5:49 p.m.
Hi,
I am having some problems with GMM. I try to run the following script:
# initializations go here
series e1 = 0
series e2 = 0
set bfgs_toler 1e-7
set lbfgs 1
matrix W = I(10)
scalar beta = 0.95
scalar lambda = 0.05
scalar sigma = 1.05
scalar gamma = 0.1
scalar phi = 0.3
scalar theta = 1
list IVs = const R_DETREND(-1) CONSUMPTION_GRO(-1) GDP_GROWTH(-1)
LN_RATIO_DETREN(-1)
# list IVs = const R_DETREND(-1) R_DETREND(-2) R_DETREND(-3)
CONSUMPTION_GRO(-1) CONSUMPTION_GRO(-2) CONSUMPTION_GRO(-3) GDP_GROWTH(-1)
GDP_GROWTH(-2) GDP_GROWTH(-3) LN_RATIO_DETREN(-1) LN_RATIO_DETREN(-2)
LN_RATIO_DETREN(-3)
gmm
e1 = (R_DETREND * beta * (((CONS_DETREND(+1) - lambda * GDP_DETREND(+1)) ^
((-1) * (sigma))) * ((CONS_DETREND - lambda * GDP_DETREND) ^ (gamma)) *
(LABOR_DETREND(+1) ^ phi) + (gamma * beta * ((CONS_DETREND(+2) - lambda *
GDP_DETREND(+2)) ^ (1 - sigma)) * ((CONS_DETREND(+1) - lambda *
GDP_DETREND(+1)) ^ (gamma - 1)) * (LABOR_DETREND(+2) ^ phi)) / (1-sigma)) -
(((CONS_DETREND - lambda * GDP_DETREND) ^ ((-1) * (sigma))) *
((CONS_DETREND(-1) - lambda * GDP_DETREND(-1)) ^ (gamma)) * (LABOR_DETREND ^
phi) + (gamma * beta * ((CONS_DETREND(+1) - lambda * GDP_DETREND(+1)) ^ (1 -
sigma)) * ((CONS_DETREND - lambda * GDP_DETREND) ^ (gamma - 1)) *
(LABOR_DETREND(+1) ^ phi)) / (1-sigma)))
e2 = (phi * ((CONS_DETREND - lambda * GDP_DETREND) ^ (1 - sigma)) *
((CONS_DETREND(-1) - lambda * GDP_DETREND(-1)) ^ gamma) * (LABOR_DETREND ^
(phi-1))) / (1-sigma) + (((CONS_DETREND - lambda * GDP_DETREND) ^ ((-1) *
sigma)) * ((CONS_DETREND(-1) - lambda * GDP_DETREND(-1)) ^ gamma) *
(LABOR_DETREND ^ phi) + (beta * gamma * beta * ((CONS_DETREND(+1) - lambda *
GDP_DETREND(+1)) ^ (1-sigma)) * ((CONS_DETREND - lambda * GDP_DETREND) ^
(gamma - 1)) * (LABOR_DETREND(+1) ^ phi)) / (1 - sigma)) * (-theta) * ((1 -
LABOR_DETREND) ^ (theta - 1)))
orthog e1 ; IVs
orthog e2 ; IVs
weights W
params beta lambda sigma gamma phi
end gmm --iterate
And get all kinds of error messages. I have tried running this with various
different starting values of the parameters and I get a different error
message each time. So far I have seen: "convergence criterion was not met",
"missing values encountered", "insufficient degrees of freedom" and
"matrix
not positive definite". Other times it gives me results but they are
identical to the starting values and the std. errors are all 0. All this
just by changing the starting values. Does anyone have an explanation for
this? And how to solve it?
6:53 p.m.
On Wed, 4 Nov 2009, Gal Wettstein wrote:
I am having some problems with GMM. I try to run the following
script:
# initializations go here
series e1 = 0
series e2 = 0
set bfgs_toler 1e-7
set lbfgs 1
matrix W = I(10)
scalar beta = 0.95
scalar lambda = 0.05
scalar sigma = 1.05
scalar gamma = 0.1
scalar phi = 0.3
scalar theta = 1
list IVs = const R_DETREND(-1) CONSUMPTION_GRO(-1) GDP_GROWTH(-1)
LN_RATIO_DETREN(-1)
# list IVs = const R_DETREND(-1) R_DETREND(-2) R_DETREND(-3)
CONSUMPTION_GRO(-1) CONSUMPTION_GRO(-2) CONSUMPTION_GRO(-3) GDP_GROWTH(-1)
GDP_GROWTH(-2) GDP_GROWTH(-3) LN_RATIO_DETREN(-1) LN_RATIO_DETREN(-2)
LN_RATIO_DETREN(-3)
gmm
e1 = (R_DETREND * beta * (((CONS_DETREND(+1) - lambda * GDP_DETREND(+1)) ^
((-1) * (sigma))) * ((CONS_DETREND - lambda * GDP_DETREND) ^ (gamma)) *
(LABOR_DETREND(+1) ^ phi) + (gamma * beta * ((CONS_DETREND(+2) - lambda *
GDP_DETREND(+2)) ^ (1 - sigma)) * ((CONS_DETREND(+1) - lambda *
GDP_DETREND(+1)) ^ (gamma - 1)) * (LABOR_DETREND(+2) ^ phi)) / (1-sigma)) -
(((CONS_DETREND - lambda * GDP_DETREND) ^ ((-1) * (sigma))) *
((CONS_DETREND(-1) - lambda * GDP_DETREND(-1)) ^ (gamma)) * (LABOR_DETREND ^
phi) + (gamma * beta * ((CONS_DETREND(+1) - lambda * GDP_DETREND(+1)) ^ (1 -
sigma)) * ((CONS_DETREND - lambda * GDP_DETREND) ^ (gamma - 1)) *
(LABOR_DETREND(+1) ^ phi)) / (1-sigma)))
e2 = (phi * ((CONS_DETREND - lambda * GDP_DETREND) ^ (1 - sigma)) *
((CONS_DETREND(-1) - lambda * GDP_DETREND(-1)) ^ gamma) * (LABOR_DETREND ^
(phi-1))) / (1-sigma) + (((CONS_DETREND - lambda * GDP_DETREND) ^ ((-1) *
sigma)) * ((CONS_DETREND(-1) - lambda * GDP_DETREND(-1)) ^ gamma) *
(LABOR_DETREND ^ phi) + (beta * gamma * beta * ((CONS_DETREND(+1) - lambda *
GDP_DETREND(+1)) ^ (1-sigma)) * ((CONS_DETREND - lambda * GDP_DETREND) ^
(gamma - 1)) * (LABOR_DETREND(+1) ^ phi)) / (1 - sigma)) * (-theta) * ((1 -
LABOR_DETREND) ^ (theta - 1)))
I'm sorry, but this is totally unreadable, and in fact has a
syntax error (unmatched trailing parenthesis in the definition of
e2).
If you'd like feedback on your model, please:
* break up these huge chunks: define some intermediate variables
as components of e1 and e2 (e.g. e1 = e1a - e1b)
* lose all the confusing parentheses, for example
x^((-1) * (sigma)) -> x^(-sigma)
(and note that '^' has higher precedence than '*', which allows
you to delete a few dozen parentheses).
* use terse variable names in this sort of expression so that the
math is clear, e.g. R_DETREND -> r, CONSUMPTION_GRO -> dc and so
on.
That said, one hint is to drop the "--iterate" flag to gmm and add
the "--verbose" flag.
Allin Cottrell
11:30 p.m.
On Wed, 4 Nov 2009, Gal Wettstein wrote:
I am having some problems with GMM...
One more tip for debugging GMM, Try calculating your "GMM
residual" outside of the GMM block and see what it looks like.
It should not contain any NAs other than for lost observations at
the beginning and/or end of the sample, associated with taking
leads and lags. I suspect this condition is not met in your case.
A construction that arises in a few places in your "e1" and "e2"
series is:
(CONS_DETREND - lambda*GDP_DETREND)^(gamma - 1)
I assume that "DETREND" means that these series are either first
differences or residuals from some sort of regession on time;
either way the series will include both positive and negative
values. Now you initialize lambda to 0.05 and gamma to 0.1.
This means that you're almost certainly trying to take negative
numbers to a fractional power. Bzzt! This would result in a
complex number and is not legal in regular floating-point
arithmetic. In gretl it will produce an NA.
Allin Cottrell
5527
days inactive
5527
days old
2 comments
2 participants
participants (2)
-
Allin Cottrell -
Gal Wettstein