Here is an example showing that the problem is that the quantity you are trying to estimate cannot be estimated. That is, it is not estimable. I load the data and run svyset:

. use diabetes1.dta, clear
. svyset secu [pweight=kwgtr], ///
>     strata(stratum) vce(linearized) singleunit(missing)

A logit model is fit:

. svy: logit diabetes ibn.white ///
>     ibn.white#(i.female i.educ4cat i.obese c.age##c.age##c.age)
  < output omitted >

The mchange commands computes marginal effects for female, but those for white are missing:

. mchange female white

svy logit: Changes in Pr(y) | Number of obs = 16088

Expression: Pr(diabetes), predict(pr)

                    |    Change    p-value
--------------------+----------------------
female              |
     female vs male |    -0.037      0.000
white               |
 white vs non-white |         .          .      <== this is the problem!

Average predictions

             |        no   diabetes
-------------+----------------------
  Pr(y|base) |     0.819      0.181

To figure out why, I add the details option to mchange which displays the output from the margins commands used by mchange. In this case,

Pairwise comparisons of predictive margins
Model VCE    : Linearized

Expression   : Pr(diabetes), predict(pr)

---------------------------------------------------------------------
                    |            Delta-method         Unadjusted
                    |   Contrast   Std. Err.     [95% Conf. Interval]
--------------------+------------------------------------------------
              white |
white vs non-white  |          .  (not estimable)
---------------------------------------------------------------------

The marginal effect cannot be estimated. In this case it is caused by a poorly specified model. See Stata the manual for more information on how margins determines if a function is estimable. 2014-09-05