On Fri, 28 Jan 2011, GREg Ory wrote:
> I need some tips on estimating panel data using nonlinear switching
> regression technique. AFAIK there are two ways of solving it.
I've had one more go at this problem, because it seems like quite
a nice showcase for what one can now do with gretl scripting. I'm
attaching a script which combines the grid-search approach with
NLS estimation. That is, I first do the grid-search, over what
you stated were acceptable values for mu and sigma, then I
initialize NLS using the best values from the grid. The script
includes a "switch" to turn fixed-effects on or off.
What I'm getting from this (with fixed effects on) is:
grid-search
(mu,sigma) = (0.40, 0.20); SSR = 25.6531, R^2 = 0.803620
NLS (unconstrained)
(mu,sigma) = (0.18, 0.03); SSR = 25.4212, R^2 = 0.806201
That is, NLS finds a slightly better fit with an "out of bounds"
mu value. NLS also shows that the sigma estimate is statistically
indistinguishable from zero, which suggests that the logistic term
may not be appropriate.
Allin Cottrell
set echo off
set messages off
function series GDV(series x, scalar m, scalar s)
series den = exp(pi*(m-x)/(s*sqrt(3)))
return 1/(1+den)
end function
open gretl_plant.gdt -q
# CHOICE here: do we want fixed effects or not?
scalar FE = 1
# scalar FE = 0
scalar SSRmin = 1e10
scalar SSRmax = 0
scalar optmu
scalar optsig
scalar badmu
scalar badsig
scalar nvals = 0
# apply scaling (though not for HHI)
Price /=100
Domestic_Sales /= 10000
Export_Sales /= 10000
Total_import /= 10000
# range and step for mu
scalar m_min = 0.28
scalar m_max = 0.501
scalar m_step = 0.01
# range and step for sigma
scalar s_min = 0.01
scalar s_max = 0.201
scalar s_step = 0.01
# regular regressors: include Cost?
list X = const Cost Domestic_Sales Export_Sales Soda PCFT Total_import
# logistic term
series Log
# dependent variable
series y
/**** GRID SEARCH ****/
loop for (mu=m_min; mu<m_max; mu+=m_step) --quiet
loop for (sig=s_min; sig<s_max; sig+=s_step) --quiet
Log = GDV(HHI, mu, sig)
y = Price - Log
if FE
panel y X --quiet
else
ols y X --quiet
endif
SSR = $ess
if SSR < SSRmin
optmu = mu
optsig = sig
SSRmin = SSR
endif
if SSR > SSRmax
badmu = mu
badsig = sig
SSRmax = SSR
endif
nvals++
endloop
endloop
printf "*** Estimated via GRID SEARCH\n\n"
printf "mu in [%g,%g), sigma in [%g,%g)\n", m_min, m_max,
s_min, s_max
printf "number of points evaluated: %d\n", nvals
printf "min(SSR): %g for mu = %.3f, sigma = %.3f\n",
SSRmin, optmu, optsig
printf "max(SSR): %g for mu = %.3f, sigma = %.3f\n",
SSRmax, badmu, badsig
Log = GDV(HHI, optmu, optsig)
y = Price - Log
if FE
panel y X
else
ols y X
endif
/**** NLS ****/
# initialize NLS from grid search results
scalar m = optmu
scalar s = optsig
printf "*** Estimated via NLS\n"
if FE
series u = $unit
list D = dummify(u)
matrix b = $coeff
matrix g = zeros(nelem(D),1)
nls Price = lincomb(X,b) + lincomb(D,g) + Log
series Log = GDV(HHI, m, s)
params b g m s
end nls -q
b |= g
vnam = varname(X) ~ "," ~ varname(D)
else
matrix b = $coeff
nls Price = lincomb(X,b) + Log
series Log = GDV(HHI, m, s)
params b m s
end nls -q
vnam = varname(X)
endif
# print out NLS results with nice formatting
matrix cf = b | {m; s}
matrix se = sqrt(diag($vcv))
vnam = vnam ~ ",mu,sigma"
cfse = cf ~ se
modprint cfse vnam
printf "SSR = %g, R-squared = %g\n\n", $ess, $rsq
if m < m_min || m > m_max
printf "NLS: mu (%g) is not in [%g,%g)\n", m, m_min, m_max
endif
if s < s_min || s > s_max
printf "NLS: sigma (%g) is not in [%g,%g)\n", s, s_min, s_max
endif
