Title

    [R] stepwise -- Stepwise estimation


Syntax

        stepwise [, options ] : command

    options           description
    -------------------------------------------------------------------------
    Model
    * pr(#)           significance level for removal from the model
    * pe(#)           significance level for addition to the model

    Model2
      forward         perform forward-stepwise selection
      hierarchical    perform hierarchical selection
      lockterm1       keep the first term
      lr              perform likelihood ratio-test instead of Wald test
    -------------------------------------------------------------------------
    * At least one of pr(#) or pe(#) must be specified.
    by and xi are allowed; see prefix.
    Weights are allowed if command allows them; see weight.
    All postestimation commands behave as they would after command without
      the stepwise prefix; see the postestimation manual entry for command.


Menu

    Statistics > Other > Stepwise estimation


Description

    stepwise performs stepwise estimation.  Typing

        . stepwise, pr(#): command

    performs backward-selection estimation for command.  The stepwise
    selection method is determined by the following option combinations:

    options                   description
    -------------------------------------------------------------------------
    pr(#)                     backward selection
    pr(#) hierarchical        backward hierarchical selection
    pr(#) pe(#)               backward stepwise

    pe(#)                     forward selection
    pe(#) hierarchical        forward hierarchical selection
    pr(#) pe(#) forward       forward stepwise
    -------------------------------------------------------------------------

    command defines the estimation command to be executed.  The following
    Stata commands are supported by stepwise.

        clogit, cloglog, glm, intreg, logistic, logit, nbreg, ologit,
        oprobit, poisson, probit, qreg, regress, scobit, stcox, stcrreg,
        streg, tobit

    stepwise expects command to have the following form:

        command_name [depvar] term [term ...] [if] [in] [weight] [,
                     command_options]

    where term is either varname or (varlist) (a varlist in parentheses
    indicates that this group of variables is to be included or excluded
    together).  depvar is not present when command_name is streg, stcox, or
    stcrreg; otherwise, depvar is assumed to be present.  For intreg, depvar
    is actually two dependent variable names (depvar1 and depvar2).

    sw is a synonym for stepwise.


Options

        +-------+
    ----+ Model +------------------------------------------------------------

    pr(#) specifies the significance level for removal from the model; terms
        with p>=pr() are eligible for removal.

    pe(#) specifies the significance level for addition to the model; terms
        with p<pe() are eligible for addition.

        +--------+
    ----+ Model2 +-----------------------------------------------------------

    forward specifies the forward-stepwise method and may be specified only
        when both pr() and pe() are also specified.  Specifying both pr() and
        pe() without forward results in backward-stepwise selection.
        Specifying only pr() results in backward selection, and specifying
        only pe() results in forward selection.

    hierarchical specifies hierarchical selection.

    lockterm1 specifies that the first term be included in the model and not
        be subjected to the selection criteria.

    lr specifies the test of term significance be the likelihood-ratio test.
        The default is the less computationally expensive Wald test; that is,
        the test is based on the estimated variance-covariance matrix of the
        estimators.


Examples

    Setup
        . sysuse auto
        . generate weight2 = weight^2

    Backward selection
        . stepwise, pr(.2): regress mpg weight weight2 displ gear turn
            headroom foreign price

    Backward selection; consider engine displacement and gear ratio together
        . stepwise, pr(.2): regress mpg weight weight2 (displ gear) turn
            headroom foreign price

    Backward selection; force weight to be included in model
        . stepwise, pr(.2) lockterm1: regress mpg weight weight2 displ gear
            turn headroom foreign price

    Backward selection; force weight and weight2 to be included in model
        . stepwise, pr(.2) lockterm1: regress mpg (weight weight2) displ gear
            turn headroom foreign price

    Backward hierarchical selection
        . stepwise, pr(.2) hierarchical: regress mpg weight weight2 displ
            gear turn headroom foreign

    Forward selection
        . stepwise, pe(.2): regress mpg weight weight2 displ gear turn
            headroom foreign price

    Forward hierarchical selection
        . stepwise, pe(.2) hierarchical: regress mpg weight weight2 displ
            gear turn headroom foreign price


Programming for stepwise

    stepwise requires that command_name follow standard Stata syntax and
    allow the if qualifier.  Furthermore, command_name must have sw or swml
    as a program property; see [P] program properties.  If command_name has
    swml as a property, command_name must save the log-likelihood value in
    e(ll) and model degrees of freedom in e(df_m).


Saved results

    stepwise saves whatever is saved by the underlying estimation command.

    Also, stepwise saves stepwise in e(stepwise).


Also see