratio statistics were calculated to test the hypothesis that the model including the PRO measure and covariates predicted the outcome equally well compared to the nested model with only covariates. We used the traditional threshold of .05 when reporting significance. All analyses were performed in Stata 13 (StataCorp, College Station, TX). The methods for patient consent, data collection, analyses, and reporting were reviewed and approved by the CCHHS institutional review board.

RESULTS

A total of 196 patients completed the initial survey. The completion rates were 98%, 90%, and 88% for the 30‐, 90‐, and 180‐day follow‐up surveys, respectively. As shown in Table 1, participants average age was 52 years, and about half were women. The majority was non‐Hispanic black, and 21% preferred to complete the survey in Spanish. Diabetes, congestive heart failure, cancer, and chronic pulmonary disease were each prevalent in at least one‐fifth of our patient cohort. Demographic characteristics were similar between the 160 patients who completed all 3 follow‐up surveys and the 36 who missed at least 1 follow‐up survey. Among the latter group, 1 withdrew at 30 days, 1 withdrew and 4 had died at 90 days, and 1 withdrew and 9 had died at 180 days.

Figure 1 shows a timeline of the follow‐up surveys and utilization events in the form of overlapping histograms. The majority of 30‐day follow‐up questionnaires were completed earlier than targeted, at a median of 17 (interquartile range [IQR] 16, 20) days after discharge. Similarly, questionnaires targeted for 90 and 180 days were completed at medians of 78 (IQR 7684) and 167 (IQR 166169) days from discharge. Fifty‐four (28%) patients experienced a first utilization event in the first 30 days following discharge. During the 60‐, 90‐, and 30‐day intervals after the first, second, and third follow‐up surveys, respectively, 63 (33%), 54 (31%), and 16 (9%) patients experienced a first utilization event.

Figure 1

A significant improvement in MSAS physical symptom severity was detected between the hospitalization and the 30‐day follow‐up (Figure 2A). Although the mean Global Physical Health score was below the national mean of 50 at every survey period, a similar improvement in the measure was noted between the hospitalization and the 30‐day follow‐up (Figure 2B). The mean Global Mental Health score was also below the national mean but remained stable throughout the study (Figure 2C). The median GSRH was stable at 2 (IQR 23) at every survey wave (Figure 2D). Of note, compared to patients who completed all 3 follow‐up surveys, patients who missed at least 1 follow‐up reported higher MSAS score (1.5 vs 1.8, P=0.03), lower Global Physical Health (36.1 vs 33.5, P=0.09), and lower Global Mental Health (44.7 vs 41.0, P=0.03) during their hospitalization. In addition, patients with complete data experienced an average of 1.2 utilization events during the study, whereas those with missing data experienced an average of 2.1 utilization events (P=0.03).

Figure 2

The MSAS physical symptom severity and Global Physical Health scores from the index hospitalizations did not identify patients with a first utilization event within 30 days. However, patients with poor Global Mental Health and GSRH in the hospital were more likely to experience a utilization event within 14 days of discharge (Figure 3). During the postdischarge period, patients scoring poorly on each of the PRO measures trended toward a greater risk of an early utilization event, but the association between utilization and MSAS was most consistently significant (Figure 3A). In general, the associations with MSAS, Global Physical Health, and GSRH were stronger with the risk of utilization events within 14 days than within 30 days (Figure 3A,B,D). The Global Mental Health score was not associated with a subsequent utilization when measured during the 180‐day postdischarge period.

Figure 3

As shown in Table 2, Cox proportional hazard models incorporating covariates preserved most of the significant associations seen in the unadjusted analyses. Global Mental Health and good relative to poor GSRH obtained during the hospitalization remained significant. MSAS obtained at each postdischarge follow‐up trended positively with utilization and was statistically significant at 90 and 180 days. Global Physical Health obtained at each postdischarge follow‐up similarly trended negatively with utilization and was significant at 180 days. Each multivariate model incorporating a PRO measure with a significant coefficient contributed to better fit of the predictive model compared to the nested model without the PRO measure.

DISCUSSION

In this longitudinal cohort study, patients, on average, reported relatively severe symptoms, low PROMIS Global Physical and Mental Health scores, and poor GSRH during the inpatient stay in an urban safety‐net hospital. Symptom severity and Global Physical Health improved, on average, by 30 days before stabilizing, but their poor levels in the hospital did not predict 30‐day hospital‐based utilization events. On the other hand, Global Mental Health and GSRH were stable through hospitalizations, and patients scoring poorly on these measures were at greater risk of utilization events within 14 days of discharge. PRO measures obtained during the 180‐day postdischarge period trended toward distinguishing populations with greater baseline risk of proximate utilization events. However, MSAS physical symptom severity and Global Physical Health were more consistently predictive of these events at statistically significant levels compared to Global Mental Health and GSRH in our relatively small sample of patients. Each of these measures selectively improved the fit‐of‐risk prediction models for hospital‐based utilization.

Some of the heterogeneity in readmission risk is explained by differences in PRO measures. Although the MSAS score and Global Physical Health assessment were reliable predictors of utilization when measured in ambulatory settings, they were less discriminating during acute hospitalizations when everyone, on average, reported severe symptoms and poor function. Our results were consistent with other studies that demonstrated the fairly rapid recovery in symptoms that follow hospitalizations,[28, 29] and these measures may become informative of utilization risk as early as 2 weeks postdischarge. GSRH and Global Mental Health (a measure of health‐related quality of life) only predicted utilizations immediately at hospital discharge. As multidimensional measures that reflect physical, social, and emotional capacity, these measures may indicate vulnerabilities in patients least able to handle the stresses of the early postdischarge period.

There is growing momentum around collecting PRO measures in routine clinical care as quality indicators that capture patient‐centered concerns.[30] Our study explored a novel application of these measures whose routine collection will likely proliferate, not solely for the purpose of helping healthcare systems identify patients at risk of unplanned resource utilization. Although multidimensional PRO measures seldom reflect conditions directly modifiable by simple interventions, we believe that the association between physical symptom burden and utilization in our data reveals a possible target for practice improvement. Hospitalists have contributed enormously to shorter lengths of stay that risk sicker and quicker discharges.[31] To mitigate its potential side effects on symptom management, a discharge plan that acknowledges physical symptoms that sometimes persist or recur beyond the hospitalization may be appropriate. This may be accomplished by ensuring that acute symptoms are resolving, giving clear instructions for symptom management at home, as now the standard of care for conditions like asthma,[32] and explicitly communicating the presence of residual symptoms to providers entrusted with continuity care. As an effective feedback measure that can drive continuous quality improvement, we believe that a technology‐based surveillance strategy that spans both the inpatient and outpatient domains is necessary.[23]

There are some notable similarities and differences between the results of our study and a recent hospital‐based study of PRO measures that used data from the Multi‐Center Hospitalist Project.[16] The Physical Component Score of the SF12 is similar to the PROMIS Global Physical Health score in that both incorporate measures of physical function, perceived health, pain, and energy level. Curiously, the SF12 Physical Component Score, but not the PROMIS Global Physical Health score, was associated with 30‐day rehospitalizations. An important difference between the measures is where the SF12 asks about limitations during the past 4 weeks the PROMIS instrument inquires about physical function in general and levels of fatigue and pain in the past 7 days. Considering most hospitalizations last <7 days, the PROMIS instrument may better reflect the declines associated with the acute illness related to the hospitalization than the SF12 score. Additionally, the discrepancy between the association between hospital‐based GSRH and utilization in our study and the absence, thereof, in Hasan et al. is noteworthy. The difference here may be explained by their use of a 0‐ to 100‐point response scale in contrast to our study's verbally labeled 5‐point scale in the PROMIS instrument. The range of rating scales for survey questions is traditionally governed by the tension between the difficulty with mapping respondents judgment on an excessively large scale on one hand, and the failure of insufficient response options to discriminate between respondents with different underlying judgment on the other.[33] We suspect the former to be a drawback of the unlabeled 100‐point response scale, and conjecture that an association might be found in the Multi‐Center Hospitalist Study data if the responses were grouped into summative categories.

We recognize several limitations in our study. The first is the generalizability of our patient population to others, not insignificantly because of the high proportion of the uninsured (around 70% during the study period) and racial/ethnic minorities among them. Although utilization patterns are clearly affected by socioeconomic status,[34] there may also be differences in the way validated PRO measures are calibrated between patients of public and private healthcare systems.[35] Another limitation is our inability to count utilization events at institutions outside of the CCHHS during our study. However, because the study was conducted prior to Cook County's Medicaid expansion demonstration program as part of the Affordable Care Act,[36] many patients established in our system faced barriers to receiving nonemergency care outside of the CCHHS supporting our assumption that few of our patients were discharged from other hospitals. Causality cannot be established in observational studies. Consequently, high prior‐symptom burden may be associated with utilizations through unmeasured variables. Measures of symptom burden are vulnerable to overendorsement and amplification.[37, 38] Inferences based on statistical significance are affected by sample size, and our conclusions may change if conducted with a larger number of participants. Our response rates were excellent through the survey waves, but we did not achieve perfect follow‐up. Worse levels of PRO responses and higher levels of utilization among censored patients biased our results toward the null. Finally, although we did not find any predominant comorbidities associated with hospital‐based utilizations in our sample, our analyses may be vulnerable to inadequate control for illness severity, which may also have biased our results.

PRO measures are likely to be useful in clinical medicine.[39] But to fully apply the powers of PROs in informing clinically and operationally relevant outcomes, we must actively develop a system for obtaining these measures in routine clinical care. The availability of patient downtime makes hospitalizations conducive to gathering patient‐generated data, and may further enhance patient‐provider communication if survey output was readily available in electronic medical records. Exploring innovative strategies for collecting PROs in the hospital and beyond remains our future work.

Despite widespread efforts to predict 30‐day rehospitalizations among discharged general medical patients,[1, 2, 3] not many strategies have incorporated patient‐reported outcome (PRO) measures in predictive models.[4] This despite the many longitudinal studies of the ambulatory population that demonstrate the higher likelihood of hospitalizations among those who score poorly on General Self‐Rated Health (GSRH),[5, 6, 7] baseline or declining Health‐Related Quality of Life,[8, 9, 10, 11, 12] psychological symptoms,[13, 14] and physical symptoms assessments.[15] One of the few existing studies that included PRO measures in 30‐day readmission models showed the predictive value of the 12‐item short form (SF12) Physical Component Score.[16] Others showed that persistent symptoms were associated with readmissions in patients with heart disease.[17, 18]

The paucity of efforts to connect PRO measures to utilization may be due to the limited availability of these measures in routine clinical records and the incomplete knowledge about how various PRO measures may fluctuate during episodes of acute illnesses and their treatments during hospitalizations. Health perception measures reflect both enduring features like self‐concept as well as dynamic features like a person's immediate health status.[19] As such, GSRH reflects the presence of chronic illnesses but is also responsive to acute events.[20, 21] Similarly, Health‐Related Quality of Life measures are dynamic as they decline around episodes of acute illness but are stable over a longer time window in their tendency to recover.[22] We do not know how fluctuations in measures of symptom burden, perceived health, and quality of life around the hospital‐to‐home transition may differentially inform readmission risk. Using a longitudinal cohort study, we addressed 2 questions: (1) How do PRO measures change when measured serially during the hospital‐to‐home transition? (2) How does the relative timing of each PRO measure variably inform the risk of subsequent utilization events including hospital readmissions?

METHODS

We conducted a longitudinal cohort study using data originally collected for a trial (ClinicalTrials.gov Identifier NCT01391026) of an intervention that was shown to have no associations with variables evaluated in this study. Patients were recruited from the John H. Stroger Hospital of Cook County, an urban safety‐net hospital that serves 128 municipalities in northeastern Illinois including the City of Chicago. Patients were eligible if they (1) were admitted to the general medical wards, the medical intensive care unit, or the cardiac care unit between May 2011 and February 2012; (2) had a clinic appointment in the Hospital's general medicine clinic (GMC) in the prior 12 months to facilitate follow‐up; and (3) were able to communicate independently in English or Spanish. Randomly selected patients were approached during their hospitalization and consenting subjects completed an in‐person questionnaire on the day of discharge. Subjects were contacted by telephone around 30, 90, and 180 days thereafter to complete follow‐up questionnaires; we began calling patients around 2 weeks prior to the target day anticipating noncontact on the first attempts. All telephone interviews were conducted by research assistants who had no clinical training and who did not give care‐related advice to patients based on their survey response. A few patients whose follow‐up survey window straddled the date of a scheduled clinic appointment were invited to complete the questionnaire in the GMC's waiting area using computer kiosks enabled with audio computer‐assisted self‐interview technology described elsewhere.[23] The Charlson Comorbidity Index was calculated inclusive of diagnostic codes assigned over 3 months preceding the index hospitalization.[24]

The following instruments were administered at each interview. The physical symptom severity portion of the Memorial Symptom Assessment Scale (MSAS) solicited the severity rank (none/a little bit/somewhat/quite a bit/very much) of 17 physical symptoms in the last week; the score was calculated by averaging the severity rank of the 12 most common symptom in the sample.[25, 26] The Patient Reported Outcomes Measurement Information System (PROMIS) Global Health Short Form is an instrument assessing GSRH (1 item), Social Activities (1 item), Global Physical Health (4 items), and Global Mental Health (4 items including a single‐item quality‐of‐life measure). Fatigue and pain for Global Physical Health, and emotional health for Global Mental Health were assessed over the past 7 days. Each of the 2 Global Health scores was standardized to a national mean of 50 and standard deviation of 10.[27]

The rate of survey completion at each follow‐up was calculated. Characteristics of participants were tabulated. Characteristics of patients censored prior to study completion were compared with patients with complete data. Box plots for MSAS physical symptom severity, and Global Physical and Mental Health scores were constructed to illustrate the comparisons of the mean scores between each consecutive survey period using t tests assuming unequal variance. A similar box plot of GSRH illustrated the comparison of the median score between consecutive surveys using the rank sum test. Hospital‐based utilization events were defined as either an emergency department visit or hospitalization at 1 of the 2 hospitals of the Cook County Health & Hospitals System (CCHHS). After accounting for patient data censored due to death (date reported by family) or withdrawal from study, Kaplan‐Meier curves showing time to first hospital‐based utilization event during each interval between surveys were drawn separately for above‐ and below‐median MSAS, Global Physical and Mental Health scores, and for poor or fair versus good, very good, or excellent GSRH assessment. The null hypothesis that the survivor functions were equal between the better and worse median quantiles or GSRH categories was tested using the log‐rank test at 14 and 30 days from survey completion. Hazard ratios for time to first utilization event within 14 days of each survey were calculated for the MSAS score, Global Physical and Mental Health as continuous variables, and GSRH response categories relative to poor using bivariate and multivariate Cox proportional hazard equations. Multivariate models incorporated the following 5 covariates: at least 1 utilization event during the study period prior to the survey, Charlson score, age, gender, and race/ethnicity category. Likelihood ratio statistics were calculated to test the hypothesis that the model including the PRO measure and covariates predicted the outcome equally well compared to the nested model with only covariates. We used the traditional threshold of .05 when reporting significance. All analyses were performed in Stata 13 (StataCorp, College Station, TX). The methods for patient consent, data collection, analyses, and reporting were reviewed and approved by the CCHHS institutional review board.

RESULTS

A total of 196 patients completed the initial survey. The completion rates were 98%, 90%, and 88% for the 30‐, 90‐, and 180‐day follow‐up surveys, respectively. As shown in Table 1, participants average age was 52 years, and about half were women. The majority was non‐Hispanic black, and 21% preferred to complete the survey in Spanish. Diabetes, congestive heart failure, cancer, and chronic pulmonary disease were each prevalent in at least one‐fifth of our patient cohort. Demographic characteristics were similar between the 160 patients who completed all 3 follow‐up surveys and the 36 who missed at least 1 follow‐up survey. Among the latter group, 1 withdrew at 30 days, 1 withdrew and 4 had died at 90 days, and 1 withdrew and 9 had died at 180 days.

Figure 1 shows a timeline of the follow‐up surveys and utilization events in the form of overlapping histograms. The majority of 30‐day follow‐up questionnaires were completed earlier than targeted, at a median of 17 (interquartile range [IQR] 16, 20) days after discharge. Similarly, questionnaires targeted for 90 and 180 days were completed at medians of 78 (IQR 7684) and 167 (IQR 166169) days from discharge. Fifty‐four (28%) patients experienced a first utilization event in the first 30 days following discharge. During the 60‐, 90‐, and 30‐day intervals after the first, second, and third follow‐up surveys, respectively, 63 (33%), 54 (31%), and 16 (9%) patients experienced a first utilization event.

Figure 1

A significant improvement in MSAS physical symptom severity was detected between the hospitalization and the 30‐day follow‐up (Figure 2A). Although the mean Global Physical Health score was below the national mean of 50 at every survey period, a similar improvement in the measure was noted between the hospitalization and the 30‐day follow‐up (Figure 2B). The mean Global Mental Health score was also below the national mean but remained stable throughout the study (Figure 2C). The median GSRH was stable at 2 (IQR 23) at every survey wave (Figure 2D). Of note, compared to patients who completed all 3 follow‐up surveys, patients who missed at least 1 follow‐up reported higher MSAS score (1.5 vs 1.8, P=0.03), lower Global Physical Health (36.1 vs 33.5, P=0.09), and lower Global Mental Health (44.7 vs 41.0, P=0.03) during their hospitalization. In addition, patients with complete data experienced an average of 1.2 utilization events during the study, whereas those with missing data experienced an average of 2.1 utilization events (P=0.03).

Figure 2

The MSAS physical symptom severity and Global Physical Health scores from the index hospitalizations did not identify patients with a first utilization event within 30 days. However, patients with poor Global Mental Health and GSRH in the hospital were more likely to experience a utilization event within 14 days of discharge (Figure 3). During the postdischarge period, patients scoring poorly on each of the PRO measures trended toward a greater risk of an early utilization event, but the association between utilization and MSAS was most consistently significant (Figure 3A). In general, the associations with MSAS, Global Physical Health, and GSRH were stronger with the risk of utilization events within 14 days than within 30 days (Figure 3A,B,D). The Global Mental Health score was not associated with a subsequent utilization when measured during the 180‐day postdischarge period.

Figure 3

As shown in Table 2, Cox proportional hazard models incorporating covariates preserved most of the significant associations seen in the unadjusted analyses. Global Mental Health and good relative to poor GSRH obtained during the hospitalization remained significant. MSAS obtained at each postdischarge follow‐up trended positively with utilization and was statistically significant at 90 and 180 days. Global Physical Health obtained at each postdischarge follow‐up similarly trended negatively with utilization and was significant at 180 days. Each multivariate model incorporating a PRO measure with a significant coefficient contributed to better fit of the predictive model compared to the nested model without the PRO measure.

DISCUSSION

In this longitudinal cohort study, patients, on average, reported relatively severe symptoms, low PROMIS Global Physical and Mental Health scores, and poor GSRH during the inpatient stay in an urban safety‐net hospital. Symptom severity and Global Physical Health improved, on average, by 30 days before stabilizing, but their poor levels in the hospital did not predict 30‐day hospital‐based utilization events. On the other hand, Global Mental Health and GSRH were stable through hospitalizations, and patients scoring poorly on these measures were at greater risk of utilization events within 14 days of discharge. PRO measures obtained during the 180‐day postdischarge period trended toward distinguishing populations with greater baseline risk of proximate utilization events. However, MSAS physical symptom severity and Global Physical Health were more consistently predictive of these events at statistically significant levels compared to Global Mental Health and GSRH in our relatively small sample of patients. Each of these measures selectively improved the fit‐of‐risk prediction models for hospital‐based utilization.

Some of the heterogeneity in readmission risk is explained by differences in PRO measures. Although the MSAS score and Global Physical Health assessment were reliable predictors of utilization when measured in ambulatory settings, they were less discriminating during acute hospitalizations when everyone, on average, reported severe symptoms and poor function. Our results were consistent with other studies that demonstrated the fairly rapid recovery in symptoms that follow hospitalizations,[28, 29] and these measures may become informative of utilization risk as early as 2 weeks postdischarge. GSRH and Global Mental Health (a measure of health‐related quality of life) only predicted utilizations immediately at hospital discharge. As multidimensional measures that reflect physical, social, and emotional capacity, these measures may indicate vulnerabilities in patients least able to handle the stresses of the early postdischarge period.

There is growing momentum around collecting PRO measures in routine clinical care as quality indicators that capture patient‐centered concerns.[30] Our study explored a novel application of these measures whose routine collection will likely proliferate, not solely for the purpose of helping healthcare systems identify patients at risk of unplanned resource utilization. Although multidimensional PRO measures seldom reflect conditions directly modifiable by simple interventions, we believe that the association between physical symptom burden and utilization in our data reveals a possible target for practice improvement. Hospitalists have contributed enormously to shorter lengths of stay that risk sicker and quicker discharges.[31] To mitigate its potential side effects on symptom management, a discharge plan that acknowledges physical symptoms that sometimes persist or recur beyond the hospitalization may be appropriate. This may be accomplished by ensuring that acute symptoms are resolving, giving clear instructions for symptom management at home, as now the standard of care for conditions like asthma,[32] and explicitly communicating the presence of residual symptoms to providers entrusted with continuity care. As an effective feedback measure that can drive continuous quality improvement, we believe that a technology‐based surveillance strategy that spans both the inpatient and outpatient domains is necessary.[23]

There are some notable similarities and differences between the results of our study and a recent hospital‐based study of PRO measures that used data from the Multi‐Center Hospitalist Project.[16] The Physical Component Score of the SF12 is similar to the PROMIS Global Physical Health score in that both incorporate measures of physical function, perceived health, pain, and energy level. Curiously, the SF12 Physical Component Score, but not the PROMIS Global Physical Health score, was associated with 30‐day rehospitalizations. An important difference between the measures is where the SF12 asks about limitations during the past 4 weeks the PROMIS instrument inquires about physical function in general and levels of fatigue and pain in the past 7 days. Considering most hospitalizations last <7 days, the PROMIS instrument may better reflect the declines associated with the acute illness related to the hospitalization than the SF12 score. Additionally, the discrepancy between the association between hospital‐based GSRH and utilization in our study and the absence, thereof, in Hasan et al. is noteworthy. The difference here may be explained by their use of a 0‐ to 100‐point response scale in contrast to our study's verbally labeled 5‐point scale in the PROMIS instrument. The range of rating scales for survey questions is traditionally governed by the tension between the difficulty with mapping respondents judgment on an excessively large scale on one hand, and the failure of insufficient response options to discriminate between respondents with different underlying judgment on the other.[33] We suspect the former to be a drawback of the unlabeled 100‐point response scale, and conjecture that an association might be found in the Multi‐Center Hospitalist Study data if the responses were grouped into summative categories.

We recognize several limitations in our study. The first is the generalizability of our patient population to others, not insignificantly because of the high proportion of the uninsured (around 70% during the study period) and racial/ethnic minorities among them. Although utilization patterns are clearly affected by socioeconomic status,[34] there may also be differences in the way validated PRO measures are calibrated between patients of public and private healthcare systems.[35] Another limitation is our inability to count utilization events at institutions outside of the CCHHS during our study. However, because the study was conducted prior to Cook County's Medicaid expansion demonstration program as part of the Affordable Care Act,[36] many patients established in our system faced barriers to receiving nonemergency care outside of the CCHHS supporting our assumption that few of our patients were discharged from other hospitals. Causality cannot be established in observational studies. Consequently, high prior‐symptom burden may be associated with utilizations through unmeasured variables. Measures of symptom burden are vulnerable to overendorsement and amplification.[37, 38] Inferences based on statistical significance are affected by sample size, and our conclusions may change if conducted with a larger number of participants. Our response rates were excellent through the survey waves, but we did not achieve perfect follow‐up. Worse levels of PRO responses and higher levels of utilization among censored patients biased our results toward the null. Finally, although we did not find any predominant comorbidities associated with hospital‐based utilizations in our sample, our analyses may be vulnerable to inadequate control for illness severity, which may also have biased our results.

PRO measures are likely to be useful in clinical medicine.[39] But to fully apply the powers of PROs in informing clinically and operationally relevant outcomes, we must actively develop a system for obtaining these measures in routine clinical care. The availability of patient downtime makes hospitalizations conducive to gathering patient‐generated data, and may further enhance patient‐provider communication if survey output was readily available in electronic medical records. Exploring innovative strategies for collecting PROs in the hospital and beyond remains our future work.

Despite widespread efforts to predict 30‐day rehospitalizations among discharged general medical patients,[1, 2, 3] not many strategies have incorporated patient‐reported outcome (PRO) measures in predictive models.[4] This despite the many longitudinal studies of the ambulatory population that demonstrate the higher likelihood of hospitalizations among those who score poorly on General Self‐Rated Health (GSRH),[5, 6, 7] baseline or declining Health‐Related Quality of Life,[8, 9, 10, 11, 12] psychological symptoms,[13, 14] and physical symptoms assessments.[15] One of the few existing studies that included PRO measures in 30‐day readmission models showed the predictive value of the 12‐item short form (SF12) Physical Component Score.[16] Others showed that persistent symptoms were associated with readmissions in patients with heart disease.[17, 18]

The paucity of efforts to connect PRO measures to utilization may be due to the limited availability of these measures in routine clinical records and the incomplete knowledge about how various PRO measures may fluctuate during episodes of acute illnesses and their treatments during hospitalizations. Health perception measures reflect both enduring features like self‐concept as well as dynamic features like a person's immediate health status.[19] As such, GSRH reflects the presence of chronic illnesses but is also responsive to acute events.[20, 21] Similarly, Health‐Related Quality of Life measures are dynamic as they decline around episodes of acute illness but are stable over a longer time window in their tendency to recover.[22] We do not know how fluctuations in measures of symptom burden, perceived health, and quality of life around the hospital‐to‐home transition may differentially inform readmission risk. Using a longitudinal cohort study, we addressed 2 questions: (1) How do PRO measures change when measured serially during the hospital‐to‐home transition? (2) How does the relative timing of each PRO measure variably inform the risk of subsequent utilization events including hospital readmissions?

METHODS

We conducted a longitudinal cohort study using data originally collected for a trial (ClinicalTrials.gov Identifier NCT01391026) of an intervention that was shown to have no associations with variables evaluated in this study. Patients were recruited from the John H. Stroger Hospital of Cook County, an urban safety‐net hospital that serves 128 municipalities in northeastern Illinois including the City of Chicago. Patients were eligible if they (1) were admitted to the general medical wards, the medical intensive care unit, or the cardiac care unit between May 2011 and February 2012; (2) had a clinic appointment in the Hospital's general medicine clinic (GMC) in the prior 12 months to facilitate follow‐up; and (3) were able to communicate independently in English or Spanish. Randomly selected patients were approached during their hospitalization and consenting subjects completed an in‐person questionnaire on the day of discharge. Subjects were contacted by telephone around 30, 90, and 180 days thereafter to complete follow‐up questionnaires; we began calling patients around 2 weeks prior to the target day anticipating noncontact on the first attempts. All telephone interviews were conducted by research assistants who had no clinical training and who did not give care‐related advice to patients based on their survey response. A few patients whose follow‐up survey window straddled the date of a scheduled clinic appointment were invited to complete the questionnaire in the GMC's waiting area using computer kiosks enabled with audio computer‐assisted self‐interview technology described elsewhere.[23] The Charlson Comorbidity Index was calculated inclusive of diagnostic codes assigned over 3 months preceding the index hospitalization.[24]

The following instruments were administered at each interview. The physical symptom severity portion of the Memorial Symptom Assessment Scale (MSAS) solicited the severity rank (none/a little bit/somewhat/quite a bit/very much) of 17 physical symptoms in the last week; the score was calculated by averaging the severity rank of the 12 most common symptom in the sample.[25, 26] The Patient Reported Outcomes Measurement Information System (PROMIS) Global Health Short Form is an instrument assessing GSRH (1 item), Social Activities (1 item), Global Physical Health (4 items), and Global Mental Health (4 items including a single‐item quality‐of‐life measure). Fatigue and pain for Global Physical Health, and emotional health for Global Mental Health were assessed over the past 7 days. Each of the 2 Global Health scores was standardized to a national mean of 50 and standard deviation of 10.[27]

The rate of survey completion at each follow‐up was calculated. Characteristics of participants were tabulated. Characteristics of patients censored prior to study completion were compared with patients with complete data. Box plots for MSAS physical symptom severity, and Global Physical and Mental Health scores were constructed to illustrate the comparisons of the mean scores between each consecutive survey period using t tests assuming unequal variance. A similar box plot of GSRH illustrated the comparison of the median score between consecutive surveys using the rank sum test. Hospital‐based utilization events were defined as either an emergency department visit or hospitalization at 1 of the 2 hospitals of the Cook County Health & Hospitals System (CCHHS). After accounting for patient data censored due to death (date reported by family) or withdrawal from study, Kaplan‐Meier curves showing time to first hospital‐based utilization event during each interval between surveys were drawn separately for above‐ and below‐median MSAS, Global Physical and Mental Health scores, and for poor or fair versus good, very good, or excellent GSRH assessment. The null hypothesis that the survivor functions were equal between the better and worse median quantiles or GSRH categories was tested using the log‐rank test at 14 and 30 days from survey completion. Hazard ratios for time to first utilization event within 14 days of each survey were calculated for the MSAS score, Global Physical and Mental Health as continuous variables, and GSRH response categories relative to poor using bivariate and multivariate Cox proportional hazard equations. Multivariate models incorporated the following 5 covariates: at least 1 utilization event during the study period prior to the survey, Charlson score, age, gender, and race/ethnicity category. Likelihood ratio statistics were calculated to test the hypothesis that the model including the PRO measure and covariates predicted the outcome equally well compared to the nested model with only covariates. We used the traditional threshold of .05 when reporting significance. All analyses were performed in Stata 13 (StataCorp, College Station, TX). The methods for patient consent, data collection, analyses, and reporting were reviewed and approved by the CCHHS institutional review board.

RESULTS

A total of 196 patients completed the initial survey. The completion rates were 98%, 90%, and 88% for the 30‐, 90‐, and 180‐day follow‐up surveys, respectively. As shown in Table 1, participants average age was 52 years, and about half were women. The majority was non‐Hispanic black, and 21% preferred to complete the survey in Spanish. Diabetes, congestive heart failure, cancer, and chronic pulmonary disease were each prevalent in at least one‐fifth of our patient cohort. Demographic characteristics were similar between the 160 patients who completed all 3 follow‐up surveys and the 36 who missed at least 1 follow‐up survey. Among the latter group, 1 withdrew at 30 days, 1 withdrew and 4 had died at 90 days, and 1 withdrew and 9 had died at 180 days.

Figure 1 shows a timeline of the follow‐up surveys and utilization events in the form of overlapping histograms. The majority of 30‐day follow‐up questionnaires were completed earlier than targeted, at a median of 17 (interquartile range [IQR] 16, 20) days after discharge. Similarly, questionnaires targeted for 90 and 180 days were completed at medians of 78 (IQR 7684) and 167 (IQR 166169) days from discharge. Fifty‐four (28%) patients experienced a first utilization event in the first 30 days following discharge. During the 60‐, 90‐, and 30‐day intervals after the first, second, and third follow‐up surveys, respectively, 63 (33%), 54 (31%), and 16 (9%) patients experienced a first utilization event.

Figure 1

A significant improvement in MSAS physical symptom severity was detected between the hospitalization and the 30‐day follow‐up (Figure 2A). Although the mean Global Physical Health score was below the national mean of 50 at every survey period, a similar improvement in the measure was noted between the hospitalization and the 30‐day follow‐up (Figure 2B). The mean Global Mental Health score was also below the national mean but remained stable throughout the study (Figure 2C). The median GSRH was stable at 2 (IQR 23) at every survey wave (Figure 2D). Of note, compared to patients who completed all 3 follow‐up surveys, patients who missed at least 1 follow‐up reported higher MSAS score (1.5 vs 1.8, P=0.03), lower Global Physical Health (36.1 vs 33.5, P=0.09), and lower Global Mental Health (44.7 vs 41.0, P=0.03) during their hospitalization. In addition, patients with complete data experienced an average of 1.2 utilization events during the study, whereas those with missing data experienced an average of 2.1 utilization events (P=0.03).

Figure 2

The MSAS physical symptom severity and Global Physical Health scores from the index hospitalizations did not identify patients with a first utilization event within 30 days. However, patients with poor Global Mental Health and GSRH in the hospital were more likely to experience a utilization event within 14 days of discharge (Figure 3). During the postdischarge period, patients scoring poorly on each of the PRO measures trended toward a greater risk of an early utilization event, but the association between utilization and MSAS was most consistently significant (Figure 3A). In general, the associations with MSAS, Global Physical Health, and GSRH were stronger with the risk of utilization events within 14 days than within 30 days (Figure 3A,B,D). The Global Mental Health score was not associated with a subsequent utilization when measured during the 180‐day postdischarge period.

Figure 3

As shown in Table 2, Cox proportional hazard models incorporating covariates preserved most of the significant associations seen in the unadjusted analyses. Global Mental Health and good relative to poor GSRH obtained during the hospitalization remained significant. MSAS obtained at each postdischarge follow‐up trended positively with utilization and was statistically significant at 90 and 180 days. Global Physical Health obtained at each postdischarge follow‐up similarly trended negatively with utilization and was significant at 180 days. Each multivariate model incorporating a PRO measure with a significant coefficient contributed to better fit of the predictive model compared to the nested model without the PRO measure.

DISCUSSION

In this longitudinal cohort study, patients, on average, reported relatively severe symptoms, low PROMIS Global Physical and Mental Health scores, and poor GSRH during the inpatient stay in an urban safety‐net hospital. Symptom severity and Global Physical Health improved, on average, by 30 days before stabilizing, but their poor levels in the hospital did not predict 30‐day hospital‐based utilization events. On the other hand, Global Mental Health and GSRH were stable through hospitalizations, and patients scoring poorly on these measures were at greater risk of utilization events within 14 days of discharge. PRO measures obtained during the 180‐day postdischarge period trended toward distinguishing populations with greater baseline risk of proximate utilization events. However, MSAS physical symptom severity and Global Physical Health were more consistently predictive of these events at statistically significant levels compared to Global Mental Health and GSRH in our relatively small sample of patients. Each of these measures selectively improved the fit‐of‐risk prediction models for hospital‐based utilization.

Some of the heterogeneity in readmission risk is explained by differences in PRO measures. Although the MSAS score and Global Physical Health assessment were reliable predictors of utilization when measured in ambulatory settings, they were less discriminating during acute hospitalizations when everyone, on average, reported severe symptoms and poor function. Our results were consistent with other studies that demonstrated the fairly rapid recovery in symptoms that follow hospitalizations,[28, 29] and these measures may become informative of utilization risk as early as 2 weeks postdischarge. GSRH and Global Mental Health (a measure of health‐related quality of life) only predicted utilizations immediately at hospital discharge. As multidimensional measures that reflect physical, social, and emotional capacity, these measures may indicate vulnerabilities in patients least able to handle the stresses of the early postdischarge period.

There is growing momentum around collecting PRO measures in routine clinical care as quality indicators that capture patient‐centered concerns.[30] Our study explored a novel application of these measures whose routine collection will likely proliferate, not solely for the purpose of helping healthcare systems identify patients at risk of unplanned resource utilization. Although multidimensional PRO measures seldom reflect conditions directly modifiable by simple interventions, we believe that the association between physical symptom burden and utilization in our data reveals a possible target for practice improvement. Hospitalists have contributed enormously to shorter lengths of stay that risk sicker and quicker discharges.[31] To mitigate its potential side effects on symptom management, a discharge plan that acknowledges physical symptoms that sometimes persist or recur beyond the hospitalization may be appropriate. This may be accomplished by ensuring that acute symptoms are resolving, giving clear instructions for symptom management at home, as now the standard of care for conditions like asthma,[32] and explicitly communicating the presence of residual symptoms to providers entrusted with continuity care. As an effective feedback measure that can drive continuous quality improvement, we believe that a technology‐based surveillance strategy that spans both the inpatient and outpatient domains is necessary.[23]

There are some notable similarities and differences between the results of our study and a recent hospital‐based study of PRO measures that used data from the Multi‐Center Hospitalist Project.[16] The Physical Component Score of the SF12 is similar to the PROMIS Global Physical Health score in that both incorporate measures of physical function, perceived health, pain, and energy level. Curiously, the SF12 Physical Component Score, but not the PROMIS Global Physical Health score, was associated with 30‐day rehospitalizations. An important difference between the measures is where the SF12 asks about limitations during the past 4 weeks the PROMIS instrument inquires about physical function in general and levels of fatigue and pain in the past 7 days. Considering most hospitalizations last <7 days, the PROMIS instrument may better reflect the declines associated with the acute illness related to the hospitalization than the SF12 score. Additionally, the discrepancy between the association between hospital‐based GSRH and utilization in our study and the absence, thereof, in Hasan et al. is noteworthy. The difference here may be explained by their use of a 0‐ to 100‐point response scale in contrast to our study's verbally labeled 5‐point scale in the PROMIS instrument. The range of rating scales for survey questions is traditionally governed by the tension between the difficulty with mapping respondents judgment on an excessively large scale on one hand, and the failure of insufficient response options to discriminate between respondents with different underlying judgment on the other.[33] We suspect the former to be a drawback of the unlabeled 100‐point response scale, and conjecture that an association might be found in the Multi‐Center Hospitalist Study data if the responses were grouped into summative categories.

We recognize several limitations in our study. The first is the generalizability of our patient population to others, not insignificantly because of the high proportion of the uninsured (around 70% during the study period) and racial/ethnic minorities among them. Although utilization patterns are clearly affected by socioeconomic status,[34] there may also be differences in the way validated PRO measures are calibrated between patients of public and private healthcare systems.[35] Another limitation is our inability to count utilization events at institutions outside of the CCHHS during our study. However, because the study was conducted prior to Cook County's Medicaid expansion demonstration program as part of the Affordable Care Act,[36] many patients established in our system faced barriers to receiving nonemergency care outside of the CCHHS supporting our assumption that few of our patients were discharged from other hospitals. Causality cannot be established in observational studies. Consequently, high prior‐symptom burden may be associated with utilizations through unmeasured variables. Measures of symptom burden are vulnerable to overendorsement and amplification.[37, 38] Inferences based on statistical significance are affected by sample size, and our conclusions may change if conducted with a larger number of participants. Our response rates were excellent through the survey waves, but we did not achieve perfect follow‐up. Worse levels of PRO responses and higher levels of utilization among censored patients biased our results toward the null. Finally, although we did not find any predominant comorbidities associated with hospital‐based utilizations in our sample, our analyses may be vulnerable to inadequate control for illness severity, which may also have biased our results.

PRO measures are likely to be useful in clinical medicine.[39] But to fully apply the powers of PROs in informing clinically and operationally relevant outcomes, we must actively develop a system for obtaining these measures in routine clinical care. The availability of patient downtime makes hospitalizations conducive to gathering patient‐generated data, and may further enhance patient‐provider communication if survey output was readily available in electronic medical records. Exploring innovative strategies for collecting PROs in the hospital and beyond remains our future work.

Over the past 2 decades, the care of the hospitalized patient has changed dramatically. Hospitalists now account for the care of more than one‐third of general medicine inpatients, and this number is likely to grow.[1] The emergence of hospital medicine has resulted in a partnership between primary care physicians (PCPs) and hospitalists wherein hospitalists focus on acute medical issues requiring hospitalization, whereas more chronic issues unrelated to the reason for hospitalization remain largely the domain of the PCP.[2, 3]

However, several evolving financial and quality incentives have already begun to blur the distinction between inpatient and outpatient care. First, as private and public payers increasingly scrutinize readmission rates, it has become clear that the responsibility for patient outcomes extends beyond the day of discharge.[4] The birth of Accountable Care Organizations and patient‐centered medical homes may further blur distinctions between what has traditionally constituted inpatient and outpatient care.[5] Bundled payments may force providers to ensure that each visit, whether hospital‐ or clinic‐based, is taken as an opportunity to enact meaningful change.[6] The Centers for Medicare and Medicaid Services (CMS) are already tracking hospital performance on institution of medical therapy for certain conditions regardless of their relatedness to the reason for hospitalization.[7]

No published literature has yet examined the attitudes of inpatient and outpatient providers regarding this issue. Through a case‐based survey conducted at 3 large urban academic medical centers, we aimed to assess opinions among hospitalists and PCPs regarding the role of hospitalists in the management of conditions unrelated to the reason for admission. Our study had 2 main objectives: (1) to determine whether surveyed physicians were more likely to rate an inpatient intervention as appropriate when it related to the reason for admission as compared to interventions unrelated to the reason for admission; and (2) to determine whether these attitudes differed between PCPs and hospitalists.

METHODS

Survey Administration

Between October 23, 2012 and November 10, 2012, 3 emails containing a link to the online survey were sent to all hospitalist and hospital‐based PCPs at the 3 study institutions. The BIDMC Committee on Clinical Investigations, to whom authority was ceded by the remaining 2 study institutions, certified this research protocol as exempt.

RESULTS

DISCUSSION

In this survey‐based study of PCPs and hospitalists across 3 Boston‐area academic medical centers, we found that: (1) physicians were more likely to see inpatient interventions as appropriate when those interventions dealt with the reason for admission as compared to interventions unrelated to the reason for admission; and (2) PCPs were more likely than hospitalists to feel that inpatient interventions were appropriate, even when they targeted chronic conditions unrelated to the reason for admission. To our knowledge, this study represents the first investigation into the attitudes of PCPs and hospitalists regarding the inpatient management of conditions unrelated to the reason for admission.

That surveyed physicians, regardless of role, were less likely to report an intervention unrelated to the reason for hospitalization as appropriateeven those with likely mortality benefitsuggests that opportunities to affect meaningful change may be missed in a healthcare system that adheres to strict inpatient and outpatient roles. For several of the cases, a change in therapy could lead to benefit soon after implementation. For example, aldosterone antagonists reduce mortality as early as 1 month after initiation in select patients.[8] If a major goal of inpatient care is to reduce 30‐day mortality, it could be argued that hospitalists should more actively adjust congestive heart failure therapy in appropriate inpatients, even when this is not their admitting diagnosis.

For some conditions, CMS is already tracking hospital performance. Since 2003, hospitals have been required to document whether a patient with congestive heart failure (either acute or chronic and regardless of the relationship to admission) was prescribed an angiotensin‐converting enzyme (ACE) inhibitor or angiotensin receptor blocker (ARB) at the time of discharge.[7] CMS has determined that the proven benefits of ACE inhibitors and ARBs confer hospital accountability for their inclusion in appropriate patients, independent of the acuity of heart failure. There are many potential therapeutic maneuvers on which health systems (and their physicians) may be graded, and accepting the view that a hospitalization provides a window of opportunity for medical optimization may allow for more fruitful interventions and more patient‐centered care.

Despite the potential benefits of addressing chronic medical issues during hospitalization, there are important limitations on what can and/or should be done in the hospital setting. Hospitalizations are a time of fluctuating clinical status, which continues beyond discharge and is often accompanied by several medication changes.[9] In our study, more than 20% of those who believed that a medication intervention was inappropriate selected I do not believe hospitalization is the right time to start this new medication as one of their explanations. Although some medication interventions have been shown in randomized controlled trials to reduce short‐term mortality, the ability to generalize these findings to the average hospitalized patient with multiple comorbidities, concurrent medication changes, and rapidly fluctuating clinical status is limited. Furthermore, there are interventions most would agree should not be dealt with in the hospital (eg, screening colonoscopy) and encounters that may be too short to allow for change (eg, 24‐hour observation). These issues notwithstanding, the average 4‐day hospitalization likely provides an opportunity for monitored change that may currently be underutilized.

Our study suggests several additional explanations for physicians' current practice and opinions. Only 6.5% of respondents who answered that an intervention was inappropriate indicated as a justification that I do not believe this is an appropriate pharmacologic intervention for this particular medical problem. This suggests that the hesitancy has little to do with a lack of benefit but instead relates to systems issues (eg, access to all pertinent records and concerns regarding follow‐up testing) and perceived limitations to what a hospitalist should and should not do without actively involving the PCP. There are likely additional concerns that the medical record and/or patient histories do not fully outline the rationale for exclusion or inclusion of particular medications. Advances in information technology that enhance information exchange and enable streamlined communication may help to address these perceived barriers. However, an additional barrier may be trust, as PCPs appear more concerned that hospitalists will not review all the pertinent records or discuss risks and benefits before enacting important medication changes. Increased attempts at communication between hospitalists and outpatient providers may help to build trust and alleviate concerns regarding the loss of information that often occurs both on admission and at discharge.

We also noted that PCPs were more likely than hospitalists to feel that inpatient interventions were appropriate, even when targeting chronic conditions unrelated to the reason for admission. It may be that PCPs, with an increasing number of problems to address per outpatient visit,[10, 11] are more open to hospitalists managing any medical problems during their patients' admissions. At the same time, with increased acuity[12, 13, 14] and shortened length of stays,[15, 16] hospitalists have only a finite amount of time to ensure acute issues are managed, leaving potentially modifiable chronic conditions to the outpatient setting. These differences aside, a minority of both PCPs and hospitalists in our study were ready to embrace the idea of hospitalists playing a larger role in the management of conditions unrelated to the reason for hospitalization.

Even though our study benefits from its multisite design, there are a number of limitations. First, although we crafted our survey with input from general medicine focus groups, our survey instrument has not been validated. In addition, the cases are necessarily contrived and do not take into account the complexities of inpatient medicine. Furthermore, though our goal was to create paired cases that isolate a management decision as being simply based on whether it was related or unrelated to the reason for admission, it is possible that other factors, not captured by our survey, influenced the responses. For example, the benefits of aspirin as part of secondary prevention are not equal to the benefits in an acute MI.[17]

In an attempt to isolate the hospitalists' role in these management decisions, respondents were instructed to assume that the decisions were being made without discussing it with the primary care physician, but that the hospitalist would communicate the details of any hospitalization at the time of discharge. They were also instructed to assume that the hospitalist has access to the patient's outpatient electronic medical record. These assumptions were made to address concerns regarding the flow of information and communication, and to simulate the ideal system from a communication and information accessibility standpoint. Had these assumptions not been placed, the responses may have differed. It is likely that PCPs and hospitalists practicing in systems without shared, accessible inpatient/outpatient medical records would be even more reluctant to enact medication changes unrelated to the reason for admission.

Along the same lines, our physician cohort consisted of several metropolitan academic physician groups, in which hospitalists have had a presence for almost 20 years. As a result, our findings may not be generalizable to other academic hospitals, community‐based hospitalist programs, or nonhospital‐based PCP practices. Finally, we do not know whether survey nonresponders differed from responders in ways that could have meaningfully affected our results.

In conclusion, our findings suggest that both PCPs and hospitalists see the management of conditions unrelated to the reason for admission as less appropriate than the management of conditions related to the reason for admission. Our findings also suggest that PCPs may be more open to this practice when compared to hospitalists. Failure to capitalize on opportunities for meaningful medical interventions, independent of patient location, suggests a possible lack of patient centeredness in the current partnership between PCPs and hospitalists. Further studies should examine existing barriers and investigate interventions designed to address those barriers, in an effort to improve both quality of care and the degree of patient‐centeredness in our current healthcare system.

Disclosures: Dr. Herzig is supported by a grant from the National Institute on Aging (K23 AG042459). Dr. Herzig had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Author contributions: study concept and design, Breu, Allen‐Dicker, Mueller, Herzig; acquisition of data, Breu, Allen‐Dicker, Mueller, Palamara, Herzig; analysis and interpretation of data, Breu, Allen‐Dicker, Hinami, Herzig; drafting of the manuscript, Breu; critical revision of the manuscript for important intellectual content, Breu, Allen‐Dicker, Mueller, Palamara, Hinami, Herzig; statistical analysis, Allen‐Dicker, Hinami, Herzig; study supervision, Breu, Herzig. This study was presented as a poster at the Society of Hospital Medicine National Meeting, Washington, DC, May 17, 2013.

Over the past 2 decades, the care of the hospitalized patient has changed dramatically. Hospitalists now account for the care of more than one‐third of general medicine inpatients, and this number is likely to grow.[1] The emergence of hospital medicine has resulted in a partnership between primary care physicians (PCPs) and hospitalists wherein hospitalists focus on acute medical issues requiring hospitalization, whereas more chronic issues unrelated to the reason for hospitalization remain largely the domain of the PCP.[2, 3]

However, several evolving financial and quality incentives have already begun to blur the distinction between inpatient and outpatient care. First, as private and public payers increasingly scrutinize readmission rates, it has become clear that the responsibility for patient outcomes extends beyond the day of discharge.[4] The birth of Accountable Care Organizations and patient‐centered medical homes may further blur distinctions between what has traditionally constituted inpatient and outpatient care.[5] Bundled payments may force providers to ensure that each visit, whether hospital‐ or clinic‐based, is taken as an opportunity to enact meaningful change.[6] The Centers for Medicare and Medicaid Services (CMS) are already tracking hospital performance on institution of medical therapy for certain conditions regardless of their relatedness to the reason for hospitalization.[7]

No published literature has yet examined the attitudes of inpatient and outpatient providers regarding this issue. Through a case‐based survey conducted at 3 large urban academic medical centers, we aimed to assess opinions among hospitalists and PCPs regarding the role of hospitalists in the management of conditions unrelated to the reason for admission. Our study had 2 main objectives: (1) to determine whether surveyed physicians were more likely to rate an inpatient intervention as appropriate when it related to the reason for admission as compared to interventions unrelated to the reason for admission; and (2) to determine whether these attitudes differed between PCPs and hospitalists.

METHODS

Survey Administration

Between October 23, 2012 and November 10, 2012, 3 emails containing a link to the online survey were sent to all hospitalist and hospital‐based PCPs at the 3 study institutions. The BIDMC Committee on Clinical Investigations, to whom authority was ceded by the remaining 2 study institutions, certified this research protocol as exempt.

RESULTS