help quadchk dialog: quadchk
-------------------------------------------------------------------------------
Title
[XT] quadchk -- Check sensitivity of quadrature approximation
Syntax
quadchk [#1 #2] [, nooutput nofrom ]
Menu
Statistics > Longitudinal/panel data > Setup and utilities > Check
sensitivity of quadrature approximation
Description
quadchk checks the quadrature approximation used in the random-effects
estimators of the following commands:
xtcloglog
xtintreg
xtlogit
xtpoisson, re with the normal option
xtprobit
xttobit
quadchk refits the model for different numbers of quadrature points and
then compares the different solutions.
#1 and #2 specify the number of quadrature points to use in the
comparison runs of the previous model. The default is to use n_q -
int(n_q/3) and n_q + int(n_q/3) points, where n_q is the number of
quadrature points used in the original estimation.
Most options supplied to the original model are respected by quadchk, but
several are not. These are or, vce(), and the maximize_options.
Option
nooutput suppresses the iteration log and output of the refitted models.
nofrom forces refitted models to start from scratch rather than starting
from the previous estimation results. Adaptive quadrature with
intmethod(aghermite) is more sensitive to starting values than
nonadaptive quadrature, intmethod(ghermite), or the default method of
adaptive quadrature, intmethod(mvaghermite). Specifying the nofrom
option can level the playing field in testing estimation results.
Remarks
As a rule of thumb, if the coefficients do not change by more than a
relative difference of 10^-4 (0.01%), the choice of quadrature points
does not significantly affect the outcome, and the results may be
confidently interpreted. However, if the results do change appreciably
-- greater than a relative difference of 10^-2 (1%) -- then you should
question whether the model can be reliably fit using the chosen
quadrature method and the number of integration points.
Two aspects of random-effects models have the potential to make the
quadrature approximation inaccurate: large group sizes and large
correlations within groups. These factors can also work in tandem,
decreasing or increasing the reliability of the quadrature. Increasing
the number of integration points increases the accuracy of the quadrature
approximation.
Examples
Setup
. webuse quad1
. xtset id
Fit random-effects (RE) probit model
. xtprobit z x1-x6
Check stability of quadrature calculation
. quadchk
Fit RE probit model using nonadaptive Gauss-Hermite quadrature
. xtprobit z x1-x6, intmethod(ghermite)
Check stability of quadrature approximation, suppressing output of models
. quadchk, nooutput
Same as above xtprobit, but increase the number of iteration points to
120
. xtprobit z x1-x6, intmethod(ghermite) intpoints(120)
Check stability of quadrature approximation, suppressing output of models
. quadchk, nooutput
Also see
Manual: [XT] quadchk
