What’s new in multiple imputation
New imputation features
- The mi impute command now supports multivariate imputation using
chained equations (ICE), mi impute chained, also known as
sequential regression multivariate imputation (SRMI).
ICE is a flexible imputation technique for imputing various types of
data. The variable-by-variable specification of ICE allows you to impute
variables of different types by choosing from several univariate
imputation methods the appropriate one for each variable. Variables can
have an arbitrary missing-data pattern. By specifying a separate model
for each variable, you can incorporate certain important characteristics,
such as ranges and restrictions within a subset, specific to each
variable.
Use any of nine univariate imputation methods to build a flexible
imputation model.
Customize prediction equations for imputed variables (for example, omit
hsgrad from the model for bmi).
Impute some variables using conditional imputation.
Allow general expressions of imputed variables in the equations of other
imputed variables (such as include bmi^2 in age’s
imputation model).
- There are four new univariate imputation methods that can also be used as
building blocks for multivariate imputation using the monotone method or
the new chained-equations method.
- mi impute truncreg imputes continuous variables with a
restricted range using the truncated regression method
- mi impute intreg imputes continuous censored variables using
the interval regression method
- mi impute poisson imputes count variables using the Poisson
method
- mi impute nbreg imputes overdispersed count variables using
the negative binomial method
- Conditional imputation is now supported with all
imputation techniques, except multivariate normal imputation (MVN), via the
conditional() option.
Conditional imputation allows you to impute variables which are defined
within a particular subset of the data. Outside that subset, the variables
are known to be constant. For example, the number of pregnancies is
relevant only to females and is always zero for males, the smoking of
high-tar cigarettes is relevant only to smokers and is always zero for
nonsmokers.
To properly impute, say, whether a person smokes high-tar cigarettes, we
should condition on the smoking status during imputation. That is, we
should impute missing values of subjects who smoke using only data on
smokers and replace missing values of subjects who do not smoke with zeros.
We would like to be able to do that even when the smoking status contains
missing values and is itself being imputed. The conditional() option
provides such capability:
- Separate imputation for different groups of the data is now available via
mi impute's new by() option.
- Imputation by drawing posterior estimates from bootstrapped samples is
now available with all imputation techniques, except MVN, via the new
bootstrap option.
- Perfect prediction is now handled during imputation of categorical data
using logistic, ordered logistic, or multinomial logistic imputation
methods when the new augment option is specified.
- mi impute is now faster in the wide, mlong, and flong styles.
New estimation and postestimation features
- Estimate the amount of simulation error in your final model, so you can
decide whether you need more imputations using mi estimate’s
new mcerror option.
- mi estimate now supports panel-data and multilevel models. Included
are xtcloglog, xtgee, xtlogit, xtmelogit,
xtmepoisson, xtmixed, xtnbreg, xtpoisson,
xtprobit, xtrc, and xtreg.
- mi estimate now supports the total command.
- Compute linear and nonlinear predictions after MI estimation using new
commands mi predict and mi predictnl.
New data-management features
- misstable summarize will now create summary variables recording
the missing-values pattern via the new generate() option.
Back to highlights
See New in Stata 18 to learn about what was added in Stata 18.
