[R] fit.mult.impute() in Hmisc

[Paul Johnson]

On Thu, Mar 31, 2011 at 2:56 PM, Yuelin Li <liy12 at mskcc.org> wrote:
> I tried multiple imputation with aregImpute() and
> fit.mult.impute() in Hmisc 3.8-3 (June 2010) and R-2.12.1.
>
> The warning message below suggests that summary(f) of
> fit.mult.impute() would only use the last imputed data set.
> Thus, the whole imputation process is ignored.
>
>  "Not using a Design fitting function; summary(fit)
>   will use standard errors, t, P from last imputation only.
>   Use vcov(fit) to get the correct covariance matrix,
>   sqrt(diag(vcov(fit))) to get s.e."
>

Hello.  I fiddled around with rms & multiple imputation when I was
preparing these notes from our R summer course. I ran into that same
thing you did, and my conclusion is slightly different from yours.

http://pj.freefaculty.org/guides/Rcourse/multipleImputation/multipleImputation-1-lecture.pdf

Look down to slide 80 or so, where I launch off into that question.
It appears to me that aregImpute will give the "right answer" for
fitters from rms, but if you want to feel confident about the results
for other fitters, you should use mitools or some other paramater
combining approach. My conclusion (slide 105) is

"Please note: the standard errors in the output based on lrm match
the std.errors estimated by MItools. Thus I conclude
sqrt(diag(cov(fit.mult.impute.object) did not give correct results"




> But the standard errors in summary(f) agree with the values
> from sqrt(diag(vcov(f))) to the 4th decimal point.  It would
> seem that summary(f) actually adjusts for multiple
> imputation?
>
> Does summary(f) in Hmisc 3.8-3 actually adjust for MI?
>
> If it does not adjust for MI, then how do I get the
> MI-adjusted coefficients and standard errors?
>
> I can't seem to find answers in the documentations, including
> rereading section 8.10 of the Harrell (2001) book  Googling
> located a thread in R-help back in 2003, which seemed dated.
> Many thanks in advance for the help,
>
> Yuelin.
> http://idecide.mskcc.org
> -------------------------------
>> library(Hmisc)
> Loading required package: survival
> Loading required package: splines
>> data(kyphosis, package = "rpart")
>> kp <- lapply(kyphosis, function(x)
> +       { is.na(x) <- sample(1:length(x), size = 10); x })
>> kp <- data.frame(kp)
>> kp$kyp <- kp$Kyphosis == "present"
>> set.seed(7)
>> imp <- aregImpute( ~ kyp + Age + Start + Number, dat = kp, n.impute = 10,
> +                      type = "pmm", match = "closest")
> Iteration 13
>> f <- fit.mult.impute(kyp ~ Age + Start + Number, fitter=glm, xtrans=imp,
> +                 family = "binomial", data = kp)
>
> Variance Inflation Factors Due to Imputation:
>
> (Intercept)         Age       Start      Number
>       1.06        1.28        1.17        1.12
>
> Rate of Missing Information:
>
> (Intercept)         Age       Start      Number
>       0.06        0.22        0.14        0.10
>
> d.f. for t-distribution for Tests of Single Coefficients:
>
> (Intercept)         Age       Start      Number
>    2533.47      193.45      435.79      830.08
>
> The following fit components were averaged over the 10 model fits:
>
>  fitted.values linear.predictors
>
> Warning message:
> In fit.mult.impute(kyp ~ Age + Start + Number, fitter = glm, xtrans = imp,  :
>  Not using a Design fitting function; summary(fit) will use
> standard errors, t, P from last imputation only.  Use vcov(fit) to get the
> correct covariance matrix, sqrt(diag(vcov(fit))) to get s.e.
>
>
>> f
>
> Call:  fitter(formula = formula, family = "binomial", data = completed.data)
>
> Coefficients:
> (Intercept)          Age        Start       Number
>    -3.6971       0.0118      -0.1979       0.6937
>
> Degrees of Freedom: 80 Total (i.e. Null);  77 Residual
> Null Deviance:      80.5
> Residual Deviance: 58   AIC: 66
>> sqrt(diag(vcov(f)))
> (Intercept)         Age       Start      Number
>  1.5444782   0.0063984   0.0652068   0.2454408
>> -0.1979/0.0652068
> [1] -3.0350
>> summary(f)
>
> Call:
> fitter(formula = formula, family = "binomial", data = completed.data)
>
> Deviance Residuals:
>   Min      1Q  Median      3Q     Max
> -1.240  -0.618  -0.288  -0.109   2.409
>
> Coefficients:
>            Estimate Std. Error z value Pr(>|z|)
> (Intercept)  -3.6971     1.5445   -2.39   0.0167
> Age           0.0118     0.0064    1.85   0.0649
> Start        -0.1979     0.0652   -3.03   0.0024
> Number        0.6937     0.2454    2.83   0.0047
>
> (Dispersion parameter for binomial family taken to be 1)
>
>    Null deviance: 80.508  on 80  degrees of freedom
> Residual deviance: 57.965  on 77  degrees of freedom
> AIC: 65.97
>
> Number of Fisher Scoring iterations: 5
>
>
>     =====================================================================
>
>     Please note that this e-mail and any files transmitted with it may be
>     privileged, confidential, and protected from disclosure under
>     applicable law. If the reader of this message is not the intended
>     recipient, or an employee or agent responsible for delivering this
>     message to the intended recipient, you are hereby notified that any
>     reading, dissemination, distribution, copying, or other use of this
>     communication or any of its attachments is strictly prohibited.  If
>     you have received this communication in error, please notify the
>     sender immediately by replying to this message and deleting this
>     message, any attachments, and all copies and backups from your
>     computer.
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>



-- 
Paul E. Johnson
Professor, Political Science
1541 Lilac Lane, Room 504
University of Kansas