simul_these/Modules/ado/plus/i/ice.hlp

{smcl}
{* 30aug2005}{...}
{hline}
help for {hi:ice}, {hi:uvis}{right:(SJ5-4: st0067_2; SJ5-2: st0067_1; SJ4-3: st0067)}
{hline}

{title:Multiple imputation by the MICE system of chained equations}

{p 8 17 2}
{cmd:ice}
{it:mainvarlist}
{cmd:using} {it:filename}[{cmd:.dta}]
{ifin}
{weight}
[{cmd:,}
{cmdab:bo:ot}[{cmd:(}{it:varlist}{cmd:)}]
{cmd:cc(}{it:varlist}{cmd:)}
{cmdab:cm:d(}{it:cmdlist}{cmd:)}
{cmdab:cy:cles(}{it:#}{cmd:)}
{cmdab:dry:run}
{cmd:eq(}{it:eqlist}{cmd:)}
{cmdab:g:enmiss(}{it:string}{cmd:)}
{cmdab:i:d(}{it:string}{cmd:)}
{cmd:m(}{it:#}{cmd:)}
{cmdab:ma:tch}[{cmd:(}{it:varlist}{cmd:)}]
{cmdab:nocons:tant}
{cmdab:nosh:oweq}
{cmd:on(}{it:varlist}{cmd:)}
{cmdab:pass:ive(}{it:passivelist}{cmd:)}
{cmdab:sub:stitute(}{it:sublist}{cmd:)}
{cmd:replace}
{cmdab:se:ed(}{it:#}{cmd:)}
{cmdab:tr:ace(}{it:filename}{cmd:)}]

{p 8 17 2}
{cmd:uvis}
{it:regression_cmd}
{it:yvar}
{it:xvarlist}
{ifin}
{weight}
{cmd:,}
{cmdab:g:en(}{it:newvarname}{cmd:)}
[{cmdab:bo:ot}
{cmdab:ma:tch}
{cmdab:nocons:tant}
{cmd:replace}
{cmdab:se:ed(}{it:#}{cmd:)}]

{p 4 4 2}
where

{p 8 8 2}
{it:regression_cmd} may be
{helpb logistic},
{helpb logit},
{helpb mlogit},
{helpb ologit},
or
{helpb regress}.

{p 4 4 2}
All weight types supported by {it:regression_cmd} are allowed; see {help weight}.


{title:Description}

{p 4 4 2}
{cmd:ice} imputes missing values
in {it:mainvarlist} by using switching regression, an iterative multivariable
regression technique. The abbreviation MICE means multiple imputation by
chained equations and was apparently coined by Steff van Buuren. {cmd:ice}
implements MICE for Stata. Sets of imputed and nonimputed variables are
stored to a new file called {it:filename}. Any number of complete imputations
may be created.

{p 4 4 2}
{cmd:uvis} (univariate imputation sampling) imputes missing values in the
single variable {it:yvar} based on multiple regression on {it:xvarlist}.
{cmd:uvis} is called repeatedly by {cmd:ice} in a regression switching mode to
perform multivariate imputation.

{p 4 4 2}
The missing observations are assumed to be missing at random (MAR) or
missing completely at random (MCAR), according to the jargon. See, for
example, van Buuren et al. (1999) for an explanation of these concepts.

{p 4 4 2}
Please note that {cmd:ice} and {cmd:uvis} require Stata 8 or later.
There have been incompatibility issues with Stata 7 and earlier.


{title:Options for ice}

{p 4 8 2}
{cmd:boot}[{cmd:(}{it:varlist}{cmd:)}] instructs that each member of
{it:varlist}, a subset of {it:mainvarlist}, be imputed with the {cmd:boot}
option of {cmd:uvis} activated. If {cmd:(}{it:varlist}{cmd:)} is omitted,
all members of {it:mainvarlist} with missing observations are imputed using
the {cmd:boot} option of {cmd:uvis}.

{p 4 8 2}
{cmd:cc(}{it:varlist}{cmd:)} prevents imputation of missing data in
{it:mainvarlist} for cases in which any member of {it:varlist} has a missing
value. "cc" signifies "complete case". Note that members of {it:varlist} are
used for imputation if they appear in {it:mainvarlist}, but not otherwise. Use
of this option is equivalent to entering {cmd:if}
{cmd:~missing(}{it:var1}{cmd:) &} {cmd:~missing(}{it:var2}{cmd:)} ..., where
{it:var1}, {it:var2}, ... denote the members of {it:varlist}.

{p 4 8 2}
{cmd:cmd(}{it:cmdlist}{cmd:)} defines the regression commands to be used for
each variable in {it:mainvarlist}, when it becomes the dependent variable in
the switching regression procedure used by {cmd:uvis} (see {hi:Remarks}).  The
first item in {it:cmdlist} may be a command, such as {cmd:regress}, or may have
the syntax {it:varlist}{cmd::}{it:cmd}, specifying that command {it:cmd}
applies to all the variables in {it:varlist}.  Subsequent items in
{it:cmdlist} must follow the latter syntax, and each item should be followed
by a comma.

{p 8 8 2}
The default {it:cmd} for a variable is {cmd:logit} when there are two distinct
values, {cmd:mlogit} when there are 3-5 and {cmd:regress} otherwise.

{p 8 18 2} Example:  {cmd:cmd(regress)} specifies that all variables are 
to be imputed by {cmd:regress}, overriding the defaults.

{p 8 18 2} Example:  {cmd:cmd(x1 x2:logit, x3:regress)} specifies that
{cmd:x1} and {cmd:x2} are to be imputed by {cmd:logit}, {cmd:x3} by
{cmd:regress} and all others by their default choices.

{p 4 8 2}
{cmd:cycles(}{it:#}{cmd:)} determines the number of cycles of regression
switching to be carried out. The default is {cmd:cycles(10)}.

{p 4 8 2}
{cmd:dryrun} does a "dry run"; that is, {cmd:ice} 
reports the prediction equations it has constructed from the various
inputs. No imputation is done, and no files are created. It is not
mandatory to specify an output file with {cmd:using} for a dry run.
Sometimes the prediction equation set-up needs to be carefully
checked before running what may be a lengthy imputation process.

{p 4 8 2}
{cmd:eq(}{it:eqlist}{cmd:)} allows one to define customized prediction
equations for any subset of variables in {it:mainvarlist}. The option,
particularly when used with {cmd:passive()}, allows
great flexibility in the possible imputation schemes. The
syntax of {it:eqlist} is {it:varname1}{cmd::}{it:varlist1}
[{cmd:,}{it:varname2}{cmd::}{it:varlist2} ...], where each
{it:varname#} (or {it:varlist#})
is a member (or subset) of {it:mainvarlist}. It is your responsibility to ensure
that each equation is sensible. {cmd:ice} places no restrictions
except to check that all variables mentioned are indeed in
{it:mainvarlist} and that an equation is not defined
for a variable specified to be passively imputed
(see the {cmd:passive()} option. Note that {cmd:eq()} takes
precedence over all default definitions and assumptions about 
the way a given variable in {cmd:mainvarlist} will be imputed.
The default, if the {cmd:passive()} and {cmd:substitute()}
options are not invoked, is that each
variable in {it:mainvarlist} with any missing data is imputed from all
the other variables in {it:mainvarlist}.

{p 4 8 2}
{cmd:genmiss(}{it:string}{cmd:)} creates an indicator variable for the
missingness of data in any variable in {it:mainvarlist} for which at least one
value has been imputed. The indicator variable is set to missing for
observations excluded by {cmd:if}, {cmd:in}, etc.  The indicator variable for
{it:xvar} is named {it:string}{it:xvar}.

{p 4 8 2}
{cmd:id(}{it:string}{cmd:)} creates a variable called {it:string} containing
the original sort order of the data. The default {it:string} is {cmd:_i}.

{p 4 8 2}
{cmd:m(}{it:#}{cmd:)} defines {it:#} as the number of imputations required
(minimum 1, no upper limit). The default is {cmd:m(1)}.

{p 4 8 2}
{cmd:match}[{cmd:(}{it:varlist}{cmd:)}] instructs that each member of
{it:varlist} be imputed with the {cmd:match} option of {cmd:uvis}.
This provides prediction matching for each member of {it:varlist}.
If {cmd:(}{it:varlist}{cmd:)} is omitted then all relevant variables are
imputed with the {cmd:match} option of {cmd:uvis}. The default, if
{cmd:match()} is not specified, is to draw from the posterior
predictive distribution of each variable requiring imputation.

{p 4 8 2}
{cmd:noconstant} suppresses the regression constant in all regressions.

{p 4 8 2}
{cmd:noshoweq} suppresses the presentation of the prediction equations.

{p 4 8 2}
{cmd:on(}{it:varlist}{cmd:)} changes the operation of {cmd:ice} in a major
way.  With this option, {cmd:uvis} imputes each member of {it:mainvarlist}
univariately on {it:varlist}. This provides a convenient way of producing
multiple imputations when imputation for each variable in {it:mainvarlist} is
to be done univariately on a set of complete predictors.

{p 4 8 2}
{cmd:passive(}{it:passivelist}{cmd:)} allows the use of "passive" imputation
of variables that depend on other variables, some of which are imputed.
The syntax of {it:passivelist} is {it:varname}{cmd::}{it:exp}
[{cmd:\}{it:varname}{cmd::}{it:exp} ...]. Notice the requirement to use
"\" as a separator between items in {it:passivelist}, rather than the usual comma;
the reason is that a comma may be a valid part of an expression.
The option is most easily explained by example. Suppose x1 is a categorical variable
with 3 levels, and that two dummy variables x1a, x1b have been created by the commands

{p 8 8 2}
     {cmd:. generate byte x1a=(x1==2)}{break}
     {cmd:. generate byte x1b=(x1==3)}

{p 8 8 2}
Now suppose that x1 is to be imputed by the {cmd:mlogit} command and is
to be treated as the two dummy variables x1a and x1b when predicting other
variables.  Use of {cmd:mlogit} is achieved by the option
{cmd:cmd(x1:mlogit)}.  When x1 is imputed, we want x1a and x1b to be updated
with new values which depend on the imputed values of x1.  This may be
achieved by specifying {cmd:passive(x1a:x1==2 \ x1b:x1==3)}. It is necessary
also to remove x1 from the list of predictors when variables other than x1 are
being imputed, and this is done by using the {cmd:substitute()} option; in the
present example, you would specify {cmd:substitute(x1:x1a x1b)}.

{p 8 8 2}
Note that although in this example x1a will take the (possibly
unintended) value of 0 when x1 is missing, {cmd:ice} is careful to
ensure that x1a (and x1b) inherit the missingness of x1 and are
passively imputed following active imputation of missing values
of x1. If this were not done, incorrect results could occur. The
responsibility of the user is to create x1a and x1b before running
{cmd:ice} such that their missing values are identical
to those of x1.

{p 8 8 2}
A second example is multiplicative interactions between variables, for
example, between x1 and x2 (e.g., x12=x1*x2); this could be entered as
{cmd:passive(x12:x1*x2)}. It would cause the interaction term
x12 to be omitted when either x1 or x2 was being imputed, since it would
make no sense to impute x1 from its interaction with x2.
{cmd:substitute()} is not needed here.

{p 8 8 2}
It should be stressed that variables to be imputed passively must already
exist and must be included in {it:mainvarlist}; otherwise, they will not be
recognized.

{p 4 8 2}
{cmd:substitute(}{it:sublist}{cmd:)} is typically used with the 
{cmd:passive()} option to represent multilevel categorical variables
as dummy variables in models for predicting other variables. See
{cmd:passive()} for more details. The syntax of {it:sublist} is
{it:varname}{cmd::}{it:dummyvarlist}
[{cmd:,}{it:varname}{cmd::}{it:dummyvarlist} ...], where {it:varname} is the
name of a variable to be substituted and {it:dummyvarlist} is the list of
dummy variables representing it.

{p 4 8 2}
{cmd:replace} permits {it:filename} to be overwritten with new data.

{p 4 8 2}
{cmd:seed(}{it:#}{cmd:)} sets the random-number seed to {it:#}.
To reproduce a set of imputations, the same random-number seed should be used. 
The default is {cmd:seed(0)}, meaning no seed is set by the program.

{p 4 8 2}
{cmd:trace(}{it:filename}{cmd:)} monitors the convergence of the imputation
algorithm. For each original variable with missing values, the mean of the
imputed values is stored as a variable in {it:filename}, together
with the cycle number at which that
mean was calculated. The results are stored only for the final imputation.
For diagnostic purposes, it is sensible to run {cmd:trace()}
with {cmd:m(1)} and many cycles, such as {cmd:cycles(100)}.
When the run is complete, it is helpful to load {it:filename}
into memory and plot the mean for each imputed
variable against the cycle number. If necessary, smoothing may be applied
to clarify any apparent pattern. Convergence is judged to have occurred
when the pattern of the imputed means is random.
The number of cycles needed for convergence is usually obvious from the appearance
of the plot.


{title:Options for uvis}

{p 4 8 2}
{cmd:gen(}{it:newvar}{cmd:)} is not optional. {it:newvar} contains original
(nonmissing) and imputed (originally missing) values of {it:yvar}.

{p 4 8 2}
{cmd:boot} invokes a bootstrap method for creating imputed values (see Remarks).
 
{p 4 8 2}
{cmd:match} creates imputations by prediction matching. The default is to draw
imputations at random from the posterior distribution of the missing values of
{it:yvar}, conditional on the observed values and the members of
{it:xvarlist}. See Remarks for further details.

{p 4 8 2}
{cmd:noconstant} suppresses the regression constant in all regressions.

{p 4 8 2}
{cmd:replace} permits {it:newvar} (see {cmd:gen(}{it:newvar}{cmd:)})
to be overwritten with new data. {cmd:replace} may not be abbreviated.

{p 4 8 2}
{cmd:seed(}{it:#}{cmd:)} sets the random-number seed to {it:#}.
See {hi:Remarks} for comments on how to ensure reproducible imputations
by using the {cmd:seed()} option.
The default is {cmd:seed(0)}, meaning no seed is set by the program.


{title:Remarks}

{p 4 4 2}
{cmd:uvis} imputes {it:yvar} from {it:xvarlist} according to the following
algorithm (see van Buuren et al. (1999, section 3.2) for further technical
details):

{p 8 12 2}
1. Estimate the vector of coefficients (beta) and the residual variance
by regressing the nonmissing values of {it:yvar} on the current "completed"
version of {it:xvarlist}. Predict the fitted values {it:etaobs} at the
nonmissing observations of {it:yvar}.

{p 8 12 2}
2. Draw at random a value (sigma_star) from the posterior distribution of the
residual standard deviation.

{p 8 12 2}
3. Draw at random a value (beta_star) from the posterior distribution of beta,
allowing, through sigma_star, for uncertainty in beta.

{p 8 12 2}
4. Use beta_star to predict the fitted values {it:etamis}
at the missing observations of {it:yvar}.

{p 8 12 2}
5. The imputed values are predicted directly from beta_star, sigma_star and
the covariates. When imputation is by linear regression ({cmd:regress}
command), this step assumes that {it:yvar} is Normally distributed, given the
covariates.  For other types of imputation, samples are drawn from the
appropriate distribution.

{p 4 4 2}
With the {cmd:match} option, step 5 is replaced by the following.
For each missing observation of {it:yvar} with prediction {it:etamis},
find the non-missing observation of {it:yvar} whose prediction
({it:etaobs}) on observed data is closest to {it:etamis}. This closest
non-missing observation is used to impute the missing value of {it:yvar}.

{p 4 4 2}
The default draw method is not robust to departures from Normality and
may produce implausible imputations. For example, if the original distribution
is skew and positive-valued, the imputed distribution will not necessarily
have the appropriate amount of skewness, nor will all the imputed values
necessarily be positive. Log transformation of positive variables may greatly
improve the appropriateness of the imputations.

{p 4 4 2}
The alternative {cmd:match} method is recommended only for continuous variables
when the Normality assumption is clearly untenable, even approximately.
It is not necessary, nor is it recommended, for binary, ordered categorical or
nominal variables. {cmd:match} may work well when the distribution of a
continuous variable is very non-Normal, but it may sometimes result in biased
imputations.

{p 4 4 2}
With the {cmd:boot} option, steps 2-4 are replaced by a bootstrap estimation of
beta_star; beta_star
is estimated by regressing {it:yvar} on {it:xvarlist} after taking a bootstrap sample
of the non-missing observations. This has the advantage of robustness since the
distribution of beta is no longer assumed to be multivariate normal.

{p 4 4 2}
Note that {cmd:uvis} will not impute observations for which a value 
of a variable in {it:xvarlist} is missing. However, all original
(missing or nonmissing) observations of {it:yvar} will be copied into
{it:newvarname} in such cases. This is a change from the first release of
{cmd:uvis} (with {cmd:mvis}). Previously, {it:newvarname} would be set to
missing whenever a value of a variable in {it:xvarlist} was missing,
irrespective of the value of {it:yvar}.

{p 4 4 2}
Missing data for ordered (or unordered) categorical covariates should
be imputed by using the {cmd:ologit} (or {cmd:mlogit}) command. In these cases,
prediction matching is done on the scale of the mean absolute difference
in the predicted class probabilities, preceded by logit transformation.

{p 4 4 2}
{cmd:ice} carries out multivariate imputation in {it:mainvarlist} using
regression switching (van Buuren et al. 1999) as follows:

{p 8 12 2}
1. Ignore any observations for which {it:mainvarlist} has only missing values,
   or if the {cmd:ccvarlist(}{it:varlist}{cmd:)} option has been specified, for
   which any member of {it:varlist} has a missing value.

{p 8 12 2}
2. For each variable in {it:mainvarlist} with any missing data, randomly order
   that variable and replicate the observed values across the missing cases.
   This step initializes the iterative procedure by ensuing that no relevant
   values are missing.

{p 8 12 2}
3. For each variable in {it:mainvarlist} in turn, impute missing values by
   applying {cmd:uvis} with the remaining variables as covariates.

{p 8 12 2}
4. Repeat step 3 {cmd:cycles()} times, replacing the imputed values with updated
   values at the end of each cycle.

{p 4 4 2}
A single imputation sample is created for each variable with any relevant
missing values.

{p 4 4 2}
Van Buuren recommends {cmd:cycles(20)} but goes on to say that 10 or even 5
iterations are probably sufficient. We have chosen a compromise default of 10.

{p 4 4 2}
"Multiple imputation" (MI) implies the creation and analysis of several
imputed datasets. To do this, one would run {cmd:ice} with {it:m} set
to a suitable number, for example 5. To obtain final estimates
of the parameters of interest and their standard errors,
one would fit a model in 
each imputation and carry out the appropriate post-MI averaging procedure
on the results from the {it:m} separate imputations. A suitable
estimation tool for this purpose is {helpb micombine}.

{title:Handling categorical variables}

{p 4 4 2}
Binary variables present no difficulty: by default, in the MICE
procedure, when such a variable is the response, it is
predicted from other variables by using logistic regression;
when it is a covariate, it is modeled in the only way possible, 
effectively as a single dummy variable. Categorical variables with 3 or
more levels may in principle be treated in different ways.
By default, in {cmd:ice} variables with 3-5 levels are modeled
using multinomial logistic regression ({cmd:mlogit} command) when
the response, and as a single linear term when a covariate. The
same behavior occurs with the ordered logistic model ({cmd:ologit}
command), requested via the {cmd:cmd()} option. The use of dummy variables
instead of a single linear term may be imposed as described under
the {cmd:passive()} option. The requisite dummy variables
must be created before {cmd:ice} is invoked. Variables with 6 or
more levels are treated as ordered and continuous, but again 
different choices may be imposed by use of the {cmd:cmd()},
{cmd:passive()} and {cmd:substitute()} options.

{p 4 4 2}
You should be aware that
unless the dataset is large, use of the {cmd:mlogit} command may produce
unstable estimates if the number of levels is too large, and
may compromise the accuracy of the imputations. It is hard to
predict when this will occur.

{p 4 4 2}
Note that due to a peculiarity of the way the {cmd:mlogit} command works,
variables with score labels cause problems to {cmd:ice} 
and {cmd:uvis} when missing data are imputed using {cmd:mlogit}. 
Score labels for such variables are removed in the file of imputed
data. See also the related comment on {hi:Postestimation prediction} in
{helpb micombine}.


{title:Further notes}

{p 4 4 2}
{cmd:ice} determines the order of imputing variables in the round
of chained equations according to the amount of missing data.
Variables with the least missingness are imputed first.

{p 4 4 2}
An important application of MI is to investigate possible models, for example
prognostic models, in which selection of influential variables is required
(Clark and Altman 2003). For example, the stability of the final model across
the imputation samples is of interest. This area of inquiry is in its infancy.

{p 4 4 2}
In survival analysis, it is recommended to include the censoring indicator 
and the log of the survival time in the variables to be used for imputation.
Van Buuren et al. (1999) give a detailed discussion of the different types
of covariate that can be included in the imputation model and discuss the
important issue of how to deal with variables which are missing completely at
random (MCAR), missing at random (MAR), and missing not at random (MNAR).

{p 4 4 2}
See also Van Buuren's web site http://www.multiple-imputation.com for further
information and software sources.


{title:Examples}

{p 4 10 2}
{cmd:. uvis regress y x1 x2 x3, gen(ym)}

{p 4 10 2}
{cmd:. ice x1 x2 x3 using imputed, m(5)}

{p 4 10 2}
{cmd:. ice x1 x2 x3 using imputed, m(5) cycles(20) cc(x4 x5)}

{p 4 10 2}
{cmd:. ice x1-x5 using imputed, m(10) boot match(x1 x2 x3) cmd(x1 x2:mlogit, x3:ologit) id(pid) seed(101) genmiss(m_)}

{p 4 10 2}
{cmd:. ice x1 x1a x1b x2 x3 x23 using imputed, m(5) cmd(x1:ologit) passive(x1a:x1==2 \x1b:x1==3 \x23=x2*x3) substitute(x1:x1a x1b)}

{p 4 10 2}
{cmd:. ice y1 y2 y3 x1 x2 x3 x4 using imputed, m(5) eq(y1:x1 x2 y2, y2:y1 x3 x4, y3:y1 y2) match(y3)}


{title:Acknowledgement}

{p 4 4 2}
I am grateful to Gillian Raab for pointing out certain issues with the prediction
matching approach, particularly that it is only useful with continuous variables.
As a result, the default imputation method has been
changed from matching to drawing from the predictive distribution. Gillian also
suggested imputing the variables in reverse order of the amount of missingness,
and selecting the imputed value at random from the set determined by the available
matching predictions. Both suggestions have been implemented in this software update. 


{title:Author}

{p 4 4 2}
Patrick Royston, MRC Clinical Trials Unit, London.{break}
patrick.royston@ctu.mrc.ac.uk


{title:References}

{p 4 8 2}
van Buuren S., H. C. Boshuizen and D. L. Knook. 1999. Multiple imputation of
    missing blood pressure covariates in survival analysis.
    {it:Statistics in Medicine} {cmd:18}:681-694. 
    Also see http://www.multiple-imputation.com.

{p 4 8 2}
Carlin  J. B., N. Li, P. Greenwood, and C. Coffey. 2003. Tools for analyzing
multiple imputed datasets. {it:Stata Journal} 3(3): 226-244.

{p 4 8 2}
Clark T. G. and D. G. Altman. 2003. Developing a prognostic model
in the presence of missing data: an ovarian cancer case-study.
{it:Journal of Clinical Epidemiology} 56: 28-37.

{p 4 8 2}
Royston P. 2004. Multiple imputation of missing values.
{it:Stata Journal} 4(3): 227-241.


{title:Also see}

{psee}
Online:  {helpb mijoin}, {helpb micombine}, {helpb mitools}, and related programs,
if installed
{p_end}