Efficiency and productivity of world health systems: where does your country stand?

Abstract

This article deals with estimating efficiency of world health systems using panel data on World Health Organization (WHO) member countries. The Stochastic Frontier (SF) approach is used for this purpose. We evaluate absolute efficiency as well as rankings and their sensitivity across alternative model specifications using both output-maximizing and cost-minimizing frameworks. We also compare productivity of health service (among countries) derived from the output-maximizing and cost-minimizing models. Results show that efficiency rankings of countries vary substantially depending on whether the goal of healthcare provision is output maximization or health cost minimization. The same applies to different SF model specifications.

Acknowledgements

Keynote address given at the 10th Annual Conference of the Italian Health Economics Association held at the University of Genova, 10 and 11 November 2005. An earlier version of this article was presented in Lugano (Switzerland) and Sydney (Australia). I thank seminar participants, especially, Massimo Filippini, Alan Woodland and Bob Bartels for their comments. Comments from two anonymous referees are gratefully acknowledged. The usual disclaimer applies.

Notes

¹ This is the standard definition of efficiency in production. See, for example, Kumbhakar and Lovell (2000, chapter 2) for details. See also Puig-Junoy (1998), Giuffrida and Gravelle (2001) and Lothgren (2000) for health applications.

² In spite of this, the WHO study is heavily criticized by many (Williams, 2000; Hollingsworth and Wildman, 2002; Gravelle et al., 2002, among others) from different directions. Some of these are related to the measure of output (health), quality of data, objective of health production, estimation method, etc.

³ Theoretically, efficiency ranking from these two models will be the same if the underlying production technology is homogeneous. More specifically, efficiency levels from one model is just a constant multiple of those obtained from the other model.

⁴ The first measure is mostly used because of its easy interpretation. The second measure is a (equally weighted) composite measure of success in five health goals (by year health, health distribution, responsiveness, responsiveness in distribution and fairness in financing) and is problematic because the individual components are often difficult to measure. Also there seems to be no a priori justification for weighting each health attribute equally.

⁵ This fact can be used as support against employing panel data methods.

⁶ Kumbhakar and Lovell (2000, p. 86) give the formula for estimating technical efficiency.

⁷ Ideally one would like to use a model that will determine which countries are maximizing healthcare services and which are minimizing healthcare costs. For example, the procedure used in Kumbhakar et al. (2007) can be used for this purpose. Another alternative is to use some exogenous information to group countries in terms of their objectives of healthcare services. Latent class modelling approach (Orea and Kumbhakar, 2004) is another possibility that can we plan to explore in a future study.

⁸ Note that there will be some sign changes in the JLMS and BC formulae for computing technical (in)efficiency. Details can be found in chapter 4 of Kumbhakar and Lovell (2000).

⁹ The data are described in the WHO report (2000), Greene (2002) and many other publications that can be obtained from the WHO website.

¹⁰ See Wang (2002) for issues regarding estimation of these models.

¹¹ We find that the mean technical efficiency score is quite robust (although the observation-specific scores are quite sensitive) to the choice of the z-variables in the μ, σ _u ² and σ _v ² functions.

¹² Removing the insignificant parameters from the translog function does not change the efficiency rankings.

¹³ Note that we are not using any panel model, and therefore, the i subscript indicates both country and time.

¹⁴ See Wang (2002) for details on this, as well as the marginal effect formulae for the conditional mean, E(u|(v − u)).

¹⁵ We used the same z-variables in both models. As mentioned before, we decided to keep only those z-variables in the μ, σ _u ² and σ _v ² functions that are statistically significant.

¹⁶ Theoretically these two measures are the same when there is no inefficiency. However, empirically they differ because the production function is estimated for computing HPP while the input requirement function is estimated to compute HPC.