Creating tables of model estimates¶
The “mtable” function¶
In conjunction with LaTeX, the output of the mtable()
function of the package can produce output as can be seen in the following example (which is LaTeX converted into a pngfile):
This LaTeX output was generated by the following code:
# First a couple of models are run:
model1 < glm((vote3=="Truman")~occup4,data=vote.48,
family="binomial")
model2 < glm((vote3=="Truman")~total.income,data=vote.48,
family="binomial")
model3 < glm((vote3=="Truman")~occup4+total.income,data=vote.48,
family="binomial")
model4 < glm((vote3=="Truman")~relig3,data=vote.48,
family="binomial")
model5 < glm((vote3=="Truman")~occup4+relig3,data=vote.48,
family="binomial")
# then an mtable object is produced, relabelled and converted
# to LaTeX:
mt145 < mtable("Model 1"=model1,
"Model 4"=model4,
"Model 5"=model5,
summary.stats=c("Nagelkerke Rsq.","Deviance","AIC","N")
)
mt145 < relabel(mt145,
"[(]Intercept[)]"="\\\\emph{Intercept}",
occup4="Occup. class",
relig3="Religion",
gsub=TRUE)
ltx.m145 < toLatex(mt145,ddigits=5)
writeLines(ltx.m145,
con="somewhere/mt145.tex")
Creating of this output involves the following steps: First, mtable()
is called
mt145 < mtable("Model 1"=model1,
"Model 4"=model4,
"Model 5"=model5,
summary.stats=c("Nagelkerke Rsq.","Deviance","AIC","N")
)
The resulting object, if printed to the Rconsole, looks like so:
print(mt145)
Calls:
Model 1: glm(formula = (vote3 == "Truman") ~ occup4, family = "binomial",
data = vote.48)
Model 4: glm(formula = (vote3 == "Truman") ~ relig3, family = "binomial",
data = vote.48)
Model 5: glm(formula = (vote3 == "Truman") ~ occup4 + relig3, family = "binomial",
data = vote.48)
==================================================================================
Model 1 Model 4 Model 5

(Intercept) 1.401*** 0.213 1.687***
(0.271) (0.126) (0.294)
occup4: Other white collar/Upper white collar 1.368*** 1.287***
(0.373) (0.381)
occup4: Blue collar/Upper white collar 2.448*** 2.385***
(0.327) (0.337)
occup4: Farmer/Upper white collar 1.826*** 2.039***
(0.413) (0.426)
relig3: Catholic/Protestant 0.877*** 0.685*
(0.243) (0.292)
relig3: Other,none/Protestant 0.975** 1.191**
(0.347) (0.441)

Nagelkerke Rsq. 0.2 0.1 0.3
Deviance 404.2 537.7 393.1
AIC 412.2 543.7 405.1
N 344 402 344
==================================================================================
Second, the result of mtable()
is “embellished” using a call to relabel()
:
mt145 < relabel(mt145,
"[(]Intercept[)]"="\\\\emph{Intercept}",
occup4="Occup. class",
relig3="Religion",
gsub=TRUE)
Third, the relabelled objected is converted into LaTeX and written to a file:
ltx.m145 < toLatex(mt145,ddigits=5)
writeLines(ltx.m145,
con="somewhere/mt145.tex")
mtable()
can also be used to generate HTMLformatted tables of estimates that can be included into wordprocessor software such as LibreOffice or MSWord.
# We get rid of the LaTeX formatting ...
mt145 < relabel(mt145,
"\\emph{Intercept}"="<em>Intercept</em>",
fixed=TRUE)
show_html(mt145)
Model 1  Model 4  Model 5  
Intercept  −1  .  401***  −0  .  213  −1  .  687*** 
(0  .  271)  (0  .  126)  (0  .  294)  
Occup. class: Other white collar/Upper white collar  1  .  368***  1  .  287***  
(0  .  373)  (0  .  381)  
Occup. class: Blue collar/Upper white collar  2  .  448***  2  .  385***  
(0  .  327)  (0  .  337)  
Occup. class: Farmer/Upper white collar  1  .  826***  2  .  039***  
(0  .  413)  (0  .  426)  
Religion: Catholic/Protestant  0  .  877***  0  .  685*  
(0  .  243)  (0  .  292)  
Religion: Other,none/Protestant  0  .  975**  1  .  191**  
(0  .  347)  (0  .  441)  
Nagelkerke Rsq.  0  .  2  0  .  1  0  .  3 
Deviance  404  .  2  537  .  7  393  .  1 
AIC  412  .  2  543  .  7  405  .  1 
N  344  402  344 
Choosing what information is presented in the table and how¶
mtable()
allows to determine what and how additional information accompanies (coefficient) estimates in model tables, i.e. whether

standard errors

pvalues

symbols for statistical significance (for regression astrologists) or

confidence intervals
are displayed and whether additional information

appears below estimates

or to their right.
Further mtable()
allows to specify which summary statistics are shown. In general, the way model estimated are represented can be further customized as described further below.
Estimates, standard errors, confidence intervals etc.¶
The chief method to determine how coefficient estmates are shown and what and how additional information is provided is by using the coef.style=
argument of mtable
with which a predefined or userprovided template is selected. To demonstrate this, we start with the ‘standard’ way in which coefficient estimates are presented:
library(memisc)
lm0 < lm(sr ~ pop15 + pop75, data = LifeCycleSavings)
lm1 < lm(sr ~ dpi + ddpi, data = LifeCycleSavings)
lm2 < lm(sr ~ pop15 + pop75 + dpi + ddpi, data = LifeCycleSavings)
mtable("Model 1"=lm0,"Model 2"=lm1,"Model 3"=lm2,
summary.stats=c("Rsquared","N"))
Model 1  Model 2  Model 3  
(Intercept)  30  .  628***  6  .  360***  28  .  566*** 
(7  .  409)  (1  .  252)  (7  .  355)  
pop15  −0  .  471**  −0  .  461**  
(0  .  147)  (0  .  145)  
pop75  −1  .  934  −1  .  691  
(1  .  041)  (1  .  084)  
dpi  0  .  001  −0  .  000  
(0  .  001)  (0  .  001)  
ddpi  0  .  529*  0  .  410*  
(0  .  210)  (0  .  196)  
Rsquared  0  .  3  0  .  2  0  .  3 
N  50  50  50 
We now require Waldstatistics instead of standard errors:
mtable("Model 1"=lm0,"Model 2"=lm1,"Model 3"=lm2,
coef.style="stat",
summary.stats=c("Rsquared","N"))
Model 1  Model 2  Model 3  
(Intercept)  30  .  628***  6  .  360***  28  .  566*** 
(4  .  134)  (5  .  080)  (3  .  884)  
pop15  −0  .  471**  −0  .  461**  
(−3  .  207)  (−3  .  189)  
pop75  −1  .  934  −1  .  691  
(−1  .  858)  (−1  .  561)  
dpi  0  .  001  −0  .  000  
(1  .  962)  (−0  .  362)  
ddpi  0  .  529*  0  .  410*  
(2  .  517)  (2  .  088)  
Rsquared  0  .  3  0  .  2  0  .  3 
N  50  50  50 
then confidence intervals below estimates
mtable("Model 1"=lm0,"Model 2"=lm1,"Model 3"=lm2,
coef.style="ci",
summary.stats=c("Rsquared","N"))
Model 1  Model 2  Model 3  
(Intercept)  30  .  628  6  .  360  28  .  566 
[15  .  724  [3  .  841  [13  .  753  
45  .  532]  8  .  879]  43  .  379]  
pop15  −0  .  471  −0  .  461  
[−0  .  766  [−0  .  753  
−0  .  175]  −0  .  170]  
pop75  −1  .  934  −1  .  691  
[−4  .  028  [−3  .  874  
0  .  160]  0  .  491]  
dpi  0  .  001  −0  .  000  
[−0  .  000  [−0  .  002  
0  .  002]  0  .  002]  
ddpi  0  .  529  0  .  410  
[0  .  106  [0  .  015  
0  .  952]  0  .  805]  
Rsquared  0  .  3  0  .  2  0  .  3 
N  50  50  50 
confidence intervals to the right
mtable("Model 1"=lm0,"Model 2"=lm1,"Model 3"=lm2,
coef.style="ci.horizontal",
summary.stats=c("Rsquared","N"))
Model 1  Model 2  Model 3  
(Intercept)  30  .  628  [15  .  724  45  .  532]  6  .  360  [3  .  841  8  .  879]  28  .  566  [13  .  753  43  .  379] 
pop15  −0  .  471  [−0  .  766  −0  .  175]  −0  .  461  [−0  .  753  −0  .  170]  
pop75  −1  .  934  [−4  .  028  0  .  160]  −1  .  691  [−3  .  874  0  .  491]  
dpi  0  .  001  [−0  .  000  0  .  002]  −0  .  000  [−0  .  002  0  .  002]  
ddpi  0  .  529  [0  .  106  0  .  952]  0  .  410  [0  .  015  0  .  805]  
Rsquared  0  .  3  0  .  2  0  .  3  
N  50  50  50 
confidence intervals and standard errors
mtable("Model 1"=lm0,"Model 2"=lm1,"Model 3"=lm2,
coef.style="ci.se.horizontal",
summary.stats=c("Rsquared","N"))
Model 1  Model 2  Model 3  
(Intercept)  30  .  628  (7  .  409)  6  .  360  (1  .  252)  28  .  566  (7  .  355) 
[15  .  724  45  .  532]  [3  .  841  8  .  879]  [13  .  753  43  .  379]  
pop15  −0  .  471  (0  .  147)  −0  .  461  (0  .  145)  
[−0  .  766  −0  .  175]  [−0  .  753  −0  .  170]  
pop75  −1  .  934  (1  .  041)  −1  .  691  (1  .  084)  
[−4  .  028  0  .  160]  [−3  .  874  0  .  491]  
dpi  0  .  001  (0  .  001)  −0  .  000  (0  .  001)  
[−0  .  000  0  .  002]  [−0  .  002  0  .  002]  
ddpi  0  .  529  (0  .  210)  0  .  410  (0  .  196)  
[0  .  106  0  .  952]  [0  .  015  0  .  805]  
Rsquared  0  .  3  0  .  2  0  .  3  
N  50  50  50 
It is also possible to change the symbols for the significance levels:
# Why would one want to have letters instead? I have no idea, but
# some Germen authors/editors seem to like it that way ...
mtable("Model 1"=lm0,"Model 2"=lm1,"Model 3"=lm2,
signif.symbols=c("<sup>a</sup>"=.05,
"<sup>b</sup>"=.01,
"<sup>c</sup>"=.001),
summary.stats=c("Rsquared","N"))
Model 1  Model 2  Model 3  
(Intercept)  30  .  628^{c}  6  .  360^{c}  28  .  566^{c} 
(7  .  409)  (1  .  252)  (7  .  355)  
pop15  −0  .  471^{b}  −0  .  461^{b}  
(0  .  147)  (0  .  145)  
pop75  −1  .  934  −1  .  691  
(1  .  041)  (1  .  084)  
dpi  0  .  001  −0  .  000  
(0  .  001)  (0  .  001)  
ddpi  0  .  529^{a}  0  .  410^{a}  
(0  .  210)  (0  .  196)  
Rsquared  0  .  3  0  .  2  0  .  3 
N  50  50  50 
Summary statistics¶
In general there is a certain variety of summary statistics available in mtable()
. What statistics these are depends on the statistical model in question and the facilities provided by the corresponding getSummary()
method (see below). If mtable()
is called without the summary.stats=
argument all available summary statistics are shown:
mtable("Model 1"=lm0,"Model 2"=lm1,"Model 3"=lm2)
Model 1  Model 2  Model 3  
(Intercept)  30  .  628***  6  .  360***  28  .  566*** 
(7  .  409)  (1  .  252)  (7  .  355)  
pop15  −0  .  471**  −0  .  461**  
(0  .  147)  (0  .  145)  
pop75  −1  .  934  −1  .  691  
(1  .  041)  (1  .  084)  
dpi  0  .  001  −0  .  000  
(0  .  001)  (0  .  001)  
ddpi  0  .  529*  0  .  410*  
(0  .  210)  (0  .  196)  
Rsquared  0  .  3  0  .  2  0  .  3 
adj. Rsquared  0  .  2  0  .  1  0  .  3 
sigma  3  .  9  4  .  2  3  .  8 
F  8  .  3  4  .  5  5  .  8 
p  0  .  0  0  .  0  0  .  0 
Loglikelihood  −137  .  8  −141  .  0  −135  .  1 
Deviance  726  .  2  824  .  7  650  .  7 
AIC  283  .  7  290  .  0  282  .  2 
BIC  291  .  3  297  .  7  293  .  7 
N  50  50  50 
So if we prefer (or our reviewer, editor, supervisor or boss) we can show some unusual statistics
mtable("Model 1"=lm0,"Model 2"=lm1,"Model 3"=lm2,
summary.stats=c("sigma","BIC","N"))
Model 1  Model 2  Model 3  
(Intercept)  30  .  628***  6  .  360***  28  .  566*** 
(7  .  409)  (1  .  252)  (7  .  355)  
pop15  −0  .  471**  −0  .  461**  
(0  .  147)  (0  .  145)  
pop75  −1  .  934  −1  .  691  
(1  .  041)  (1  .  084)  
dpi  0  .  001  −0  .  000  
(0  .  001)  (0  .  001)  
ddpi  0  .  529*  0  .  410*  
(0  .  210)  (0  .  196)  
sigma  3  .  9  4  .  2  3  .  8 
BIC  291  .  3  297  .  7  293  .  7 
N  50  50  50 
For more examples regarding the content of the results of mtable()
refer to the example code in the help page given by ?mable
.
Exporting mtable results¶
Of course you would not only like to see your table of estimates on screen but also to include it into a documement that reports your findings. memisc()
supports exporting tables of model estimates (i.e. results of mtable()
) into LaTeX documents and into formats that can be read in by wordprocessing software: tabdelimited format and HTML. There is no direct way to export model tables into a wordprocessing format yet, mainly because there is no (simple) file format standard (OpenDocument may be an emerging standard, but it is not easy to create files in this format  at least not as easy as HTML). These various options of exporting the results of mtable()
are discussed in the following. To this purpose, we return to the example from above and ‘embellish’ a bit by changing the coefficient labels:
lm0 < lm(sr ~ pop15 + pop75, data = LifeCycleSavings)
lm1 < lm(sr ~ dpi + ddpi, data = LifeCycleSavings)
lm2 < lm(sr ~ pop15 + pop75 + dpi + ddpi, data = LifeCycleSavings)
mt123 < mtable("Model 1"=lm0,"Model 2"=lm1,"Model 3"=lm2,
summary.stats=c("Rsquared","N"))
mt123 < relabel(mt123,
"(Intercept)" = "Constant",
pop15 = "Percentage of population under 15",
pop75 = "Percentage of population over 75",
dpi = "Real percapita disposable income",
ddpi = "Growth rate of real percapita disp. income"
)
mt123
Calls:
Model 1: lm(formula = sr ~ pop15 + pop75, data = LifeCycleSavings)
Model 2: lm(formula = sr ~ dpi + ddpi, data = LifeCycleSavings)
Model 3: lm(formula = sr ~ pop15 + pop75 + dpi + ddpi, data = LifeCycleSavings)
================================================================================
Model 1 Model 2 Model 3

Constant 30.628*** 6.360*** 28.566***
(7.409) (1.252) (7.355)
Percentage of population under 15 0.471** 0.461**
(0.147) (0.145)
Percentage of population over 75 1.934 1.691
(1.041) (1.084)
Real percapita disposable income 0.001 0.000
(0.001) (0.001)
Growth rate of real percapita disp. income 0.529* 0.410*
(0.210) (0.196)

Rsquared 0.3 0.2 0.3
N 50 50 50
================================================================================
LaTeX Format¶
The first format to export mtable()
results into is TeX/LaTeX simply because this is the format in which the author of this package usually writes his papers. This is achieved by, first, applying the function toLatex()
to the results of mtable()
, which tranlates them into a character string containing TeX/LaTeX code and, second, using writeLines()
to send this character string into a text file. This was already shown at the beginning, but we take a closer look at it here. Continuing the example immediately above, we call toLatex()
to see the TeX/LaTeX code:
toLatex(mt123)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Calls:
% Model 1: lm(formula = sr ~ pop15 + pop75, data = LifeCycleSavings)
% Model 2: lm(formula = sr ~ dpi + ddpi, data = LifeCycleSavings)
% Model 3: lm(formula = sr ~ pop15 + pop75 + dpi + ddpi, data = LifeCycleSavings)
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{tabular}{lD{.}{.}{3}cD{.}{.}{3}cD{.}{.}{3}}
\toprule
&\multicolumn{1}{c}{Model 1}&&\multicolumn{1}{c}{Model 2}&&\multicolumn{1}{c}{Model 3}\\
\midrule
Constant&30.628^{***}&&6.360^{***}&&28.566^{***}\\
&(7.409)&&(1.252)&&(7.355)\\
Percentage of population under 15&0.471^{**}&&&&0.461^{**}\\
&(0.147)&&&&(0.145)\\
Percentage of population over 75&1.934&&&&1.691\\
&(1.041)&&&&(1.084)\\
Real percapita disposable income&&&0.001&&0.000\\
&&&(0.001)&&(0.001)\\
Growth rate of real percapita disp. income&&&0.529^{*}&&0.410^{*}\\
&&&(0.210)&&(0.196)\\
\midrule
Rsquared&0.3&&0.2&&0.3\\
N&50&&50&&50\\
\bottomrule
\end{tabular}
After formatting with LaTex, this may look like this:
It should be noted that toLatex()
is a generic function and the “memisc” package ‘only’ defines a method for “mtable” objects. Alternatively one could use the call
format(mt123,target="LaTeX")
or even call the internal formatting function itself:
mtable_format_latex(m123)
By default, the TeX/LaTeX code created this way uses the macros \toprule
, \midrule
, and \bottomrule
provided by the LaTeX package “booktabs”. If you do not like this package (why shouldn’t anyone?) you can resolve this dependency by calling toLatex()
, format()
, or mtable_format_latex()
with the optional argument useBooktabs=FALSE
. Another default dependency is the LaTeX package “dcolumn”, which is used to make sure that floating point numbers are aligned on their decimal dots. This dependency can be resolved by the optional argument useDcolumn=FALSE
. For more aspects of LaTeX output that can be customized see the help page ?mtable_format_latex
.
Until version 0.97 of “memisc”, creating a text file with TeX/LaTeX code had to be done in two steps, first creating a text string with the code and then writing the text string into the file. Since version 0.98 both steps can be done with a single function call, using write.mtable()
. That is, to write the LaTeX formatted mtable()
result in mt123
one can simple call:
write.mtable(mt123,format="LaTeX",
file="mt123.tex")
Text file format (tabdelimited and “CSV”)¶
Tabdelimited format is of course bestsuited for exporting data frames or matrices into files, but since the results of mtable()
have a tabular structure, they can also be exported into this format. Tabdelimited format is quite simple and can be read by a wide variety of software. However, this simplicity also means that only the cell contents are exported while embellishments of the contents (e.g. horizonal alignment of cells and cell borders) are not.
The following code exports the mtable()
result named mt123
into a text file in tabdelimited format:
write.mtable(mt123,file="mt123.txt")
After opening this file with LibreWriter, using its “covert text into table” tool, and some manual tweaking the result looks like this:
(The original text file is available here while the LibreOffice file is available here.) Obviously some further tweaking (such as commaoriented column tabulators) is needed to make this work in a publication.
It is also posslibe export an mtable()
result into “CSV” format and import it into some spreadsheet software. This would be done so:
write.mtable(mt123,file="mt123.csv",colsep=",")
After opening this file with LibreCalc and some tweaking of the format, the result looks like this:
(Here is the original CSV file.)
HTML (for webpages, knitr reports and word processors)¶
Of course, having to tweak the format of mtable()
results by hand is frustrating, so in order to make easier to get wellformatted tables, version 0.98 of “memisc” provides for exporting the results of mtable()
into HTML. HTML is the format of websites, but it can also imported into contemporary word processing software with little loss in formatting. Further, if one is using the “Rmarkdown” and “knitr” packages to produced HTMLformatted reports, it is convenient to have HTML versions of mtable()
results.
To get a file in HTML format that contains the results of mtable()
one can again use the function write.mtable()
, yet in this case with the option format="HTML"
, or directly the function write_html()
:
write_html(mt123,file="mt123.html")
The file “mt123.html” generated that way can be included into your favourite wordprocessing software, e.g. LibreOffice. This is how the table would look like after including it into LibreOffece (and setting the columns to “optimal width”):
and this how it would look like after including into Word:
To view results of mtable()
in HTML in interactive sessions with RStudio, one can simply call show_html()
as in
show_html(mt123)
Of course, this document is not an interactive sesssion but produced using the “knitr” package. In this context, show_html()
used inside an Rchunk with chunk option results='asis'
and the options setting options(html_viewer="stdout")
. With the following code one could even make sure that all results of mtable()
in a knitr document are printed in HTML format:
knit_print.mtable <function(x,...)
knitr::asis_output(format_html(x))
This “trick” was used previously in this document where the different display options of coefficient estmates were discussed. Thus after this trick, in a knitr document we get
mt123
Model 1  Model 2  Model 3  
Constant  30  .  628***  6  .  360***  28  .  566*** 
(7  .  409)  (1  .  252)  (7  .  355)  
Percentage of population under 15  −0  .  471**  −0  .  461**  
(0  .  147)  (0  .  145)  
Percentage of population over 75  −1  .  934  −1  .  691  
(1  .  041)  (1  .  084)  
Real percapita disposable income  0  .  001  −0  .  000  
(0  .  001)  (0  .  001)  
Growth rate of real percapita disp. income  0  .  529*  0  .  410*  
(0  .  210)  (0  .  196)  
Rsquared  0  .  3  0  .  2  0  .  3 
N  50  50  50 
Adapting mtable to new model classes and other tricks  the API of mtable¶
mtable()
is designed to be easily adapted to all kind of model classes: If there is a model class like, say, “modcls” then all that is needed to get mtable()
to report estimates of instances of this model class is to define a function getSummary.modcls()
, i.e. a method function of objects of class “modcls”. This function needs to return a list with components

“coef”: A matrix or array with coefficients and additional information. The rows should refer to coefficients, the columns should contain the estimates, standard errors, pvalues, lower and upper confidence interval limits. The columns should be labelled “est”, “se”, “stat”, “p”, “lwr”, and “upr”.
For singleequation models, this component should be a matrix. For multipleequation models, it should be a threedimensional array, with the third dimension corresponding to the equations.

“sumstat”: A vector (with named elements) containing model summary statistics, such goodnessoffit statistics etc.

“contrasts”: A list of the contrast functions, their names, or contrast matrices; one for each (ordered or unordered) factors present in the model.

“xlevels”: A list with the names of the levels of the factors present in the model.

“call”: A language object describing the call that created the model estimates.
To illustrate, here are the results of the method function getSummary.glm()
(which is called when the generic function getSummary()
is applied to an object of class “glm”):
# Model from the first example at the beginning ...
model1 < glm((vote3=="Truman")~occup4,data=vote.48,
family="binomial")
getSummary(model1) # or getSummary.glm(model1)
$coef
est se stat p lwr upr
(Intercept) 1.40 0.271 5.17 2.29e07 1.932 0.87
occup4Other white collar 1.37 0.373 3.67 2.42e04 0.638 2.10
occup4Blue collar 2.45 0.327 7.48 7.36e14 1.807 3.09
occup4Farmer 1.83 0.413 4.42 9.85e06 1.016 2.64
$sumstat
phi LR df p logLik
1.00e+00 7.01e+01 3.00e+00 4.11e15 2.02e+02
deviance Aldrich.Nelson McFadden Cox.Snell Nagelkerke
4.04e+02 1.69e01 1.48e01 1.84e01 2.46e01
AIC BIC N
4.12e+02 4.28e+02 3.44e+02
$contrasts
$contrasts$occup4
[1] "contr.treatment"
$xlevels
$xlevels$occup4
[1] "Upper white collar" "Other white collar" "Blue collar"
[4] "Farmer"
$call
glm(formula = (vote3 == "Truman") ~ occup4, family = "binomial",
data = vote.48)
The definition of this function is available in here and can be used as a starting point for other such method functions.
A demonstration: “mtable” and sandwich estimators of variance¶
The modularity of mtable()
through the use of the generic getSummary()
function allows other kinds of extensions, e.g. adapting it to the use of “sandwich” estimators of standard errors. This can be achieved, first, by defining yet another method function of getSummary()
, e.g. the one defined in the R available here. As a second step, one marks model estimation results such that this newly defined method function is applied to them by attaching the appropriate class attribute. For example to get sandwich estimators of standard errors for “lm”” or “glm”” objects one can attach the classes “lm_sandwich” or “glm_sandwich”, respectively, as in the following example:
library(memisc)
library(sandwich)
library(lmtest)
source("getSummaryglmsandwich.R")
data(Mandible)
fm1 < lm(length ~ age, data=Mandible, subset=(age <= 28))
fm1.sw < fm1
class(fm1.sw) < c("lm_sandwich",class(fm1))
mtable(
"Conventional"=fm1,
"Sandwich"=fm1.sw,
summary.stats=c("Rsquared","N"))
Conventional  Sandwich  
(Intercept)  −11  .  953***  −11  .  953*** 
(0  .  976)  (1  .  010)  
age  1  .  773***  1  .  773*** 
(0  .  048)  (0  .  054)  
Rsquared  0  .  9  0  .  9 
N  158  158 
Alternatively, one can use an appropriate getSummary=
argument to mtable()
like in this example:
fm1 < lm(length ~ age, data=Mandible, subset=(age <= 28))
fm2 < lm(length ~ age+I(scale(age)^2), data=Mandible, subset=(age <= 28))
# Using 'conventional' i.e. modelbased standard errors:
mtable(fm1,fm2,summary.stats=c("Rsquared","N"))
fm1  fm2  
(Intercept)  −11  .  953***  −11  .  303*** 
(0  .  976)  (1  .  060)  
age  1  .  773***  1  .  754*** 
(0  .  048)  (0  .  049)  
I(scale(age)^2)  −0  .  362  
(0  .  236)  
Rsquared  0  .  9  0  .  9 
N  158  158 
# Using sandwichbased standard errors:
mtable(fm1,fm2,
summary.stats=c("Rsquared","N"),
getSummary=getSummary.lm_sandwich)
fm1  fm2  
(Intercept)  −11  .  953***  −11  .  303*** 
(1  .  010)  (1  .  348)  
age  1  .  773***  1  .  754*** 
(0  .  054)  (0  .  063)  
I(scale(age)^2)  −0  .  362  
(0  .  258)  
Rsquared  0  .  9  0  .  9 
N  158  158 
Since version 0.98 of “memisc” we can also put mtable results together etc.
mtfm12.conv < mtable("Model 1"=fm1,
"Model 2"=fm2,
summary.stats=c("Rsquared","N"))
mtfm12.sndw < mtable("Model 1"=fm1,
"Model 2"=fm2,
summary.stats=c("Rsquared","N"),
getSummary=getSummary.lm_sandwich)
c(Conventional=mtfm12.conv,Sandwich=mtfm12.sndw)
Conventional  Sandwich  
Model 1  Model 2  Model 1  Model 2  
(Intercept)  −11  .  953***  −11  .  303***  −11  .  953***  −11  .  303*** 
(0  .  976)  (1  .  060)  (1  .  010)  (1  .  348)  
age  1  .  773***  1  .  754***  1  .  773***  1  .  754*** 
(0  .  048)  (0  .  049)  (0  .  054)  (0  .  063)  
I(scale(age)^2)  −0  .  362  −0  .  362  
(0  .  236)  (0  .  258)  
Rsquared  0  .  9  0  .  9  0  .  9  0  .  9 
N  158  158  158  158 