prodest only estimated correctly elasticities once and now it gives only negative coefficients

[Jenniffer Solorzano]

Dear Gabriele Rovigatti ,

I am estimating a production function on COMPUSTAT data for the U.S. and using prodest:

prodest y, method(op) free(lcogs) proxy(i) state(k) poly(3) reps(50)

y = ln(revenue)
lcogs = ln(cost of goods sold)
i =ln(investment)
k = ln(cpaital)

How is it possible that the first time I obtain:

And now that I had to rerun it again because I close the do file for a moment, I get this:



. prodest y, method(op) free(lcogs) proxy(i) state(k) poly(3) reps(50)
.........10.........20.........30.........40...... ...50


op productivity estimator Cobb-Douglas PF

Dependent variable: revenue Number of obs = 199521
Group variable (id): id Number of groups = 19442
Time variable (t): year
Obs per group: min = 1
avg = 10.3
max = 61

------------------------------------------------------------------------------
year | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
lcogs | -1.131983 .0763131 -14.83 0.000 -1.281554 -.9824126
k | 1.125322 . . . . .
------------------------------------------------------------------------------
Wald test on Constant returns to scale: Chi2 = 174.01
p = (0.00)



Same data, number of observations, everything, I did not touch anything. Now, I have tried more times and keep getting the same or similar coefficient values, like -1.13 for the lcogs variable. Why was it able to give me correct results once but now every time I repeat it, those are pretty similar but very bad values. Appreciate the help. Thanks.

[Gabriele Rovigatti]

Dear Jenniffer,

I am sorry for getting back to you so late: I wasn't noticed the new post and the tag. Next time you'd need any help, I'd recommend to send a direct email to my personal account ([email protected]) on top of posting here on Statalist for future reference.

Now, the issue that you are describing is incredible, and I have never seen any similar one - so I'd ask you to send me (or to post here) a minimal reproducible example of it, as soon as you figure out what is going on. While the instability of point estimates with ACF-corrected models is well-known, such a behavior with plain Olley-Pakes or Levinsohn-Petrin methods rings an alarm bell for me, as it is rather unusual to see even small changes depending on the seed set.

However, you might try to change the optimizer (e.g., move to Newton-Raphson, opt(nr) ) and see whether you retrieve the original estimates.

I look forward to receiving any news,

Gabriele