Signalling compliance
III. Stochastic frontier models
A. Fixed Effects model
The simplest panel data model in an efficiency measurement context is the fixed
effects model, such as that proposed by Schmidt and Sickles, and applied by Evans 25
et al. The time invariant fixed effects model is simple because it requires no 26
assumption with respect to the distribution of πi nor does it require the assumption
That is, the first order derivatives πy/ππ±1 and πy/ππ±2are positive, and the second order derivatives 24
π2y/ππ±12 and π2y/ππ±22 = Ξ²11 and Ξ²12 are negative. Schmidt, P., and R. Sickles. Supra n. 17 25
Evans D., A. Tandon, C. Murray, and J. Lauer. Supra n. 15 26
that the πi are uncorrelated with the πit and the other regressors. The πi values are treated as state-specific constants so that the basic model in (1) becomes
(2)
where Ξ± represents the frontier intercept and the πi capture the degree of non-
compliance. The fixed effects model can be estimated using the βwithinβ estimator, which corresponds to the Corrected Ordinary Least Squares (COLS) estimator where the parameters in (2) are estimated by OLS and the intercept is shifted up to the observation of the best performing state so that all other residuals are non- positive. Here, the state with the maximum Ξ±i is assumed to be fully compliant (when
πi = 0) and every other state-specific Ξ±i is measured as a deviation from this
benchmark. The differences in πi amongst states should, therefore, be interpreted as
relative rather than actual non-compliance.
(3)
There is therefore an implicit overestimation of compliance in the fixed effects model; it is unlikely that the best performing state in the sample is fully compliant. This would not be problematic if all the analysis was seeking to do was to provide a ranking of all states with respect to compliance. But this is not all that is required. What is additionally sought is a more precise estimate of compliance. Since there is no way of testing the extent of compliance-overestimation, the revealed fixed effects estimates must be interpreted with this limitation in mind.
Nevertheless, since there is no requirement to make any assumptions with respect to
πi, estimates of the πi and the parameters in the fixed effects model are consistent as yit =(Ξ±βui)+X / itΞ²+vit yit =Ξ±i +X / itΞ²+vit
N β β and/or T β β. At the same time however, because the fixed effects model
creates no space for other time invariant variables to enter it, whatever the model may gain in robustness it may well lack in precision. Specifically, the fixed effects, that
is the Ξ±i and ultimately the πi, will be capturing not only the variation in time
invariant non-compliance across states but will also be capturing any time invariant
heterogeneity. A problem emphasised by Greene and others. The degree to which 27
this problem is indeed problematic is dependent on the nature of the data under investigation: the degree of heterogeneity. In this analysis the data cover the whole world, encompassing quite different disease and population environments. Time invariant heterogeneity is, therefore, likely to play an influential role in a stateβs ability to βproduceβ health and ignoring it in the modelling process will likely result in
erroneous estimates of πi. As a result, the fixed effects model can be treated as a
base model against which others, that can accommodate time invariant heterogeneity, may be compared.
B. Heterogeneous Random Effects model
In the foregoing fixed effects model it was assumed that the πi were fixed but were
allowed to be correlated with the regressors and with the πit. In the random effects
model the alternate assumption is made: the πi are randomly distributed with a
constant mean and variance, but are assumed to be uncorrelated with the regressors
and with the πit. The πit are assumed to be symmetric, have zero mean and constant
variance. Pitt and Lee first applied the time invariant random effects model to a panel
data version of stochastic frontier analysis. By assuming that the 28 πi are random,
Greene, W.H. βReconsidering Heterogeneity in Panel Data Estimators of the Stochastic Frontier 27
Model.β (2005); Greene, W.H. Supra n. 21; Kumbhakar, S.C., and K. Lovell. Stochastic Frontier Analysis. 2000
Pitt, M., and L. Lee. βThe Measurement and Sources of Technical Inefficiency in the Indonesian 28
rather than fixed, the random effects model provides an entrance for time invariant covariates.
As noted above, with respect to measuring compliance across states worldwide, there
are important variables, (πβs) besides inputs, (π±βs) that could influence the position or
the shape of the frontier and/or the efficiency distribution. There are, however, a
number of ways in which this heterogeneity (the πβs) can enter the model: i) as
additional shift parameters in the production function; ii) in the conditional mean of the πi; and/or iii) in the variance of either or both parts of the combined error term,
Ο2π and/or Ο2π. Fortunately, this choice need not be dilemmatic. The previous
chapters have provided a theoretical basis upon which to make the decision. In Chapter 3, it was argued that a stateβs ability to fulfil the right to health is not solely determined by the inputs it has at its disposal, those being financial and human resources. In addition, other environmental factors over which the state has little to
no control may also facilitate or impede a state in its efforts to do so. These πβs have
already been identified as mortality density and population density and, since they are additional determinants of ability, they should appear as shift parameters in the production function. As such, the standard stochastic frontier model in (1) becomes
(4)
where πβi is the vector of time invariant environmental variables. These πβs are state-
specific and can shift the frontier by changing Ξ± or can change the shape of the
frontier by influencing both Ξ± andΞ².
yit =Ξ± +X/
itΞ²+z
/
The random effects model can be estimated with estimators based on OLS, which
require no assumption on the distribution of the πi, (although the πi are still required
to be non-negative). However, if an assumption on the distribution is tenable, maximum likelihood estimation (MLE) is feasible, which makes estimation more precise since it can exploit distributional information which OLS estimators cannot.
In this case, the distribution of the error components in (4) remains to be
determined. Given the non-negativity of πi the choices are generally limited to half-
normal, truncated-normal, exponential, or gamma. For simplicity, here the normal- half-normal distribution is applied to the model so that
1. πit ~ iid N(0, Ο2π) the random error term is assumed to be normally distributed
2. πi ~ iid N+(0, Ο2π) the non-compliance term is non-negative and half-
normal. 29
The next step is to obtain estimates of state-specific non-compliance. Because the process for estimating the parameters only produces an estimate of the combined
error term, πi, Jondrow et al.βs conditional mean estimator (JLMS) is applied to 30
separate non-compliance from the combined error term.
Whilst some authors have noted that the assumption in ii. (that the mean value of πi is zero) is a 29
significant restriction to the stochastic frontier analysis, it is used and treated here as a first-step upon which more sophisticated extensions can be built.
Jondrow, J., K. Lovell, I. Materov, and P. Schmidt. βOn the Estimation of Technical Inefficiency in 30