Abstract
A host of different factors affect health and longevity, ranging from genetic endowments to public policy. Physicians have a substantial influence on patients’ health and health-related costs, but we know little about the extent of this influence beyond clinical decisions, such as adequate diagnosis and treatment. This paper demonstrates that the health management styles of primary care physicians significantly affect the health outcomes of their patients. Using data on the population of statin users in Denmark and matching patients to their primary care physicians, we show that the physician’s ability to facilitate adherence with prescription medications has significant positive effects on patient outcomes and health costs, even after controlling for observable and unobservable patient characteristics. Policy interventions aimed at improving this aspect of physicians’ health management styles have important implications for patient outcomes and healthcare costs.
I. Introduction
There are large, persistent differences in patient outcomes across physicians and health facilities (Epstein and Nicholson 2009; Currie, MacLeod, and Van Parys 2016; Molitor 2018). The root causes of these differences are not well understood. Variations in clinical quality across physicians are unlikely to account fully for the gaps in outcomes, and neither are disparities in patient backgrounds. A possible contributing factor is that the patients of different physicians pursue different health behaviors. It is an open question whether these behaviors can be affected by the health management style of the physician. Recent work suggests that this is likely the case.1
There are several reasons why it is important to better understand if physicians’ health management styles can affect patient health outcomes and health costs. First, insight into what features of the healthcare delivery process affect patient outcomes is crucial in the design of contracts between payers and providers. As societies struggle to address growing healthcare costs, quality contracts are becoming increasingly popular, and with them the managerial approaches taken by healthcare systems. A quality contract between the payer and the healthcare provider usually specifies that a portion of the provider’s financial compensation is dependent on adequate performance according to a set of health quality metrics taken over the patient population. In outpatient and primary care settings, these quality metrics include routine screening and vaccinations, as well as proper maintenance of chronic conditions and avoidance of preventable hospitalizations and rehospitalizations. Practitioners have argued that despite attempts to adjust such metrics to the patient mix, providers who disproportionately work with patients whose health behaviors are harder to manage are at a great disadvantage.
Second, we need to understand better how different dimensions of physician management skills relate to patient outcomes. While it is widely recognized that multiple dimensions of firms’ management matter for firm output (Bloom and Reenen 2007, 2010) and hospital performance (Bloom et al. 2015), it has yet to be studied in the case of outpatient-based healthcare provision. The role of health management skills in the primary care and outpatient settings is particularly relevant because patients have substantial control over the provider they see and the actions they take conditional on receiving a clinical recommendation by that provider.
Third, a key question in the literature is whether cross-physician differences in patient outcomes are due to factors controlled by the physician or due to patient selection. For example, if a physician’s practice happens to be geographically located in an area where most people adhere with medication and engage in proper health maintenance, the practice will likely score high on any patient health outcomes metrics. This is particularly important because in some settings provider-specific quality metrics, or “report cards,” have been made public. These report cards are based on past patient outcomes for common conditions. On the one hand, this may facilitate efficient patient allocation across providers and encourage better quality among physicians who lag behind their peers. On the other, the literature suggests that patients who are able to shop for providers, either because they are healthier or because they are better informed, would select out of lower-quality providers after quality information becomes public (Cutler, Huckman, and Landrum 2004; Santos, Gravelle, and Propper 2017). If patient selection and sorting account for some of the differences in observed provider quality, making quality metrics public will likely result in any remaining “good” patients abandoning practices with a suboptimal patient mix.
We use Danish registry panel data on the population of statin users and their primary care physicians (PCP) to investigate how physician health management styles impact patient health outcomes. We construct a measure of health management style for each PCP in Denmark based on their patients’ adherence with prescribed statin medication therapy.2 This measure captures the physician’s ability to inspire the patient to adhere to that therapy, perhaps by clearly communicating its benefits and encouraging and responding appropriately to patient feedback about their experience with the medication. We consider this metric a proxy measure of the physician management style and show how it affects patient hospitalizations for cardio vascular diseases (CVDs) and associated costs.
This study makes several contributions to the existing literature. First, we demonstrate that the physician’s ability to facilitate patient adherence with prescribed therapy has sizeable effects on patient outcomes and healthcare costs. Second, we show that the physician’s health management skill maintains its strong explanatory power even after we control for patient selection into practices and unobserved patient heterogeneity. Overall, the estimates suggest that investing in improving physician health management skills would have substantial positive effects on health outcomes and result in reductions in health cost for CVDs. Increasing the health management skills of the physician by one standard deviation could lead to health outcomes improvement equivalent to 8.5 percent of the total estimated therapeutic effect of statin therapy based on data from clinical trials (Scandinavian Simvastatin Survival Study 1996).
Third, we demonstrate that differences in established measures of clinical quality (constructed using outcomes related to acute Ambulatory Care Sensitive Conditions) across PCPs correlate strongly with patient CVD-related health outcomes, but are in fact due to patient selection. Denmark, along with many other European countries, does not have a public physician (or hospital) quality reporting system, so our findings are not affected by actions that may be taken by physicians to influence the type of patients selecting into their practice. We uncover differential sorting of patients across physicians based on unobservable patient characteristics that positively correlate with health outcomes. Because of the institutional setting, this sorting appears to be demand-driven. The introduction of incentive contracts tied to such measures of clinical quality would likely introduce supply-side sources of selection in addition to the demand-side selection we document here, substantially exacerbating this positive patient selection. Making these clinical quality metrics public would increase patient selection even more. To our knowledge, this is the first study to investigate explicitly whether physicians’ health management skills affect health outcomes, conditional on controlling for physician clinical quality and observed and unobserved patient characteristics.
II. Background and Institutional Setting
A. Background
A large body of work documents differences in clinical treatment styles across physicians (Grytten and Sorensen 2003; Epstein and Nicholson 2009; Burke, Fournier, and Prasad 2010; Phelps and Mooney 1993; Phelps 1995). The reasons behind the variation in clinical treatment choices across doctors are subject to an active debate in health economics (for example, Chandra, Cutler, and Song 2011), but the physician influence on health outcomes is not restricted to the choice of clinical therapy. Notably, at least one study in economics and a number of studies in medicine report that physicians’ communication styles also affect patients’ adherence decisions (Koulayev, Simeonova, and Skipper 2017; Cooper 2009; Sleath et al. 2000; Haynes, McDonald, and Garg 2002; see also reviews by DiMatteo 2004; Meyers et al. 2019). This is important because recent adjustments to physician compensation schemes have introduced quality metrics into the contracts in some healthcare systems. Quality is generally assessed using patient health outcomes, such as avoidable hospitalizations and rehospitalizations and good maintenance of chronic conditions like diabetes and hypertension, as well as more routine metrics, such as vaccinations and recommended recurring checkups. In practice, what contributes to good performance on these quality metrics is not well understood. The correct clinical diagnosis and choice of treatment is obviously an important first step. Once this has been established, and especially in the management of chronic diseases in an outpatient setting, the patient is an equally important partner in the healthcare process. The physician’s ability to respond to individualized patients’ needs, such as medication side effects, and to effectively communicate clinical decisions to the patient is very likely an additional and important determinant of health outcomes.
We analyze the physician’s ability to inspire and manage optimal patient health investment, which we term the physician’s health management skill. We argue that in the case of preventive medicine, this dimension of the physician’s treatment style is relevant, and if maximizing population health is the goal, then pay-for-performance schemes should also provide incentives along this dimension. Prior research has shown that primary care providers vary significantly in the average medication adherence of their patients (Koulayev, Simeonova, and Skipper 2017). Though intuitive, the link between patient adherence and healthcare costs and outcomes is harder to identify. An important source of bias is the association between unobserved patient health status, and its progression, and medication adherence. When individuals feel worse, or perceive their health as deteriorating, they are more likely to follow the doctor’s orders. As recognized by Encinosa, Bernard, and Dor (2010), among others, this biases the estimated coefficients from simple regression models. Using an instrumental variables approach, studies find large positive effects of adherence on health outcomes and health cost reductions.
B. Institutional Setting
The Danish healthcare system can be divided into two main sectors: the primary healthcare sector and the hospital sector. The primary healthcare service sector deals with treatment and care from PCPs, specialists, physiotherapists, and dentists, among others. Furthermore, the primary sector also includes preventive health schemes and preventive healthcare. The hospital sector deals with conditions that are more complex and require more advanced treatment. Admission in nonacute cases requires a referral from the primary sector.
Denmark has universal and tax-financed health insurance run by the government. All individuals residing in Denmark are given a social security number. The social security number ensures free access and treatment at PCPs and specialists, as well as free in-hospital stays. All services provided to an individual are registered via the social security number, and all expenses are covered by the national health insurance.
1. Primary care physicians
The Danish public health insurance provides visits and services at the PCP free of charge. In Denmark, PCPs serve as gatekeepers to the rest of the healthcare system in the sense that they refer to specialists and hospital admissions. There are approximately 3,500 PCPs in Denmark working from 2,200 different practices. In order to get reimbursed by the national insurance, the physician needs to acquire a clinic-ID (ydernummer). The number of clinic licenses is controlled by the government, based on factors such as the population density in different areas.
The PCPs are responsible for a large portion of a patient’s medication therapy. The physician has no financial incentives to choose specific medication brands. First-choice medication recommendations are issued by the national health authorities, but practitioners can choose a different therapy if they consider it more appropriate. Prescription drugs are sold at government licensed pharmacies only. All information about purchases is registered in a database at the Danish Medicines Agency (DMA). It is important to note that an individual’s choice set of PCPs is limited in Denmark. A patient can choose any PCP, as long as the PCP’s practice is located within 15 km of the patient’s home.3 The patient needs to be enlisted with a PCP in order to visit, and changing to a different PCP costs a fee of 150 DKK4 and can be done only if the new doctor is open for new patient intake. This restricts the possibility of changing PCPs, as well as the possibility of choosing PCPs not in the individual’s choice set.
An institutional feature of great importance to this study is that it is very difficult for physicians to selectively turn away individual patients. When a clinic list has reached 1,600 patients (per physician), the physician can apply to the local government to stop the intake of new patients.5 However, if working below this capacity, the physician has to take in patients who wish to be listed with them. According to the collective agreement made between the government and the Danish PCPs, a physician can discontinue the physician–patient relationship only if the patient acts violently or threatening or misbehaves in any other way during the clinical encounter.
2. Hospitals
There are five regions in Denmark in charge of operating a total of 54 hospitals. The funding is partly state grants, which are activity and demography based, and partly funded by the municipalities. Out-patient and in-patient care are both free of charge for the individual. There is a group of specialized privately operated hospitals, where patients are covered by public insurance if waiting lists at public hospitals exceed two months.
III. Empirical Strategy
The main goal of the paper is to highlight one of the mechanisms through which physicians can impact the health of their patients. To this end, we leverage the observed link between physicians and patients in the registry data. Conceptually, we can divide the factors affecting individual health investment into physician- and patient-driven. The physician’s ability to properly diagnose and prescribe clinical therapy is one important physician-driven input. The physician’s health management skill enters the health production function through the patient-driven inputs. One potential channel is its influence on the patient’s own investment in health through patient adherence with medication.
To have a setting where we expect physician patient health management skills related to patient medication adherence to be relevant, we focus on the prevention of CVD-related hospitalizations. Cardiovascular disease is the leading cause of death and hospitalization in most developed economies. The PCP has a central role as a first-line healthcare provider. Importantly, modern treatment has a pharmacological component that is both crucial in the maintenance of these chronic conditions and also allows for the construction of measures of patient health behavior outside of the physician’s office.
Our empirical strategy is implemented in several steps. First, we calculate adherence rates for patients who were prescribed lipid-lowering drugs to reduce cholesterol levels. Next, we use information on all other patients at a practice to estimate a time-constant measure of health management skills, relying on logic similar to the teacher value-added literature (Chetty, Friedman, and Rockoff 2014). That is, we leverage the average behavior of patients with the same physician to measure an individual physician-specific component in individual patient adherence. We refer to these systematic differences across physicians as physician health management skill.6 In the absence of patient sorting across physicians, our estimated physician health management skill is a consistent estimator of a physician’s impact on patient health. The assumption that patients do not sort across the physicians is a strong one. We check its validity by assessing whether this particular dimension is sensitive to the inclusion of observed and unobserved patient heterogeneity. Ultimately, this strategy allows us to investigate whether differences in health management skill across physicians matter for patient health or whether the relationship between a physician’s skills in health management and patient health outcomes is driven by patient composition and sorting.
A. Estimating Health Management Skills
Previous research shows that there is persistent heterogeneity in medication adherence rates of patients across physicians (for example, Simeonova 2013; Koulayev, Simeonova, and Skipper 2017). We focus on adherence with the major cholesterol-lowering drug group, statins.7
The effectiveness of lipid-lowering drugs in reducing the risk of fatal and nonfatal events has been documented in numerous clinical trials (see for example, Scandinavian Simvastatin Survival Study 1996; Sacks et al. 1996; Shepherd et al. 1995). In some trials (Scandinavian Simvastatin Survival Study 1996) reductions in mortality are found after approximately one year in the treatment group, but other studies find effects on nonfatal outcomes of treatment after only a month (O’Driscoll, Green, and Taylor 1997; Ratchford et al. 2011). Some of the most widely prescribed statins are simvastatin (brand name Zocor), atorvastatin (brand name Lipitor), and fluvastatin (brand name Lescol). Statins are the most widely prescribed drug class targeting a chronic condition. More than 10 percent of the Danish population are prescribed statins (more than 25 percent in the age group over 50). Because statins are the only drug group targeting high cholesterol, the indication of use is very clear. This is convenient because we are looking to establish the link between adherence with medication and health outcomes for CVD.8 Statins have two characteristics that are suitable for our type of study. First, as treatment with statins is chronic, they must be taken all the time, once the need is established; second, they are indicated for a specific condition, which is clearly related to a particular health outcome. Further, patients are typically instructed to take one pill a day,9 which makes it easier to measure adherence10 with treatment using claims data.
We begin by constructing patient specific leave-one-out adherence rates. Using data on all individuals with at least two statin prescription claims in Denmark, we calculate the leave-one-out average adherence within each physician
1
ADH–itj is the individual-specific average adherence rate with physician j at time t not including the focal (index) individual. mjt is the number of patients at physician j at time t with at least two statin claims. We then residualize ADH–itj by regressing it on patient and physician observables, as well as a patient fixed effect to obtain the residual, 
Finally, the health management skill, 
, is constructed by averaging 
over time within physician j:
2
B. Validating our Measure of Health Management Skill
It is important to notice that 
only truly reflects physician health management skills if patients do not select into physicians on the basis of the physician’s ability to inspire adherence with statin medications in their patients. Hence, it is important for our analysis to verify that this derived metric is not reflecting patient selection. Given the Danish institutional setting, the choice of PCP is up to each patient and likely endogenous with some physician characteristics. As information about alternative physicians within the patient’s choice set is very limited, one might not be as worried about sorting into the “good” physicians, as much as the sorting out of the “bad” physicians. Hence, we are interested in finding situations in which we can identify separations of patients from physicians where the switching is less likely to be related to the health status of the patient or the physician health management style. We identify two different types of patient–physician separations in an attempt to tease out the potential impact of patient composition from the true underlying physician health management skill.
The two types of separations we identify are:
i. Separations due to clinic closures. These are patients who are forced to change physician because their physician goes out of business due to retirement or residential relocation.
ii. Separations due to patient residential relocation. These patients shift physician because they relocate residentially and cannot stay affiliated with their PCP.
When physicians close their clinic, patients affiliated with the clinic are forcibly separated from their physician and are effectively spread out across other local physicians.11 We believe that this is the setting with the least scope for sorting out of clinics due to bad physician quality. Other studies—most recently Fadlon and Van Parys (2020), who use PCP relocations and retirements to study the effects of practice styles on patient healthcare utilization—have used this type of separations for identification.12
Leveraging these quasi-exogenous separations, we examine whether the changes in health management skills that patients are subsequently exposed to lead to changes in adherence even in a setting with limited scope for physician selection.
C. Estimating the Impact of Health Management Skill on Health Outcomes
In an ideal research setting, we would be able to randomly assign patients to different physicians and estimate how different physicians’ health management styles affect patient outcomes. In the absence of such an experiment, we rely on rich registry data and “natural experiments” in physician–patient matching precipitated by residential relocations of patients and primary care clinic closures in Denmark.
To assess the impact of the health management style on patient health outcomes we estimate a model of risk of CVD-related hospitalization as a function of the physician management skill. We estimate the model
3
The experiment we try to mimic is to see how similar individuals would react to different physicians. In Equation 3, β1 is identified from both between- and within-patient variation in the quality indexes. Hence, the coefficients reflect differences in quality dimensions across the entire population of patients and physicians; however, we are primarily interested in the estimates that are derived from within-individual variation in exposure to physician management skills. The thought experiment is to assign the same patient to different physicians and track the evolution of her health outcomes across those physicians. To do this, in some models we estimate Equation 3 but also control for time-invariant individual heterogeneity through the individual fixed effect, αi
4
The primary outcome variables of interest are hospitalization due to CVD and healthcare expenditures due to CVD hospitalizations. As 
is constant across time and within physician, β1 in Equation 4 is estimated off variation in physicians within the same patient across time. That is, we only identify the coefficient from individuals seeing different physicians over time, and we will use the separations described previously as natural instruments of separations to assess the robustness of our results.
In addition to the challenge of sorting between patients and physicians, we need to include sufficient controls for baseline CVD hospitalization risk in order to allow for a physician value-added interpretation. We achieve this as our regressions include highly detailed patient characteristics as well as the past CVD and other comorbidities that have been demonstrated to predict all-cause and CVD-specific mortality.
IV. Data
We use the full population of adult individuals in Denmark who have at least two statin claims13 between January 1, 2004 and July 1, 2008. In addition to the prescription data, we have data on hospitalization and the primary diagnoses (by ICD-10) for hospitalizations, as well as healthcare expenditures. The prescriptions and hospitalization data are augmented with detailed individual-level economic and demographic information from several different registries.
Our main sample is constructed by observing the initial individual purchases of lipid-lowering prescription drugs. We then calculate the adherence as a proportion of days covered through a period of six months. The outcome measures are the probability of CVD-related hospitalizations and the associated hospitalization expenditures within the subsequent 12 months.
This approach offers the possibility to evaluate the relationship between short-term fluctuations in adherence and short-term health outcomes, as well as strictly separating the periods when adherence and hospitalization outcomes are measured. This is important because adherence may respond to a hospital stay either mechanically, through patients receiving new directions or new therapies during their in-patient stay, or as a behavioral response to a negative update on their health status that triggers (temporarily) improved adherence.
Figure 1 explains the sample construction. For Individual A, we observe adherence for six months initiated at the first pharmacy claim. We also observe whether the individual is hospitalized in the subsequent 12 months. We measure adherence as Proportion of Days Covered.14 Starting from the day of initiation of treatment, we measure the fraction of days within a six-month period the individual is covered with any statin. The individual is allowed to keep excess medication for one period, and because we focus on adherence on the intensive margin, we drop individuals when there is no claim in a period and no excess medication from previous periods. Patients who completely discontinue treatment in this fashion are nonpersistent or could have discontinued the therapy for reasons unknown.15 The fact that we focus on the intensive margin of adherence has implications for the interpretation of the results. Namely, our estimates capture the effects of differences in patient adherence with statins on health outcomes, conditional on continued statin intake. The effects we find are identified from marginal changes in the adherence rate, not from drastic changes in the therapy initiated by the physician or the patient. We exclude 0.25 percent of the patient observations because the individual dies in a period and where we observe a statin claim or excess medication has been kept from a previous period.
Sample Construction Linking Adherence to Outcomes over Time
Table 1 displays the means and standard deviations of the main demographic and outcome variables in the full sample and in the different subsamples. We separate out patients who changed physicians over the observation period, and within that sample we identify two distinct subsamples: those who changed physician because they relocated to a different residential address and those who changed physician because the original clinic closed. A priori one would have different expectations regarding the causes of these shifts relative to the rest of physician–patient separations. It is reasonable to expect that patients who initiate a separation without any observable external event tend to switch based on personal preferences, but individuals who relocate are forced to change physicians. The set of individuals affiliated with primary care clinics that close due to, for example, physician retirement, long-term sickness, or geographic relocation of the physician,16 are also effectively forced to initiate a switch. Thus, for these two sets of physician–patient separations, we can plausibly assume that the change in physician is not the result of patient selection out of the care of the provider. About 17 percent of the sample change physicians during the observation period for any reason. Less than one-half of those changes are precipitated by a clinic closure or patient residential relocation.
Descriptive Statistics by Subgroup Measured at Sample Entry
Statin users in Denmark are more likely to be male and married, on average, in their early- to mid-60s, with annual incomes roughly equivalent to 40,000 dollars. About 40 percent of them have completed only primary education, 50 percent have finished vocational training or some college, and only about 6 percent have a university education. Only 5 percent of the sample are born outside of Denmark.
The prescriber ID we observe in the data pertains to the clinic and not the individual prescribing physician within the clinic. However, in our sample, 46 percent of primary care clinics have only one physician associated with them, so that the clinic ID will uniquely identify the prescribing physician. In cases where the clinic is operated by more than one physician, we interpret any measure of quality as representing an average treatment style of the clinic. As a robustness check, we conduct the analysis separately for clinics with only one physician (see below).
For the patients who switch providers during the observation window, we assign the health management metric of the new physician based on all years for all of their patients other than the index individual i. That is, in t = 0 where we observe the new physician–patient match, we assign the time-constant health management skill of the new physician. We have assigned the health management skill values of the original physician at t = –2,–1 and the value of the new physician on t = 0,1,2.
Figure 2 shows the standardized distributions of physician health management skill. The standard deviation in the nonstandardized measure is 0.005.
Kernel Density Estimate of the Standardized Health Management Skill Measure for the Entire Sample, as Well as the Two Subsamples Experiencing Relocations or Practice Closures
Notes: The kernel used is Gaussian, and the bandwidth is 0.001. The standard deviation in the entire population is 0.005. For the relocation and the practice closure samples, the standard deviations are 0.01 and 0.017, respectively.
V. Validating the Metric of Health Management Skill
In this section, we seek to validate our measure of physician health management skills. To do this we establish that individual adherence is affected by the physician health management skill and is not the result of patient sorting. In particular, we show that individual adherence with statins improves upon switching to a new physician who has better health management skills, or it decreases when the patient encounters a new physician with worse health management skills. This finding is robust to controlling for individual-specific time-invariant unobservable characteristics. We also show that physician health management exhibits mean reversion upon changes, indicating that our findings are not the result of patient sorting.
To establish that individual adherence is affected by physician health management skills, we estimate models similar to Equations 3 and 4 with adherence as the outcome. These results are reported in Table 2. The first two columns in Panel A report results based on the entire population. Column 1 shows how an increase in average health management skill of one standard deviation increases individual adherence by about one percentage point, and this effect is not sensitive to the inclusion of individual characteristics in Column 2.
Individual-Level Adherence Regressed on Health Management Skills
Having shown that health management skill affects individual health behavior, we utilize the quasi- exogenous patient–physician matches formed after patients relocate or physicians close their practice to verify the metric of health management skill. These new matches are less likely to suffer from selection. We start off by showing in an event-study-type setting that the forced change of PCP leads to an immediate change in health management style, to which adherence responds accordingly. We also show that health management skill exhibits mean reversion, in the sense that individuals who experienced lower levels of physician health management pre-closure (pre-relocation) on average experience the largest increases in physician health management skill post-closure (post-relocation). It is important for the validity of our estimates that there is no assortative matching between patients and physicians. If, for instance, high-adhering patients systematically choose particular physicians, these physicians would seemingly have a high level of health management skill. Hence, we would like to see that the physician health management skill exhibits mean reversion properties in the event of exogenous physician switches, such that individuals who had high-type physicians do not systematically switch to new high-type physicians.
Analyzing changes in health management skills in a standard event study framework is problematic. When switching to a new physician in these types of settings the change is not guaranteed to be uniformly either positive or negative, as a switch from high to low levels for some individuals might be offset by low- to high-level switches for other individuals. Hence, the changes in health management skills might balance out on average. To overcome this issue, we conduct separate analyses based on the level of health management skill of the patient’s pre-closure physician.17
Figures 3 and 4 present the evolution in adherence for patients who relocate and individuals experiencing a clinic closure, respectively. The graphs depict the evolution in adherence for individuals who pre-separation have providers with physician health management skill in the lowest quartile (Panel A), and physician health management skill in the highest quartile (Panel B).18 These graphs show that physician health management has substantial impact on patient adherence.19 In Panel A of Figure 3, we see that individuals who prior to relocating have a physician with health management skills in the lowest quartile, on average see an increase in their adherence by 4.3 percentage points, or 5.4 percent relative to the mean. On the contrary, Panel B of Figure 3 shows that individuals in the highest quartile of physician health management pre-relocation on average experience a decrease in adherence of 2.8 percentage points, or 3.3 percent relative to the mean.
Residential Relocators
Notes: Physician health management skill and adherence for individuals relocating by quartile of pre-closure level of physician health management skill. Individuals are measured relative to the last period when they encounter the old physician (t = –1).
Physician Health Management Skill and Adherence for Individuals Experiencing a Clinic Closure
Notes: Physician health management skill and adherence for individuals experiencing a clinic closure by quartile of pre-closure level of physician health management skill. Individuals are measured relative to the last period when they encounter the old physician (t = –1).
Figure 4 shows that the same pattern exists for individuals who move to a new physician due to a clinic closure. Panel A of Figure 4 shows that individuals who prior to closure have a physician in the lowest quartile of health management skills on average see an increase of 8.8 percentage points in adherence. This corresponds to 11.1 percent relative to the mean of adherence. Similar to those who relocate their geographic residence, patients experiencing a closure of a PCP office that was in the highest quartile of physician health management skills experience a reduction in adherence of 1.4 percentage points or 1.6 percent to the mean of adherence. Online Appendix Tables A1 and A2 also report the implied changes in health management skills. It is clear that coming from a lower (higher) level of health management skill on average results in increases (decreases) in health management skill post-separation. In Online Appendix Figures A1 and A2, which are figures similar to Figure 3 and 4 also including the change in health management skill, show how the change in health management skill is sudden and persistent.20 In Figures 3 and 4, we observe that the adherence rates of those belonging to the fourth quartile of pre-HMS return to their pre-separation levels after around two years. This may be related to differences in post-separation search behavior between the two pre-HMS quartiles. We note that there is a dip in t – 1 in adherence for both residential relocators and those who are separated from their physician due to clinic closures. While we do not have any definitive explanation for this, other research has shown that there is a decrease in the probability of having visits close to clinic closures (see Simonsen et al. 2019). This would mechanically decrease adherence. In sensitivity analyses we exclude t – 1, and the results are similar.21
Having shown that changing provider leads to changes in health management skills that lead to changes in adherence, we leverage the quasi-exogenous changes in provider to validate the impact of health management skill on patient adherence. Columns 3 and 4 in Table 2 estimate models similar to Columns 1 and 2 for the entire subpopulation of residential relocators. The specifications in Columns 5 and 6 repeat the exercise for the subgroup experiencing a clinic closure. The impact of one standard deviation increase in physician health management skill on adherence is 0.8 percentage points and 0.7 percentage points in the group of residential relocators and individuals experiencing a clinic closure, respectively. The estimates are not affected by the inclusion of individual characteristics. In Panel B, we control for time-invariant individual unobserved heterogeneity (individual fixed effects) and find very similar results.
To further support our finding that physician health management skill exhibits mean reversion, we analyze how the change in health management skill due to separations is associated with the health management levels before the change. While Figures 2 and 3 illustrated how individuals in opposite quartiles of the initial health management levels experienced opposite evolutions in health management skill, Figure 5 illustrates the case for the entire distribution of health management skill before switches happen. To construct this figure, we collapse the data to the periods immediately before [t ∈ (−2, −1)] and after [t ∈ (0, 1, 2)] and calculate the change in health management skill. We then group the data into 20 equally sized bins based on the initial (pre-switch) health management skill and calculate the mean change in health management within each of these bins. We do this separately for separations due to residential relocators (Panel A) and for individuals experiencing a clinic closure (Panel B). Additionally, we fit a line through the bins to estimate the relationship between the initial pre-switch health management skills and the change. We find that, based on relocations, one standard deviation increase in pre-shift health management skill is associated with a change in health management skill of 1.36 of a standard deviation. Similarly, for individuals experiencing a clinic closure, one standard deviation higher pre-switch health management skill is associated with a change of 1.16 of a standard deviation.
Mean Reversion in Physician Health Management Skill
Notes: Mean reversion in physician health management skill. The figure plots the change in health management skill relative to pre-shift health management skill. Pre-shift health management skill is grouped into 20 equally sized bins, within which the mean change is calculated. The coefficient (highly significant at the 1 percent level) is the fitted line through these 20 points.
VI. Results—Health Management Skills and Patient Outcomes
In this section, we estimate the impact of physician health management skills on patient health outcomes. After presenting results using the entire population, we conduct robustness analyses on the subgroups switching physicians due to either residential relocation or clinic closure. Table 3 shows the results from empirical models of hospitalization risk of CVD-related admissions,22 measured as a binary variable, as a function of physician health management skills. The health management style variable is standardized with mean zero and unit standard deviation, and coefficients should be interpreted as the impact of a one standard deviation increase in the skill dimension on the outcome of interest. The time window for hospitalization is within one year after the end of the six-month period during which the adherence measure is calculated.
Association between Physician Health Management Skills and Cardiovascular Disease–Related Hospitalization Risk—Linear Probability Regressions.
The regression in Column 1 of Table 3, Panel A includes only calendar year fixed effects and geographic region fixed effects. Taking the results at face value, Column 1 shows that one standard deviation increase in physician health management skill decreases the probability of a CVD-related hospitalization in the subsequent year by 0.07 percent. Relative to the outcome mean of 5.7 percent, this is a relative reduction of 1.3 percent. Controlling for individual-level socio-demographic information in Column 2 and health status (captured here through the Charlson comorbidity index) only marginally reduces the estimate of the impact of health management skills.
Panel B reports the same specifications as in Panel A, but adds patient fixed effects to control for time-invariant individual-level unobserved heterogeneity. The estimates of the associations between physician health management skills and CVD-related hospitalizations are mostly unchanged after the inclusion of the patient-level fixed effects.
The models in Table 4 are identical to those presented in Table 3, but the outcome is now the log of CVD-related hospitalization expenditures in Danish crowns (DKK) incurred in the year after the end of the six-month period during which the health management skill measure was calculated. The pattern is the same as for the hospitalization risk. Consistently across all specifications, a one standard deviation increase in physician health management skill is associated with a 0.24–0.3 percent decrease in CVD-related hospitalization expenditures in the next year. In Online Appendix Tables A3 and A4, we redo Tables 3 and 4, respectively, for single-physician clinics only. The results are similar.
Association between Physician Health Management Skills and Cardiovascular Disease–Related Hospitalization Expenditures
To get an idea about the economic significance of the estimates, we make some simple assumptions about costs and benefits to extrapolate an estimate of the willingness-to-pay for a unit standard deviation increase in health management skill. On the cost side, we assume that the only component is the medication cost. A one standard deviation increase in health management skill will increase adherence by 0.9 percentage points (in a six-month period). The price of (generic) statin treatment is around DKK 1 per day, implying drug costs resulting from increased adherence of DKK 1.62 per six months. In terms of benefits, we conservatively include only hospitalization expenditures and lost productivity. The 0.28 percent reduction in hospitalization costs at a mean of DKK 4,050 is equivalent to DKK 11.4. The mean hospitalization length for the CVD admissions in question is 6.5 days. Using the mean income from Table 1, this is equivalent to an expected income loss of DKK 2.4 per six months. In sum, the willingness-to-pay for a one standard deviation increase in health management skill would be DKK 24.4 per patient–year.23 If we could come up with an intervention that would increase the health management skill of all PCPs in Denmark by one standard deviation, the willingness-to-pay would be at least DKK 33.6 million (around $5.6 million) per year. This is obviously a very conservative lower bound estimate of the true willingness-to-pay, as we do not consider factors like sick leave after hospital discharge, informal care, and changes in the quality of life. Further, this calculation is based only on the benefits from improved health management skills for statin users, though the improvement in these skills will likely affect health outcomes for other chronic conditions as well.
A. Subgroup Analysis—Quasi-Exogenous Separations: Residential Relocations and Clinic Closures
We now present a set of sensitivity analyses to assess whether the relationship between health management skills and CVD outcomes are due to patient selection. Again, we focus on two particular groups of patients: patients who relocated their residence and thus had to change physician and patients whose original physician practice closed.
It is unlikely that patients would move residence because they want to switch physicians or that practices would close because of the quality of the patient load—this is why estimates based on these two subsamples are less likely to be affected by selective patient–physician matching. The sample size is reduced considerably to approximately 10 percent of the original for two reasons. First, we discard those individuals who stay with the same physician for the entire observation window; second, we focus on the time period close to the switches (results are similar if we include time periods further away from the switches; see Online Appendix Tables A5 and A6). We assign the old physician’s health management skill to the months during which the patient is visiting that physician and the new physician’s skill for the months after the change. Table 5 reports results. Panel A gives the estimates from the general pooled regressions, and Panel B presents the corresponding estimates from models including patient fixed effects.
Associations between Physician Health Management Skills and Cardiovascular Disease Hospitalization Risk—Linear Probability Regressions
For the two groups of switchers where the separation from the physician is plausibly due to non–health related reasons, we see comparable estimates to those of the full sample. Considering switches due to clinic closures in Columns 2 and 3, we find that the estimates are qualitatively the same as those for the entire sample (Column 1). For those who switch physician due to residential relocation (Column 2), the point estimate is numerically larger at −0.114, but it is not statistically significantly different.
We report the results for the hospitalization expenditures in Table 6, Panels A and B, and the results are comparable to those of Table 4. For the sample of all movers, we find that one standard deviation increase in health management skills decreases CVD-related hospitalization expenditures for the physician’s patients in the next year by 0.275 percent.
Associations between Physician Health Management Skills and Cardiovascular Disease Hospitalization Expenditures
B. Health Management Skills and Clinical Quality
Having shown that there is a robust association between the health management skills of the physician and patient health, we might worry that our measure of physician skills is merely reflecting the clinical quality of the physician or their ability to make an adequate diagnosis and prescribe the correct treatment. If clinical quality is highly correlated with health management skills, and we are not controlling for it, then this could lead to misleading interpretation of the estimates.
To explore this, we turn to the literature on clinical quality in primary care. Hospitalizations due to ambulatory care sensitive conditions (ACSC) are widely accepted and frequently used metrics used to assess the quality of primary healthcare (see for example, Harrison et al. 2014; Oster and Bindman 2003; Fadlon and Van Parys 2020; Johnson et al. 2012). These are hospitalizations that are avoidable in the sense that proper outpatient care would prevent them. In a U.S. setting, Oster and Bindman (2003) find that African Americans, patients covered by Medicaid, and uninsured patients constitute a disproportionate share of emergency department visits that accrue to ACSCs.24 We identify the subset of ACSCs that are related to urinary tract infections, bacterial pneumonia, chronic obstructive pulmonary disease, and dehydrations; see Online Appendix Table A8 for a complete list of ICD-10 codes. We intentionally only keep the subset of ACSCs that are related to acute conditions, as these are presumably unlikely to be affected by aspects of the physician’s health management style related to chronic disease management and chronic medication intake, such as statins.25 While the ability to properly diagnose and treat acute illness episodes is likely highly correlated with the ability to treat chronic disease, acute conditions such as the ones we consider here generally do not require maintaining long-term medication adherence from the patient.
As with the health management skill, we calculate leave-one-out means of the included admissions on a patient–physician level and combine the admissions into one index. Next, we replicate the analysis from Table 5, with the inclusion of this new quality metric derived from avoidable hospital admissions (for the acute ACSCs listed above); see Table 7. Note that a higher ACSC (“better quality”) is associated with lower rates of CVD admissions, which suggests that clinical quality associated with acute conditions extends to the treatment of chronic conditions as well. Strikingly, adding this quality metric to the regressions does not affect the impact of the health management skill variable in any meaningful way. In our preferred specification in Column 3 using pooled data, we find that one standard deviation change in ACSC-based quality has similar effects on CVD hospitalization as one standard deviation change in physician management skills. Further, we notice that controlling for time-invariant unobserved patient characteristics drastically reduces the association from the ACSC quality metric and CVD outcomes (the coefficient becomes close to zero and insignificant); see Panel B of Table 7. Similar to the results in Table 5, adding individual fixed effects does not affect the estimated coefficient of health management skill in any meaningful way.
Association between Physician Health Management Skills, Ambulatory Care Sensitive Conditions Quality, and Cardiovascular Disease–Related Hospitalization Risk—Linear Probability Regressions
Table 8 reports results for healthcare expenditures, and the conclusion is similar. Notably, the correlations of health management skill and ACSC with healthcare costs are very similar in the pooled sample, but the coefficient on the ACSC-based quality metric is positive and very close to zero when we account for unobserved time-invariant patient characteristics.
Association between Physician Health Management Skills, Ambulatory Care Sensitive Conditions Quality, and Cardiovascular Disease–Related Hospitalization Risk—Linear Probability Regressions
VII. Health Management Skills and Physician Characteristics and Utilization
We have established that the physician’s health management skills affect patient health outcomes. Our setting provides a unique opportunity to investigate whether health management skills are associated with observable provider characteristics.26 Previous research has demonstrated that the gender of the treating physician is correlated with patient outcomes. For example, investigating records on hospitalized Medicare beneficiaries in the United States, Tsugawa et al. (2017) showed that patients who were treated by a female physician had significantly lower mortality and readmission rates. In another study on Canadian data, Wallis et al. (2017) found that among 25 different surgical procedures, 30-day readmission rates were statistically significantly lower when the surgeon was female (however, the difference was small). As both are studies of associations, it is not fully understood what the underlying mechanism is, that is, whether the findings are driven by a differential patient mix, though they do control for some patient characteristics. In a meta-analytic review, Roter, Hall, and Aoki (2002) found that female PCPs engage in more patient-centered communication and had visits of longer duration compared to their male colleagues. If this is the case, we would expect female gender of the PCP to be positively correlated with our measure of health management skills. To investigate this, we estimate the following physician-level regression of the standardized health management skill metric on practice level characteristics:
5
is the estimated health management skill from Equation 2, Xjt are practice level characteristics, rj are municipality fixed effects, and θjt is an error term. Xjt includes an indicator for whether there is at least one female in the practice, whether there are any immigrant physicians in the practice, and the mean age of physicians working in the practice. Of the practices, 48.9 percent have a female physician employed, and 9.2 percent have a physician with an immigrant background. The mean age of the physicians is 54.0 years, where the mean age in practices that do not employ females is 56.3, and it is 51.5 in practices that do employ at least one female. Table 9 reports the results from estimating Equation 8. All models include year and municipality dummies, and standard errors are clustered at the municipality level.
Associations between Practice Health Management Skills and Provider Characteristics
Our results reveal an interesting pattern. In Column, 1 we find that practices with female physicians have better health management skills on average. Having a female in the practice is associated with approximately 16 percent of a standard deviation higher health management skill. The estimate is statistically significant and the positive relationship aligns with the aforementioned literature on gender differences. Further controlling for physician immigrant status, the correlation between health management skills and female gender remains positive and statistically significant. However, when we also control for provider age, the coefficient drops to approximately 2 percent and is insignificant at conventional levels. However, the coefficient on age is highly significant and implies that an increase in provider age of one year is associated with a decrease in health management skills of 3 percent of a standard deviation. Perhaps unsurprisingly, relatively younger practices are more likely to have female physicians, and younger practices are also more likely to have high levels of health management skills, conditional on patient characteristics. The difference in average health management skills between provider genders disappears when we control for provider age.
This is not only the case at the mean, as we can see in Figure 6 below (see Online Appendix Figure A7 for single physician practices only). Here we present local linear regressions of the standardized health management skills on provider age. The relationship is presented for males and females separately. The negative relationship is evident in both groups, and they are never statistically different from each other. These findings are consistent with the fact that from 1977 to 2017 the female share of the Danish PCP workforce increased from 10 to 50 percent.27 That is, the female PCPs in our sample are systematically younger than male PCPs. Hence, we feel confident concluding that in a setting that evaluates health management skills in primary care, where the patient mix is adequately controlled for, there is no differential return to female doctors, after we include controls for the age of the physician. Further, we interpret our findings that more recent (and perhaps more up-to-date) training of physicians is an important factor in physician health management skills.
Local Linear Regressions of Communication Quality and Age by Provider Gender
Finally, as a last set of results, we investigate how the health management skill is associated with physician utilization intensity. There exists a large and developed literature on the determinants of healthcare utilization, with Finkelstein, Gentzkow, and Williams (2016) being one of the more notable examples. In a hospital setting, Doyle, Ewer, and Wagner (2010) investigate how physician treatment styles differ between residents from a high-ranking university (presumably higher quality) and residents trained at a lower-ranking hospital (presumably lower quality). They find that the residents from the higher-ranked university have 10–25 percent lower in-hospital utilization costs per patient, with no associated differences in health outcomes. To the best of our knowledge this has not been studied in the primary care setting.
In Figure 7, we show how physician per patient utilization intensity (measured as the sum of reimbursed claims) varies across providers in our data. As a measure of provider utilization, we calculate a patient weighted average of primary care expenditures.
Distribution of Log of Mean per Patient Healthcare Utilization for Providers
Notes: Kernel density estimate of the distribution of log of mean provider utilization. The figure shows the variation in (log of) mean primary care expenditures. The measure is based on all patients with at least one interaction with the provider and not just those with a statin claim.
In Table 10, we present evidence that physician utilization correlates with health management skill. Here we estimate a series of linear probability models of provider utilization on health management skill. In all models (Columns 1–4), we include year and regional dummies. In addition, we also include the (acute) ACSC-based clinical quality metric. In Columns 2 and 3 we include measures of patient composition and provider characteristics respectively. In Column 4 we include all of these simultaneously. Our measure for patient composition is average income, age, sex, and marital status for all patients at the clinic. This is intended to capture the effect of having a practice situated in an affluent area. Included provider characteristics are age, immigrant status, and sex. We see that a one standard deviation increase in health management skill is significantly associated with approximately a 0.5 percentage point increase in provider utilization.
Linear Regressions of Provider Utilization on Health Management Skill and Ambulatory Care Sensitive Conditions Measures
After establishing that the health management skill and provider utilization are positively correlated, we now ask whether there are different correlations between utilization and health management skill across the distribution of health management skill. It is theoretically possible that the relationship is monotonically positive—either exponentially or linearly increasing. Alternatively, if high quality doctors are able to use less procedures, lab tests, and other types of additional checks without compromising the health of their patients (as is the case in Doyle, Ewer, and Wagner 2010) we might expect an inverse U-shape in the relationship. Below we present evidence of the latter relationship. In Figure 8, we plot the fraction of providers in each utilization quartile by the quartile of the health management skill. The share of providers in Utilization Quartiles 3 and 4 follows an inverse U-shape in terms of the HMS quartiles—it increases through HMS Quartiles 1 to 3, but falls again in the fourth quartile of the HMS. Similarly, the share of lowest-utilization providers follows a U-shape.
Distribution of Provider Utilization
Notes: This figure presents the fraction of providers in each quartile of health management skill (HMS, the bars) by utilization quartile (x-axis).
More formally, we also perform linear regression of the probability of being in Utilization Quartiles 1–4. We sequentially regress the probability of being in either of the utilization quartiles on the health management quartiles with HMS Quartile 1 as the left-out category. As can be seen from Table 11, providers in HMS Quartiles 2 and 3 are less likely to be in Utilization Quartile 1 compared with providers in HMS Quartile 4. Similarly, providers in HMS Quartile 4 are less likely to be in the highest utilization quartile compared with HMS Quartiles 2 and 3. This is consistent with Doyle, Ewer, and Wagner (2010), who show an inverse relationship between physician quality and the intensity of utilization.
Linear Regression of Provider Utilization Quartile on Health Management Skill Quartiles
VIII. Conclusions
The full extent of the contribution that physicians have to patient health, while intuitive, is not well understood. As quality contracts become increasingly popular across various healthcare systems, it is important to highlight what facets of individual physicians’ health management styles have meaningful impact on health outcomes and to what extent they vary across physicians. The physician’s ability to correctly diagnose and treat common conditions is one of the central tenets of quality contracts. But the link between these skills and patient outcomes is at best tenuous. Critics have emphasized that unobserved patient-specific characteristics are important and under-researched contributors to the variability of patient health outcomes conditional on physician clinical skill.
This research uses the population of Danish statin users between 2004 and 2008 and shows that the physician’s health management skills, as proxied by that physician’s average patient adherence with prescribed therapy, remain predictive of health outcomes even after accounting for individual patient heterogeneity. Further, investigating potential observable correlates of physician health management skills reveals that younger physicians have on average more adherent patients. We find no substantial difference in health management skills between male and female physicians after we control for physician age and the patient mix. Our research demonstrates that interventions aiming at improving physicians’ health management skills as they relate to patient adherence with prescribed therapy will have positive impacts on patient health outcomes.
Acknowledgments
Simeonova and Skipper gratefully acknowledge support from the Danish Council for Independent Research. The authors have nothing to disclose. Data are drawn from a restricted access data source and hence cannot be made publicly available. The data used in the paper are administrative Danish data maintained by Statistics Denmark, which is kept in a secure server. However, the data can be accessed remotely from within Danish universities and research institutions. If a researcher at a university or other research institution outside Denmark wishes to use these data, this may be accomplished by visiting a Danish research institution or by cooperating with researchers or research assistants working in Denmark.
Footnotes
↵1. Using Danish population registry data, Koulayev, Simeonova, and Skipper (2017) demonstrate substantial differences in average patient adherence with prescribed medication across different physicians. Simeonova (2013) shows that racial gaps in chronic heart failure survival rates persist even after individual physicians’ unobserved characteristics are accounted for, but the gaps disappear among patients who mostly adhere with their prescribed medication. Analyzing the results from a randomized experiment, Alsan, Garrick, and Graziani (2019) show that Black men in Oakland are much more likely to complete preventive screening tests and to communicate with their doctors when seen by a Black (male) physician, though they report no difference in patient perceptions of physicians’ clinical abilities. Meyers et al. (2019) also demonstrate substantial variability in AIDS-specific medication adherence associated with provider and practice characteristics.
↵2. Adherence measures the intensive side of following prescribed therapy, namely taking as much of the medication as prescribed over a pre-determined period. If a patient stops taking statins altogether, they becomes nonpersistent and are excluded from the sample. Our measures and estimated effects are based on deviations from adequate therapy compliance conditional on maintaining the patient on the therapy. Patients who completely discontinue the therapy likely experience different, potentially much more severe, health consequences.
↵3. Five kilometers in some nonrural areas.
↵4. USD $1 is approximately DKK 6.
↵5. If the physician is more than 60 years old, they can apply to stop the intake of new patients at a lower threshold.
↵6. Although some physician characteristics likely evolve over time, for example, prescribing of medications and use of procedures, we model the physician health management skill as a fixed characteristic. We do this out of two reasons. First, the time window we consider is short—only five years. Second, patient adherence is unobservable for physicians in Denmark. Thus, the physician does not receive feedback on their effort to inspire adherence, if they consciously exert such effort. The opportunity for physician learning is very limited.
↵7. There are two forms of cholesterol found in the blood: high-density lipoprotein (HDL) and low-density-lipoprotein (LDL). The first form is commonly denoted the “good cholesterol” because it transports the harmful cholesterol (LDL) out of arteries, and high blood level concentrations of HDL is recommended. The latter form (LDL)—or “bad cholesterol”—could potentially combine with other fats to create blockage of arteries and veins. Atherosclerosis, which is the thickening of artery walls, is partly due to high levels of cholesterol and is broadly recognized as an important and modifiable risk factor for CVD. Plaque causes the actual clogging and is often a result of a cumulative buildup of lipids—small fatty particles penetrating the walls from the blood to arteries, with a speed conditional on the concentration of cholesterol in one’s blood. Lipid-lowering drugs reduce the buildup of plaque, thus effectively reducing the likelihood of clogging. However, lipid buildup is irreversible, once present in veins/arteries, and pharmacological treatment only prevents further buildup. Hence, for the treatment to be effective, a continuous intake of medication is necessary, and pharmacological treatment, once initiated, is considered permanent.
↵8. For example, beta blockers, another drug used in chronic conditions related CVD, are used for multiple conditions (for example, congestive heart failure, cardiac arrhythmias, and hypertension) and could also be used to relieve anxiety on an as-needed basis.
↵9. This is the case for 98 percent if the patients in our sample.
↵10. Adherence is calculated as proportion of days covered, also known as the medication possession ratio. For details, see the next section.
↵11. When a clinic closes, the patients are free to choose a physician from their choice set that is open for intake of new patients.
↵12. Markussen, Røed, and Røgeberg (2013) and Godøy and Dale-Olsen 2018 use separations due to closures to assess the impact of physicians on sickness absence. Finkelstein, Gentzkow, and Williams 2016 and Laird and Nielsen (2016) use residential relocators to identify place and provider fixed effects.
↵13. The first claim is the initiation. We drop individuals who never have more than one claim.
↵14. Also sometimes referred to as a “medication possession ratio.”
↵15. Note that we cannot study nonpersistence or the related phenomenon of primary nonadherence, which happens when a patient receives a prescription for a drug but never fills it. We only observe medication claims that were filled, and we do not observe electronic prescriptions by providers.
↵16. The only other reason why a clinic might close is that a physician’s license is revoked. Effectively this never happens.
↵17. We do this by analyzing changes in adherence when an individual separates from a provider from different quartiles of health management skills. The quartiles are calculated within the respective samples. In the relocator sample, the standard deviation of the health management skill is 0.009, while it for the clinic closure sample it is 0.018.
↵18. Similar graphs for individuals with pre-physician health management in the second and third quartile are presented in the appendix.
↵19. To measure the change in adherence, we estimate regression models of the form Adhiτ = θiτ + ηiτ for each quartile of the health management skill of the patient’s pre-closure physician. The results are available in Online Appendix Tables A1 (relocators) and A2 (practice closures). Similar regressions are estimated with the outcome being health management skill. These results are also available in the Online Appendix.
↵20. Note that in these figures there is a slight fluctuation in pre-closure health management skills due to patients entering and leaving the sample when they stop using statins. In Online Appendix Figures A5 and A6, we present graphs that are constructed by restricting the sample to those that have only one physician in the periods prior to the closure and are in our sample for the four periods prior and four periods after (six months each) experiencing the patient–physician separation. From these it is even more evident that the shift indeed leads to an immediate change in physician health management skills.
↵21. Results available upon request.
↵22. In identifying the CVD-related admissions, we rely on the approach by Pittmann et al. (2011). The full set of ICD-10 codes included in our definition of admissions related to CVD can be found in Online Appendix Table A7.
↵23. 2 × (13.8 – 1.62) = DKK 24.4.
↵24. Harrison et al. (2014) study the impact of a national primary care pay for performance scheme rolled out by the English National Health Service. As part of the performance scheme, reductions in a subset of the ACSCs were incentivized on the physician side, and the rates of these hospitalizations declined in comparison to nonincentivized ACSC hospitalizations.
↵25. Our results are robust to including asthma. Results available upon request.
↵26. As we only have characteristics on the practice level, we show that our findings are robust to limiting the sample to single-physician practices, where we know the specifics of the provider characteristics. In Online Appendix Table A9, we replicate our finding using only single physician practices, where we know for a fact whether the patients see a female or male physician.
↵27. See https://www.rm.dk/api/NewESDHBlock/DownloadFile?agendaPath=%5C%5CRMAPPS0221.onerm.dk%5CCMS01-EXT%5CESDH%20Data%5CRM_Internet%5CDagsordener%5CSU_for_almen_praksis%202017%5C21-09-2017%5CAaben_dagsorden&appendixId=177205 (in Danish, accessed December 5, 2023).
- Received April 2020.
- Accepted November 2021.
References
Jump to section
- Article
- Abstract
- I. Introduction
- II. Background and Institutional Setting
- III. Empirical Strategy
- IV. Data
- V. Validating the Metric of Health Management Skill
- VI. Results—Health Management Skills and Patient Outcomes
- VII. Health Management Skills and Physician Characteristics and Utilization
- VIII. Conclusions
- Acknowledgments
- Footnotes
- References
- Figures & Data
- Supplemental
- Info & Metrics
- References