Jeph Herrin <[email protected]> noticed that -adjust- has been superseeded by
the -margins- command in Stata 11:
After getting an error with -adjust-, I discovered that it
had been deprecated in version 11. The online help says:
"adjust has been superseded by margins. margins can do everything
that adjust did and more."
However, as far as I can tell, this is not the case. I could use
-adjust- with models that did not include factor variables; however,
with -margins- there seems to be no option to find marginal effects
for independent variables that are not i.vars.
Is this true? Or is there a way to fool -margins- into working with
non-factor variables? Apologies if this has come up before, I could
not find anything on the list or FAQ.
If it is true, this is a complete nusiance. I do not normally
include dichotomous variables as factor variables, and am not
interested in starting now. Moreover, I have estimation results
from -xtmelogit- models that ran for days; am I going to have
to re-estimate with -i.female- instead of -female- to get marginal
effects? (-adjust- gives errors). That, or calculate my own
marginal effects from scratch.
It isn't clear what Jeph means by "marginal effects" related to the -adjust-
command.
Martin Weiss <[email protected]> gave two examples using the -dydx()-
option of -margins- to compute marginal effects using a -logit- model fit.
The first used the 'i.' operator on a factor variable (named 'race') in the
-logit- model fit, the second treated 'race' as a standard variable.
Peter Lachenbruch <[email protected]> noted that Martin's
examples were fitting two different models. Note -margins- computed the
marginal effect of 'race' differently, depending of whether it was specfied as
a factor variable (with the 'i.' operator) or not.
Jeph replied to Martin's post using a regular variable in the -margins-
varlist, such as
. margins smoke
factor smoke not found in e(b)
r(111);
The variables in -margins- varlist specification are interpreted as factor
variables, identifying the predictive margins of interest. Thus the above
example is requesting the predictive margins for the -smoke- factor variable.
I imagine that Jeph might be referring to the -by()- option of -adjust-. If
that is the case, then we just need to translate between -adjust- and
-margins- syntax.
Here are a few examples based on Martin's first post to this thread (I've
included the Stata output from these examples following my signature):
. webuse lbw
. logit low age lwt race smoke ptl ht ui
Compute the mean prediction, Pr(low), over each level of smoking status,
adjusting for the mean of each predictor within smoking status:
. adjust, pr by(smoke)
translates to
. margins, over(smoke) at((means) *)
Compute the mean prediction, Pr(low), over each level of smoking status,
adjusting for the overall mean of each predictor (except smoking status):
. adjust age lwt race ptl ht ui, pr by(smoke)
translates to
. margins, over(smoke) at((omeans) * (asobserved) smoke)
Note that Jeph will get the same results from -margins- when the -logit-
fit treated -smoke- (and -race-) as factor variables.
Verkuilen, Jay <[email protected]> posted an example where -adjust-
allows variables in the -by()- option that were not in the model fit.
The -over()- option of -margins- allows this too.
Philip Ender <[email protected]> used Jay's -logit- model fit to show that
-margins- with the -if- clause produces similar point and standard error
estimates, but that the confidence intervals are different.
The -pr- confidence intervals from -adjust- are computed by transforming the
end-points of the CI limits from the linear prediction.
-margins- computes the CI limits using the normal approximation is valid.
--Jeff
[email protected]
*** Stata output from the preceeding examples:
. webuse lbw, clear
(Hosmer & Lemeshow data)
. logit low age lwt race smoke ptl ht ui
Iteration 0: log likelihood = -117.336
Iteration 1: log likelihood = -102.61679
Iteration 2: log likelihood = -102.17476
Iteration 3: log likelihood = -102.17373
Iteration 4: log likelihood = -102.17373
Logistic regression Number of obs = 189
LR chi2(7) = 30.32
Prob > chi2 = 0.0001
Log likelihood = -102.17373 Pseudo R2 = 0.1292
------------------------------------------------------------------------------
low | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
age | -.0332327 .0357713 -0.93 0.353 -.1033431 .0368777
lwt | -.0120948 .0066158 -1.83 0.068 -.0250617 .000872
race | .4462621 .2150445 2.08 0.038 .0247827 .8677415
smoke | .9255414 .3980923 2.32 0.020 .1452947 1.705788
ptl | .5397383 .3469187 1.56 0.120 -.1402099 1.219686
ht | 1.799337 .6872194 2.62 0.009 .4524118 3.146262
ui | .7148045 .4634311 1.54 0.123 -.1935037 1.623113
_cons | -.1029665 1.284609 -0.08 0.936 -2.620754 2.414821
------------------------------------------------------------------------------
.
. // Compute the mean prediction, Pr(low), over each level of smoking status,
. // adjusting for the mean of each predictor within smoking status:
. adjust, pr by(smoke)
-------------------------------------------------------------------------------
Dependent variable: low Equation: low Command: logit
Variables left as is: age, lwt, race, ptl, ht, ui
-------------------------------------------------------------------------------
----------------------
smoked |
during |
pregnancy | pr
----------+-----------
0 | .22079
1 | .395668
----------------------
Key: pr = Probability
. margins, over(smoke) at((means) *)
Adjusted predictions Number of obs = 189
Model VCE : OIM
Expression : Pr(low), predict()
over : smoke
at : 0.smoke
age = 23.42609 (mean)
lwt = 130.9043 (mean)
race = 2.095652 (mean)
smoke = 0 (mean)
ptl = .1217391 (mean)
ht = .0608696 (mean)
ui = .1304348 (mean)
1.smoke
age = 22.94595 (mean)
lwt = 128.1351 (mean)
race = 1.459459 (mean)
smoke = 1 (mean)
ptl = .3108108 (mean)
ht = .0675676 (mean)
ui = .1756757 (mean)
------------------------------------------------------------------------------
| Delta-method
| Margin Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
smoke |
0 | .2207901 .0417458 5.29 0.000 .1389698 .3026104
1 | .3956685 .0609789 6.49 0.000 .2761519 .515185
------------------------------------------------------------------------------
.
. // Compute the mean prediction, Pr(low), over each level of smoking status,
. // adjusting for the overall mean of each predictor (except smoking status):
. adjust age lwt race ptl ht ui, pr by(smoke)
-------------------------------------------------------------------------------
Dependent variable: low Equation: low Command: logit
Covariates set to mean: age = 23.238095, lwt = 129.82011, race = 1.8465608,
ptl = .1957672, ht = .06349206, ui = .14814815
-------------------------------------------------------------------------------
----------------------
smoked |
during |
pregnancy | pr
----------+-----------
0 | .214918
1 | .408544
----------------------
Key: pr = Probability
. margins, over(smoke) at((omeans) * (asobserved) smoke)
Predictive margins Number of obs = 189
Model VCE : OIM
Expression : Pr(low), predict()
over : smoke
at : 0.smoke
age = 23.2381 (omean)
lwt = 129.8201 (omean)
race = 1.846561 (omean)
ptl = .1957672 (omean)
ht = .0634921 (omean)
ui = .1481481 (omean)
1.smoke
age = 23.2381 (omean)
lwt = 129.8201 (omean)
race = 1.846561 (omean)
ptl = .1957672 (omean)
ht = .0634921 (omean)
ui = .1481481 (omean)
------------------------------------------------------------------------------
| Delta-method
| Margin Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
smoke |
0 | .2149182 .0435287 4.94 0.000 .1296035 .3002328
1 | .4085436 .0662837 6.16 0.000 .2786299 .5384573
------------------------------------------------------------------------------
<end>
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
