- Email: [email protected]
The changing of the guards
Journal of Health Economics 32 (2013) 1230–1239
Contents lists available at ScienceDirect
Journal of Health Economics journal homepage: www.elsevier.com/locate/econbase
The changing of the guards Can family doctors contain worker absenteeism?夽 Simen Markussen, Knut Røed ∗ , Ole Røgeberg Ragnar Frisch Centre for Economic Research, Norway
a r t i c l e
i n f o
Article history: Received 31 August 2012 Received in revised form 8 October 2013 Accepted 9 October 2013 Available online 24 October 2013 JEL classification: H55 I13
a b s t r a c t Using administrative data from Norway, we examine the extent to which family doctors influence their clients’ propensity to claim sick-pay. The analysis exploits exogenous switches of family doctors occurring when physicians quit, retire, or for other reasons sell their patient lists. We find that family doctors have significant influence on their clients’ absence behavior, particularly on absence duration. Their influence is stronger in geographical areas with weaker competition between physicians. We conclude that it is possible for family doctors to contain sick-pay expenditures to some extent, and that there is a considerable variation in the way they perform this task. © 2013 Elsevier B.V. All rights reserved.
Keywords: Sick-pay GP practice styles Absence certification Gatekeepers Family doctors
1. Introduction Publicly mandated sick-pay is an important part of the social safety net in many countries. Given the relatively high replacement ratios in these programs, they entail the risk of excessive use and shirking. To counter this risk, paid sickness spells require certification by a physician in virtually all countries (OECD, 2010, p. 139). This has remained the case, despite a trend in OECD countries toward raising the powers of the benefit-granting institutions in relation to long-term disability benefit claims. The GPs thus remain important gatekeepers of the welfare states’ public purses. The rationale behind this policy is that physicians can, better than others, verify the presence and severity of a medical condition, and thus decide whether or not a person is eligible for sick-pay. At the same time, however, these physicians operate businesses in a competitive environment, and may be tempted to attract and retain patients by giving easier access to services and treatments – such
夽 This research has received financial support from the Norwegian Research Council through the program “Sickness, absence, work and health” (grant nos. 187924 and 201416). Data made available by Statistics Norway have been essential. Thanks to seminar participants in Bergen and Oslo, and to two anonymous referees for a number of useful comments and suggestions. ∗ Corresponding author at: Ragnar Frisch Centre for Economic Research, Gaustadalléen 21, 0349 Oslo, Norway. Tel.: +47 22958813. E-mail address: [email protected] (K. Røed). 0167-6296/$ – see front matter © 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.jhealeco.2013.10.005
as sickness benefits – that are desired by (potential) clients. In this way, market pressures and the physicians’ own economic interests may erode their gatekeeper role. An additional role conflict may arise if the physician interprets his/her duty as being an advocate for their patient; see, e.g., Reiso et al. (2000). There is a theoretical literature discussing the efficacy of physicians as gatekeepers in relation to absenteeism, as well as to costly specialist treatments and drugs prescription. Focus has been on the information advantage physicians have over patients regarding the appropriate medical treatment (e.g., Scott, 2000; Brekke et al., 2007), the information advantage patients have over doctors regarding private health (e.g., Stone, 1984), and the tension in the physician’s role as both advocate of the patient and health-assessor on behalf of the insurance provider (e.g., Blomqvist, 1991). On the basis of a theoretical model and qualitative interview data from a sample of general practitioners, Carlsen and Nyborg (2009) argue that there are neither theoretical nor empirical grounds for expecting general practitioners to fulfill the role of effective gate-keeping in relation to sick-pay, and conclude that “the gate is wide open”. Empirical evidence also indicates that it is difficult for physicians to contain sick-pay expenditures. Englund and Svärdsudd (2000) report that Swedish GPs find it hard to overrule their patients’ requests for sick-leave certificates. Even in cases where GPs do not see sick-leave as recommendable, they still issue sickness certificates in as much as 87% of the cases where the patient asks for it. Based on focus group interviews of Norwegian GPs,
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
Nilsen et al. (2011) conclude that “trust in the patient’s own story and self-judgment” is deemed crucial, and that sick-leave certificates should be considered “mainly patient-driven”. On the other hand, Markussen et al. (2011) report a large variation in Norwegian GPs’ observed tendency to certify sick-leave, suggesting that the physicians’ priorities or attitudes may be important after all. And empirical evidence related to other aspects of the GPs gatekeeper role also indicates significant variation in strictness. Wilkin (1992), for example, identifies a large variation in the frequency with which physicians refer their patients to specialist treatment, and Grytten and Sørensen (2003) identify a similarly large variation in physicians’ use of (expensive) laboratory tests. Based on a reform in the funding scheme for certain types of elective surgery, Dusheiko et al. (2006) show that GPs tend to have a more restrictive referral practice when money not spent on surgery alternatively can be channeled back to their own practice. A fundamental problem with empirically identifying the influence of physicians on their patients’ behavior is that patients are not allocated randomly to physicians; consequently, differences in observed absence rates across physicians’ patient groups may reflect sorting as well as causality. In particular, patients demanding a (questionable) sickness certificate may seek out an accommodative (“lenient”) physician. While this identification problem can be reduced by controlling for observed patient characteristics, it is typically difficult to ascertain that no unobserved sorting remains. And to our knowledge, no empirical evaluation of family doctors’ influence on their clients’ absence behavior has convincingly solved this selection problem. An important contribution of the present paper is that we examine the influence that family doctors have on their clients’ absence behavior by studying how the absence rates of a patient group shifts when the entire group is moved exogenously from one physician’s list to another. This allows us to observe the absence rates of a large set of patient groups under two different GPs, largely avoiding the selection problem. Our empirical basis is administrative data from Norway that allows us to link all citizens to their family doctors as well as to their individual records on employment and physician-certified sick leaves. Norway is a particularly well suited country for investigating the role of physicians, since everyone in Norway has been assigned a GP who is their primary contact point with the public health care and sickness insurance systems, regardless of whether they have actually consulted the doctor or not. In the main, these GPs are the ones responsible for certifying sick-pay for workers. The system has the additional feature that when a GP moves to another job/district, retires or closes the practice for other reasons, the list of patients is sold to another GP and the patients are collectively moved to this new GP. Further strengthening the identification strategy is a key principle of the system which states that GPs have no direct opportunities for shaping their patient list: If an exiting GP’s list is split between two GPs, the split is random and performed by a central organization rather than the GPs themselves. The main research question we seek to answer in the present paper is the following: Does a family doctor have a significant influence on the level of sickness absence in his/her group of patients? The more extreme variants of the gate-keeping-does-not-work-andmay-even-be-impossible argument would predict that the level of absenteeism is independent of which physician the patient is referred to, and we would consequently not expect an exogenous switch of GP to cause changes in claim behavior. If, on the other hand, there are substantial and persistent differences between physicians in terms of their “strictness” or “leniency” which are not fully neutralized by “physician shopping”, we would expect an exogenous switch of GP to entail noticeable changes.
1231
Our paper is also motivated by a more methodological issue: If the family doctor has a direct impact on a worker’s benefit claims, this would potentially imply that indicators for physician strictness can be used as exogenous instruments to answer questions about the causal effect of becoming a benefit claimant. This strategy was pursued by Markussen (2012), who investigated the impact of absenteeism on subsequent earnings growth using physician-strictness indicators computed in Markussen et al. (2011) as instrumental variables. Similar identification strategies have also been used by others: Duggan (2005) examined the impact of new/expensive antipsychotic drugs on subsequent health care spending, using psychiatrists’ propensity to prescribe expensive drugs as instrumental variable. French and Song (2009) and Maestas et al. (2011) investigated the impact of disability insurance receipt on subsequent labor supply, using the health examiners’ observed allowance rates as an instrumental variable. Our paper sheds light on the potential power of these empirical approaches by assessing the influence that gate keepers really have on these kinds of decisions when all individual and environmental factors are appropriately controlled for. In particular, we offer a reexamination of the physician-strictness indicators computed by Markussen et al. (2011) as our data to some extent encompass the same group of physicians. The key finding of our paper is that family doctors really do have a significant impact on their patients’ benefit claims. Comparing patient groups who were subject to an exogenous switch of family doctor with similar groups who were not subject to such a switch, we find that the changes in absenteeism – positive or negative – were significantly larger for the groups subject to a switch. Measured by the standard deviation in the groups’ sick-leave changes over the next two years, we find that the switches induced a 26% increase in absence variability. The family doctors exert much more influence on the duration of sick-leaves than on their occurrence. We also find that the influence of the family doctor is larger in areas where there is little competition between doctors, and thus a smaller pool of doctors that patients can choose from. Comparing our results to those reported by Markussen et al. (2011) for the same group of physicians, we find a large degree of concordance. Given the completely different identification strategies used, we interpret this as further evidence that there really are significant differences between physicians in the way they fulfill their gatekeeper role. Family doctors do indeed have some powers to contain social insurance costs, but they use them to varying degrees. 2. Institutional setting and data In Norway, workers receive 100% wage compensation from the first day of sickness absence and for up to one year (up to an income ceiling of approximately 500,000 NOK p.a. (2012)). The first 16 days are paid for by the employer, after which the costs are covered by general (payroll) taxation. A sick leave certificate issued by a physician is normally required for all spells lasting more than three days.1 On a typical working day, 6–7% of Norwegian employees are absent from work due to sickness, and almost 90% of these absences are certified by a physician. OECD has repeatedly pointed out that Norway’s rate of sick-leave is among the highest in the world, and recommended that independent checks on the GPs’ sickness assessments are performed more systematically; see, e.g., OECD (2012). Reducing absenteeism has also been made a major priority in a tripartite agreement between the employer associations, the employee associations, and the state (the inclusive workplace
1 Some firms have agreed to accept self-reported sickness claims for up to eight days.
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
2 The supply of physicians with vacant list-slots varies significantly from region to region, particularly between rural and urban areas. In some parts of the country, patients may in practice have no choice other than accepting their assigned physician. 3 It is common to use graded absence certificates in Norway, implying that the worker is only partially absent from the workplace, see Markussen et al. (2012) for details. If, for example, a worker has a 50% absence spell for 10 days, we count this as 5 days of absence.
0
Number of switches 15 10 20 5
(a) Number of switches per month
July 2002
July 2003
July 2004
July 2005
July 2006
Percent 10
15
(b) Size-distribution of the switches
5
agreement), and in the approved strategy documents, physicians are attributed a key role. New comprehensive guidelines have been developed for the physicians’ sick-listing practices and for their dialog with employers and the social insurance administration. Norway has, since May 2001, practiced a family (panel) doctor system, whereby each citizen is assigned a single GP who receives a capitation fee from the social security administration. Sickness absence certificates can in principle be issued by any authorized physician, but will normally be issued by the family doctor, except in emergency cases and when the patient is hospitalized or subject to intensive specialist treatment. Norwegian workers are free to choose their family doctor insofar as the desired doctor has vacant patient slots, but unless they make an active choice, they will be assigned a “default” GP, based on their residential address. When a physician retires or moves to a new job (or a new location), the whole patient list will be sold to a new GP, but – as we return to below – patients may of course self-select to other physicians at these occasions.2 The panel doctor determines a maximum number of patients (with an upper limit of 2500). The data we use in the present paper are collected from Norwegian administrative registers made available to us by Statistics Norway. They cover the complete population during the period from May 2001 through September 2008, with individual level (encrypted) panel data information concerning patient-GPlinkages, employment spells, physician-certified absence spells, demographic information (gender, age, nationality), and educational attainment. Using these data, we construct a monthly outcome measure indicating – for each employee – the number of physician-certified sick-pay days, adjusted for grading.3 Note that it is the timing of the certification that determines to which month the recorded absence days are attributed, and not the timing of the absence itself. Longer absence spells will typically consist of several certification periods, and hence be distributed across a number of months (the average length of a certification period is approximately 19 days). The population we are interested in consists of workers who are forced to change family doctor because their original GP sells the whole practice to a new doctor. The resultant new patientGP matches are exogenous with respect to each patient’s behavior and characteristics, as a core design principle of the system is that GPs have no direct influence over the composition of their list. In particular, it is not possible for a new doctor to “cream skim” by taking over only a selected subset of the previous doctor’s list. It is the municipality, not the doctor that puts vacant panel doctor practices on the market and allocates them to new physicians. The incoming physicians must then negotiate prices with the outgoing ones (if they do not agree, the prices are determined by a committee under the auspices of the national physicians’ association). The basic idea behind our empirical approach is to examine the frequency of sick-pay claims before and after physician switches in order to assess whether there are changes in sick-pay patterns that are attributable to these switches. Since the frequency of sickpay claims obviously changes over time for a number of reasons (changes in employment status, cyclical and seasonal fluctuations, time-trends, aging, pure randomness), we need a control group of
0
1232
0
200
400 Number of employed clients
600
800
Fig. 1. The number of exogenous switches, their timing and the number of employed patients involved. Note: The months indicated in panel (a) are the last months before the switch (e.g., a July-bar indicates the shifts occurring between July and August).
physicians with stable patient lists to represent the counterfactual case of remaining with the same GP. A potential problem with this approach is that some GP switches may be anticipated. If, for instance, a doctor is approaching retirement, the patient may well be aware of this and start looking for a new GP even before the switch occurs. This would mean that the transferred list contains only a selected sub-group of original list members. To address this concern, we will interpret the client-list that applied 12 months before the actual switch as defining the “treated” population. We also disregard any subsequent physician switches, since these may occur as an endogenous response to the behavior of the new physician. This implies that we use a completely balanced panel to study the frequency of sick-pay claims before and after the time of the physician switches. As we return to below, this creates a sort of attenuation bias when we evaluate the influence of the new physician relative to the old one. A second sample restriction is to include only persons who were employed from 12 months prior to the switch to 24 months after, since employment is a precondition for sick-pay. The three-year employment condition may be disputable if family doctors’ certification practices also affect employment, as indicated by the results reported by Markussen et al. (2012). Within the relatively short time-window used in the present analysis, however, we consider this impact to be negligible. We nevertheless relax the three-year employment restriction in a robustness analysis reported below. Formally, we denote the set of patients moving collectively from physician m to physician n at time s and observed in month t by .4 We can refer to m as the outgoing (patient-supplying) GP, {Ps }m,n t and n as the incoming (patient-receiving) GP. The treatment sample is constructed by identifying all GP induced switches that involved at least 95% of the listed patients, and took place in the period from May 2002 through September 2006.5 This time period is selected to ensure that we can follow the list-members absence behavior from one year before to two years after the switch. The treatment group is defined as all employees who were on an outgoing physician’s list exactly one year before a switch. Fig. 1, panel (a) shows the frequency of exogenous physician switches included in our analysis by calendar month. Although the
4 Note that physician shifts in Norway always occur at the border between two months. 5 In addition, we require that the switch moved at least 150 persons (and at least 50 employed persons).
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
328 98,190 38.8 42.6 13.6 10.2 3.79
switches are spread out over time, there is a disproportionally large fraction of them occurring before/after the summer and Christmas holidays. Panel (b) shows the size distribution of the employed list-members actually included in our analysis. A typical switch involves around 200–400 persons satisfying our employment criterion. Our focus on complete list transfers also ensures that most of the patient-receiving physicians (92.3%) are really “new”, in the sense that they did not have any patients at all in the municipality prior to the exogenous switch. For each treated patient group, we select a corresponding (employed) control patient group belonging to a physician who did not quit. Formally, we denote the control group as {Ps }m,m , t where s refers to the timing of the physician switch occurring in the treatment group for which the group is a control. Hence, we may think of the control groups as experiencing “placebo” physician switches (with n = m) at exactly the same time as the treatment groups. The purpose of the “placebo” switches in this context is to ensure that the “time-dimension” in the before-after-comparisons is exactly the same in the treatment and control groups, both in relation to calendar time (generated by, e.g., seasonal and cyclical fluctuations) and in relation to the three year period for which we have conditioned on employment. Control groups are matched to the treatment groups one year before the physician switches occur in the treatment groups. They are selected among the patient groups who were not subject to an exogenous physician switch during our observation window. The matching procedure is designed to ensure as similar patient composition as possible in treatment and control groups, particularly in terms of past absence behavior. More specifically, for each treatment group we identify all patient groups of similar sizes (±100 patients), whose employed members had absence rates that did not deviate more than 0.5 percentage point from those of the treatment group in both the year just prior to matching and in the year before that (this is done to ensure both similarity in both absence levels and trends). Among these, we select the group that is closest to the treatment group in terms of average absence rate in the two-year period prior to matching.6 Finally, we ensure that each treatmentcontrol group pair is of exactly the same size by randomly removing patients from the larger of the two. Table 1 provides some summary statistics for the treated and non-treated workers, whereas Fig. 2 compares their absence behavior for as much as five years backwards in time; i.e., prior to the physician switch. In total, we have data for 328 treatment and control groups, respectively. On average, each group consists of 299 employed patients (after the groups have been shaved to ensure equal sizes). Although we have not used individual characteristics
6 We have chosen to use past absence behavior as the key matching criterion here since we consider it important to ensure similar baseline absence levels/trends in treatment and control groups. Propensity score matching is not appropriate in this case, since it turns out that patient group characteristics do not contribute much to explaining the exogenous physician switches.
.04
328 98,190 39.2 44.6 13.8 10.4 3.72
Absence rate .02 .03
Treatment group
.01
Control group
0
Number of groups Number of persons Mean age Percent females Mean years of schooling Percent non-natives Local unemployment rate at the time of matching
.05
Table 1 Descriptive statistics patients groups.
1233
-5 years
-4 years
-3 years
-2 years
-1 year
Years before physician switch
Control
Treatment
Fig. 2. Absenteeism 1–5 years prior to the time of exogenous physician switch in treatment and control groups. Note: Absence data for the pre-matching period cover long-term absence (>16 days) only; hence, they are not directly comparable to the data used in the outcome period.
(apart from past absenteeism) as matching criteria, the numbers reported in Table 1 indicate that the treatment and control groups are similar, both in terms of observed characteristics like age, gender, education, and nationality, and in terms of local labor market conditions. Their past absence behavior also follow a similar pattern; conf. Fig. 2.7 As pointed out above, some patients in any group {Ps }m,n may t be exposed to GPs other than m and n, since individuals are free to choose their GP at any time. The sampling scheme ensures that we follow all patients associated with a physician m from one year before the exogenous (genuine or placebo) switch, and interpret them as treated by this particular switch regardless of any subsequent transitions to other doctors. This implies that the estimated effects of the switches can be given a sort of intention-to-treat interpretation. What we can hope to identify from our analysis is thus not the impacts of moving from physician m to physician n, but rather the impact of being exogenously split from physician m and offered physician n as a new default choice. The impacts will include any effects that the exogenous switch has on subsequent patient-initiated switches. Table 2 shows some descriptive statistics for the physicians involved in our analysis. Unsurprisingly, there are some significant differences between the three groups of incoming, outgoing, and staying (control) physicians. Incoming physicians are generally younger and less educated than outgoing and staying physicians. And, although the average age of outgoing and staying physicians are similar, the underlying age distributions are different. In particular, since GP-practices are typically sold early in careers (as young GPs move on to specialization or more attractive jobs) or late (as they retire), the outgoing physicians have a bimodal age distribution with spikes around the retirement age and around the late thirties (not shown). It may also be noted from Table 2 that the distributions of the municipalities’ overall physician coverage (total patient slots relative to population sizes) were very similar for the treatment and control physicians. Fig. 3 illustrates the degree of “non-compliance” for the treatment and the control groups; e.g., the fractions of clients that at
7 The apparently rising absence trends revealed by Fig. 2 are primarily caused by the employment condition imposed from the time of matching and onwards. As we move backwards in time, an increasing proportion of the populations are likely to be non-employed, and thus not under risk of being absent.
1234
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
Table 2 Descriptive statistics physicians (standard deviations in parentheses).
Percent female Age Percent with specialist education Average GP capacity in municipality (#list-slots/#inhabitants) Population-physician-ratio in municipality (#inhabitants/#physicians) Number (N)
different points in time have self-selected away from their initial or assigned physician. Not surprisingly, patient-initiated switching is much more common in the treated patient groups–especially just after the exogenous switch. This illustrates the point that a new doctor may be under stronger competitive pressure than an established one. 3. Empirical analysis As a patient group is exogenously moved from one GP to another, there are at least four different mechanisms that may induce changes in the group-members’ sick-leave behavior. First, if the new GP has a different practice style than the old one, this may induce an upwards or downwards shift in the group’s overall benefit claim propensity. With the term “practice style”, we primarily have in mind the physicians’ understanding of their gatekeeper role and their predisposition toward using absence certificates as part of a medical treatment. We realize, however, that differences in the use of absence certificates may also arise from differences in other aspects their clinical practices and patient management that affect their patients’ health. Second, there may be a non-familiarity-effect, resulting from the fact that the new GP has less information and less personal ties to the patients. This effect is also of unknown sign; the new doctor may be more lenient than the old one if it takes time to notice patient-specific patterns indicative of shirking, and stricter if a settled GP-patient relationship introduces a familiarity/friendship that makes the GP more inclined to adhere to the patient’ wishes. Third, it is probable that a new doctor is subject to stronger competitive pressures than more established ones, e.g., as a result of less loyal list-members; confer the large fraction of patients who self-select out the list after a takeover
.6
Fraction remaining with initial physician .7 .8 .9
1
Fraction of initial patient population remaining with the same (or asigned) physician
1 year before
Switch
1 year after Control
Treatment
Fig. 3. Fraction of initial patient population (12 months before the switch) that stays with initial or new assigned physician.
Control group Staying phys.
Treatment group Outgoing phys.
Treatment group Incoming phys.
29.6 48.2 (9.3) 44.5 1.11 (0.05) 1164 (138) 328
32.6 47.3 (12.2) 26.2 1.11 (0.06) 1140 (143) 328
41.2 36.0 (7.7) 3.7 1.11 (0.06) 1139 (143) 328
(Fig. 2). This effect may be particularly prominent just after a switch. And finally, a shift of physician may in some cases disrupt ongoing medical treatments and/or cooperative efforts with employer and employee to facilitate work-resumption, most likely contributing to increased absenteeism in the short run. As noted above, absence certificates can be issued by any authorized physician, making it possible for dissatisfied patients to circumvent their own family doctor. Also, in cases requiring consultations outside office hours or emergency room visits, the family doctor will normally not be available. These mechanisms are likely to imply that differences observed in absence patterns in relation to a forced shift of family doctor are smaller than what the differences in the two physicians’ practice styles would imply. As a measure for differences in practice styles between physician pairs, we thus expect our intention-to-treat estimators to be considerably attenuated. Taken together, it is not clear how exogenous switch of physicians affect the average level of sick-pay certification. Fig. 4, panel (a), shows the number of certified sick-pay days per month in the treatment and control groups before and after the (genuine or placebo) switch. There are no clear signs of a treatment effect on average absenteeism, with a possible exception for a slightly higher absence propensity just around the time of the exogenous physician switch. Somewhat surprisingly, this effect seems to be caused by a decline in the control group’s absence, rather than by an increase in the treatment group’s absence. This phenomenon arises, however, from the fact that a disproportionally large fraction of physician switches occur in holiday periods (see Fig. 1), implying that the seasonal pattern is correlated with the timing of the switches. In panel (b), we report the time pattern in absence rates that remains after having controlled for seasonality (by means of 12 calendar month dummy variables) and individual patient characteristics (dummy variables for gender, age, education, and immigration status) within a linear regression framework. From these graphs, there seems to be a pattern of increased absenteeism in the treatment groups relative to the control groups in the first few months after the physician switch. Note that the apparently rising absence trends in both groups are primarily related to the way we have constructed the data in the form of a three-year employment condition (from one year before to two years after the physician switch). As we move toward the end of this conditioning period, its implications for absence behavior become weaker and weaker, implying that long-term absenteeism becomes more and more likely. The appropriate test for the existence and strength of physician influence on sick-pay claims is, however, not the extent to which physician switches affect the average absence rate, but the extent to which they induce excess changes (positive or negative) in sickpay patterns. When the new GP is more lenient than the old one, the frequency of benefit claims will rise, ceteris paribus; when the new GP is stricter than the old one, it will decline. In both cases,
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
(b) Controlled for ind.char. and seasonal variation
1.7 1.6 1.5 1.4
1.4
1.5
1.6
1.7
1.8
1.8
1.9
1.9
(a) Observed
12 months before
1235
Switch
12 months after
12 months before
Control
Switch
12 months after
Treatment
Fig. 4. Average number of sick-pay days per month before and after physician switch.
it will induce a change over and above what can be expected to occur for other reasons (randomness, business cycles, time-trends, etc.). Hence, to evaluate the physicians’ overall influence, we need a statistical model that allows the impacts on the affected groups’ absenteeism to take different directions for different groups. Let yimt be the number of physician-certified sick-pay days prescribed to individual i, belonging to group m in month t. To identify the effects of interest, we formulate the linear model yimt = ˛m + m,n I(t ≥ s) + Xit ˇ + Mt + εit .
physician-switch effects is 26% larger for the treatment groups than for the control groups. This is further illustrated in Fig. 5, where we have shown box-plots for the two estimated physician-switch distributions. The distribution is clearly wider among treatment
Table 3 Main results (P-values in parentheses). I. Group-fixed effects (˛m ) Difference in mean between treatment and control group
(1)
Here, ˛m is a patient-group fixed effect, I(.) is an indicator variable taking the value 1 after the (genuine or placebo) physician switch has occurred, m,n is a patient-group-specific change in the absence level following this switch, Xit is the vector of individual control variables referred to above, i.e., dummies for gender, age (32 categories), education (10 categories for level and 10 categories for type), and immigration status (6 categories, based on origin country), and Mt is a vector of seasonal dummy variables (one for each month of the year). Our main interest lies in comparing the estimated changes associated with genuine physician switches with those associated with placebo switches; i.e., in comparing the distribution of m,n given that n = / m, with the corresponding distribution given that n = m. The estimation results are presented in Table 3 and in Figs. 5 and 6. The first part of Table 3 shows that there were no significant differences between the control and treatment groups prior to the physician switches, either in the mean or in the spread of the group-fixed effects. The second part of the table presents the estimated physician switch effects. They indicate that there was a small, but statistically insignificant, increase in average absenteeism in the treatment group relative to the control group. The most conspicuous pattern revealed in Table 3, however, is the highly significant rise in the variation across the treatment groups’ absence behavior. The standard deviation of the estimated
Ratio of standard deviations treatment and control group
II. Physician switch effects ( m,n ) Mean switch effects control group (Mean(ˆ m,m )) Mean switch effects treatment group (Mean(ˆ m,n )|m = / n) Difference in mean between treatment and control group Standard deviation of switch effects control group (St.Dev(ˆ m,m )) Standard deviation of switch effects treatment group (St.Dev(ˆ m,n )|m = / n) Ratio of standard deviations treatment and control group
0.0008 (P = 0.9808) 1.02 (P = 0.7248)
0.1067 0.1268 0.0201 (P = 0.4927) 0.3307 0.4151 1.256*** (P = 0.0000)
Mean deviation of switch effects control group absolute
0.2666
Mean deviation of switch effects absolute treatment group
0.3162
Difference in mean absolute deviations between treatment and control group
0.0496***
Mean ˆ m,m − Mean(ˆ m,m ) |m = / n Mean ˆ m,n − Mean(ˆ m,n ) |m = / n
(P = 0.0070) Number of patients Number of physicians *
Significant at the 10 percent level. Significant at the 5 percent level. *** Significant at the 1 percent level.
**
98,190 984
1236
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239 Treatment
-2
-1
0
1
2
Control
Fig. 5. Box plots for the distributions of estimated physician switch effects for the control and treatment groups. Note: The shaded boxes show the first, second, and third quartiles. The upper and lower horizontal lines indicate the maximum and minimum values observed within the inter-quartile ranges (IQR = 1.5 (Q3 − Q1)). The dots indicate observations outside IQR.
than among control groups.8 We have investigated whether the excess variation in the treatment groups is associated with particular medical diagnosis. This does not seem to be the case: The estimated switch-effects follow the same pattern within each of the three major diagnosis groups (musculoskeletal diseases, mental disorders, and “other”). The group-fixed effects (˛m )may be interpreted as combined measures of the patient groups’ absence propensities and of their initial physicians’ practice styles. In addition they absorb some random variation in absenteeism, implying that groups which by chance happened to have high absence propensity in the pre-switch period are attributed high group-fixed effects. On average, we will expect groups which by chance had high absence levels in the pre-switch period to experience a decline in the post-switch period, ceteris paribus, since it is unlikely that the chance-generated high absenteeism will repeat itself. This will generate a pattern of regression toward the mean in the estimated switch-effects; i.e., the estimated switch effects will tend to be smaller (or more negative) the higher was the initially estimated group-fixed effect. This argument applies to both control and treatment groups. However, in the treatment groups there will be an additional source of regression toward the mean, namely that patient groups who used to have high absenteeism due to a particularly lenient doctor before the switch will tend to be subject to a less lenient physician after the switch (and vice versa). Hence, genuine physician influences will show up in the form of a stronger regression-toward-the-mean-tendency in the treatment groups than in the control groups. Fig. 6 shows cross plots of the estimated physician switch effects against the (initial) group fixed effects. There is indeed a much stronger regression-toward-the-mean-tendency in the treatment groups than in the control groups. Based on the regression lines displayed in Fig. 6, we find that while one additional monthly absence day attributed to the group-fixed effects imply a negative switch effect of 0.37 days in the control group, it implies a negative switch effect of 0.56 days in the treatment group. The difference is statistically significant at all conventional levels (P = 0.003). As explained in Section 2, absence certificates can be issued by any authorized physician (not only the family doctor), and Norwegian workers are also free to select a new family doctor if they
8 It may be noted from Figs. 5 and 6 that there is one relatively «extreme» (negative) physician shift effect. If this “outlier” is dropped from the analysis, the estimated treatment effect on the standard deviation is reduced from 26% to 20%.
are not satisfied with the one they have. It is thus important to bear in mind that the physician-impacts identified in our analysis do not represent the pair-wise differences in the physicians’ strictness/leniency per se, but the differences in the influence they manage to exert on worker-absenteeism, given that employees are free to choose another doctor. Seen from this perspective, we would expect the switch-effects identified in our analysis to represent a lower bound on the true differences in practice styles between physicians. The easier it is for a patient to find a new doctor, the more difficult it will be for a physician to influence the patients absence behavior, as employees with, say, a greater propensity to claim sickleave more easily can locate an accommodative physician. Hence, we would expect that high competition between GP’s in a geographical area crowds out their gatekeeping capabilities, and thus renders the switch effects smaller. Since we have data on local variations in the overall GP capacity (see Table 2), we can to some extent test this hypothesis. To do this, we divide the estimated switcheffects into two groups based on local GP capacity; distinguished by whether the vacant GP-slots constitute more or less than 10% of the overall capacity. We then find that the estimated GP influence indeed is much larger in areas where doctors are in short-supply (139 groups) than in areas with larger vacant capacity (189 groups); see Table 4. While the standard deviation of the estimated switch effects is 51% larger for treatments than for controls in the former group it is only 10% larger in the latter group.9 Given that we have access to some physician-characteristics, such as gender, age, education, and the desired (max) number of patients, we can also investigate the extent to which our estimated switch-effects vary with the changes in physician characteristics. In particular, we could perhaps expect that physicians with a relatively large number of desired patients, relative to the actual list-length, would be more willing to obey the patients’ demands than physicians whose patient lists are fully booked. We have not been able to identify any robust patterns that confirm or reject this hypothesis, however. A main problem with the empirical strategy seems to be that there is little observed variation in the variables of interest. The analysis presented so far has been based on data for individuals who were in employment from one year before to two years after genuine/placebo physician switches. As pointed out above, the employment condition may bias our results if physicians also have an impact on their patients’ employment propensity. To check this, we have also estimated the model on an alternative sample of employees, where we only condition on employment at the time of matching; i.e., one year before the physician switches. This approach is of course also disputable, since, e.g., lenient physicians who contribute to raise non-employment among their patients will be observed with low absenteeism simply because they have removed these patients from the risk-set. It may nevertheless provide information on how important the employment condition is for our results. As it turns out, the results are not very sensitive to the employment condition. We still find a considerable pattern of excess absence changes in the treatment groups. The standard deviation for the treatment groups’ switch effects is 18% larger than for the control groups (P = 0.0029).10
9 Unsurprisingly, we also find that the actual fraction of patient-driven physician shifts in the period after the exogenous switch is much larger in areas with ample vacant capacity (not shown). 10 In a previous (Discussion Paper) version of this paper (Markussen et al., 2013), we have also applied a more comprehensive outcome variable, including both sickpay and longer-term disability benefits. The standard deviation for the treatment groups’ shift effects was then 25% larger than for the control groups (P = 0.002).
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
1 Physician shift effect -1 -.5 0 .5 -1.5 -2 -2.5
-2.5
-2
-1.5
Physician shift effect -1 -.5 0 .5
1
1.5
1.5
2
(b) Treatment
2
(a) Control
1237
0
.5
1 1.5 2 Group fixed effect
2.5
3
0
.5
1 1.5 2 Group fixed effect
2.5
3
Fig. 6. Cross-plot of estimated physician switch effects against group-fixed effects.
Physicians may fulfill their gatekeeper role by seeking to avoid issuing (unneeded) absence certificates at all and/or by trying to limit the number of absence days in each spell. To assess which of these channels that is most important for the physician switch effects, we examine the pattern of “incidence” and “duration” around the time of the genuine and placebo physician switches. More specifically, we compute the group-specific changes from the 12-month period before to the two first 12-month periods after the switch for the following two statistics: (i) the fraction of employed
clients that obtain any absence certificate at all, and (ii) the average number of absence days for the absent clients. The results of this exercise are presented in Table 5. For incidence, we find no noticeable differences between the development in the control and the treatment groups, either in the mean change or in the standard deviation. For the number of days per absent client, we also find almost no difference between treatment and control groups in the average change. The variation in these changes, however, was significantly larger for the treatment groups. Hence,
Table 4 Main results by the new physician’s competitive situation (P-values in parentheses). Low competition between physicians
High competition between physicians
0.0024 (P = 0.9631) 1.088 (P = 0.3211)
−0.0004 (P = 0.9916) 0.963 (P = 0.6012)
0.1290 0.1423 0.0133 (P = 0.7705)
0.0903 0.1154 0.0251 (P = 0.5121)
0.2974 0.4478 1.5054*** (P = 0.0000)
0.3530 0.3903 1.1056 (P = 0.1696)
/ n Mean absolute deviation of switch effects control group Mean ˆ m,m − Mean(ˆ m,m ) |m =
0.2519
0.2774
Mean absolute deviation of switch effects treatment group
0.3298
0.3062
Difference in mean absolute deviations between treatment and control group
0.0779*** (P = 0.0075)
0.02884 (P = 0.2240)
Number of patients Number of physicians
42,652 417
55,538 567
I. Group-fixed effects (˛m ) Difference in mean between treatment and control group Ratio of standard deviations treatment and control group
II. Physician switch effects ( m,n ) Mean switch effects control group (Mean(ˆ m,m )) Mean switch effects treatment group (Mean(ˆ m,n )|m = / n) Difference in mean between treatment and control group Standard deviation of switch effects control group (St.Dev(ˆ m,m )) Standard deviation of switch effects treatment group (St.Dev(ˆ m,n )|m = / n) Ratio of standard deviations treatment and control group
Mean ˆ m,n − Mean(ˆ m,n ) |m = / n
Note: Low (high) competition is defined as less (more) than 10% vacant patient slots. Significant at the 10 percent level. ** Significant at the 5 percent level. *** Significant at the 1 percent level. *
1238
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
Table 5 Changes in average sick-leave incidence and duration in control and treatment groups from the 12-month period prior to the switch (P-values in parentheses).
I. Incidence (fraction of employees with absence spell) Mean change in control groups Mean change in treatment groups Differences in mean (treatment-control) Standard deviation of change in control groups Standard deviation of change in treatment groups Ratio of standard deviations (treatment/control)
II. Duration (number of absence days) Mean change in control groups Mean change in treatment groups Differences in mean (treatment-control) Standard deviation of change in control groups Standard deviation of change in treatment groups Ratio of standard deviations (treatment/control) *
. . .to month 1–12 after the switch
. . .to month 13–24 after the switch
0.0031 0.0045 0.0014 (P = 0.6701)
0.0059 0.0063 0.0004 (P = 0.9236)
0.0408 0.0408 1.0000 (P = 0.9800)
0.0450 0.0447 0.9933 (P = 0.9078)
2.2013 2.3419 0.1406 (P = 0.8618)
6.9396 8.0444 1.1048 (P = 0.2498)
9.5337 11.0745 1.1616*** (P = 0069)
11.1053 13.3582 1.2029*** (P = 0.0009)
Significant at the 10 percent level. Significant at the 5 percent level. *** Significant at the 1 percent level.
1 .5 0 -.5
Markussen et al. (2011) computed physician-strictness indicators for all family doctors in Norway, based on their sickness absence certification propensity after controlling for the observed characteristics of all their employed clients. Even though they used a multivariate hazard rate model, the identification of physicianstrictness was almost entirely based on the cross-sectional variation in absence-certification practices. In contrast, our identification strategy relies entirely on the longitudinal variation in certification practices as one physician replaces another one for a fixed group of individuals. It is thus of some interest to examine the degree of concordance between these two strategies for identifying essentially the same thing; namely the physicians’ strictness with respect to the certification of health related social insurance benefits. We would not expect a perfect correlation: Not only have we used different statistical approaches and identification strategies, but the method used in the present paper – though highly robust with respect to the existence of unobserved patient sorting – also suffers from a significant attenuation problem. Given our use of group-fixed effects, we can obviously not offer strictness indicators for each physician; we can only offer estimates for the differences in strictness between the pairs of outgoing and incoming physicians. However, based on the set of strictness indicators computed for each physician in Markussen et al. (2011), it is possible to compute similar difference-indicators for the same physician-pairs. We have thus accessed the estimated strictnessindicators computed in Markussen et al. (2011) for each of the physician-pairs appearing in the present dataset (177 pairs in total), and computed the differences between the two doctors involved in each switch. A comparison between pair-wise differences-in-strictness estimates reported in the present paper and those obtained by Markussen et al. (2011) is presented in Fig. 7. There is obviously
Physician differences estimated from exogenous switches
4. Comparison with previous findings
-1
based on these findings, it appears that the differences in strictness/leniency between physicians do not stem from differences in the willingness to issue absence certificates, but rather from differences in the number of prescribed absence days for each absentee.
1.5
**
-.04
-.02
0
.02
.04
Pairwise differences between physicians (from Markussen et al. 2011)
Fig. 7. Comparison of pair-wise differences-in-strictness estimates reported in this paper with corresponding estimates computed from the results in Markussen et al. (2011).
a strong positive correlation between these two estimators. The correlation coefficient is 0.49 (P = 0.0000). The rankings of the two physicians involved in each switch are concordant in 66% of the cases. This is far from perfect. However, given the extremely different ways in which these estimators have been obtained, we interpret the high correlation as a confirmation that there really are significant differences in the physicians’ fulfillment of their gatekeeper role. 5. Conclusion General Practitioners have been given an important costcontaining role in many welfare states. They certify sick leaves, disability benefits, and prescription drugs, and refer patients to specialist treatment. Previous research has cast doubt on whether GPs are at all able to act as gatekeepers, particularly in relation to sickness absence. We study this using Norwegian register data that makes it possible to examine groups of workers moving in
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
bulk from one GP to another, switches that can be interpreted as exogenous changes of GP. We find that physicians do differ: The same employees have substantially and significantly different sick-pay propensities when connected to two different GPs. While physicians have limited influence on absence incidence, they have large influence on the total number of sick-pay days. This suggest that it may be difficult for a family doctor to outright reject a client’s request for an absence certificate, but that there is considerable more scope for discrepancy in the determination of the number of days and/or in the degree of absence. We take this as evidence that gatekeeping is possible to some extent. The current Norwegian policy strategy of focusing on the GPs’ sick-listing practices in order to reduce worker absenteeism may thus have some merit – provided, of course, that it succeeds in changing the most lenient GPs’ behavior. However, we also find that family doctors have less influence over their patients’ absence behavior in areas where it is easy to find a new/additional doctor. Hence, the gatekeeper role of family doctors may to some extent be undermined by easy access to other physicians. From a methodological point of view, the finding that physicians have a significant impact on their clients’ use of social insurance also suggest that the physicians’ observed “practice styles” may constitute a potentially useful instrument for their clients’ social insurance claims, making it possible to identify causal effects of such claims on, e.g., subsequent employment and earnings paths. References Blomqvist, Å., 1991. The doctor as double agent: information asymmetry, health insurance, and medical care. Journal of Health Economics 10, 411–432. Brekke, K.R., Nuscheler, R., Straume, O.R., 2007. Gatekeeping in health care. Journal of Health Economics 26, 149–170. Carlsen, N., Nyborg, K., 2009. The Gate is Open: Primary Care Physicians as Social Security Gatekeepers. Memorandum 07/2009. Department of Economics, University of Oslo, Oslo.
1239
Duggan, M., 2005. Do new prescription drugs pay for themselves? The case of second-generation antipsychotics. Journal of Health Economics 24, 1–31. Dusheiko, M., Gravelle, H., Jacobs, R., Smith, P., 2006. The effect of financial incentives on gatekeeping doctors: evidence from a natural experiment. Journal of Health Economics 25, 449–478. Englund, L., Svärdsudd, K., 2000. Sick-listing habits among general practitioners in a Swedish county. Scandinavian Journal of Primary Health Care 18, 81–86. French, E., Song, J., 2009. The Effect of Disability Insurance Receipt on Labor Supply. Working Paper. Federal Reserve Bank of Chicago. Grytten, J., Sørensen, R., 2003. Practice variation and physician-specific effects. Journal of Health Economics 22, 403–418. Maestas, N., Mullen, K., Strand, A., 2011. Does Disability Insurance Receipt Discourage Work? Using Examiner Assignments to Estimate Causal Effects of SSDI Receipt. Rand Working Paper, WR-853-2. Markussen, S., 2012. The effects of sick-leaves on earnings. Journal of Population Economics 25, 1287–1306. Markussen, S., Røed, K., Røgeberg, O.J., Gaure, S., 2011. The anatomy of absenteeism. Journal of Health Economics 30, 277–292. Markussen, S., Mykletun, A., Røed, K., 2012. The case for presenteeism—evidence from Norway’s Sickness Insurance Program. Journal of Public Economics 96, 959–972. Markussen, S., Røed, K., Røgeberg, O.J., 2013. The Changing of the Guards: Can Physicians Contain Social Insurance Costs? IZA Discussion Paper No. 7122. Nilsen, S., Werner, E.L., Maeland, S., Eriksen, H.R., Magnussen, L.H., 2011. Considerations made by the general practitioner when dealing with sick-listing of patients suffering from subjective and composite health complains. Scandinavian Journal of Primary Health Care 29, 7–12. O.E.C, D., 2010. Sickness Disability and Work: Breaking the Barriers—Key Trends and Outcomes in Sickness and Disability. OECD, Paris. O.E.C, D., 2012. OECD Economic Surveys Norway. OECD, Paris. Reiso, H., Nygård, J.F., Brage, S., Gulbrandsen, P., Tellnes, G., 2000. Work ability assessed by patients and their GPs in new episodes of sickness certification. Family Practice 17, 139–144. Scott, A., 2000. Economics of general practice. Handbook of Health Economics, vol. 1. Elsevier science, Amsterdam (Chapter 22). Stone, D., 1984. The Disabled State. Temple University Press, Philadelphia. Wilkin, D., 1992. Patterns of referral: explaining variation. In: Roland, M., Coulter, A. (Eds.), Hospital Referrals, Oxford General Practice Series 22. Oxford University Press, Oxford.
Contents lists available at ScienceDirect
Journal of Health Economics journal homepage: www.elsevier.com/locate/econbase
The changing of the guards Can family doctors contain worker absenteeism?夽 Simen Markussen, Knut Røed ∗ , Ole Røgeberg Ragnar Frisch Centre for Economic Research, Norway
a r t i c l e
i n f o
Article history: Received 31 August 2012 Received in revised form 8 October 2013 Accepted 9 October 2013 Available online 24 October 2013 JEL classification: H55 I13
a b s t r a c t Using administrative data from Norway, we examine the extent to which family doctors influence their clients’ propensity to claim sick-pay. The analysis exploits exogenous switches of family doctors occurring when physicians quit, retire, or for other reasons sell their patient lists. We find that family doctors have significant influence on their clients’ absence behavior, particularly on absence duration. Their influence is stronger in geographical areas with weaker competition between physicians. We conclude that it is possible for family doctors to contain sick-pay expenditures to some extent, and that there is a considerable variation in the way they perform this task. © 2013 Elsevier B.V. All rights reserved.
Keywords: Sick-pay GP practice styles Absence certification Gatekeepers Family doctors
1. Introduction Publicly mandated sick-pay is an important part of the social safety net in many countries. Given the relatively high replacement ratios in these programs, they entail the risk of excessive use and shirking. To counter this risk, paid sickness spells require certification by a physician in virtually all countries (OECD, 2010, p. 139). This has remained the case, despite a trend in OECD countries toward raising the powers of the benefit-granting institutions in relation to long-term disability benefit claims. The GPs thus remain important gatekeepers of the welfare states’ public purses. The rationale behind this policy is that physicians can, better than others, verify the presence and severity of a medical condition, and thus decide whether or not a person is eligible for sick-pay. At the same time, however, these physicians operate businesses in a competitive environment, and may be tempted to attract and retain patients by giving easier access to services and treatments – such
夽 This research has received financial support from the Norwegian Research Council through the program “Sickness, absence, work and health” (grant nos. 187924 and 201416). Data made available by Statistics Norway have been essential. Thanks to seminar participants in Bergen and Oslo, and to two anonymous referees for a number of useful comments and suggestions. ∗ Corresponding author at: Ragnar Frisch Centre for Economic Research, Gaustadalléen 21, 0349 Oslo, Norway. Tel.: +47 22958813. E-mail address: [email protected] (K. Røed). 0167-6296/$ – see front matter © 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.jhealeco.2013.10.005
as sickness benefits – that are desired by (potential) clients. In this way, market pressures and the physicians’ own economic interests may erode their gatekeeper role. An additional role conflict may arise if the physician interprets his/her duty as being an advocate for their patient; see, e.g., Reiso et al. (2000). There is a theoretical literature discussing the efficacy of physicians as gatekeepers in relation to absenteeism, as well as to costly specialist treatments and drugs prescription. Focus has been on the information advantage physicians have over patients regarding the appropriate medical treatment (e.g., Scott, 2000; Brekke et al., 2007), the information advantage patients have over doctors regarding private health (e.g., Stone, 1984), and the tension in the physician’s role as both advocate of the patient and health-assessor on behalf of the insurance provider (e.g., Blomqvist, 1991). On the basis of a theoretical model and qualitative interview data from a sample of general practitioners, Carlsen and Nyborg (2009) argue that there are neither theoretical nor empirical grounds for expecting general practitioners to fulfill the role of effective gate-keeping in relation to sick-pay, and conclude that “the gate is wide open”. Empirical evidence also indicates that it is difficult for physicians to contain sick-pay expenditures. Englund and Svärdsudd (2000) report that Swedish GPs find it hard to overrule their patients’ requests for sick-leave certificates. Even in cases where GPs do not see sick-leave as recommendable, they still issue sickness certificates in as much as 87% of the cases where the patient asks for it. Based on focus group interviews of Norwegian GPs,
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
Nilsen et al. (2011) conclude that “trust in the patient’s own story and self-judgment” is deemed crucial, and that sick-leave certificates should be considered “mainly patient-driven”. On the other hand, Markussen et al. (2011) report a large variation in Norwegian GPs’ observed tendency to certify sick-leave, suggesting that the physicians’ priorities or attitudes may be important after all. And empirical evidence related to other aspects of the GPs gatekeeper role also indicates significant variation in strictness. Wilkin (1992), for example, identifies a large variation in the frequency with which physicians refer their patients to specialist treatment, and Grytten and Sørensen (2003) identify a similarly large variation in physicians’ use of (expensive) laboratory tests. Based on a reform in the funding scheme for certain types of elective surgery, Dusheiko et al. (2006) show that GPs tend to have a more restrictive referral practice when money not spent on surgery alternatively can be channeled back to their own practice. A fundamental problem with empirically identifying the influence of physicians on their patients’ behavior is that patients are not allocated randomly to physicians; consequently, differences in observed absence rates across physicians’ patient groups may reflect sorting as well as causality. In particular, patients demanding a (questionable) sickness certificate may seek out an accommodative (“lenient”) physician. While this identification problem can be reduced by controlling for observed patient characteristics, it is typically difficult to ascertain that no unobserved sorting remains. And to our knowledge, no empirical evaluation of family doctors’ influence on their clients’ absence behavior has convincingly solved this selection problem. An important contribution of the present paper is that we examine the influence that family doctors have on their clients’ absence behavior by studying how the absence rates of a patient group shifts when the entire group is moved exogenously from one physician’s list to another. This allows us to observe the absence rates of a large set of patient groups under two different GPs, largely avoiding the selection problem. Our empirical basis is administrative data from Norway that allows us to link all citizens to their family doctors as well as to their individual records on employment and physician-certified sick leaves. Norway is a particularly well suited country for investigating the role of physicians, since everyone in Norway has been assigned a GP who is their primary contact point with the public health care and sickness insurance systems, regardless of whether they have actually consulted the doctor or not. In the main, these GPs are the ones responsible for certifying sick-pay for workers. The system has the additional feature that when a GP moves to another job/district, retires or closes the practice for other reasons, the list of patients is sold to another GP and the patients are collectively moved to this new GP. Further strengthening the identification strategy is a key principle of the system which states that GPs have no direct opportunities for shaping their patient list: If an exiting GP’s list is split between two GPs, the split is random and performed by a central organization rather than the GPs themselves. The main research question we seek to answer in the present paper is the following: Does a family doctor have a significant influence on the level of sickness absence in his/her group of patients? The more extreme variants of the gate-keeping-does-not-work-andmay-even-be-impossible argument would predict that the level of absenteeism is independent of which physician the patient is referred to, and we would consequently not expect an exogenous switch of GP to cause changes in claim behavior. If, on the other hand, there are substantial and persistent differences between physicians in terms of their “strictness” or “leniency” which are not fully neutralized by “physician shopping”, we would expect an exogenous switch of GP to entail noticeable changes.
1231
Our paper is also motivated by a more methodological issue: If the family doctor has a direct impact on a worker’s benefit claims, this would potentially imply that indicators for physician strictness can be used as exogenous instruments to answer questions about the causal effect of becoming a benefit claimant. This strategy was pursued by Markussen (2012), who investigated the impact of absenteeism on subsequent earnings growth using physician-strictness indicators computed in Markussen et al. (2011) as instrumental variables. Similar identification strategies have also been used by others: Duggan (2005) examined the impact of new/expensive antipsychotic drugs on subsequent health care spending, using psychiatrists’ propensity to prescribe expensive drugs as instrumental variable. French and Song (2009) and Maestas et al. (2011) investigated the impact of disability insurance receipt on subsequent labor supply, using the health examiners’ observed allowance rates as an instrumental variable. Our paper sheds light on the potential power of these empirical approaches by assessing the influence that gate keepers really have on these kinds of decisions when all individual and environmental factors are appropriately controlled for. In particular, we offer a reexamination of the physician-strictness indicators computed by Markussen et al. (2011) as our data to some extent encompass the same group of physicians. The key finding of our paper is that family doctors really do have a significant impact on their patients’ benefit claims. Comparing patient groups who were subject to an exogenous switch of family doctor with similar groups who were not subject to such a switch, we find that the changes in absenteeism – positive or negative – were significantly larger for the groups subject to a switch. Measured by the standard deviation in the groups’ sick-leave changes over the next two years, we find that the switches induced a 26% increase in absence variability. The family doctors exert much more influence on the duration of sick-leaves than on their occurrence. We also find that the influence of the family doctor is larger in areas where there is little competition between doctors, and thus a smaller pool of doctors that patients can choose from. Comparing our results to those reported by Markussen et al. (2011) for the same group of physicians, we find a large degree of concordance. Given the completely different identification strategies used, we interpret this as further evidence that there really are significant differences between physicians in the way they fulfill their gatekeeper role. Family doctors do indeed have some powers to contain social insurance costs, but they use them to varying degrees. 2. Institutional setting and data In Norway, workers receive 100% wage compensation from the first day of sickness absence and for up to one year (up to an income ceiling of approximately 500,000 NOK p.a. (2012)). The first 16 days are paid for by the employer, after which the costs are covered by general (payroll) taxation. A sick leave certificate issued by a physician is normally required for all spells lasting more than three days.1 On a typical working day, 6–7% of Norwegian employees are absent from work due to sickness, and almost 90% of these absences are certified by a physician. OECD has repeatedly pointed out that Norway’s rate of sick-leave is among the highest in the world, and recommended that independent checks on the GPs’ sickness assessments are performed more systematically; see, e.g., OECD (2012). Reducing absenteeism has also been made a major priority in a tripartite agreement between the employer associations, the employee associations, and the state (the inclusive workplace
1 Some firms have agreed to accept self-reported sickness claims for up to eight days.
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
2 The supply of physicians with vacant list-slots varies significantly from region to region, particularly between rural and urban areas. In some parts of the country, patients may in practice have no choice other than accepting their assigned physician. 3 It is common to use graded absence certificates in Norway, implying that the worker is only partially absent from the workplace, see Markussen et al. (2012) for details. If, for example, a worker has a 50% absence spell for 10 days, we count this as 5 days of absence.
0
Number of switches 15 10 20 5
(a) Number of switches per month
July 2002
July 2003
July 2004
July 2005
July 2006
Percent 10
15
(b) Size-distribution of the switches
5
agreement), and in the approved strategy documents, physicians are attributed a key role. New comprehensive guidelines have been developed for the physicians’ sick-listing practices and for their dialog with employers and the social insurance administration. Norway has, since May 2001, practiced a family (panel) doctor system, whereby each citizen is assigned a single GP who receives a capitation fee from the social security administration. Sickness absence certificates can in principle be issued by any authorized physician, but will normally be issued by the family doctor, except in emergency cases and when the patient is hospitalized or subject to intensive specialist treatment. Norwegian workers are free to choose their family doctor insofar as the desired doctor has vacant patient slots, but unless they make an active choice, they will be assigned a “default” GP, based on their residential address. When a physician retires or moves to a new job (or a new location), the whole patient list will be sold to a new GP, but – as we return to below – patients may of course self-select to other physicians at these occasions.2 The panel doctor determines a maximum number of patients (with an upper limit of 2500). The data we use in the present paper are collected from Norwegian administrative registers made available to us by Statistics Norway. They cover the complete population during the period from May 2001 through September 2008, with individual level (encrypted) panel data information concerning patient-GPlinkages, employment spells, physician-certified absence spells, demographic information (gender, age, nationality), and educational attainment. Using these data, we construct a monthly outcome measure indicating – for each employee – the number of physician-certified sick-pay days, adjusted for grading.3 Note that it is the timing of the certification that determines to which month the recorded absence days are attributed, and not the timing of the absence itself. Longer absence spells will typically consist of several certification periods, and hence be distributed across a number of months (the average length of a certification period is approximately 19 days). The population we are interested in consists of workers who are forced to change family doctor because their original GP sells the whole practice to a new doctor. The resultant new patientGP matches are exogenous with respect to each patient’s behavior and characteristics, as a core design principle of the system is that GPs have no direct influence over the composition of their list. In particular, it is not possible for a new doctor to “cream skim” by taking over only a selected subset of the previous doctor’s list. It is the municipality, not the doctor that puts vacant panel doctor practices on the market and allocates them to new physicians. The incoming physicians must then negotiate prices with the outgoing ones (if they do not agree, the prices are determined by a committee under the auspices of the national physicians’ association). The basic idea behind our empirical approach is to examine the frequency of sick-pay claims before and after physician switches in order to assess whether there are changes in sick-pay patterns that are attributable to these switches. Since the frequency of sickpay claims obviously changes over time for a number of reasons (changes in employment status, cyclical and seasonal fluctuations, time-trends, aging, pure randomness), we need a control group of
0
1232
0
200
400 Number of employed clients
600
800
Fig. 1. The number of exogenous switches, their timing and the number of employed patients involved. Note: The months indicated in panel (a) are the last months before the switch (e.g., a July-bar indicates the shifts occurring between July and August).
physicians with stable patient lists to represent the counterfactual case of remaining with the same GP. A potential problem with this approach is that some GP switches may be anticipated. If, for instance, a doctor is approaching retirement, the patient may well be aware of this and start looking for a new GP even before the switch occurs. This would mean that the transferred list contains only a selected sub-group of original list members. To address this concern, we will interpret the client-list that applied 12 months before the actual switch as defining the “treated” population. We also disregard any subsequent physician switches, since these may occur as an endogenous response to the behavior of the new physician. This implies that we use a completely balanced panel to study the frequency of sick-pay claims before and after the time of the physician switches. As we return to below, this creates a sort of attenuation bias when we evaluate the influence of the new physician relative to the old one. A second sample restriction is to include only persons who were employed from 12 months prior to the switch to 24 months after, since employment is a precondition for sick-pay. The three-year employment condition may be disputable if family doctors’ certification practices also affect employment, as indicated by the results reported by Markussen et al. (2012). Within the relatively short time-window used in the present analysis, however, we consider this impact to be negligible. We nevertheless relax the three-year employment restriction in a robustness analysis reported below. Formally, we denote the set of patients moving collectively from physician m to physician n at time s and observed in month t by .4 We can refer to m as the outgoing (patient-supplying) GP, {Ps }m,n t and n as the incoming (patient-receiving) GP. The treatment sample is constructed by identifying all GP induced switches that involved at least 95% of the listed patients, and took place in the period from May 2002 through September 2006.5 This time period is selected to ensure that we can follow the list-members absence behavior from one year before to two years after the switch. The treatment group is defined as all employees who were on an outgoing physician’s list exactly one year before a switch. Fig. 1, panel (a) shows the frequency of exogenous physician switches included in our analysis by calendar month. Although the
4 Note that physician shifts in Norway always occur at the border between two months. 5 In addition, we require that the switch moved at least 150 persons (and at least 50 employed persons).
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
328 98,190 38.8 42.6 13.6 10.2 3.79
switches are spread out over time, there is a disproportionally large fraction of them occurring before/after the summer and Christmas holidays. Panel (b) shows the size distribution of the employed list-members actually included in our analysis. A typical switch involves around 200–400 persons satisfying our employment criterion. Our focus on complete list transfers also ensures that most of the patient-receiving physicians (92.3%) are really “new”, in the sense that they did not have any patients at all in the municipality prior to the exogenous switch. For each treated patient group, we select a corresponding (employed) control patient group belonging to a physician who did not quit. Formally, we denote the control group as {Ps }m,m , t where s refers to the timing of the physician switch occurring in the treatment group for which the group is a control. Hence, we may think of the control groups as experiencing “placebo” physician switches (with n = m) at exactly the same time as the treatment groups. The purpose of the “placebo” switches in this context is to ensure that the “time-dimension” in the before-after-comparisons is exactly the same in the treatment and control groups, both in relation to calendar time (generated by, e.g., seasonal and cyclical fluctuations) and in relation to the three year period for which we have conditioned on employment. Control groups are matched to the treatment groups one year before the physician switches occur in the treatment groups. They are selected among the patient groups who were not subject to an exogenous physician switch during our observation window. The matching procedure is designed to ensure as similar patient composition as possible in treatment and control groups, particularly in terms of past absence behavior. More specifically, for each treatment group we identify all patient groups of similar sizes (±100 patients), whose employed members had absence rates that did not deviate more than 0.5 percentage point from those of the treatment group in both the year just prior to matching and in the year before that (this is done to ensure both similarity in both absence levels and trends). Among these, we select the group that is closest to the treatment group in terms of average absence rate in the two-year period prior to matching.6 Finally, we ensure that each treatmentcontrol group pair is of exactly the same size by randomly removing patients from the larger of the two. Table 1 provides some summary statistics for the treated and non-treated workers, whereas Fig. 2 compares their absence behavior for as much as five years backwards in time; i.e., prior to the physician switch. In total, we have data for 328 treatment and control groups, respectively. On average, each group consists of 299 employed patients (after the groups have been shaved to ensure equal sizes). Although we have not used individual characteristics
6 We have chosen to use past absence behavior as the key matching criterion here since we consider it important to ensure similar baseline absence levels/trends in treatment and control groups. Propensity score matching is not appropriate in this case, since it turns out that patient group characteristics do not contribute much to explaining the exogenous physician switches.
.04
328 98,190 39.2 44.6 13.8 10.4 3.72
Absence rate .02 .03
Treatment group
.01
Control group
0
Number of groups Number of persons Mean age Percent females Mean years of schooling Percent non-natives Local unemployment rate at the time of matching
.05
Table 1 Descriptive statistics patients groups.
1233
-5 years
-4 years
-3 years
-2 years
-1 year
Years before physician switch
Control
Treatment
Fig. 2. Absenteeism 1–5 years prior to the time of exogenous physician switch in treatment and control groups. Note: Absence data for the pre-matching period cover long-term absence (>16 days) only; hence, they are not directly comparable to the data used in the outcome period.
(apart from past absenteeism) as matching criteria, the numbers reported in Table 1 indicate that the treatment and control groups are similar, both in terms of observed characteristics like age, gender, education, and nationality, and in terms of local labor market conditions. Their past absence behavior also follow a similar pattern; conf. Fig. 2.7 As pointed out above, some patients in any group {Ps }m,n may t be exposed to GPs other than m and n, since individuals are free to choose their GP at any time. The sampling scheme ensures that we follow all patients associated with a physician m from one year before the exogenous (genuine or placebo) switch, and interpret them as treated by this particular switch regardless of any subsequent transitions to other doctors. This implies that the estimated effects of the switches can be given a sort of intention-to-treat interpretation. What we can hope to identify from our analysis is thus not the impacts of moving from physician m to physician n, but rather the impact of being exogenously split from physician m and offered physician n as a new default choice. The impacts will include any effects that the exogenous switch has on subsequent patient-initiated switches. Table 2 shows some descriptive statistics for the physicians involved in our analysis. Unsurprisingly, there are some significant differences between the three groups of incoming, outgoing, and staying (control) physicians. Incoming physicians are generally younger and less educated than outgoing and staying physicians. And, although the average age of outgoing and staying physicians are similar, the underlying age distributions are different. In particular, since GP-practices are typically sold early in careers (as young GPs move on to specialization or more attractive jobs) or late (as they retire), the outgoing physicians have a bimodal age distribution with spikes around the retirement age and around the late thirties (not shown). It may also be noted from Table 2 that the distributions of the municipalities’ overall physician coverage (total patient slots relative to population sizes) were very similar for the treatment and control physicians. Fig. 3 illustrates the degree of “non-compliance” for the treatment and the control groups; e.g., the fractions of clients that at
7 The apparently rising absence trends revealed by Fig. 2 are primarily caused by the employment condition imposed from the time of matching and onwards. As we move backwards in time, an increasing proportion of the populations are likely to be non-employed, and thus not under risk of being absent.
1234
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
Table 2 Descriptive statistics physicians (standard deviations in parentheses).
Percent female Age Percent with specialist education Average GP capacity in municipality (#list-slots/#inhabitants) Population-physician-ratio in municipality (#inhabitants/#physicians) Number (N)
different points in time have self-selected away from their initial or assigned physician. Not surprisingly, patient-initiated switching is much more common in the treated patient groups–especially just after the exogenous switch. This illustrates the point that a new doctor may be under stronger competitive pressure than an established one. 3. Empirical analysis As a patient group is exogenously moved from one GP to another, there are at least four different mechanisms that may induce changes in the group-members’ sick-leave behavior. First, if the new GP has a different practice style than the old one, this may induce an upwards or downwards shift in the group’s overall benefit claim propensity. With the term “practice style”, we primarily have in mind the physicians’ understanding of their gatekeeper role and their predisposition toward using absence certificates as part of a medical treatment. We realize, however, that differences in the use of absence certificates may also arise from differences in other aspects their clinical practices and patient management that affect their patients’ health. Second, there may be a non-familiarity-effect, resulting from the fact that the new GP has less information and less personal ties to the patients. This effect is also of unknown sign; the new doctor may be more lenient than the old one if it takes time to notice patient-specific patterns indicative of shirking, and stricter if a settled GP-patient relationship introduces a familiarity/friendship that makes the GP more inclined to adhere to the patient’ wishes. Third, it is probable that a new doctor is subject to stronger competitive pressures than more established ones, e.g., as a result of less loyal list-members; confer the large fraction of patients who self-select out the list after a takeover
.6
Fraction remaining with initial physician .7 .8 .9
1
Fraction of initial patient population remaining with the same (or asigned) physician
1 year before
Switch
1 year after Control
Treatment
Fig. 3. Fraction of initial patient population (12 months before the switch) that stays with initial or new assigned physician.
Control group Staying phys.
Treatment group Outgoing phys.
Treatment group Incoming phys.
29.6 48.2 (9.3) 44.5 1.11 (0.05) 1164 (138) 328
32.6 47.3 (12.2) 26.2 1.11 (0.06) 1140 (143) 328
41.2 36.0 (7.7) 3.7 1.11 (0.06) 1139 (143) 328
(Fig. 2). This effect may be particularly prominent just after a switch. And finally, a shift of physician may in some cases disrupt ongoing medical treatments and/or cooperative efforts with employer and employee to facilitate work-resumption, most likely contributing to increased absenteeism in the short run. As noted above, absence certificates can be issued by any authorized physician, making it possible for dissatisfied patients to circumvent their own family doctor. Also, in cases requiring consultations outside office hours or emergency room visits, the family doctor will normally not be available. These mechanisms are likely to imply that differences observed in absence patterns in relation to a forced shift of family doctor are smaller than what the differences in the two physicians’ practice styles would imply. As a measure for differences in practice styles between physician pairs, we thus expect our intention-to-treat estimators to be considerably attenuated. Taken together, it is not clear how exogenous switch of physicians affect the average level of sick-pay certification. Fig. 4, panel (a), shows the number of certified sick-pay days per month in the treatment and control groups before and after the (genuine or placebo) switch. There are no clear signs of a treatment effect on average absenteeism, with a possible exception for a slightly higher absence propensity just around the time of the exogenous physician switch. Somewhat surprisingly, this effect seems to be caused by a decline in the control group’s absence, rather than by an increase in the treatment group’s absence. This phenomenon arises, however, from the fact that a disproportionally large fraction of physician switches occur in holiday periods (see Fig. 1), implying that the seasonal pattern is correlated with the timing of the switches. In panel (b), we report the time pattern in absence rates that remains after having controlled for seasonality (by means of 12 calendar month dummy variables) and individual patient characteristics (dummy variables for gender, age, education, and immigration status) within a linear regression framework. From these graphs, there seems to be a pattern of increased absenteeism in the treatment groups relative to the control groups in the first few months after the physician switch. Note that the apparently rising absence trends in both groups are primarily related to the way we have constructed the data in the form of a three-year employment condition (from one year before to two years after the physician switch). As we move toward the end of this conditioning period, its implications for absence behavior become weaker and weaker, implying that long-term absenteeism becomes more and more likely. The appropriate test for the existence and strength of physician influence on sick-pay claims is, however, not the extent to which physician switches affect the average absence rate, but the extent to which they induce excess changes (positive or negative) in sickpay patterns. When the new GP is more lenient than the old one, the frequency of benefit claims will rise, ceteris paribus; when the new GP is stricter than the old one, it will decline. In both cases,
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
(b) Controlled for ind.char. and seasonal variation
1.7 1.6 1.5 1.4
1.4
1.5
1.6
1.7
1.8
1.8
1.9
1.9
(a) Observed
12 months before
1235
Switch
12 months after
12 months before
Control
Switch
12 months after
Treatment
Fig. 4. Average number of sick-pay days per month before and after physician switch.
it will induce a change over and above what can be expected to occur for other reasons (randomness, business cycles, time-trends, etc.). Hence, to evaluate the physicians’ overall influence, we need a statistical model that allows the impacts on the affected groups’ absenteeism to take different directions for different groups. Let yimt be the number of physician-certified sick-pay days prescribed to individual i, belonging to group m in month t. To identify the effects of interest, we formulate the linear model yimt = ˛m + m,n I(t ≥ s) + Xit ˇ + Mt + εit .
physician-switch effects is 26% larger for the treatment groups than for the control groups. This is further illustrated in Fig. 5, where we have shown box-plots for the two estimated physician-switch distributions. The distribution is clearly wider among treatment
Table 3 Main results (P-values in parentheses). I. Group-fixed effects (˛m ) Difference in mean between treatment and control group
(1)
Here, ˛m is a patient-group fixed effect, I(.) is an indicator variable taking the value 1 after the (genuine or placebo) physician switch has occurred, m,n is a patient-group-specific change in the absence level following this switch, Xit is the vector of individual control variables referred to above, i.e., dummies for gender, age (32 categories), education (10 categories for level and 10 categories for type), and immigration status (6 categories, based on origin country), and Mt is a vector of seasonal dummy variables (one for each month of the year). Our main interest lies in comparing the estimated changes associated with genuine physician switches with those associated with placebo switches; i.e., in comparing the distribution of m,n given that n = / m, with the corresponding distribution given that n = m. The estimation results are presented in Table 3 and in Figs. 5 and 6. The first part of Table 3 shows that there were no significant differences between the control and treatment groups prior to the physician switches, either in the mean or in the spread of the group-fixed effects. The second part of the table presents the estimated physician switch effects. They indicate that there was a small, but statistically insignificant, increase in average absenteeism in the treatment group relative to the control group. The most conspicuous pattern revealed in Table 3, however, is the highly significant rise in the variation across the treatment groups’ absence behavior. The standard deviation of the estimated
Ratio of standard deviations treatment and control group
II. Physician switch effects ( m,n ) Mean switch effects control group (Mean(ˆ m,m )) Mean switch effects treatment group (Mean(ˆ m,n )|m = / n) Difference in mean between treatment and control group Standard deviation of switch effects control group (St.Dev(ˆ m,m )) Standard deviation of switch effects treatment group (St.Dev(ˆ m,n )|m = / n) Ratio of standard deviations treatment and control group
0.0008 (P = 0.9808) 1.02 (P = 0.7248)
0.1067 0.1268 0.0201 (P = 0.4927) 0.3307 0.4151 1.256*** (P = 0.0000)
Mean deviation of switch effects control group absolute
0.2666
Mean deviation of switch effects absolute treatment group
0.3162
Difference in mean absolute deviations between treatment and control group
0.0496***
Mean ˆ m,m − Mean(ˆ m,m ) |m = / n Mean ˆ m,n − Mean(ˆ m,n ) |m = / n
(P = 0.0070) Number of patients Number of physicians *
Significant at the 10 percent level. Significant at the 5 percent level. *** Significant at the 1 percent level.
**
98,190 984
1236
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239 Treatment
-2
-1
0
1
2
Control
Fig. 5. Box plots for the distributions of estimated physician switch effects for the control and treatment groups. Note: The shaded boxes show the first, second, and third quartiles. The upper and lower horizontal lines indicate the maximum and minimum values observed within the inter-quartile ranges (IQR = 1.5 (Q3 − Q1)). The dots indicate observations outside IQR.
than among control groups.8 We have investigated whether the excess variation in the treatment groups is associated with particular medical diagnosis. This does not seem to be the case: The estimated switch-effects follow the same pattern within each of the three major diagnosis groups (musculoskeletal diseases, mental disorders, and “other”). The group-fixed effects (˛m )may be interpreted as combined measures of the patient groups’ absence propensities and of their initial physicians’ practice styles. In addition they absorb some random variation in absenteeism, implying that groups which by chance happened to have high absence propensity in the pre-switch period are attributed high group-fixed effects. On average, we will expect groups which by chance had high absence levels in the pre-switch period to experience a decline in the post-switch period, ceteris paribus, since it is unlikely that the chance-generated high absenteeism will repeat itself. This will generate a pattern of regression toward the mean in the estimated switch-effects; i.e., the estimated switch effects will tend to be smaller (or more negative) the higher was the initially estimated group-fixed effect. This argument applies to both control and treatment groups. However, in the treatment groups there will be an additional source of regression toward the mean, namely that patient groups who used to have high absenteeism due to a particularly lenient doctor before the switch will tend to be subject to a less lenient physician after the switch (and vice versa). Hence, genuine physician influences will show up in the form of a stronger regression-toward-the-mean-tendency in the treatment groups than in the control groups. Fig. 6 shows cross plots of the estimated physician switch effects against the (initial) group fixed effects. There is indeed a much stronger regression-toward-the-mean-tendency in the treatment groups than in the control groups. Based on the regression lines displayed in Fig. 6, we find that while one additional monthly absence day attributed to the group-fixed effects imply a negative switch effect of 0.37 days in the control group, it implies a negative switch effect of 0.56 days in the treatment group. The difference is statistically significant at all conventional levels (P = 0.003). As explained in Section 2, absence certificates can be issued by any authorized physician (not only the family doctor), and Norwegian workers are also free to select a new family doctor if they
8 It may be noted from Figs. 5 and 6 that there is one relatively «extreme» (negative) physician shift effect. If this “outlier” is dropped from the analysis, the estimated treatment effect on the standard deviation is reduced from 26% to 20%.
are not satisfied with the one they have. It is thus important to bear in mind that the physician-impacts identified in our analysis do not represent the pair-wise differences in the physicians’ strictness/leniency per se, but the differences in the influence they manage to exert on worker-absenteeism, given that employees are free to choose another doctor. Seen from this perspective, we would expect the switch-effects identified in our analysis to represent a lower bound on the true differences in practice styles between physicians. The easier it is for a patient to find a new doctor, the more difficult it will be for a physician to influence the patients absence behavior, as employees with, say, a greater propensity to claim sickleave more easily can locate an accommodative physician. Hence, we would expect that high competition between GP’s in a geographical area crowds out their gatekeeping capabilities, and thus renders the switch effects smaller. Since we have data on local variations in the overall GP capacity (see Table 2), we can to some extent test this hypothesis. To do this, we divide the estimated switcheffects into two groups based on local GP capacity; distinguished by whether the vacant GP-slots constitute more or less than 10% of the overall capacity. We then find that the estimated GP influence indeed is much larger in areas where doctors are in short-supply (139 groups) than in areas with larger vacant capacity (189 groups); see Table 4. While the standard deviation of the estimated switch effects is 51% larger for treatments than for controls in the former group it is only 10% larger in the latter group.9 Given that we have access to some physician-characteristics, such as gender, age, education, and the desired (max) number of patients, we can also investigate the extent to which our estimated switch-effects vary with the changes in physician characteristics. In particular, we could perhaps expect that physicians with a relatively large number of desired patients, relative to the actual list-length, would be more willing to obey the patients’ demands than physicians whose patient lists are fully booked. We have not been able to identify any robust patterns that confirm or reject this hypothesis, however. A main problem with the empirical strategy seems to be that there is little observed variation in the variables of interest. The analysis presented so far has been based on data for individuals who were in employment from one year before to two years after genuine/placebo physician switches. As pointed out above, the employment condition may bias our results if physicians also have an impact on their patients’ employment propensity. To check this, we have also estimated the model on an alternative sample of employees, where we only condition on employment at the time of matching; i.e., one year before the physician switches. This approach is of course also disputable, since, e.g., lenient physicians who contribute to raise non-employment among their patients will be observed with low absenteeism simply because they have removed these patients from the risk-set. It may nevertheless provide information on how important the employment condition is for our results. As it turns out, the results are not very sensitive to the employment condition. We still find a considerable pattern of excess absence changes in the treatment groups. The standard deviation for the treatment groups’ switch effects is 18% larger than for the control groups (P = 0.0029).10
9 Unsurprisingly, we also find that the actual fraction of patient-driven physician shifts in the period after the exogenous switch is much larger in areas with ample vacant capacity (not shown). 10 In a previous (Discussion Paper) version of this paper (Markussen et al., 2013), we have also applied a more comprehensive outcome variable, including both sickpay and longer-term disability benefits. The standard deviation for the treatment groups’ shift effects was then 25% larger than for the control groups (P = 0.002).
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
1 Physician shift effect -1 -.5 0 .5 -1.5 -2 -2.5
-2.5
-2
-1.5
Physician shift effect -1 -.5 0 .5
1
1.5
1.5
2
(b) Treatment
2
(a) Control
1237
0
.5
1 1.5 2 Group fixed effect
2.5
3
0
.5
1 1.5 2 Group fixed effect
2.5
3
Fig. 6. Cross-plot of estimated physician switch effects against group-fixed effects.
Physicians may fulfill their gatekeeper role by seeking to avoid issuing (unneeded) absence certificates at all and/or by trying to limit the number of absence days in each spell. To assess which of these channels that is most important for the physician switch effects, we examine the pattern of “incidence” and “duration” around the time of the genuine and placebo physician switches. More specifically, we compute the group-specific changes from the 12-month period before to the two first 12-month periods after the switch for the following two statistics: (i) the fraction of employed
clients that obtain any absence certificate at all, and (ii) the average number of absence days for the absent clients. The results of this exercise are presented in Table 5. For incidence, we find no noticeable differences between the development in the control and the treatment groups, either in the mean change or in the standard deviation. For the number of days per absent client, we also find almost no difference between treatment and control groups in the average change. The variation in these changes, however, was significantly larger for the treatment groups. Hence,
Table 4 Main results by the new physician’s competitive situation (P-values in parentheses). Low competition between physicians
High competition between physicians
0.0024 (P = 0.9631) 1.088 (P = 0.3211)
−0.0004 (P = 0.9916) 0.963 (P = 0.6012)
0.1290 0.1423 0.0133 (P = 0.7705)
0.0903 0.1154 0.0251 (P = 0.5121)
0.2974 0.4478 1.5054*** (P = 0.0000)
0.3530 0.3903 1.1056 (P = 0.1696)
/ n Mean absolute deviation of switch effects control group Mean ˆ m,m − Mean(ˆ m,m ) |m =
0.2519
0.2774
Mean absolute deviation of switch effects treatment group
0.3298
0.3062
Difference in mean absolute deviations between treatment and control group
0.0779*** (P = 0.0075)
0.02884 (P = 0.2240)
Number of patients Number of physicians
42,652 417
55,538 567
I. Group-fixed effects (˛m ) Difference in mean between treatment and control group Ratio of standard deviations treatment and control group
II. Physician switch effects ( m,n ) Mean switch effects control group (Mean(ˆ m,m )) Mean switch effects treatment group (Mean(ˆ m,n )|m = / n) Difference in mean between treatment and control group Standard deviation of switch effects control group (St.Dev(ˆ m,m )) Standard deviation of switch effects treatment group (St.Dev(ˆ m,n )|m = / n) Ratio of standard deviations treatment and control group
Mean ˆ m,n − Mean(ˆ m,n ) |m = / n
Note: Low (high) competition is defined as less (more) than 10% vacant patient slots. Significant at the 10 percent level. ** Significant at the 5 percent level. *** Significant at the 1 percent level. *
1238
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
Table 5 Changes in average sick-leave incidence and duration in control and treatment groups from the 12-month period prior to the switch (P-values in parentheses).
I. Incidence (fraction of employees with absence spell) Mean change in control groups Mean change in treatment groups Differences in mean (treatment-control) Standard deviation of change in control groups Standard deviation of change in treatment groups Ratio of standard deviations (treatment/control)
II. Duration (number of absence days) Mean change in control groups Mean change in treatment groups Differences in mean (treatment-control) Standard deviation of change in control groups Standard deviation of change in treatment groups Ratio of standard deviations (treatment/control) *
. . .to month 1–12 after the switch
. . .to month 13–24 after the switch
0.0031 0.0045 0.0014 (P = 0.6701)
0.0059 0.0063 0.0004 (P = 0.9236)
0.0408 0.0408 1.0000 (P = 0.9800)
0.0450 0.0447 0.9933 (P = 0.9078)
2.2013 2.3419 0.1406 (P = 0.8618)
6.9396 8.0444 1.1048 (P = 0.2498)
9.5337 11.0745 1.1616*** (P = 0069)
11.1053 13.3582 1.2029*** (P = 0.0009)
Significant at the 10 percent level. Significant at the 5 percent level. *** Significant at the 1 percent level.
1 .5 0 -.5
Markussen et al. (2011) computed physician-strictness indicators for all family doctors in Norway, based on their sickness absence certification propensity after controlling for the observed characteristics of all their employed clients. Even though they used a multivariate hazard rate model, the identification of physicianstrictness was almost entirely based on the cross-sectional variation in absence-certification practices. In contrast, our identification strategy relies entirely on the longitudinal variation in certification practices as one physician replaces another one for a fixed group of individuals. It is thus of some interest to examine the degree of concordance between these two strategies for identifying essentially the same thing; namely the physicians’ strictness with respect to the certification of health related social insurance benefits. We would not expect a perfect correlation: Not only have we used different statistical approaches and identification strategies, but the method used in the present paper – though highly robust with respect to the existence of unobserved patient sorting – also suffers from a significant attenuation problem. Given our use of group-fixed effects, we can obviously not offer strictness indicators for each physician; we can only offer estimates for the differences in strictness between the pairs of outgoing and incoming physicians. However, based on the set of strictness indicators computed for each physician in Markussen et al. (2011), it is possible to compute similar difference-indicators for the same physician-pairs. We have thus accessed the estimated strictnessindicators computed in Markussen et al. (2011) for each of the physician-pairs appearing in the present dataset (177 pairs in total), and computed the differences between the two doctors involved in each switch. A comparison between pair-wise differences-in-strictness estimates reported in the present paper and those obtained by Markussen et al. (2011) is presented in Fig. 7. There is obviously
Physician differences estimated from exogenous switches
4. Comparison with previous findings
-1
based on these findings, it appears that the differences in strictness/leniency between physicians do not stem from differences in the willingness to issue absence certificates, but rather from differences in the number of prescribed absence days for each absentee.
1.5
**
-.04
-.02
0
.02
.04
Pairwise differences between physicians (from Markussen et al. 2011)
Fig. 7. Comparison of pair-wise differences-in-strictness estimates reported in this paper with corresponding estimates computed from the results in Markussen et al. (2011).
a strong positive correlation between these two estimators. The correlation coefficient is 0.49 (P = 0.0000). The rankings of the two physicians involved in each switch are concordant in 66% of the cases. This is far from perfect. However, given the extremely different ways in which these estimators have been obtained, we interpret the high correlation as a confirmation that there really are significant differences in the physicians’ fulfillment of their gatekeeper role. 5. Conclusion General Practitioners have been given an important costcontaining role in many welfare states. They certify sick leaves, disability benefits, and prescription drugs, and refer patients to specialist treatment. Previous research has cast doubt on whether GPs are at all able to act as gatekeepers, particularly in relation to sickness absence. We study this using Norwegian register data that makes it possible to examine groups of workers moving in
S. Markussen et al. / Journal of Health Economics 32 (2013) 1230–1239
bulk from one GP to another, switches that can be interpreted as exogenous changes of GP. We find that physicians do differ: The same employees have substantially and significantly different sick-pay propensities when connected to two different GPs. While physicians have limited influence on absence incidence, they have large influence on the total number of sick-pay days. This suggest that it may be difficult for a family doctor to outright reject a client’s request for an absence certificate, but that there is considerable more scope for discrepancy in the determination of the number of days and/or in the degree of absence. We take this as evidence that gatekeeping is possible to some extent. The current Norwegian policy strategy of focusing on the GPs’ sick-listing practices in order to reduce worker absenteeism may thus have some merit – provided, of course, that it succeeds in changing the most lenient GPs’ behavior. However, we also find that family doctors have less influence over their patients’ absence behavior in areas where it is easy to find a new/additional doctor. Hence, the gatekeeper role of family doctors may to some extent be undermined by easy access to other physicians. From a methodological point of view, the finding that physicians have a significant impact on their clients’ use of social insurance also suggest that the physicians’ observed “practice styles” may constitute a potentially useful instrument for their clients’ social insurance claims, making it possible to identify causal effects of such claims on, e.g., subsequent employment and earnings paths. References Blomqvist, Å., 1991. The doctor as double agent: information asymmetry, health insurance, and medical care. Journal of Health Economics 10, 411–432. Brekke, K.R., Nuscheler, R., Straume, O.R., 2007. Gatekeeping in health care. Journal of Health Economics 26, 149–170. Carlsen, N., Nyborg, K., 2009. The Gate is Open: Primary Care Physicians as Social Security Gatekeepers. Memorandum 07/2009. Department of Economics, University of Oslo, Oslo.
1239
Duggan, M., 2005. Do new prescription drugs pay for themselves? The case of second-generation antipsychotics. Journal of Health Economics 24, 1–31. Dusheiko, M., Gravelle, H., Jacobs, R., Smith, P., 2006. The effect of financial incentives on gatekeeping doctors: evidence from a natural experiment. Journal of Health Economics 25, 449–478. Englund, L., Svärdsudd, K., 2000. Sick-listing habits among general practitioners in a Swedish county. Scandinavian Journal of Primary Health Care 18, 81–86. French, E., Song, J., 2009. The Effect of Disability Insurance Receipt on Labor Supply. Working Paper. Federal Reserve Bank of Chicago. Grytten, J., Sørensen, R., 2003. Practice variation and physician-specific effects. Journal of Health Economics 22, 403–418. Maestas, N., Mullen, K., Strand, A., 2011. Does Disability Insurance Receipt Discourage Work? Using Examiner Assignments to Estimate Causal Effects of SSDI Receipt. Rand Working Paper, WR-853-2. Markussen, S., 2012. The effects of sick-leaves on earnings. Journal of Population Economics 25, 1287–1306. Markussen, S., Røed, K., Røgeberg, O.J., Gaure, S., 2011. The anatomy of absenteeism. Journal of Health Economics 30, 277–292. Markussen, S., Mykletun, A., Røed, K., 2012. The case for presenteeism—evidence from Norway’s Sickness Insurance Program. Journal of Public Economics 96, 959–972. Markussen, S., Røed, K., Røgeberg, O.J., 2013. The Changing of the Guards: Can Physicians Contain Social Insurance Costs? IZA Discussion Paper No. 7122. Nilsen, S., Werner, E.L., Maeland, S., Eriksen, H.R., Magnussen, L.H., 2011. Considerations made by the general practitioner when dealing with sick-listing of patients suffering from subjective and composite health complains. Scandinavian Journal of Primary Health Care 29, 7–12. O.E.C, D., 2010. Sickness Disability and Work: Breaking the Barriers—Key Trends and Outcomes in Sickness and Disability. OECD, Paris. O.E.C, D., 2012. OECD Economic Surveys Norway. OECD, Paris. Reiso, H., Nygård, J.F., Brage, S., Gulbrandsen, P., Tellnes, G., 2000. Work ability assessed by patients and their GPs in new episodes of sickness certification. Family Practice 17, 139–144. Scott, A., 2000. Economics of general practice. Handbook of Health Economics, vol. 1. Elsevier science, Amsterdam (Chapter 22). Stone, D., 1984. The Disabled State. Temple University Press, Philadelphia. Wilkin, D., 1992. Patterns of referral: explaining variation. In: Roland, M., Coulter, A. (Eds.), Hospital Referrals, Oxford General Practice Series 22. Oxford University Press, Oxford.