RESEARCH/Original article
An economic analysis of a nurse-led telephone triage service
Journal of Telemedicine and Telecare 2014, Vol. 20(6) 330–338 ! The Author(s) 2014 Reprints and permissions: sagepub.co.uk/journalsPermissions.nav DOI: 10.1177/1357633X14545430 jtt.sagepub.com
Jessica L Navratil-Strawn1, Ronald J Ozminkowski2 and Stephen K Hartley3
Summary Telephone nurse lines help callers to select the most appropriate site and level of care for acute conditions. We examined whether compliance with nurse recommendations was associated with lower average health care expenditure, and identified the employer characteristics associated with higher than average savings. Telephone calls to a nurse-led help line made by commercial health plan members who worked for large employers were identified. The callers’ intention before calling and the nurse recommendation regarding site/level of care were recorded. Compliance was determined using medical claims during a 30-day post-call observation period and was based on adherence to nurse recommendations. A total of 132,509 calls during 2012 were identified for the study. Nurse recommendations were that 31% of the callers seek a higher level of care than mentioned at the start of the call, 25% use a lower level of care and 44% pursue their originally intended level of care. After regression-based adjustment, the average medical expenditures were compared between compliers and non-compliers. Overall, 57% of callers were compliant with nurse recommendations. The average expenditures were $328 lower among compliant callers. A logistic regression analysis identified employer characteristics positively associated with achieving higher than average savings. These were having a low employee-to-dependent ratio, a headquarters in the Western region of the US, a low prospective health risk score, and participation by the employer in a targeted communication campaign. Compliance with the triage recommendations resulted in lower average health care expenditures, and several characteristics were positively associated with achieving higher savings. Accepted: 27 June 2014
Introduction Providing patients with guidance over the telephone has become an important service offered by many health care organizations. Telephone triage services can offer patients information and education, and can help direct patients to the most appropriate level or site of care for their acute problems.1 This guidance could reduce the burden on the health care system. The opportunity for such savings is evidenced in previous studies which showed that some patients may not be utilizing health care services appropriately. Examples include patients who visit an emergency department but who do not have an urgent care need.2 To provide the best results, telephone triage programmes should be staffed with experienced and well-trained nurses, and structured protocols or clinical algorithms should be used when providing recommendations to patients about where to receive services for their acute problems.3 In particular, nurses who provide triage advice should have good communication skills, so that they can effectively gather information about a patient’s problem, structure the call in an efficient manner, and check the caller’s understanding of the advice given.4
A variety of organizations use telephone triage services, including employers, commercial health plans, and government health programmes such as Medicaid.1 Approximately 100 million Americans had access to telephone triage services in 2001. Potential benefits that telephone triage provide include improved quality of care, better ability to manage demand, and reduction in healthcare costs.1 Patient satisfaction with telephone triage services tends to be high, with about 90% or more of individuals indicating in surveys that they were satisfied using triage services.5 Realization of benefits from telephone triage services is in part contingent upon patient compliance with the nurse’s recommendations.6 Prior research suggests that the majority of patients adhere to treatment recommendations they receive from telephone triage nurses, 1
Optum Consumer Solutions Group, Golden Valley, Minnesota, USA Optum Consumer Solutions Group, Ann Arbor, Michigan, USA 3 Optum Consumer Solutions Group, Phoenix, Arizona, USA 2
Corresponding author: Jessica L Navratil-Strawn, Optum Consumer Solutions Group, 6300 Olson Memorial Hwy, Golden Valley, MN 55427, USA. Email: [email protected]
Navratil-Strawn et al. although overall estimates of compliance vary among past studies, ranging from 55% to 88%.6 Differences in reported rates of patient compliance may be due to a variety of factors, such as study population or study design. For example, studies gauging compliance through selfreported data (e.g. a follow-up survey) have been found to report higher rates of adherence than studies using an independent assessment (e.g. medical claims) to measure compliance.6 In addition, some studies have shown that adherence varies based on type of referral. For example, higher adherence has been reported among patients who received recommendations to seek emergency services than among those advised to make a physician appointment. Miscommunication or misunderstanding between the nurse and the patient may contribute to noncompliance with the nurse’s recommendation.7 Other reasons for noncompliance may include a change in health status or symptoms, lack of trust, or inability to access health care.7 By guiding patients toward an appropriate level of care in a timely manner, telephone triage programmes are expected to help reduce health care costs. Two studies have reported reductions in health care spending among users of triage programmes (by about 39% and 16%).6 These studies compared a patient’s actual health services utilization following a call to what the caller’s original intent had been, but did not take into account administrative costs of running a telephone triage service. Expenses associated with running a telephone triage programme may be high, and include staffing costs for nurses, software costs, and facilities costs such as rent, furniture and utilities.1 Other studies have estimated the return on investment (ROI) associated with telephone triage programmes by calculating how much money is saved in health care costs for each dollar spent on the programme. O’Connell et al. reported an ROI of $1.70.8 More recently, Navratil-Strawn et al. reported an ROI of $1.59 for a triage programme used by a national population of patients.9 This suggests that telephone triage programmes can be cost-effective to operate. Previous studies which evaluated benefits from triage programmes in the US have been restricted to a specific geographic region or to an elderly population, and thus have limited generalizability. Also, to the best of our knowledge, no previous studies have described the employer-level characteristics associated with above average savings for a nurse triage programme. We have therefore examined the employer-level characteristics associated with achieving higher than average savings for employers who contracted for insurance with a large US health plan.
Methods The Optum NurseLine is a 24-hour telephone service that provides coaching and personalized support. It also provides evidence-based information about medical
331 conditions, specifics about treatment options, navigation to higher-quality providers and connections to other health resources. The service is available to approximately 8 million people who are insured by a national managed care organization through their employers. About 46% of the calls involve triage, where the nurse recommends a course of action for an acute health problem. Based on an evaluation of the caller’s symptoms, the nurse may confirm the caller’s original intention or recommend a different level of care for the caller’s problem. The unit of observation for the present study was a triage call. Calls placed during 2012 by members who worked for 295 different large employers were eligible for inclusion in the study. Calls from members who were not continuously enrolled with medical coverage for a 6-month period before the call or a 30-day observation period following the call were excluded. Also, calls from people with missing demographic information or who were no longer eligible due to employment status were excluded. Repeat calls placed less than 30 days apart by the same member were excluded to ensure accurate measurement of the member’s compliance and outcome for each call. However, calls placed more than 30 days apart by the same member were eligible for inclusion. Finally, calls from people who had no health care expenditure in the 3-month period following the call were excluded. Demographic measures included the caller’s age, gender (assessed via eligibility records) and two variables measuring location (whether the caller resided in a rural or an urban location, and the national census region). Socioeconomic variables included imputations of the caller’s race and income. We only had access to administrative and call data, so these variables were coded based on the zip code where the caller resided. Race was categorized as high, medium or low, depending on the percentage of minority residents in the caller’s zip code. Income was inferred as high, medium-high, medium or low based on whether the median income in the caller’s zip code area was in the highest, second-highest, third-highest or lowest quartile, according to US Census records. A binary variable was created to account for the impact of having one or more claims for prescription medicines in the 6 months before the call. Health status measures included a prospective risk score,10 which used information about diagnoses observed in the claims data to generate a score, centred around 1.0, to estimate whether allowed charges would be higher or lower than average in the following year. The risk score can be viewed as a predicted health status for the following year. Other health status measures accounted for differences in the supply of health care services in the areas where callers lived, because these are known to influence health care utilization and expenditures.11 Therefore, measures based on the number of primary care physicians, specialists and hospital beds in the caller’s zip code of residence were included. The number of physicians and specialists were calculated per 100,000 residents, while
332 the number of hospital beds was calculated per 1000 residents in the caller’s hospital service area.12 We also accounted for the type of the health complaint made by the caller during the call. The calls were categorized based on the caller’s chief health care complaint diagnostic category (e.g. circulatory), and were also categorized as occurring during a weekday or weekend, since weekend calls may be more urgent. The nurse’s recommended level or site of service was captured. Possible choices included the emergency room, urgent care clinic, doctor’s office visit or self-treatment at home. Finally, a variable was used to describe whether a call was from a repeat caller (a member who in 2012 called the help line two or more times that were greater than 30 days apart).
Assignment to care paths When a caller contacted the help line, a nurse captured the caller’s intended site or level of care before engaging with nurse. Approximately 75 registered nurses, with over 10 years of clinical experience each, provided the service during the study period. The subsequent recommendation from the nurse on the call was also noted, based on an assessment of the caller’s symptoms and condition. For quality control purposes, supervisors coach nurses who make consistently higher or lower recommendations than their peers for several months. The medical directors also review the presenting diagnoses to see if there are any patterns that indicate the need for further education. The nurse pay is not tied to the recommendations provided. Therefore, there is no incentive to make any recommendation outside what the clinical algorithms and evidence based guidelines suggest. After the nurse interaction, post-call behaviour was determined by analysing the caller’s medical claims data to see which site of care had been chosen. The recommended movement in care levels was categorized into five possible care paths. Two care paths were categorized as referrals to a lower level of care, when people were advised to seek a less immediate and intensive level of care than originally intended. These two lower levels of care were different from each other based on whether the caller’s intention had been to visit an emergency room. Thus we distinguished between callers who originally intended to visit an emergency room but who were directed by the nurse to a lower level of care (such as self-care, making an appointment to see a medical doctor, or going to urgent care), versus callers who did not say they intended to use the emergency room. Another two care paths were categorized as referrals to a higher level of care. These people were encouraged by the nurse to seek a more immediate and intensive level of care than they told the nurse they were thinking about at the start of the call. These two care paths were distinguished by whether the nurse recommended going to the emergency room or not.
Journal of Telemedicine and Telecare 20(6) Finally, one care path was categorized as confirm and prepare, when the nurse confirmed the caller’s intended course of action and helped prepare the caller for that action. Callers were classified as ‘‘compliers’’ if they received care that was consistent with the nurse’s recommendation, and they were classified as non-compliers if they did not receive care that was consistent with the nurse’s recommendation. Compliance was determined using medical claims data during the early part of the 30-day observation period following the call. Compliance with a recommendation of emergency room treatment required presence of claims-based evidence of such a visit within 2 days of the call. Compliance with a recommended visit to a doctor’s office required claims-based evidence of such a visit occurring within 14 days of the call. Compliance with a recommendation of self-care required absence of claimsbased evidence of an emergency room visit or an urgent care visit within 2 days of the call and the absence of claims-based evidence of a doctor’s office visit within 14 days of the call. As it can be difficult to identify urgent care visits using claims data, callers were considered to be compliant with a recommendation to go to urgent care if they had claimsbased evidence of an urgent care clinic visit or doctor’s office visit within 2 days of the call. Finally, callers were considered to be compliant with a recommendation to call their doctor if they had claims-based evidence of a doctor’s office visit within 14 days of the call, or an absence of any claims for 14 days after the call.
Analyses Chi-square and Student’s t-tests were used to test for differences in the categorical and continuous variables that were measured for compliers and non-compliers prior to any statistical adjustment. Then propensity score weighting13–16 was used to equalize case-mix differences between compliers and non-compliers in each of the five care path categories so that more accurate measures of the effect of compliance could be obtained. To determine a propensity score weight, the demographic, socioeconomic, supply side and other health status predictor variables noted above were first used in a logistic regression model to estimate the effect of these variables on complying or not with nurse advice. (One logistic regression was used for each of the five care paths.) The propensity score (i.e. the predicted probability of following the nurse’s recommendation or not) was obtained for each caller from the logistic regression results. This predicted probability was then used to construct a weighting variable for subsequent statistical analyses. The value of the weight for a caller who complied was defined as 1.0 divided by his or her predicted probability of complying with the nurse recommendation. For non-compliers the weighting variable was defined as 1.0 divided by his or her predicted probability of
Navratil-Strawn et al. non-compliance. Once the weights were obtained they were applied to data from the 30-day post call observation period to help adjust for case-mix differences between compliers and non-compliers. To do this five propensityscore-weighted general linear regression (GLM) models were estimated, one for each care path. Variables adjusted for in the GLM regressions were demographics (age, gender, proximity to urban area, and region), health status (prior utilization counts for physician visits, emergency room visits, inpatient visits, risk score and membermonths), socioeconomic (income and minority status), supply side (numbers of hospital beds, specialists and primary care physicians in the caller’s area of residence) and call activity (chief complaint, day of call, nurse recommendation and repeat calls). Expenditures associated with each care path were calculated using medical and pharmacy claims data for the 30-day observation period following the call, and savings were estimated from each GLM regression as the difference in average expenditures between compliers and non-compliers. Savings were calculated for both an unweighted, GLM-adjusted sample and a propensity score-weighted, GLM-adjusted sample. Savings occurred when the regression-adjusted complier expenditures were less than the regression-adjusted non-complier expenditures. Total programme savings were estimated by summing weighted, GLM-adjusted total savings or losses for all five care paths. In a separate analysis, logistic regression modelling was used to identify the characteristics associated with an employer who achieved higher than average savings. The sample members for this analysis included employers who paid for the telephone triage service during 2012. The dependent variable for this analysis was coded as 1 if the employer’s savings were higher than average or 0 if savings were average or lower than average. Employers with missing demographic, socioeconomic or occupational information were excluded, making the sample size 295 employers. The logistic regression analysis produced an odds ratio which indicated the likelihood that a specific employer-level characteristic was associated with higher than average savings. Odds ratios greater than 1.0 indicated a higher propensity to achieve savings. The independent variables used in this logistic regression analysis were employer-level characteristics which measured the gender percentages of its employees, the percentages of employees living in high minority areas, and participation by the employer in a targeted communication programme that was meant to increase the use of the help line service. Other independent variables measured whether the employer had other care coordination programmes, such as wellness or disease management programmes. Additional independent variables included the average value of the prospective risk score for the employees of the firm, the industry type associated with the employer, the region where the employer headquarters was located, a ratio measuring the number of employees to dependents in the employer’s health plan and employer size.
333
Results A total of 188,689 triage calls were identified. From these, 13,008 were excluded due to inability to match call records from the help line database to eligibility, followed by another 11,110 calls that were excluded as repeat calls, then 3191 calls were excluded due to lack of medical coverage, 21,536 calls were excluded due to no health care expenditures in the 3 months following the call, and finally 7335 calls were excluded due to missing demographic information. This left a total of 132,509 remaining calls (70%). Of these, over half of the care recommendations made by the nurse were different from the caller’s original intent, and nurses were more likely to recommend that callers sought a higher level of care rather than a lower level of care. Overall, nurses recommended that a caller should seek a higher level of care 31% of the time, a lower level of care 25% of the time, and agreed with the caller’s original intention 44% of the time. For the higher level of care paths, nurses recommended an emergency room visit for 5% of calls and a non-emergency visit for 26% of calls. For the lower level of care paths, nurses recommended a lower level of care from an emergency room visit for 12% of calls and a lower level of care from a non-emergency visit for 13% of calls. Of the calls retained in the final sample, 75,276 were made by compliant callers and 57,233 were made by noncompliant callers (overall compliance 57%). Within the various care paths, a number of demographic and clinical characteristics differed significantly between compliant and non-compliant callers. These included the chief complaint, prior 6-month health care utilization measures, whether prescription expenses were incurred, geographical region, imputed income level, the prospective risk score, day of the call, gender and age (Tables 1–3). As these differences could affect health care costs, propensity score weighting was performed to adjust for the casemix differences between compliant and non-compliant callers. After weighting, the prospective 3-month cost risk score measure remained significantly different between compliant and non-compliant callers in 3 out of the 5 care paths, but all other differences in clinical and demographic characteristics were removed (data not shown). Generalized linear modelling was then performed to further control for the remaining case-mix differences between compliers and non-compliers. Within each care path, the average 30-day weighted, GLM-adjusted medical and prescription expenditures were compared between compliant and non-compliant callers (Table 4). In the weighted sample, compliant callers had lower average expenditures than non-compliant callers for four out of the five care paths. Specifically, for compliers who were redirected to lower care rather than using an emergency room, the average savings were $1862 (P < 0.001). For compliers who were redirected to a lower level of care from a non-emergency room setting, the average savings were $620 (P < 0.001). Compliers who were not directed to a different level or site of care had average
334
Journal of Telemedicine and Telecare 20(6)
Table 1. Characteristics of groups receiving a recommendation to seek a lower level of care. Redirect to lower care (from emergency room)
Redirect to lower care (from non-emergency room)
Complier (N ¼ 9401)
Complier (n ¼ 9339)
Non-complier (N ¼ 7886)
P-value
Non-complier (n ¼ 7835)
P-value
Age category, years (%)
An economic analysis of a nurse-led telephone triage service
Journal of Telemedicine and Telecare 2014, Vol. 20(6) 330–338 ! The Author(s) 2014 Reprints and permissions: sagepub.co.uk/journalsPermissions.nav DOI: 10.1177/1357633X14545430 jtt.sagepub.com
Jessica L Navratil-Strawn1, Ronald J Ozminkowski2 and Stephen K Hartley3
Summary Telephone nurse lines help callers to select the most appropriate site and level of care for acute conditions. We examined whether compliance with nurse recommendations was associated with lower average health care expenditure, and identified the employer characteristics associated with higher than average savings. Telephone calls to a nurse-led help line made by commercial health plan members who worked for large employers were identified. The callers’ intention before calling and the nurse recommendation regarding site/level of care were recorded. Compliance was determined using medical claims during a 30-day post-call observation period and was based on adherence to nurse recommendations. A total of 132,509 calls during 2012 were identified for the study. Nurse recommendations were that 31% of the callers seek a higher level of care than mentioned at the start of the call, 25% use a lower level of care and 44% pursue their originally intended level of care. After regression-based adjustment, the average medical expenditures were compared between compliers and non-compliers. Overall, 57% of callers were compliant with nurse recommendations. The average expenditures were $328 lower among compliant callers. A logistic regression analysis identified employer characteristics positively associated with achieving higher than average savings. These were having a low employee-to-dependent ratio, a headquarters in the Western region of the US, a low prospective health risk score, and participation by the employer in a targeted communication campaign. Compliance with the triage recommendations resulted in lower average health care expenditures, and several characteristics were positively associated with achieving higher savings. Accepted: 27 June 2014
Introduction Providing patients with guidance over the telephone has become an important service offered by many health care organizations. Telephone triage services can offer patients information and education, and can help direct patients to the most appropriate level or site of care for their acute problems.1 This guidance could reduce the burden on the health care system. The opportunity for such savings is evidenced in previous studies which showed that some patients may not be utilizing health care services appropriately. Examples include patients who visit an emergency department but who do not have an urgent care need.2 To provide the best results, telephone triage programmes should be staffed with experienced and well-trained nurses, and structured protocols or clinical algorithms should be used when providing recommendations to patients about where to receive services for their acute problems.3 In particular, nurses who provide triage advice should have good communication skills, so that they can effectively gather information about a patient’s problem, structure the call in an efficient manner, and check the caller’s understanding of the advice given.4
A variety of organizations use telephone triage services, including employers, commercial health plans, and government health programmes such as Medicaid.1 Approximately 100 million Americans had access to telephone triage services in 2001. Potential benefits that telephone triage provide include improved quality of care, better ability to manage demand, and reduction in healthcare costs.1 Patient satisfaction with telephone triage services tends to be high, with about 90% or more of individuals indicating in surveys that they were satisfied using triage services.5 Realization of benefits from telephone triage services is in part contingent upon patient compliance with the nurse’s recommendations.6 Prior research suggests that the majority of patients adhere to treatment recommendations they receive from telephone triage nurses, 1
Optum Consumer Solutions Group, Golden Valley, Minnesota, USA Optum Consumer Solutions Group, Ann Arbor, Michigan, USA 3 Optum Consumer Solutions Group, Phoenix, Arizona, USA 2
Corresponding author: Jessica L Navratil-Strawn, Optum Consumer Solutions Group, 6300 Olson Memorial Hwy, Golden Valley, MN 55427, USA. Email: [email protected]
Navratil-Strawn et al. although overall estimates of compliance vary among past studies, ranging from 55% to 88%.6 Differences in reported rates of patient compliance may be due to a variety of factors, such as study population or study design. For example, studies gauging compliance through selfreported data (e.g. a follow-up survey) have been found to report higher rates of adherence than studies using an independent assessment (e.g. medical claims) to measure compliance.6 In addition, some studies have shown that adherence varies based on type of referral. For example, higher adherence has been reported among patients who received recommendations to seek emergency services than among those advised to make a physician appointment. Miscommunication or misunderstanding between the nurse and the patient may contribute to noncompliance with the nurse’s recommendation.7 Other reasons for noncompliance may include a change in health status or symptoms, lack of trust, or inability to access health care.7 By guiding patients toward an appropriate level of care in a timely manner, telephone triage programmes are expected to help reduce health care costs. Two studies have reported reductions in health care spending among users of triage programmes (by about 39% and 16%).6 These studies compared a patient’s actual health services utilization following a call to what the caller’s original intent had been, but did not take into account administrative costs of running a telephone triage service. Expenses associated with running a telephone triage programme may be high, and include staffing costs for nurses, software costs, and facilities costs such as rent, furniture and utilities.1 Other studies have estimated the return on investment (ROI) associated with telephone triage programmes by calculating how much money is saved in health care costs for each dollar spent on the programme. O’Connell et al. reported an ROI of $1.70.8 More recently, Navratil-Strawn et al. reported an ROI of $1.59 for a triage programme used by a national population of patients.9 This suggests that telephone triage programmes can be cost-effective to operate. Previous studies which evaluated benefits from triage programmes in the US have been restricted to a specific geographic region or to an elderly population, and thus have limited generalizability. Also, to the best of our knowledge, no previous studies have described the employer-level characteristics associated with above average savings for a nurse triage programme. We have therefore examined the employer-level characteristics associated with achieving higher than average savings for employers who contracted for insurance with a large US health plan.
Methods The Optum NurseLine is a 24-hour telephone service that provides coaching and personalized support. It also provides evidence-based information about medical
331 conditions, specifics about treatment options, navigation to higher-quality providers and connections to other health resources. The service is available to approximately 8 million people who are insured by a national managed care organization through their employers. About 46% of the calls involve triage, where the nurse recommends a course of action for an acute health problem. Based on an evaluation of the caller’s symptoms, the nurse may confirm the caller’s original intention or recommend a different level of care for the caller’s problem. The unit of observation for the present study was a triage call. Calls placed during 2012 by members who worked for 295 different large employers were eligible for inclusion in the study. Calls from members who were not continuously enrolled with medical coverage for a 6-month period before the call or a 30-day observation period following the call were excluded. Also, calls from people with missing demographic information or who were no longer eligible due to employment status were excluded. Repeat calls placed less than 30 days apart by the same member were excluded to ensure accurate measurement of the member’s compliance and outcome for each call. However, calls placed more than 30 days apart by the same member were eligible for inclusion. Finally, calls from people who had no health care expenditure in the 3-month period following the call were excluded. Demographic measures included the caller’s age, gender (assessed via eligibility records) and two variables measuring location (whether the caller resided in a rural or an urban location, and the national census region). Socioeconomic variables included imputations of the caller’s race and income. We only had access to administrative and call data, so these variables were coded based on the zip code where the caller resided. Race was categorized as high, medium or low, depending on the percentage of minority residents in the caller’s zip code. Income was inferred as high, medium-high, medium or low based on whether the median income in the caller’s zip code area was in the highest, second-highest, third-highest or lowest quartile, according to US Census records. A binary variable was created to account for the impact of having one or more claims for prescription medicines in the 6 months before the call. Health status measures included a prospective risk score,10 which used information about diagnoses observed in the claims data to generate a score, centred around 1.0, to estimate whether allowed charges would be higher or lower than average in the following year. The risk score can be viewed as a predicted health status for the following year. Other health status measures accounted for differences in the supply of health care services in the areas where callers lived, because these are known to influence health care utilization and expenditures.11 Therefore, measures based on the number of primary care physicians, specialists and hospital beds in the caller’s zip code of residence were included. The number of physicians and specialists were calculated per 100,000 residents, while
332 the number of hospital beds was calculated per 1000 residents in the caller’s hospital service area.12 We also accounted for the type of the health complaint made by the caller during the call. The calls were categorized based on the caller’s chief health care complaint diagnostic category (e.g. circulatory), and were also categorized as occurring during a weekday or weekend, since weekend calls may be more urgent. The nurse’s recommended level or site of service was captured. Possible choices included the emergency room, urgent care clinic, doctor’s office visit or self-treatment at home. Finally, a variable was used to describe whether a call was from a repeat caller (a member who in 2012 called the help line two or more times that were greater than 30 days apart).
Assignment to care paths When a caller contacted the help line, a nurse captured the caller’s intended site or level of care before engaging with nurse. Approximately 75 registered nurses, with over 10 years of clinical experience each, provided the service during the study period. The subsequent recommendation from the nurse on the call was also noted, based on an assessment of the caller’s symptoms and condition. For quality control purposes, supervisors coach nurses who make consistently higher or lower recommendations than their peers for several months. The medical directors also review the presenting diagnoses to see if there are any patterns that indicate the need for further education. The nurse pay is not tied to the recommendations provided. Therefore, there is no incentive to make any recommendation outside what the clinical algorithms and evidence based guidelines suggest. After the nurse interaction, post-call behaviour was determined by analysing the caller’s medical claims data to see which site of care had been chosen. The recommended movement in care levels was categorized into five possible care paths. Two care paths were categorized as referrals to a lower level of care, when people were advised to seek a less immediate and intensive level of care than originally intended. These two lower levels of care were different from each other based on whether the caller’s intention had been to visit an emergency room. Thus we distinguished between callers who originally intended to visit an emergency room but who were directed by the nurse to a lower level of care (such as self-care, making an appointment to see a medical doctor, or going to urgent care), versus callers who did not say they intended to use the emergency room. Another two care paths were categorized as referrals to a higher level of care. These people were encouraged by the nurse to seek a more immediate and intensive level of care than they told the nurse they were thinking about at the start of the call. These two care paths were distinguished by whether the nurse recommended going to the emergency room or not.
Journal of Telemedicine and Telecare 20(6) Finally, one care path was categorized as confirm and prepare, when the nurse confirmed the caller’s intended course of action and helped prepare the caller for that action. Callers were classified as ‘‘compliers’’ if they received care that was consistent with the nurse’s recommendation, and they were classified as non-compliers if they did not receive care that was consistent with the nurse’s recommendation. Compliance was determined using medical claims data during the early part of the 30-day observation period following the call. Compliance with a recommendation of emergency room treatment required presence of claims-based evidence of such a visit within 2 days of the call. Compliance with a recommended visit to a doctor’s office required claims-based evidence of such a visit occurring within 14 days of the call. Compliance with a recommendation of self-care required absence of claimsbased evidence of an emergency room visit or an urgent care visit within 2 days of the call and the absence of claims-based evidence of a doctor’s office visit within 14 days of the call. As it can be difficult to identify urgent care visits using claims data, callers were considered to be compliant with a recommendation to go to urgent care if they had claimsbased evidence of an urgent care clinic visit or doctor’s office visit within 2 days of the call. Finally, callers were considered to be compliant with a recommendation to call their doctor if they had claims-based evidence of a doctor’s office visit within 14 days of the call, or an absence of any claims for 14 days after the call.
Analyses Chi-square and Student’s t-tests were used to test for differences in the categorical and continuous variables that were measured for compliers and non-compliers prior to any statistical adjustment. Then propensity score weighting13–16 was used to equalize case-mix differences between compliers and non-compliers in each of the five care path categories so that more accurate measures of the effect of compliance could be obtained. To determine a propensity score weight, the demographic, socioeconomic, supply side and other health status predictor variables noted above were first used in a logistic regression model to estimate the effect of these variables on complying or not with nurse advice. (One logistic regression was used for each of the five care paths.) The propensity score (i.e. the predicted probability of following the nurse’s recommendation or not) was obtained for each caller from the logistic regression results. This predicted probability was then used to construct a weighting variable for subsequent statistical analyses. The value of the weight for a caller who complied was defined as 1.0 divided by his or her predicted probability of complying with the nurse recommendation. For non-compliers the weighting variable was defined as 1.0 divided by his or her predicted probability of
Navratil-Strawn et al. non-compliance. Once the weights were obtained they were applied to data from the 30-day post call observation period to help adjust for case-mix differences between compliers and non-compliers. To do this five propensityscore-weighted general linear regression (GLM) models were estimated, one for each care path. Variables adjusted for in the GLM regressions were demographics (age, gender, proximity to urban area, and region), health status (prior utilization counts for physician visits, emergency room visits, inpatient visits, risk score and membermonths), socioeconomic (income and minority status), supply side (numbers of hospital beds, specialists and primary care physicians in the caller’s area of residence) and call activity (chief complaint, day of call, nurse recommendation and repeat calls). Expenditures associated with each care path were calculated using medical and pharmacy claims data for the 30-day observation period following the call, and savings were estimated from each GLM regression as the difference in average expenditures between compliers and non-compliers. Savings were calculated for both an unweighted, GLM-adjusted sample and a propensity score-weighted, GLM-adjusted sample. Savings occurred when the regression-adjusted complier expenditures were less than the regression-adjusted non-complier expenditures. Total programme savings were estimated by summing weighted, GLM-adjusted total savings or losses for all five care paths. In a separate analysis, logistic regression modelling was used to identify the characteristics associated with an employer who achieved higher than average savings. The sample members for this analysis included employers who paid for the telephone triage service during 2012. The dependent variable for this analysis was coded as 1 if the employer’s savings were higher than average or 0 if savings were average or lower than average. Employers with missing demographic, socioeconomic or occupational information were excluded, making the sample size 295 employers. The logistic regression analysis produced an odds ratio which indicated the likelihood that a specific employer-level characteristic was associated with higher than average savings. Odds ratios greater than 1.0 indicated a higher propensity to achieve savings. The independent variables used in this logistic regression analysis were employer-level characteristics which measured the gender percentages of its employees, the percentages of employees living in high minority areas, and participation by the employer in a targeted communication programme that was meant to increase the use of the help line service. Other independent variables measured whether the employer had other care coordination programmes, such as wellness or disease management programmes. Additional independent variables included the average value of the prospective risk score for the employees of the firm, the industry type associated with the employer, the region where the employer headquarters was located, a ratio measuring the number of employees to dependents in the employer’s health plan and employer size.
333
Results A total of 188,689 triage calls were identified. From these, 13,008 were excluded due to inability to match call records from the help line database to eligibility, followed by another 11,110 calls that were excluded as repeat calls, then 3191 calls were excluded due to lack of medical coverage, 21,536 calls were excluded due to no health care expenditures in the 3 months following the call, and finally 7335 calls were excluded due to missing demographic information. This left a total of 132,509 remaining calls (70%). Of these, over half of the care recommendations made by the nurse were different from the caller’s original intent, and nurses were more likely to recommend that callers sought a higher level of care rather than a lower level of care. Overall, nurses recommended that a caller should seek a higher level of care 31% of the time, a lower level of care 25% of the time, and agreed with the caller’s original intention 44% of the time. For the higher level of care paths, nurses recommended an emergency room visit for 5% of calls and a non-emergency visit for 26% of calls. For the lower level of care paths, nurses recommended a lower level of care from an emergency room visit for 12% of calls and a lower level of care from a non-emergency visit for 13% of calls. Of the calls retained in the final sample, 75,276 were made by compliant callers and 57,233 were made by noncompliant callers (overall compliance 57%). Within the various care paths, a number of demographic and clinical characteristics differed significantly between compliant and non-compliant callers. These included the chief complaint, prior 6-month health care utilization measures, whether prescription expenses were incurred, geographical region, imputed income level, the prospective risk score, day of the call, gender and age (Tables 1–3). As these differences could affect health care costs, propensity score weighting was performed to adjust for the casemix differences between compliant and non-compliant callers. After weighting, the prospective 3-month cost risk score measure remained significantly different between compliant and non-compliant callers in 3 out of the 5 care paths, but all other differences in clinical and demographic characteristics were removed (data not shown). Generalized linear modelling was then performed to further control for the remaining case-mix differences between compliers and non-compliers. Within each care path, the average 30-day weighted, GLM-adjusted medical and prescription expenditures were compared between compliant and non-compliant callers (Table 4). In the weighted sample, compliant callers had lower average expenditures than non-compliant callers for four out of the five care paths. Specifically, for compliers who were redirected to lower care rather than using an emergency room, the average savings were $1862 (P < 0.001). For compliers who were redirected to a lower level of care from a non-emergency room setting, the average savings were $620 (P < 0.001). Compliers who were not directed to a different level or site of care had average
334
Journal of Telemedicine and Telecare 20(6)
Table 1. Characteristics of groups receiving a recommendation to seek a lower level of care. Redirect to lower care (from emergency room)
Redirect to lower care (from non-emergency room)
Complier (N ¼ 9401)
Complier (n ¼ 9339)
Non-complier (N ¼ 7886)
P-value
Non-complier (n ¼ 7835)
P-value
Age category, years (%)
