Dear all,

I'm on the research, based on stochastic frontier approach (SFA), Battese and Coelli time-varying model (1995). The study has the purpose of estimating technical efficiency and analyzing the inefficiency determinants (z variables, exogenous variables).

I have 6 independent variables, 4 dummy variables (4 out of 6 are converted to logarithms, and the others are not), and a time trend for stochastic frontier (x variables), while 4 exogenous variables and 2 dummy variables for inefficiency function (z variables).

lny = b0 + b1lnx1 + b2lnx2 + b3lnx3 + b4lnx4 + b5x5 + b6x6 + b7dum1 + b8dum2 + b9dum3 +b10dum4 + t

U = a0 + a1z1 + a2z2 + a3z3 + a4z4 + a5dum1 +a6dum2

The code is as followed:

sfpanel output lnx1 lnx2 lnx3 lnx4 x5 x6 dum1 dum2 dum3 dum4 t, model(bc95) d(t) emean(z1 z2 z3 z4 dum1 dum2)

Through organizing, cleaning, and scaling the data set, I run the code and got results.

But I found something critical that the estimation results are different depending on the order of exogenous variables. Even if I put the same z variables in, if I change the order, the result was different. One critical example, if I put the z variables randomly, the estimation result came out as all missing values (st. dev., z, p>|z|, and 95% Conf. Interval.) except coefficients; while, the other trying gives the very clear and great estimation result.

For your understanding, the results of

emean(z1 z2 z3 z4 dum1 dum2) and

emean(z4 z1 z2 z3 dum1 dum2)

are different.

Can I have confidence in this result?

Does anyone know the reason and know how to solve the problem?

Thank you very much for your attention.