Confidemce-interval options for parmest and parmby
Syntax
options Description ------------------------------------------------------------------------- eform Exponentiate estimates and confidence limits dof(numscalar) Scalar degrees of freedom for calculating confidence limits level(numlist) Confidence level(s) for calculating confidence limits clnumber(numbering_rule) Numbering rule for naming confidence limit variables mcompare(method) Multiple-comparison method mcomci(method) Multiple-comparison method for confidence limits only bmatrix(matrix_expression) Matrix from which parameter estimates will be extracted vmatrix(matrix_expression) Matrix from which parameter variances will be extracted dfmatrix(matrix_expression) Matrix from which parameter degrees of freedom will be extracted -------------------------------------------------------------------------
where numscalar is
# | scalar_name
and numbering_rule is
level | rank
and method is
noadjust | bonferroni | sidak
Description
These options allow the user to change the selection of confidence limits and P-values in the output dataset (or resultsset) created by parmest or parmby.
Options
eform specifies that the estimates and confidence limits in the output dataset are to be exponentiated, and the standard errors multiplied by the exponentiated estimates. This option is usually used if the estimated parameters were calculated on a log scale, as is done by logit and logistic with odds ratios, by glm and binreg with risk ratios, by stcox with hazard ratios, or by regress with geometric mean ratios. Note that, if the user wants exponentiated confidence intervals in the output dataset, then the eform option must be specified for parmby or parmest, whether or not the eform or equivalent option was specified for the estimation command.
dof(numscalar) specifies the scalar degrees of freedom for t-distribution-based confidence limits and P-values. If dof() is positive, then confidence limits and P-values for all parameters are calculated using the t-distribution with dof() degrees of freedom. If dof() is zero, then confidence limits are calculated using the standard Normal distribution. If dof() is absent (or missing or negative), then confidence limits are calculated from the standard Normal or t-distribution, as follows. If the dfmatrix() option specifies a valid degrees of freedom matrix (see below), then the degrees of freedom are extracted from the specified matrix. Otherwise, if there is a non-missing scalar estimation result e(df_r), then the degrees of freedom for all parameters is set to the value of e(df_r). Otherwise, the confidence limits and P-values are calculated using the standard Normal distribution.
level(numlist) specifies the confidence levels, in percent, for the confidence limit variables created in the output dataset. These levels do not have to be integers, but must be at least 0 and strictly less than 100. For each level xx, parmest and parmby calculate a lower xx percent confidence limit variable with a default name of form minxx and an upper xx percent confidence level with a default name of form maxxx. The numbering of the confidence limit variable names can be changed using the clnumber option (see below), and the names of the confidence limits can be changed using the rename option (see parmest_varmod_opts). The default is level(95), or another single number set by set level. Note that the level() option used by parmby or parmest is not affected by any level() option specified for the estimation command. (See [U] 20 Estimation and postestimation commands.)
clnumber(numbering_rule) specifies the rule used to number the names of the confidence limit variables created in the output dataset. The numbering_rule may be level or rank, and is set in default to level if the clnumber() option is not specified. For each confidence level xx specified by the levels() option, parmest and parmby calculate a lower xx percent confidence limit with the default name minyy, and an upper xx percent confidence limit with the default name maxyy, where the number yy depends on the confidence level xx according to a rule specified by the numbering_rule of the clnumber() option. If the numbering_rule is rank, then the number yy is the rank, in ascending order, of the confidence level xx in the set of confidence levels specified by the levels() option. For instance, if the user specifies levels(90 95 99) clnumber(rank), then the 90 percent confidence limits are named min1 and max1, the 95 percent confidence limits are named min2 and max2, and the 99 percent confidence limits are named min3 and max3. If the numbering rule is level (the default), then the number yy is equal to the confidence level xx. For instance, if the user specifies levels(90 95 99) clnumber(lewel), then the 90 percent confidence limits are named min90 and max90, the 95 percent confidence limits are min95 and max95, and the 99 percent confidence limits are min99 and max99. If the confidence level xx contains a decimal point, then the decimal point is replaced with "_" in the variable names minxx and maxxx. If the confidence level xx contains "e-" (because of very small e-format confidence levels), then the "e-" is replaced with "em" in the variable names minxx and maxxx. Therefore, if the user specifies level(95 99.99) clnumber(level), then the output dataset contains 95% lower and upper confidence limits in variables min95 and max95, and 99.99% lower and upper confidence limits in variables min99_99 and max99_99. The option clnumber(rank) is useful if the confidence levels contain many numbers after the decimal point, which may be the case if the user specifies Bonferroni-corrected or Sidak-corrected confidence limits.
mcompare(method) specifies a multiple-comparison method used to adjust the generated confidence limits and P-values. This method may be noadjust (the default, indicating no adjustment), bonferroni (indicating the Bonferroni adjustment), and sidak (indicating the Sidak adjustment). The adjustments, if requested, are calculated for the total number of parameters estimated (in the case of parmest), or for the number of parameters estimated for the by-group (in the case of parmby). For this reason, the mcompare() and mcompci() options are not used very often with parmest and parmby, and are more likely to be used with parmcip in a derived resultsset, containing subsets of parameters from multiple models.
mcomci(method) specifies a multiple-comparison method used to adjust the generated confidence limits only, and not the generated P-values. If the user wants to generate adjusted P-values without adjusting the confidence limits, or to generate adjusted P-values using a different method from the one used for adjusting the confidence limits, then the user is advised to use the qqvalue package, which can be downloaded from SSC. Note that the mcompci() and mcompare() options do not affect the names of the generated variables containing the confidence limits, only their values.
bmatrix(matrix_expression) specifies the matrix from which the parameter estimates will be extracted. If not set by the user, then it is set to e(b) for most estimation commands, or to e(b_mi) if the most recent estimation command is mi estimate, or to e(est) if the most recent estimation command is one of the superseded Stata 8 survey commands svymean, svyratio or svytotal, and the command was specified with the available option instead of the complete option. The matrix specified must have one row, and one column per estimated parameter. The column names and equations of the matrix are used as the source for the parameter names and equations in the output dataset.
vmatrix(matrix_expression) specifies the matrix from which the parameter variances will be extracted. If not set by the user, then it is set to e(V) for most estimation commands, or to e(V_mi) if the most recent estimation command is mi estimate, or to e(V_db) if the most recent estimation command is one of the superseded Stata 8 survey commands svymean, svyratio or svytotal, and the command was specified with the available option instead of the complete option. The matrix specified must have as many columns as the matrix specified by bmatrix(), and must either have one row (from which the variances will then be extracted), or have as many rows as columns (in which case the variances will be extracted from the diagonal).
dfmatrix(matrix_expression) specifies the matrix from which the parameter degrees of freedom will be extracted, if no dof() option has been specified by the user. If neither dof() nor dfmatrix() has been specified by the user, then, for most estimation commands, the degrees of freedom for all parameters are extracted from the scalar e(df_r) if this result is not missing, and the standard Normal distribution is used otherwise. However, dfmatrix() is set in default to e(df_mi) if the most recent estimation command is mi estimate, or to e(_N_psu)-e(_N_str) if the most recent estimation command is one of the superseded Stata 8 survey commands svymean, svyratio or svytotal, and the command was specified with the available option instead of the complete option. The matrix specified must have one row, and must have either one column (from which degrees of freedom will be extracted for all parameters), or as many columns as the matrix specified by bmatrix(). Note that dfmatrix() is ignored if the user specifies dof().
Selection of distribution and degrees of freedom
parmest and parmby calculate confidence intervals and P-values from the parameter estimates and standard errors, using either the standard Normal distribution or the t-distribution for all parameters. If the t-distribution is used, then the degrees of freedom may or may not be the same for all parameters. The distribution, and degrees of freedom, are selected as follows:
1. By first preference, the dof() option is used, if specified by the user.
2. By second preference, the dfmatrix() option is used, if specified either by the user or by default.
3. By third preference, the degrees of freedom are specified by the scalar estimation result e(df_r), if that result is present.
4. If none of the above possibilities are available, then the standard Normal distribution is used.
Note that the user can force the use of the standard Normal distribution by specifying dof(0), or force the use of e(df_r) (if present) by specifying dof(e(df_r)).
If the t-distribution is used, then the degrees of freedom for each parameter are stored in the output dataset in the variable dof, and the t-test statistics are stored in the variable t. If the standard Normal distribution is used, then the output variable dof is not created, and the z-test statistics are stored in the output variable z.
Author
Roger Newson, Imperial College London, UK. Email: r.newson@imperial.ac.uk
Also see
Manual: [U] 20 Estimation and postestimation commands
Help: [U] 20 Estimation and postestimation commands parmest, parmby, parmest_outdest_opts, parmest_varadd_opts, parmest_varmod_opts, parmby_only_opts, parmest_resultssets qqvalue if installed