Physician workload and treatment choice: the case of primary care

Physician workload and treatment choice: the case of primary care Adi Alkalay Clalit Health Services Amnon Lahad School of public health Hebrew University of Jerusalem Alon Eizenberg Department of Economics Hebrew University of Jerusalem Ity Shurtz Department of Economics Hebrew University of Jerusalem February 2017 1 Introduction In many industries and economic activities capacity is fixed or modifiable at a high cost while demand tends to be uncertain. This is typical to industries like airlines, hotels and car rentals. A common practice in those industries is to adjust prices to reflect the cost of serving the marginal consumer. By contrast, in professional services industries such as banking, legal services and healthcare, firms tend to refrain from price adjustments and prices are typically kept highly stable. How do professional service providers then deal with uncertain demand when capacity is fixed? In this study we examine whether instead of using price adjustments professional service providers adjust service quality in response to demand fluctuations. Concretely, we study the quality-of-service response to capacity constraints that arise in the case of primary care. This is a particularly interesting case because primary care provision is thought to be associated with improved population health and lower health care costs (Starfield et al. (2005)). 1 In fact, primary care is probably the most important context in which this issue arises. 2 Internal preliminary document, do not cite or quote. aloneiz@mscc.huji.ac.il ity.shurtz@huji.ac.il 1 Scott (2000) notes that GPs make many different types of decisions that influence the amount, type and location of care received by patients. These include decisions to refer to a specialist or other health professional, prescribe medication, arrange follow-up, and order tests. 2 Anand et al. (2011) for exmple notes that A major difficulty in improving productivity in such customerintensive services is the sensitivity of the service quality provided to the speed of service: as the service speed increases, the quality of service inevitably declines... Primary health-care practice in the United States epitomizes this problem 1

In primary care, capacity determines the number of patients a physician sees in a given amount of time, i.e. the physician s workload. Thus, we examine how primary care physicians respond to workload. We use a unique detailed administrative data from eleven clinics of a large Israeli HMO in the period 2011-2014 to study this issue. A common measure of workload, in the primary care setting, is the number of patients a physician sees in a given amount of time or equivalently, the mean visit length. Based on this notion, the measure of workload we use in this study is the physician s mean daily length of office visits. In order to study the causal relationship between workload and physician behavior, we take an instrumental variables approach. In essence, the approach relies on absences of fellow physicians at the clinic as a source of exogenous increase in the physicians workload. We proxy the amount of additional workload that an absence of a colleague creates by the share of the absent colleague s patients of all patients seen in the clinic in a given day. We use both a discrete indicator for absence and the share of the absent colleague s patients as two alternative instruments. We analyze how physicians treat their own patients and exclude from the analysis other physicians patients. We examine the effect of workload on referrals to the emergency room. Controlling for observables, the OLS estimates are negative, small and statistically significant, suggesting that less workload is correlated with more emergency room referrals. The instrumental variable approach reveals, however, that workload has no statistically significant effect on referrals to the emergency room. When we examine the effect of workload on prescription of antibiotics using OLS, the estimates are negative, significant and quite small. With the IV approach the results are much larger and significant. Thus, we find that when facing more workload physicians do tend to prescribe more antibiotics. One in four patients would not receive antibiotics with additional five minutes per visit. With reference to policy implications, even if primary care quality is sensitive to physician workload, are there reasons to think that this would result in inefficiencies and should have policy implications? Arguably, the answer to this question is yes. Healthcare providers may not be aware of the price, in terms of quality, of their limited capacity. Additionally, providers may not internalize all the costs that are associated with their limited capacity. Another related policy issue is the so-called primary-care crunch the shortage in primary care physicians in the United States. It is often argued that physician shortage induces increased workload on primary care physicians that results in lower quality healthcare. One related strand of literature looks at crowding in the healthcare system and its effects on delivery of treatment. Using operational data from a hospital emergency department, Batt and Terwiesch (2012) find that workload induced slow-down occurs and that care providers adjust their clinical behavior to accelerate the service. Kc and Terwiesch (2009) show that system load increase service rate and reduction of care quality. Kim et al. (2016) study admission to intensive care units (ICU) and find that ICU congestion can have a significant impact on ICU admission decisions and patient outcomes. Powell et al. (2012) find that physician workload 2

reduces that share of severe patients and consequently hospital reimbursement. Another strand of literature that is related to our work studies the impact of workload on worker productivity. Tan and Netessine (2014) use data from restaurants to examine to effect of workload, defined as the number of tables handled, on performance that is measured by sales and meal duration. They find that surprisingly workload is associated with more sales effort. Thus staff reduction may lead to higher sales and lower labor costs. Perdikaki et al. (2012) study the relationship between store traffic, labor, and sales performance. They find that converting the incoming traffic into sales declines with shoppers traffic. Chatain & Eisenberg (2014) study a legal services provider and find that service quality increase with the amount of available resources. The remainder of the paper is structured as follows. Section 2 describes the data; section 3 describes the empirical strategy; section 4 reports the main results; and section 5 concludes. 2 Data and Variables We use a detailed administrative database that covers all the primary care visits in eleven clinics in the Jerusalem area of Clalit Health Services the largest of four HMO s that provide the vast majority of health insurance in the country and deliver most of its primary care in the period 2011-2014. These data include information on visit characteristics such as visit time, visit length, and regular physician identity. They also include patient characteristics such as gender, age, country of origin and chronic conditions. Finally, detailed summary of the visit is recorded including diagnoses, prescriptions, referrals, laboratory tests, imaging and so on. In Clalit Health Services, patients are enrolled with a regular primary care physician. Normally, a primary care visit is scheduled with this physician. However, there are exceptions to this routine. If patients need urgent care, outside of their physician s office hours or when their physician is absent, they are typically referred to one of their physicain s colleagues at the clinic. In our analysis we restrict our sample to visits in which physicians see their regular patients in days in which they see at least twelve of their patients. By doing so we aim to capture the behavior of physicians treating only their own patients in a regular day. Table 1 provides descriptive statistics. The sample includes close to 1.2 million visits by about 120,000 patients. The number of physicians in the sample is 98. With respect to patient characteristics, the mean patient age is about fifty; in sixty percent of the visits patients are women; most patients are native Israeli. About thirty percent are smokers and about thirty percent are obese. Forty percent have hypertension and almost fifty percent have hyperlipidemia. Seventeen percent of the patients suffer from ischemic heart disease. Office visits are on average eleven minutes long. Patients are referred to the emergency room in one in every hundred visits. Antibiotics are prescribed in one in twelve visits. 3

3 Emprical strategy 3.1 Workload To analyze the effect of workload, one needs to establish a measure of workload. One common measure of workload in the primary care setting is the number of patients a physician sees per-hour or, equivalently, the mean visit length. 3 This is the measure of workload we use here. Thus, we define the main explanatory variable workload as the daily mean of a physician s visit lengths. For example, a physician that had overall office visit time of two hours and in that time saw ten patients has a workload of twelve minutes per-patient. An issue that arises when using this measure of workload is that it has, by construction, a measurement error. We return to this issue in the next section. 3.2 Identification Consider the following empirical model of the relationship between workload and physician behavior (1) y = α + βworkload + Xγ + ɛ where y is an outcome and X is a rich set of controls. Analyzing this model using OLS may provide biased estimates. Intuitively, our workload measure is based on the count of the number of patients the physician sees per hour and ignores possible random shocks to workload that are not captured by this measure, e.g. workload that arises because of patient type. Thus, arguably the measure of workload that we use falls into a measurement error framework where the econometrician observes workload with some noise. (2) workload = workload + u In such cases attenuation bias may arise and the estimates of ˆβ may be biased towards zero. Additionally, omitted variables may create a correlation between the measure of workload and the outcomes we measure. For example, a local infection may increase the number of patients the physician sees per hour and also the probability of prescribing antibiotics. We address these issues by taking an instrumental variable approach, as we describe in detail in the next section. 3.3 The IV approach To identify the causal effect of workload on physician behavior we use absence of colleagues at the clinic as a source of variation in workload. In Clalit Health Services, when a colleague 3 For example, when a physician sees six patients per hour it is analogous to having ten minutes per-patient on average. 4

at the clinic is absent, her patients are referred to other physicians at the clinic. The absent physician s patients create increased workload for the physicians who are present at the clinic. We use this source of variation in workload to study its effects on physician behavior. Our implementation of this approach builds on the regularity in weekdays of work at the clinic. Physicians have fixed days and hours during the week in which they schedule appointments and see patients. We exploit this regularity and define an absence as a day that satisfies two conditions. The first, it is a day in which a physician treats zero of her patients, namely, the physician is not present at the clinic in that day. The second, the physician has worked (and has seen at least 5 of her patients) in the two weeks before and after that same weekday. This ensures that it is one of the physician s regular days. With physician days of absence at the clinic defined, we calculate, for the entire clinic, the share of the missing physician s patients of all the patients that were seen in the clinic in that day. Thus, we use absences to create two alternative instruments. One instrument is based on an intensive margin notion. We proxy the amount of workload that an absence creates by the share of missing physician s patients of all patients seen in a given day in a clinic. The second instrument, that is based on an extensive margin notion, is an indicator for an absence of a physician at the clinic. 4 Results 4.1 First stage We first illustrate graphically the source of variation we use in our instrumental variable approach. Figure 1 depicts the relation between our proxy of the amount of workload that the absence of a colleague at a clinic inflicts on the other physicians in the clinic and our measure of workload mean visit length. We cut the share of the absent physician patients into bins of 0.2 percentage points. For each bin we calculate the corresponding measure of workload the mean visit length and display them in the figure. We also regress our workload measure on share of the absent physician patients. The solid line displays the relation between workload and absence as predicted by the regression. As the figure shows, an increase of ten percentage points in the share of the absent physician patients is associated with a decrease of about one minute in mean visit length. In order to illustrate the effect of days with absences on physician workload at the clinic, we analyze an event study model. Specifically, let D st be an indicator that equals one when at least one physician is absent from the clinic. Suppose for example that a physician was absent in January 5 th 2013 in clinic 5, then D 5,1/5/2013 = 1. Next, define τ st, the event relative time, as the number of days that elapsed since the absence. Thus, in our example τ 5,1/5/2013 = 0, 5

τ 5,1/4/2013 = 1 and τ 5,1/7/2013 = 2. We analyze a statistical model of the form: (3) workload jst =α + Doc id j β 1 + T ime t β 2 + γ 1 τ k +... + γ k+1 τ 0 + γ k+2 τ 1 +... + γ 2k+1 τ k 1 + ɛ jst where Doc id j is a vector of physician fixed effects, T ime t is a set of dummy variables for month of year and day of the week. τ k τ k 1, the objects of interest, are indicators that capture the effect of the event on workload. Specifically, our hypothesis is that in the periods before the absence, the effect of these indicators is not significantly different from zero; at the time of the event the effect is negative and after the event the effect is not significantly different from zero. Figure 2 displays the estimates of τ 7 τ 6. As the figure shows, in the seven days before the event, the effect of the event is insignificantly different form zero. At the time of the event, the mean visit length at the clinic length drops by about a third of a minute. In the first day after the event, the effect on mean visit length at the clinic is negative and significant, yet it is much smaller. 4 insignificantly different from zero. In the second day following the event and the following days the effect is Table 2 displays the first stage results of the two instrumental variables we use in the analysis. As Column 1 of panel (a) of the table shows, our first IV, the share of absent physician s patients of all patients, has a negative effect on workload with a point estimate of about minus nine. With additional controls for time and physician fixed effects, the estimate become roughly minus five, implying that an increase of ten percentage points in the Share of absent physician s patients is associated with decrease of 0.4 minutes in mean visit length. One threat to our instrumental variable approach is that an interaction between the physician s patient pool and the colleagues absence may arise. For example, if the easier cases are deterred by the physician s workload and decide to return on a different day. We try to assess the existence of such a selection. To do so, examine the sensitivity of the first stage to a rich set of patient characteristics: patient age, gender and chronic conditions. 5 for visits of an administrative nature. 6 We also account As column 3 of panel (a) shows, adding patient level controls does not change the first stage results, supporting the view that they indeed reflect variation in workload and they are not an artifact of changes in patient characteristics. Turning to our second instrument in panel (b) of the table, the effect of an absent physician in the clinic is a decrease of 0.27 minutes in mean visit length. When we add in Columns 2 physician and time fixed effects the estimates decrease to 0.3 minutes. As in the case of the first instrument, adding patient level controls in column 3 does not change the estimates. 4 Note that in the day following the event, the physician that was absent may exhibit increased workload and that could explain the effect in τ = 1 5 We provide a list of the chronic conditinos in appendix XX 6 We include dummy variables for visit whose main reason is: issue a medical certificate, prescription renewal, filling out forms and an administrative visit. 6

To further examine if the absence (and the physician s increased workload) creates deterrence to the physican s own patients, we examine if the number of the physicians own patients is affected by a colleagues absence (much in the spirit of McCrary (2008)). Concretely, the idea behind this examination is that if the number of patients decreases in a day of absence we may worry that the patients that decide to leave the clinic are systematically different from those who choose to stay in the clinic (and presumably their visit is less urgent). To implement the exercise, we run an event study analysis, that is analogous to the analysis summarized in Figure 2, with the number of the physician s own patients per hour as the dependent variable. Figure 3 displays the results of this analysis. As the figure shows, there appears to be a decrease of about 0.05 in the number of patients a physician sees per hour in days of absence. As the number of patients per hour a physician sees is on average roughly 3, this result indicates a decrease of about 1.6 percent. To further examine this effect, we remove administrative visits the same visits we control for in the first stage regression and repeat the analysis. The results, displayed in Figure 4, show that with this refinement, the decrease in the number of patients disappears. Namely, the number of clinical visits of her own patients the physician has, in a day of absence is not statistically different from this number in other days. This result further alleviates the concerns that absence affects the outcomes we measure in channels other than through its effect on workload. 4.2 The effect of workload on physician behavior We turn to analyze the effect of workload on physician behavior. We focus on two outcomes: referrals to the emergency room and subscription of antibiotics. 4.2.1 The effect of workload on referrals to the emergency room The first three columns of Table 3 display the OLS results from a model similar to the one in Equation 1. The outcome variable is a dummy for emergency room referral. I.e., y equals one if the patient was referred to the ER following the visit and zero otherwise. As the table shows, the results of the OLS regression are positive and significant and they are quite robust to the inclusion of the control variables. However, when we use the instrumental variables in columns (4)-(9), the results become statistically insignificant. 4.2.2 The effect of workload on antibiotics prescription Table 4 is similar to Table 3, with a dummy variable for a visit that resulted in a prescription for antibiotics as the outcome variable. As the table shows, the OLS estimates, in the first three columns, are positive and significant. Namely, longer office visits are associated with prescribing more antibiotics. However, the sign of the coefficient changes when we use the instrument. Using the first instrument the results are statistically significant and the estimates 7

remain statistically significant when we include the controls in columns (5)-(6). Interpreting these results, an increase of one minute (about 10%) in mean visit length is associated with a 0.4% decrease in antibiotics reflecting a decrease of 5% and an elasticity of 0.5. The estimates are negative yet statistically insignificant when we use the second instrument. Overall, these result suggests that when workload increases, physicians tend to prescribe more antibiotics. 5 Conclusions In this study we find that, perhaps surprisingly, increased workload does not affect the tendency of physicians to refer patients to the emergency room. However, workload is associated with prescribing more antibiotics. These results show that at a higher workload physicians change their practice style and the treatment they provide. These changes should be accounted for when optimal workload and size of workforce in healthcare are set. References Krishnan S Anand, M Fazil Pac, and Senthil Veeraraghavan. Quality-speed conundrum: Tradeoffs in customer-intensive services. Management Science, 57(1):40 56, 2011. Robert J Batt and Christian Terwiesch. Doctors under load: An empirical study of statedependent service times in emergency care. Working Paper, The Wharton School, 1, 2012. Diwas S Kc and Christian Terwiesch. Impact of workload on service time and patient safety: An econometric analysis of hospital operations. Management Science, 55(9):1486 1498, 2009. Song-Hee Kim, Carri W Chan, Marcelo Olivares, and Gabriel J Escobar. Association among icu congestion, icu admission decision, and patient outcomes. Critical Care Medicine, 44(10): 1814 1821, 2016. Justin McCrary. Manipulation of the running variable in the regression discontinuity design: A density test. Journal of econometrics, 142(2):698 714, 2008. Olga Perdikaki, Saravanan Kesavan, and Jayashankar M Swaminathan. Effect of traffic on sales and conversion rates of retail stores. Manufacturing & Service Operations Management, 14 (1):145 162, 2012. Adam Powell, Sergei Savin, and Nicos Savva. Physician workload and hospital reimbursement: Overworked physicians generate less revenue per patient. Manufacturing & Service Operations Management, 14(4):512 528, 2012. Anthony Scott. Chapter 22 economics of general practice. volume 1, Part B of Handbook of Health Economics, pages 1175 1200. Elsevier, 2000. doi: http://dx.doi.org/10. 1016/S1574-0064(00)80035-9. URL http://www.sciencedirect.com/science/article/ pii/s1574006400800359. Barbara Starfield, Leiyu Shi, and James Macinko. Contribution of primary care to health systems and health. Milbank quarterly, 83(3):457 502, 2005. 8

Tom Fangyun Tan and Serguei Netessine. When does the devil make work? an empirical study of the impact of workload on worker productivity. Management Science, 60(6):1574 1593, 2014. 9

Figure 1: Share of missing physician s patients and workload Mean visit length 8 9 10 11 12 0.1.2.3.4 Share of absent physician s patients Note: The figure plots the mean visit length for bins of share of absent physician patients. The superimposed line is the predicted relation between workload and absence. 10

Figure 2: Workload around days of a colleague s absence Mean visit length.4.3.2.1 0.1 7 6 5 4 3 2 1 0 1 2 3 4 5 6 Days relative to absence Note: The figure plots the coefficients and standard errors from the event study model described in Equation 3. The dependent variable is mean visit length. standard errors are clustered at the physician-day level. 11

Figure 3: Number of own patients per hour around days of a colleague s absence Number of own patients per hour.1.05 0.05.1 7 6 5 4 3 2 1 0 1 2 3 4 5 6 Days relative to absence Note: The figure plots the coefficients and standard errors from the event study model described in Equation 3. The dependent variable is the number of a physician s own patient per day. The analysis was conducted at the physician day level. 12

Figure 4: Number of own patients per hour around days of a colleague s absence Number of own patients per hour.1.05 0.05.1 7 6 5 4 3 2 1 0 1 2 3 4 5 6 Days relative to absence Note: The figure plots the coefficients and standard errors from the event study model described in Equation 3. The dependent variable is the number of a physician s own patient per day excluding administrative vistis. The analysis was conducted at the physician day level. 13

Table 1: Summary statistics of visits data Patient characteristics Mean age 50.60 Capitated patient weight 1.79 Share women 0.58 Share born in Israel 0.59 Share smokers 0.31 Share obese 0.28 Share hypertension 0.38 Share hyperlipidemia 0.49 Share ischemic heart disease 0.17 Office visits characteristics Visit length 11.03 Share referrals to ER 0.01 Share antibiotics 0.08 Share Painkiller 0.05 Number of patients 119,131 Number of physicians 98 Observations 1,184,648 Note. The table includes all patient visits in the eleven clinics used in this study in the period 2011-2014. 14

Table 2: The effect of absences on workload (1) (2) (3) Panel (a): IV 1 Share of absent physician s patients of all patients -9.69-5.32-5.32 (0.44) (0.31) (0.31) Panel (b): IV 2 Seeing absent physician s patients -0.27-0.31-0.31 (0.05) (0.04) (0.04) Year-month, day & physician FE No Yes Yes Patient age, gender & condition controls No No Yes Observations 1,183,539 1,183,539 1,180,968 Note: All columns report estimates of effect of absence on workload. All specifications include a constant. The Year-month fixed effects consists of a dummy variable for each of 48 calendar months in our data; the day fixed effects consist of a dummy for every weekday. Standard errors clustered by physician-day are reported in parentheses. One or two asterisks indicate significance at 5% or 1%, respectively. 15

Table 3: The effect of workload on emergency room referrals OLS IV 1 IV 2 16 (1) (2) (3) (4) (5) (6) (7) (8) (9) Mean visit length 0.00019 0.00027-0.00068-0.00013-0.00044-0.00045-0.00250-0.00055-0.00048 (0.00004) (0.00005) (0.00012) (0.00043) (0.00073) (0.00073) (0.00162) (0.00123) (0.00123) Year-month, day & physician FE No Yes Yes No Yes Yes No Yes Yes Patient age, gender & condition controls No No Yes No No Yes No Yes Yes Observations 1,183,539 1,183,539 1,180,968 1,183,539 1,183,539 1,180,968 1,183,539 1,183,539 1,180,968 Note: All columns report estimates of effect of workload on emergency room referrals. All specifications include a constant. The Year-month fixed effects consists of a dummy variable for each of 48 calendar months in our data; the day fixed effects consist of a dummy for every weekday. Standard errors clustered by physician-day are reported in parentheses. One or two asterisks indicate significance at 5% or 1%, respectively.

Table 4: The effect of workload on antibiotics prescriptions OLS IV 1 IV 2 17 (1) (2) (3) (4) (5) (6) (7) (8) (9) Mean visit length -0.00109-0.00073-0.00068-0.00556-0.00430-0.00423-0.00626-0.00345-0.00365 (0.00010) (0.00012) (0.00012) (0.00163) (0.00212) (0.00212) (0.00453) (0.00330) (0.00330) Year-month, day & physician FE No Yes Yes No Yes Yes No Yes Yes Patient age, gender & condition controls No No Yes No No Yes No Yes Yes Observations 1,183,539 1,183,539 1,180,968 1,183,539 1,183,539 1,180,968 1,183,539 1,183,539 1,180,968 Note: All columns report estimates of effect of workload on antibiotics prescription. All specifications include a constant. The Year-month fixed effects consists of a dummy variable for each of 48 calendar months in our data; the day fixed effects consist of a dummy for every weekday. Standard errors clustered by physician-day are reported in parentheses. One or two asterisks indicate significance at 5% or 1%, respectively.