Difficult-to-interpret logit results

Greetings-

In working with some large-scale survey data, I have some results I
find puzzling, and wonder if others may be able to shed light on them.

I am using Stata/SE 8.2 for Linux, and the svylogit command. The data
are from AddHealth (http://www.cpc.unc.edu/addhealth) and we have
weighted them and set up the survey structure accordingly.

One of our independent variables is a five-level factor corresponding
to political orientation, ranging from "very liberal" (pol1b) to "very
conservative" (pol5b). These are dummy variables, coded 1 if the
respondent identified as such, 0 otherwise, and they are mutually
exclusive.

My understanding is that the results should be consistent regardless
of which of the five levels is omitted as the reference category. But
in our case this is not happening.

First, using pol1b ('very conservative') as the reference category:

.. svylogit negchgfed pol2b pol3b pol4b pol5b, subpop(post911)

Survey logistic regression

pweight: gswgt3_2 Number of obs =13771
Strata: region Number of strata =4
PSU: psuscid Number of PSUs =132
Population size =21084487
Subpopulation no. of obs = 10699 F( 4, 125) =91.37
Subpopulation size = 16273782 Prob > F =0.0000

------------------------------------------------------------------------------
negchgfed | Coef. Std. Err. t P>|t| [95% Conf.Interval]
-------------+----------------------------------------------------------------
pol2b | -2.418374 .1494214 -16.18 0.000 -2.714029 -2.122718
pol3b | -2.706085 .1473055 -18.37 0.000 -2.997554 -2.414616
pol4b | -2.723853 .1465713 -18.58 0.000 -3.013869 -2.433836
pol5b | -1.955683 .2054996 -9.52 0.000 -2.362299 -1.549067
_cons | 1.573888 .1335266 11.79 0.000 1.309683 1.838093
------------------------------------------------------------------------------

Note that pol2b ('somewhat conservative') is significant at p<.001 and
is negative in sign.

Now, running the same regression using pol2b as the reference category:

.. svylogit negchgfed pol1b pol3b pol4b pol5b, subpop(post911)

Survey logistic regression

pweight: gswgt3_2 Number of obs =13771
Strata: region Number of strata =4
PSU: psuscid Number of PSUs =132
Population size =21084487
Subpopulation no. of obs = 10699 F( 4, 125) =59.10
Subpopulation size = 16273782 Prob > F =0.0000

------------------------------------------------------------------------------
negchgfed | Coef. Std. Err. t P>|t| [95% Conf.Interval]
-------------+----------------------------------------------------------------
pol1b | -.5594381 .1730002 -3.23 0.002 -.9017487-.2171276
pol3b | -1.106123 .0759412 -14.57 0.000 -1.256386-.9558604
pol4b | -1.123891 .0928584 -12.10 0.000 -1.307627-.9401548
pol5b | -.3557215 .1747169 -2.04 0.044 -.7014287-.0100143
_cons | -.0260738 .0654523 -0.40 0.691 -.1555824.1034347
------------------------------------------------------------------------------

Note that, now, pol1b is significant and also negative in sign! My
understanding is that the first model shows that, relative to "very
conservative" respondents, "somewhat conservative" respondents are
significantly less likely to exhibit negchfed, but that the
secondmodel shows that, relative to "somewhat conservative"
respondents, "very conservative" respondents are significantly less
likely to exhibit negchgfed. Any thoughts on where the model or my
interpretation of it is wrong?

Thanks in advance.

--
----------------------------------------------------------------------
Andrew J Perrin - andrew_perrin@xxxxxxx - http://www.unc.edu/~aperrin
Assistant Professor of Sociology; Book Review Editor, _Social Forces_
University of North Carolina - CB#3210, Chapel Hill, NC 27599-3210 USA
.

Relevant Pages

  • If these guys are Americas allies; then imagine Americas enemies
    ... Survey Finds Suspicion of U.S. Pervasive in Islamic World ... An average of more than 75% of respondents across the four countries ... from terrorist attacks". ...
    (soc.culture.iranian)
  • ,
    ... Survey Finds Suspicion of U.S. Pervasive in Islamic World ... An average of more than 75% of respondents across the four countries ... from terrorist attacks". ...
    (soc.culture.iranian)
  • Re: Trust in God and Trust in Man: Religious Beliefs, Trusting Beliefs and Online Behaviour
    ... Usenet as a pre-test; the main survey will use a stratified sample. ... unlikely to be known by respondents. ... The web site is a general shopping web site. ...
    (uk.religion.christian)
  • Research on Gmail
    ... I tried to contact the group moderator to get permission to post the ... I have hosted a survey on survey monkey, ... None of the members of sci.stat.math will be identified individually. ... respondents, chosen by a raffle, will win a $50 cash prize. ...
    (sci.stat.math)
  • Re: Difficult-to-interpret logit results
    ... > In working with some large-scale survey data, ... > of which of the five levels is omitted as the reference category. ...
    (sci.stat.consult)