Hello Statalisters,
Following a model described in Stewart and Swaffield (1999) "Low Pay
Dynamics and Transition Probabilities", I am using the heckprob command
to analyse the reliance on a social program in period t given that there
was reliance on that program in t-1.
I am puzzled by the fact that some variables that have a significant
impact when I look at the output from the heckprob command lose some of
their significance once I compute the marginal effects using mfx
compute, predict(pmargin). Can someone clarify this?
As an example, here are partial results (look at what happens to the
variable "ue"--that variable is not in the selection equation):
heckprob y1 female age96 patner96 kid 96 ue + other variables, sel(y2=
female age93 age2 partner93 +other variables)
| Robust
| Coef. Std. Err. z P>|z| [95% Conf.
Interval]
-------------+----------------------------------------------------------
------
y1
female | -.1430339 .131485 -1.09 0.277 -.4007398
.114672
age96 | -.0000943 .0060446 -0.02 0.988 -.0119415
.011753
partner96 | .2497955 .1521505 1.64 0.101 -.0484141
.548005
kid96 | .1101823 .1250257 0.88 0.378 -.1348636
.3552282
ue | .0671154 .0148375 4.52 0.000 .0380345
.0961964
(output omitted)
------------------------------------------------------------------------
---------
y2
female | -.4244642 .1414786 -3.00 0.003 -.7017572
-.1471712
age93 | .0293956 .0359399 0.82 0.413 -.0410454
.0998366
age2 | -.0381842 .0447901 -0.85 0.394 -.1259712
.0496027
partner93 | .0090001 .1188815 0.08 0.940 -.2240033
.2420035
(ouput omitted)
mfx compute
Marginal effects after heckprob
y = Pr(intense9698=1) (predict)
= .03427384
------------------------------------------------------------------------
------
variable | dy/dx Std. Err. z P>|z| [ 95% C.I. ]
X
---------+--------------------------------------------------------------
------
female* | -.010812 .00934 -1.16 0.247 -.029121 .007497
.474138
age96 | -7.16e-06 .00046 -0.02 0.988 -.000908 .000893
41.3409
partn~96*| .0172949 .01105 1.57 0.118 -.004364 .038954
.714587
kid96* | .0081943 .00864 0.95 0.343 -.008741 .02513
.610360
ue | .0050975 .00215 2.37 0.018 .000875 .00932
10.2234
(output omitted)
Also, I was wondering if the example given in the section on heckprob in
the reference manual (Stata 7 p.37) should not be comparing pcond with
phat instead of pmarg.
The example includes the following lines:
heckprob y1 x1, sel(y2=x1 x2) no log
predict pmarg
probit y1 x1 if y2==1
predict phat
Isn't phat a conditional probability equivalent to Pr(y1=1|y2=1) and
thus should be compared with pcond?
Thanks a lot for your help!
Anne Motte
Chercheure/Researcher
Social Research and Demonstration Corporation
50 O'Connor St. Suite 1400
Ottawa On K1P 6L2
tel.: (613) 482-1896
fax: (613) 237-5045
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
