What's New in Stata 10: Longitudinal and multi-level models (xt)

New command: xtset

A new command, xtset replaces the i(groupvar) and t(timevar) (but they can still be used). The advantage is that you only have to use xtset once for your dataset, rather than every time you run an analysis. For example: 

xtset idcode year
       panel variable:  idcode (unbalanced)
        time variable:  year, 70 to 88, but with gaps
                delta:  1 unit

xtset also allows the delta option sets the periodicity of timevar, most frequently used with %tc (time clock) where time is recorded in ms but the actual data was not collected that frequently. For example, for hourly data, one would declare delta((60*60*1000)) or delta(hours).

New Models: xtmelogit and xtmepoisson

Mixed models for dichotomous (xtmelogit) and count (xtmepoisson) outcomes!

The data for this example are from a survey of women in Bangladesh about their contraceptive use. The level 1 variables are:
   c_use: contraceptive use
   age: woman's age (mean centered)
   child1: one child
   child2: two children
   child3: three or more children
   urban: dummy for urban vs. rural

use http://www.stata-press.com/data/r10/bangladesh.dta
(Bangladesh Fertility Survey, 1989)

xtset district
       panel variable:  district (unbalanced)

xtmelogit c_use urban age child1 child2 child3 || district: urban, cov(un)

Refining starting values: 

Iteration 0:   log likelihood = -1215.8594  (not concave)
Iteration 1:   log likelihood = -1204.0802  
Iteration 2:   log likelihood = -1199.7943  

Performing gradient-based optimization: 

Iteration 0:   log likelihood = -1199.7943  
Iteration 1:   log likelihood = -1199.4697  
Iteration 2:   log likelihood = -1199.3158  
Iteration 3:   log likelihood =  -1199.315  
Iteration 4:   log likelihood =  -1199.315  

Mixed-effects logistic regression               Number of obs      =      1934
Group variable: district                        Number of groups   =        60

                                                Obs per group: min =         2
                                                               avg =      32.2
                                                               max =       118

Integration points =   7                        Wald chi2(5)       =     97.50
Log likelihood =  -1199.315                     Prob > chi2        =    0.0000

------------------------------------------------------------------------------
       c_use |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
       urban |   .8157872   .1715519     4.76   0.000     .4795516    1.152023
         age |   -.026415    .008023    -3.29   0.001    -.0421398   -.0106902
      child1 |    1.13252   .1603285     7.06   0.000     .8182819    1.446758
      child2 |   1.357739   .1770522     7.67   0.000     1.010724    1.704755
      child3 |   1.353827   .1828801     7.40   0.000     .9953881    1.712265
       _cons |   -1.71165   .1605617   -10.66   0.000    -2.026345   -1.396954
------------------------------------------------------------------------------

------------------------------------------------------------------------------
  Random-effects Parameters  |   Estimate   Std. Err.     [95% Conf. Interval]
-----------------------------+------------------------------------------------
district: Unstructured       |
                   sd(urban) |   .8162856   .1975237      .5080068     1.31164
                   sd(_cons) |   .6242944   .1035136       .451079    .8640248
           corr(urban,_cons) |   -.796473   .1151556     -.9361775   -.4394905
------------------------------------------------------------------------------
LR test vs. logistic regression:     chi2(3) =    58.42   Prob > chi2 = 0.0000

Note: LR test is conservative and provided only for reference.

For random intercept models, you will still want to use xtlogit since it will generally run faster. The parameter estimates for the same model running xtlogit and xtmelogit will not be identical because the default number of integration points for xtlogit is 12 while the default for xtmelogit is 7. You can set the number of integration points for either procedure, and with equal numbers of integration points you will receive identical estimates.

Other Improvements:

• Linear fixed effects model (xtreg ..., fe) now accepts aweights, fweights, and pweights.
• The robust and cluster(varname) have been replaced with vce(robust) and vce(cluster varname) for all xt commands. The vce(...) option also accepts other arguments which allow for a variety of vce types as well. The options vary based on which estimation command is being used.
• There have been various improvements to the numerical methods for estimating non-linear models using xt commands (e.g., xtlogit).
  • The default integration method for non-linear models is now mvaghermite (mean and variance adaptive Gauss-Hermite) as opposed to aghermite (mode and curvature adaptive Gauss-Hermite). Mvaghermite is performed first on every, and then on alternate iterations. aghermite is performed on the first iteration only.
  • The number of quadrature points that can be specified has increased (intpoints(#), up to 195 points).
  • Also the output of quadchk has been improved.
• xtdes has been renamed xtdescribe (although xtsdes will continue to function).