Evaluating the impact of eligibility for free care on the use of general practitioner (GP) services: A difference-in-difference matching approach

Abstract

In Ireland, approximately 30% of the population (‘medical card patients’) are entitled to free general practitioner (GP) care while the remaining 70% (‘private patients’) must pay the full cost of each visit. Previous research has analysed the effect of this system on GP visiting patterns using regression methods, but to date, no attempt has been made to apply techniques from the treatment evaluation literature to this issue. Treatment evaluation techniques are commonly employed when observations are not randomly assigned to treatment and control groups; this is certainly the case here, as the primary criterion for medical card eligibility is an income below a specified income threshold (and individuals may also be granted medical cards for other reasons such as chronic ill-health). In this paper, previous Irish research, which has analysed the effect of medical card eligibility on GP visiting using regression methods, is extended to consider the use of difference-in-difference matching methods, which control for non-random selection into treatment and control groups, as well as differences in time-invariant unobserved characteristics between individuals in both groups. The results are largely consistent with earlier results using pooled cross-sectional and panel data, and confirm that medical card eligibility exerts a significant effect on GP visiting, even after controlling for observed and unobserved differences in characteristics between medical card and private patients.

Introduction

In Ireland, approximately 30% of the population (medical card patients) are entitled to free general practitioner (GP) care, while the remaining 70% (private patients) must pay the full cost. Eligibility for a medical card is decided primarily on the basis of an income means test, but individuals may also be granted a medical card on the basis of age (since July 2001, all over 70s are automatically entitled to a medical card), particular health needs or participation in certain government-sponsored employment and training schemes. Previous research in Ireland using regression methods has confirmed that even after controlling for differences in socio-economic and health status characteristics between medical card and private patients, medical card patients have significantly higher GP visiting rates than private patients (see Madden et al., 2005, Nolan, 2007a, Nolan, 1991, Nolan, 1993, Nolan and Nolan, 2008, Tussing, 1983, Tussing, 1985). The purpose of this paper is to extend this research examining the effect of medical card eligibility on GP visiting by applying techniques from the treatment evaluation literature to this issue. Previous research has analysed the issue using regression methods, but to date, no attempt has been made to apply techniques from the treatment evaluation literature to this subject. A further extension is afforded by the use of longitudinal data, which allows us to follow the same individuals through time, to compare those who change their medical card status over the period with those who do not and to further control for differences in time-invariant characteristics between medical card and private patients. Matching methods using longitudinal data have been employed before, primarily in the evaluation of labour market interventions (see section Propensity score matching), but this is the first time that such methods have been applied to the issue of interest here.

The evaluation problem is essentially one of missing data; individuals are either in the treatment group or control group, but never both. Constructing the counterfactual is thus the central problem facing those involved in the evaluation of a particular treatment, i.e., how would an individual behave if they had not received the treatment? Experimental data, whereby individuals are randomly assigned to the treatment and control groups, are ideal. Then averaging over the full sample gives an unbiased estimate of the effect of the treatment. However, the costs and ethical considerations surrounding experimental studies mean that they are rarely employed. A well-known exception was the RAND Health Insurance Experiment in the USA in the 1970s where individuals were randomly assigned to a number of different health insurance plans, which differed in the degree to which co-payments were levied on the use of various health services; see Keeler (1992).

The standard means of isolating the independent effect of treatment is to control for observable differences between treated and control observations using regression methods. However, the imposition of functional form assumptions, as well as the possibility of insufficient common support (i.e., for any set of values of the independent variables, there may be insufficient numbers of both treated and control observations) means that such methods are not without problems (see LaLonde, 1986). Alternative solutions to the sample selection problem such as the Heckman sample selection estimator or the use of instrumental variables rely heavily on the identification of suitable instruments, i.e., variables that affect the probability of receiving the treatment, but not the subsequent outcome of interest. Given the data available for this study, neither the Heckman sample selection nor instrumental variables estimators are considered here (see section Data for further discussion).

In this paper, the propensity score matching method is used. This method matches treatment and control observations that are similar in terms of observed characteristics, and thereby produces unbiased estimates of the effect of the treatment on the outcome of interest. Essentially, the outcomes of individuals who are similar pre-treatment, but who differ only in their exposure to the treatment, are compared. Matching estimators have been widely applied in labour economics, in particular in the evaluation of labour market initiatives such as employment and training schemes (see for example, Bryson et al., 2002, Conniffe et al., 2000, Dehejia and Wahba, 1999; Heckman, Ichimura, & Todd, 1997; Lechner & Vazquez-Alvarez, 2003). Matching methods match treatment and control observations on the basis of observable characteristics only, i.e., it is assumed that there are no unobservable differences in characteristics between treatment and control observations. As the data employed in this study are longitudinal, matching methods are combined with a difference-in-difference approach (see also Aassve et al., 2007, Blundell and Costas Dias, 2000, Garcia-Gomez and Lopez-Nicolas, 2006, Görg and Strobl, 2006; Heckman et al., 1997; Lechner & Vazquez-Alvarez, 2003). By taking first differences of the outcome of interest, this approach controls for unobserved time-invariant differences in characteristics between treatment and control observations and as such, provides a more reliable estimate of the effect of the treatment on the outcome of interest.

The section Context and previous Irish research provides more detail on medical card eligibility in Ireland, and previous research in the area. The section Propensity score matching describes the methodology employed, while the section Data introduces the data set and explains how the treatment and control groups are constructed. The section Empirical results discusses empirical results and the section Summary and conclusions summarises and concludes.