Title

                  Best fit model selection


Syntax

        bfit regress depvar indepvars [if] [in] [, corder(#) sort(bic|aic)
               options]

        bfit logit depvar indepvars [if] [in] [, corder(#) sort(bic|aic)
               options]

        bfit poisson depvar indepvars [if] [in] [, corder(#) sort(bic|aic)
               options]

    options                       Description
    -------------------------------------------------------------------------
      quantiles(numlist)          estimate specified quantiles

      vce(vcetype [, vceoptions]) vcetype may be bootstrap, analytic, or
                                    none.
                                  analytic is the default when quantiles() is
                                  not specified.  bootstrap is the default
                                  when quantiles() is specified.
                                  vceoptions vary over vcetype and are
                                  discussed below.

INCLUDE help shortdes-coeflegend
    -------------------------------------------------------------------------
    gpsvars and cvars may contain time-series operators; see fvvarlist.

Description

    bfit subcmd sorts a set of fitted candidate regression models by an
    information criterion, puts the best-fitting model in ereturn, and
    displays a table showing the ranking of the models fitted. The Bayesian
    information criterion (BIC) is the default and the Akaike information
    criterion (AIC) may optionally be specified as the ranking criterion.

    bfit subcmd sorts a set of fitted candidate regression models by bfit
    regress fits the candidate linear-regression models by ordinary least
    squares. bfit mlogit fits the candidate mulitinomial-logit models by
    maximum likelihood. bfit poisson fits the candidate poisson regression
    models by maximum likelihood.

    bfit subcmd sorts a set of fitted candidate regression models by For each
    subcmd, the candidate models are a series of polynomials in indepvars.
    The smallest of the candidate models includes only the first variable
    specified in indepvars. The largest of the candidate models is a
    fully-interacted polynomial of the order specified in corder(). See
    Methods and formulas in !! for details on the set of candidate models.

    Cattaneo, Drukker, and Holland (2012) provides an introduction to this
    command.

Options

    corder(#) specifies the maximum order of the covariate polynomial. The
        default is 2 which specifies a fully-interacted second-order
        polynomial.

    sort() specifies the information criterion by which the candidate models
        are to be sorted. sort(bic), the default, sorts the fitted candidate
        models by the Bayesian information criterion. sort(aic) sorts the
        fitted candidate models by the Akaike information criterion.

    coptions are passed to the estimation command. The allow options depend
        on the estimation command invoked by the subcommand. For example,
        base() may be specified only the with logit subcommand.  See regress,
        mlogit, and poisson for the allowable command options.


Examples

    ---------------------------------------------------------------------------
    Setup
        . use spmdata

    Model selection with logit
        . bfit logit w pindex eindex

    Model selection with logit up to a third-order model
        . bfit logit w pindex eindex, corder(3)

    Model selection with logit, and AIC selection
        . bfit logit w pindex eindex, sort(aic)

    Model selection with regress
        . bfit regress spmeasure pindex eindex

Saved results

    bift saves the following in r():

    Macros         
      r(subcmd)           regress, logit, or poisson
      r(bmodel)           Name of selected model in estimates store
      r(bvlist)           Variables in selected model
      r(sortby)           bic or aic

    Matrices       
      e(S)                Results for each model fit

    The matrix r(S) has 7 columns with the following model-specific
    information in each row:
        Column 1 contains the names of the model in estimates store
        Column 2 contains the number of observations in the sample
        Column 3 contains the value of the log-likelihood function for the
          constant-only model
        Column 4 contains the value of the log-likelihood function
        Column 5 contains the degrees of freedom in the model
        Column 6 contains the AIC
        Column 7 contains the BIC


References

    Cattaneo, M. D., D. M. Drukker, and A. Holland. 2012.  Estimation of
        multivalued treatment effects under conditional independence.
        Working paper, University of Michigan, Department of Economics, 
        http://www-personal.umich.edu/~cattaneo/papers/Cattaneo-Drukker-Holla
        > nd_2012_STATA.pdf.


Authors

    Matias D. Cattaneo, University of Michigan, Ann Arbor, MI.
        cattaneo@umich.edu.

    David M. Drukker, StataCorp, College Station, TX.  ddrukker@stata.com.

    Ashley D. Holland, Grace College, Winona Lake, IN.  hollana@grace.edu.