The main premise of the SFA approach is a recognition that whether allDMUs are efficient or not is an empirical question that can and should be statistically tested against the data, while allowing for a statistical error. To enable such testing, the SFA approach provides a framework where production relationship is estimated also as aconditional average (of outputs given inputs and other factors, in the case of productionfunction) but the total deviation from the regression curve is decomposed into two terms
- statistical noise and inefficiency. Both of these terms are unobserved by a researcher butwith relatively mild assumptions the different approaches within SFA allow the analyst to stimate them for the sample as a whole (e.g., representing an industry) or for each individual
DMU.
In short yes, though there are pros and cons to both:
- unlike in SFA/BSFA in DEA the frontier is implicite, i.e., you do not specify it explicitly just "the general" requirements of it; this is often viewed as a strength of the DEA but I see it more as a weakness; the frontier is what you measure inefficiency against and if you are unaware of its economic characteristics (e.g., if they are in line with the underlying economic theory) this gives you a very limited understanding of what you actually measure;
- in DEA you can do multiple output production analysis without the need to use cost/profit functions; although it is possible to do in SFA/BSFA it is much more troublesome; you need to use distance functions (which use may be questionable from a statistical and inferential viewpoint) or turn to BSFA models which unfortunately are quite complex to estimate;
- DEA is non-stochastic; now many extensions to DEA have been proposed to mitigate this; StoNED is one method worth noting; but then again it is like reaching the right pocket with your left hand.
The aforementioned are key differences between DEA and SFA/BSFA. As for SFA vs. Bayesian SFA (BSFA), the key difference is that inefficiencies in BSFA must have a prior distribution specified, which location parameter (set in a way to have a meaningful economic interpretation) does have some, rather minor, impact on the (posterior) mean inefficiency you get. This is because inefficiencies are treated as latent variables and in Bayesian inference all unobserved quantities in the model need to have a prior. In practice the difference is that in "classic" SFA your inefficiency estimates will be higher (thus estimated efficiency lower) and there will be larger differences between them across DMUs. From BSFA viewpoint this is because the implied prior inefficiency in "classic" SFA is very high. So classic SFA usually finds more inefficiency in the sample with bigger inefficiency differences between DMUs than there actually is (by "borrowing" some of the distortion form the symmetric error). You can find a user-friendly implementation of some well-researched BSFA models here:
https://www.mathworks.com/matlabcentral/fileexchange/64147-kamilmakiela-bsfamk1
Hope this helps,
Kamil
- Badges
- Science topic
- Similar topics
- Psychology