Attributing practice variation by its sources: the case of varicose veins treatments in the Netherlands

When medical decisions are based on guidelines, we expect clinically comparable patients to receive comparable treatments. However, health services research has shown that clinical practice is highly variable across regions, and this variation is often unrelated to differences in population need [1]. This phenomenon is referred to as unwarranted “medical practice variation”.

Unwarranted practice variation in a health care system is an indicator for substandard quality of care and for inefficiencies and inequities: the identification of its sources is important for policy makers, professionals, and insurers alike. Existing measures of practice variation primarily rely on data gathered on treated patients at the final stage of the patient journey, generally treatment performed by a specialist. This method neglects to take into account other potential sources of variation, such initial contacts in primary care and/or initial patient preferences regarding seeking treatment. Therefore, it likely overestimates the magnitude of variation caused by the specialist as well as discounts the responsibility of the general practitioners (GPs) in weeding out patients without legitimate medical needs. GPs represent the first point of contact in health care in the Netherlands. All Dutch residents are required by law to register at one of several hundred private GP offices and obtain a GP’s referral for any initial visit with a specialist. Once a GP referral is obtained, the patient visits the specialist for an initial consult at which point the specialist determines whether and what type of treatment is necessary.

In the present study we propose a new method that allows for the comparison of the magnitude of regional variation in treatments, by looking at two different stages of the patient journey: (1) regional variation measured at the point of the initial visit with the specialist, and (2) at the point of treatment. We take the treatment of varicose veins as an example to illustrate our method.

Varicose veins are dilated superficial veins in the lower extremities [2]. Although in majority of the cases it is asymptomatic, sometimes pressure and pain in the leg, and occasionally a rupture of the vein, can occur. Hence, varicose veins treatments are often considered an elective procedure where patient demand can be highly sensitive to patient characteristics (e.g. being female, having high socio-economic status (SES)) [3]. Research from the United Kingdom has shown large regional variation not only in the volumes of varicose veins operations but also in volume of referrals by primary care trusts [4, 5]. This suggests that at least some of the variation may be present prior to reaching the treating physician’s clinic.

The outline of our paper is as follows. In Sect. 1, we introduce the current methodology used to assess practice variation, we provide a brief overview on the treatment of varicose veins and on Dutch health care system. In Sect. 2, we describe our data and present our proposed methods. In Sect. 3, we present our results, we discuss our findings and conclude.

Previous scientific research on medical practice variation in treatments has almost exclusively evaluated the magnitude of variation in care at the end-stage of the patient journey, usually at the specialists’ office [9]. Even though several papers have mentioned the need to assess the share of variation originating from events that occur before the patient reaches the specialists office (e.g. patients’ preferences to seek help, GP’s willingness to refer the patient to the specialist) which could explain a significant share of the variation, to the best of our knowledge no empirical research has addressed the issue.

We test whether the CVUs of the probability of receiving treatment, the probability of having an initial visit, and conversely, the probability of receiving treatment conditional on an initial visit, are statistically different. If the magnitude of dispersion is larger at the stage of the initial visit, then we can confirm that practice variation was reduced by the specialists in the clinics. And vice-versa, if we find that coefficient of variation was lower at the stage of an initial visit than at the stage of the treatment, we can conclude that the clinics amplified practice variation further.

Our results demonstrate that the CVU, or the measure of dispersion, in the regional probability of an initial visit was significantly larger (0.87–0.96) than at the point of treatment both conditional (0.14–0.25) and unconditional on an initial visit (0.65–0.74), suggesting that practice variation was present before the patient reached the specialist. This suggests that at least part of the variation stems from decisions made before the patient reaches the specialist. Furthermore, when comparing the dispersion in the probability of treatment and the dispersion in the probability of treatment given an initial visit, we notice that the latter is smaller. This indicates that specialists were able to reduce the variation in treatment acting as a ‘gate-keeper’, a role that is traditionally associated with primary care.

We observed that the specialist’s decision meaningfully diminished the amount of regional variation. However, the variation did not entirely disappear at any stage, even after we controlled for patient needs and patient preferences. Our data did not allow us to correct for all need factors and probably only a small share of differences in preferences. Hence, a significant share of the treatment variation remained ‘unexplained’. Factors such as religion, ethnicity, education, household income, previous health care use, proximity to the closest GP or specialist, and the financial health of the provider may also play an important role. Some of these variables (e.g. income, SES) were substituted with regional averages, which may limit their explanatory power in our regressions. Similarly, although our analysis demonstrated the presence of regional dispersion at the initial visit stage, this may not be fully attributed to decisions made by the general practitioner. We expect that much of the observed variations may find their origin in variations in the patients’ willingness to seek care. Primary care physicians are assumed to act as a gatekeeper to guard against the clinically unnecessary use of the collectively insured care. As there is no evidence of underuse from a clinical perspective, we assume that much of the variation in use results from non-clinical (i.e. esthetical) demands for care, which suggests that the GP’s gatekeeping role might have been ignored. It remains unclear to what extend regional variations in willingness to refer are the cause for variation. Our administrative dataset did not allow us to utilize expressed patient needs and preferences variables such that clinical data or survey data would have allowed. Therefore, our needs and preference variables were only proxies with presumably high level of correlation with the actual needs and preferences of the patient. If we assume that the proxy variables used in our model are adequately correlated with the true expressed variables, they can be used to indicate the direction of correlation with some ‘random’ measurement error. Therefore, using proxies, although imperfect, would lead to attenuation bias, and therefore an underestimation of the true effect. If, however, particular preferences did not correlate with our proxy measures, then our analysis could suffer from unmeasured confounding and biased results in an unpredictable direction. It would however not affect the validity of the method that we have introduced.

Our method can be extended by disaggregating and decomposing the results. As the average contribution of both steps may not be equal in every region, analysis per region may clarify the contribution of region-specific referral patterns and individual clinics in our findings. While the point estimates of contributions by region may be interesting, this would require additional attention with respect to the computation of confidence intervals, as bootstrapping would occur on a regional level. If the number of observations become low and confidence intervals wide, then our frequentist approach might result in unstable estimates. Shrinkage approaches, like empirical Bayes, might be appropriate but would require further calculations. They were, therefore, beyond the scope of this study.

In addition, our method could be used to decompose the overall contribution to variation originating from each step of the patient journey when using a model that is linear in the parameters (e.g. ANOVA or linear regression). For such decompositions one may use the law of total covariance, which states that if X,Y,Z are random variables then:

cov(X,Y) = E(cov(X,Y∣Z) + cov(E(X∣Z),E(Y∣Z)).

For example, one could use the probabilities of treatment, initial visit, and treatment conditional on initial visits as a continuous dependent variable and decompose using to the above equality. Likewise, the contribution of both stages to the total can be computed using non-linear (multiplicative) decomposition techniques, like the Oaxaca-decomposition. Such decompositions are complex to interpret, and therefore require future exploration.

In the present paper we offer a methodology for attributing practice variation by patient need, patient preferences, and patient referral along different stages of the patient journey. Our application demonstrates that regional variation in the treatment of varicose veins in the Netherlands is mostly driven by patient and GP decisions leading up to the initial visit to the specialist office. Based on our findings we conclude that any attempt to detect deviations from the norm in the treatment styles of healthcare providers, researchers must take into account the decisions of the patient to seek care and gatekeeping GPs or other referring physicians before reaching the specialist office. The method is easily adaptable to other diagnoses, health care settings and lower levels of aggregation.