The effects of copayment in primary health care: evidence from a natural experiment.

Objective: Evaluate the effects of the ‘euro per prescription’ on primary health care services (number of doctor visits), through a retrospective cohort study of health care users in Catalonia (Spain). This policy, implemented in Catalonia on 23 June 2012, only lasted six months. This policy was introduced to improve budgetary imbalances in Spain and boost the regional and national governments’ budgets. Methods: We used a retrospective cohort, composed of individuals who had had contact with primary healthcare services between January 1, 2005 and December 31, 2012. The econometric specification followed is a hurdle model. Results: Our results show that from October 2012 onwards there was a decrease in the average number of overall visits, particularly for individuals aged 65 years or more. However, this decline cannot be entirely attributed to the introduction of the euro per prescription policy as in October of that same year the Spanish government introduced its pharmaceutical copayment for pensioners. Conclusions: The policies appraised in this paper reveal a clear deterrent effect among vulnerable individuals such as those with the highest probability of being unemployed and/or those individuals with chronic conditions.


1.-Introduction
The Spanish economy sank into recession in the first quarter of 2009 following a fall in the gross domestic product (GDP) for two consecutive quarters. Although In this paper we are interested in one of these policies in particular, colloquially known and henceforth referred to as, the 'Euro per Prescription' (Law 5/2012 of March 20 [12]) policy. On 23 June, 2012 the Catalan government adopted the 'euro per prescription' which would continue until 31 December 2012, when it was suspended after the Spanish government appealed against it in the Spanish Constitutional Court.
Previous to this new policy, prescriptions had always been free; however, this new policy established and implemented a fee of one euro per prescription. The fee was applicable to all prescription-only medicines (NB: each prescription can only contain one medicine) which cost more than €1.67. An annual limit of €62 per person was set and the 'euro per prescription' did not apply to anyone receiving non-contributory subsidies or the minimum guaranteed income, to those in unsubsidized unemployed, to the disabled or to any treatments resulting from workplace accidents or occupational disease.
This policy was, in fact, a 'natural experiment' since (1) the intervention is not undertaken for the purposes of research, and (2) the variation in exposure and outcomes is analyzed using methods that attempt to make causal inferences [3].
While there is existing literature analysing the effect of pharmaceutical copayment on health care demand, i.e. consumption and prescriptions (for review, [11]; and e.g. [1,6,7,17]), little has been done on the effect of pharmaceutical copayment on the number of visits to primary care doctors (e.g. [28]). Papers looking at number of visits, usually takes into account copayment in insurance, not only medicines (e.g. [20,26,27]). A recent paper by García-Gomez et al [6] examined the same policy as we are looking at, the 'Euro per Prescription', and they found that consumption increased in the 2 months previous to the introduction of the measure, and fell with the introduction of the 'co-payment policy. Puig-Junoy et al [17] found similar results, i.e. an increase in prescriptions in the short-term, but they did not use individual patient data.
Regarding visits to doctors, Winkelmann [28] analised the 1997 German health care reform and the differences-in-differences estimates indicated that increased co-payments reduced the number of doctor visits by about 10% on an average. Our objective in this paper is to evaluate the effects of the 'euro per prescription' in primary health care services (number of doctor visits), through a retrospective cohort study of health care users in Catalonia (Spain). Our main contribution to the existing literature is firstly, the outcome of interest, i.e. number of doctor visits, as little has been done on the relationship between this outcome and pharmaceutical copayment. Secondly, the specific copayment policy which we are analysing, the 'Euro per Prescription'. Thirdly, the use of individual monthly data for January 2005 to December 2012.
The article is organized as follows. First, we explain the method. Then, we explain and discuss the results of the model. Finally, we present our conclusions.

Data setting
To evaluate the effects of the 'euro per prescription' we used a (general The Catalan public healthcare system guarantees universal and free healthcare to all the citizens of Catalonia. The system is characterized by a separation of the funding (from the Catalan public budget) and the provision and management of healthcare services. Catalonia is divided into seven health regions of which an ABS is a territorial division. All the residents of the area covered by the ABS are assigned to the provider responsible for that particular ABS.
The IAS manages all the ABSs that provide health care to the region of 'La Selva Interior', Girona (ABS Anglès; ABS Breda-Hostalric; and ABS Cassà de la Selva).
According to the Catalan Institute of Statistics (IDESCAT [8]), in 2012 the region's population was made up of 32,860 men and 32,702 women (0.87% and 0.85%, respectively, of the entire Catalan population). The area is mainly rural (or semiurban), with many towns scattered throughout the district as well as having a number of farms, estates and small far away villages. While the region has 144 municipalities (3.70% of all Catalonia), it only has 5 municipalities of more than 5,000 inhabitants and only one with a little more than 10,000. The median of the population density was 85.5 hab/km2 in 2012, and the average population density 176.2 hab/km2 (compare with 235.8 in in the whole of Catalonia) (IDESCAT).

Statistical methods
The methods that study natural experiments have the same validity threats that experimental methods (i.e. randomized controlled trials) do. The main difference is the absence of randomization. In fact, in natural experiment studies there is no general solution for the presence of selection bias i.e. the problem of In our case, however, we did not have a control group (unlike [26,28]) as all the individuals in the cohorts were exposed to the intervention. For this reason, we used a quasi-experimental design (time-interrupted time series) organized as a mixed design (panel data), with the individuals observed repeatedly over time.
With this design, the history of each individual prior to the intervention is used to construct the counterfactual. In other words, each individual becomes their own control. In fact, this type of design is equivalent to a difference-in-difference method [10] (a method which deals with unobserved factors, or 'selection on unobservables').
In particular we specified the following two generalized linear models (GLM): (1) where the subscript i denoted individual and t month (from January 2005 to December 2012), Y the dependent variable (doctor visits), µ the (conditional) mean, g() and h() (appropriate) link and variance functions, respectively, and φ a dispersion parameter. We considered visits to both prescribers, the only potentially affected by the prescriptions' copayment, and to non-prescribers.
With η we denoted an additive linear predictor composed of D, a dummy variable, equal to 0 until the time of the intervention and thereafter 1; α was a time effect (constructed from t, t=1,…,96) in order to control for a long-term trend. Month K denoted one of eleven seasonal dummies (k=2,…,12 -January was taken as reference category) in order to control for seasonality, and up to nine explanatory variables: sex (men -reference category -women); age group (under 15 years of age -reference category -; 15-34 years, 35-44, 45-54, 55-64, 65-74, 75 and older); the presence of chronic conditions: hypertension, diabetes mellitus type II, obesity, dyslipidemia-hypercholesterolemia and hypertriglyceridemia, and the quintiles of the probability of being unemployed (and also of being unemployed for more than a year). Chronic conditions were coded as 0 without the chronic condition (reference category) 1 with the chronic condition. δ, y and β's denoted unknown parameters associated with the dummy, the seasonal dummies and the explanatory variables, respectively.
Note that with the exception of gender, all the explanatory variables were timevarying. In addition to age, individuals can enter hypertensive status and/or develop diabetes, and may enter or exit the status of being obese, dyslipidemic, etc. In addition, the probability of being unemployed varied, logically, during the study period.
It is known that besides being modulated by sex and age, health care utilization is related not only to need variables (which ought to affect use i.e. health status), and approached here by the set chronic conditions, but also to non-need variables (which ought not to affect use i.e. socioeconomic conditions) [13,23].
For this reason, we included, as explanatory variable in the models (1), the probability of being unemployed (and unemployed for over a year).

Estimation of the probabilities of being unemployed
Using data from the Spanish and the Catalan Health Surveys (ENSE and ESCA, respectively) corresponding to 2006 and 2011 1 , the probability for an individual of being unemployed (or unemployed for over a year) was estimated using the following GLM with a binomial response (i.e. logistic regression): where Y denoted the event of being unemployed (or unemployed for over a year). Then, using the linear predictor estimated in (2) (i.e. the right-hand side of (2)), we predicted these two sets of probabilities (for being unemployed and for being unemployed for over a year) that would correspond to the individuals in the IAS cohort. To do this, we first divided the cohort into two sub-periods: [2005][2006][2007][2008][2009] and 2010-2012. We used the first sub-period to predict the probabilities associated with the ESCA and ENSE 2006 and the second to predict those associated with the ESCA and ENSE 2011. We assigned a zero probability to individuals under 16 and to those over 65 or to those who were over 65 during the study period. Then we stacked the probabilities corresponding to the two subperiods and estimated the following GLM with Gaussian response (i.e. a linear regression): where Prob_ESCA denoted the predicted probability of being unemployed (or unemployed for over a year) corresponding to ESCA for the individual i in the month t. Prob_ENSE denoted the predicted probability corresponding to ENSE and Unempl_rate the unemployment rate of the municipality where the individual i resided, for the sex of the individual i, and for the year corresponding to the month t (these rates were obtained from IDESCAT).
Thus, we lastly calibrated the probability of being unemployed (and unemployed for over a year), obtaining a single time-varying variable per individual.

Hurdle model
In fact, the use of medical care involves a twofold decision process [4] i.e. the decision to seek care (made by the individual or the 'principal', in a principal agent framework) and the frequency of visits (determined by the physician; the 'agent', in a principal agent framework [4,9,15,25]).
For these reasons, model (1) was estimated using a two-part econometric model, known as a hurdle model [4 ,15 ,25], specified in such a way as to gather together the two decision processes theoretically involved in the use of medical care.
The first part of the decision process was modeled using a binomial link (a logistic regression, in particular): where the subscript i denoted individual and t month, µ denoted the (conditional) mean, the subscript 1 denoted the first part of the decision process, Y the dependent variable, φ was a dispersion parameter and η denoted the additive linear predictor of (1).
Note that we assumed that the dispersion parameter was equal to the unit. This assumption was made because the information available in this first part did not allow the simultaneous identification of the parameters associated to the conditional mean and the parameters associated to the conditional variance [4].
In the second part, the distribution of use (conditional to some use) was modeled as a truncated negative binomial: where Γ(.) denoted the gamma function and the subscript 2 denoted the second part of the decision-making process.
It is important to point out that although we specified model (4) into two parts, these entered the likelihood function multiplicatively and, therefore, we estimated the two parts jointly.

Random effects
Note that some of the coefficients in (1), (2) and (3) had subscripts. In fact, we specified a random coefficient panel data models. In mixed-models terminology, we allowed (some of the) coefficients to be random effects [14] i.e. to be different for the various levels we have considered. Thus, we allowed the intercept to be different for each individual, β0i, capturing characteristics that were individual specific (i.e. individual heterogeneity). In this case, we assumed random effects were identical and independent Gaussian random variables with constant variance. The time effect varied by year-month, assuming a random walk of order 1 (i.e. independent increments) for the Gaussian random effects vector (although we also assumed a constant variance) (R-INLA project [19]). Finally, the coefficient of interest, i.e. the effect of the intervention, varied per month and also per sex, age group, chronic conditions and quintiles of probability of being unemployed. Here, we also assumed identical and independent Gaussian random effects with constant variance.
For the estimation of the models, we followed the Integrated Nested Laplace Approximation (INLA) approach (Rue et al. [22]) within a (pure) Bayesian framework. All analyses have been made with the free software R (version 3.0.2) [18], available through the INLA library (R-INLA project [22]).

3.-Results
In Fig. 1 we show the temporal evolution of the total visits to prescribers in the IAS cohort (Fig. 1) The results of estimating the effect of the intervention are shown in Table 1 (visits to prescribers) and 2 (visits to non-prescribers). In the case of the visits to prescribers, the implementation of the 'euro per prescription' showed a 4.44% reduction in the average number of visits per individual, although it was not statistically significant until December 2012. Note that, in fact the effect was only statistically significant for women, with a reduction in the range of 2.81%-4.67% between July and September 2012, and with a sharper reduction from November 2012 (a 5.49% drop in November and 7.52% in December). We found differences among the age groups. In particular, the effect was only significant for those individuals over 55 years of age. Individuals who were 65 years or more experienced the greatest reduction, largely in October and November 2012. In December of 2012 the reduction was significant only for the 55 to 64-year-old age group. In individuals under the age of 55 the effect was not statistically significant (Table 1).
It seems that the reduction in the average number of visits per individual occurred only in those with chronic conditions, with October 2012 being the exception although not statistically significant ( Table 1). Note that while for hypertensive individuals the decline in visits occurred mainly from July (4.15%) to September 2012 (4.69%), for the remainder of those with chronic conditions it had remained about the same from July 2012 onwards, albeit with peaks in September 2012, but was also much more moderated than that of hypertension (maximum of 3.08% for obese).
When stratifying by quintiles for the probability of being unemployed (Table 1), only for individuals located in the fifth quintile, (i.e. those who were more likely to be unemployed), was there a statistically significant reduction in the average number of visits. Nevertheless, this was only from July to September 2012 and the decline was very similar in that three-month period. We also found that the decrease was very similar for the fifth quintile of the probability of being unemployed for over a year.
With respect to the visits to non-prescribers (Table 2) Table   2). This reduction is greater than the average visits by individual to the prescriber.
Note that, in fact, the effect was only statistically significant for women in September 2012 (14.42%), whereas in December 2012 it was statistically significant for both males and females (12.85% and 15.81%, respectively).
However, there was a significant increase in individual visits (both male and female) to the non-prescriber in October 2012 (19.13%). We also found differences among the age groups. In particular, the effect was significant for  Pensioners, who previously paid no medication fees (except for civil servant retirees) had to pay 10% of the cost of their medication, with ceilings set at €8.00 per month for those with earnings less than €18,000per year, €18 per month for those with earnings between €18,000 and €100,000, and €60 per month for those with earnings more than €100,000per year.
In fact, according to figures from the Catalan Government thanks to the application of the 'Euro per Prescription' in just six months the Government had earned EUR 45.7m and reduced expenditure on drugs by 5.9%. More in detail, while during the two months before the policy implementation there was a monthly increase of 7.79€ daily doses, once the policy was introduced there was a significant monthly reduction of 7.54€ daily doses for the 7 months that was in place [16]. Taking into account it was a randomized controlled trial, in a Swedish study Elofsson et al. [5] found that for those who assessed their financial situation to be poor, the probability of foregoing care was 10 times greater than among those who assessed their financial situation to be fair or good. However, among women avoiding physician visits was also associated with chronic disease. This paper could have several limitations. First, as explained above, several political measures coincide over time, so it is difficult to attribute the effects we have found to only one of them. However, we believe that by allowing the time effect to vary by year-month, it allowed us to identify in which month the measures took effect. This could make it possible to attribute the effect to those policies that are temporarily closest to it. for details).
A third limitation is related to data availability. The study is performed on a region of Girona province as, at that time, it was the only population-based cohort (i.e. constituted by the entire population) that existed in Catalonia. Moreover, postpolicy data was not available as the cohort was closed in 2012. The IAS was absorbed by another health provider (the ICS) and the database feed was discontinued.

Conflicts of Interest
There are no conflicts of interest for any of the authors. All authors freely disclose any actual or potential conflict of interest including any financial, personal or other relationships with other people or organizations within three years of beginning the submitted work that could inappropriately influence, or be perceived to influence, their work.