R help - May 2011 - Analog of least significant difference error bars for proportions

Dear R-list,

	In the R-book, p.464, Michael Crawley recommends that error
bars for bar plots of normally distributed continuous response
variables with categorical explanatory variables be given by
1/2 of the least significant difference, where the least significant
difference is defines as

qt(0.975,degrees_of_freedom)*standard_error_of_the_difference.

The idea is that the above quantity visually conveys whether or not
the means are different more realistically than do standard errors.

	I have analyzed proportions with categorical variables using
the glm function with a binomial error model. I wish to plot a bar
graph with the height of the bars the proportions. Is there a way
to define error bars analogous to the least significant difference bars
described above that can convey the overlap of proportions?
The experimentalists with whom I work just love error bars. I would  
like to
make them as meaningful as possible.

Thanks and best wishes,
Rich
------------------------------------------------------------
Richard A. Friedman, PhD
Associate Research Scientist,
Biomedical Informatics Shared Resource
Herbert Irving Comprehensive Cancer Center (HICCC)
Lecturer,
Department of Biomedical Informatics (DBMI)
Educational Coordinator,
Center for Computational Biology and Bioinformatics (C2B2)/
National Center for Multiscale Analysis of Genomic Networks (MAGNet)
Room 824
Irving Cancer Research Center
Columbia University
1130 St. Nicholas Ave
New York, NY 10032
(212)851-4765 (voice)
friedman at cancercenter.columbia.edu
http://cancercenter.columbia.edu/~friedman/

I am a Bayesian. When I see a multiple-choice question on a test and I  
don't
know the answer I say "eeney-meaney-miney-moe".

Rose Friedman, Age 14

On 24/05/11 11:23, Richard Friedman wrote:
> Dear R-list,
>
>     In the R-book, p.464, Michael Crawley recommends that error
> bars for bar plots of normally distributed continuous response
> variables with categorical explanatory variables be given by
> 1/2 of the least significant difference, where the least significant
> difference is defines as
>
> qt(0.975,degrees_of_freedom)*standard_error_of_the_difference.
>
> The idea is that the above quantity visually conveys whether or not
> the means are different more realistically than do standard errors.
>
>     I have analyzed proportions with categorical variables using
> the glm function with a binomial error model. I wish to plot a bar
> graph with the height of the bars the proportions. Is there a way
> to define error bars analogous to the least significant difference bars
> described above that can convey the overlap of proportions?
> The experimentalists with whom I work just love error bars. I would 
> like to
> make them as meaningful as possible.
(1) The errbar() function in the Hmisc package will allow you to set any
``spread'' that you wish on your error bars.

(2) In respect of maximal meaningfulness:  The naive viewer tends to
interpret error bars by concluding that if the ranges of two pairs of
error bars do not overlap then the two quantities being estimated are
``significantly different''.  Hence it strikes me that you might want to
imitate what is done for the notches in boxplots, which are designed
to make such an interpretation roughly correct.

 From the help on boxplot.stats():
> The notches (if requested) extend to |+/-1.58 IQR/sqrt(n)|. This seems 
> to be based on the same calculations as the formula with 1.57 in 
> Chambers /et al./ (1983, p. 62), given in McGill /et al./ (1978, p. 
> 16). They are based on asymptotic normality of the median and roughly 
> equal sample sizes for the two medians being compared, and are said to 
> be rather insensitive to the underlying distributions of the samples. 
> The idea appears to be to give roughly a 95% confidence interval for 
> the difference in two medians. 
     cheers,

         Rolf Turner


	[[alternative HTML version deleted]]

Dear Rolf (and List),

	Thank you for your help on error bars.
I fear that neither of the suggestions quite answer my immediate need.
1. Notches will not work because I have more than 2 levels.
2. The errbar function will  useful once I know the error bars to put  
in.

I thing I have figured it out but I would greatly appreciate
feedback (positive or negative) from the list:

For 2 or more levels with ordinary ANOVA the least significant
error bars are given by

(qt(0.975,degrees_of_freedom)sqrt((s1^2+s2^2)/sqrt(n))/2

Am I correct that for 3 levels the error bars are given by

(qt(0.975,degrees_of_freedom)sqrt((s1^2+s2^2+s3^2)/sqrt(n))/2

where the argument of the first square root is the standard error
of the sample mean?

If I am correct, then an analogous express would seem to hold
where the normal approximation is a good approximation to the
binomial distribution.

for 2 samples

z(..475)sqrt(theta1(1-theta1)+theta2(1-theta2))/2

and for 3 samples

z(..475)sqrt(theta1(1-theta1)+theta2(1-theta2)+theta3(1-theta3))/2

Does this sound right?

Thanks and best wishes,
Rich



Date: Tue, 24 May 2011 12:03:16 +1200
From: Rolf Turner <rolf.turner at xtra.co.nz>
To: Richard Friedman <friedman at cancercenter.columbia.edu>
Cc: r-help at r-project.org
Subject: Re: [R] Analog of least significant difference error bars	for
	proportions
Message-ID: <4DDAF5C4.3080901 at xtra.co.nz>
Content-Type: text/plain

On 24/05/11 11:23, Richard Friedman wrote:
> Dear R-list,
>
>    In the R-book, p.464, Michael Crawley recommends that error
> bars for bar plots of normally distributed continuous response
> variables with categorical explanatory variables be given by
> 1/2 of the least significant difference, where the least significant
> difference is defines as
>
> qt(0.975,degrees_of_freedom)*standard_error_of_the_difference.
>
> The idea is that the above quantity visually conveys whether or not
> the means are different more realistically than do standard errors.
>
>    I have analyzed proportions with categorical variables using
> the glm function with a binomial error model. I wish to plot a bar
> graph with the height of the bars the proportions. Is there a way
> to define error bars analogous to the least significant difference  
> bars
> described above that can convey the overlap of proportions?
> The experimentalists with whom I work just love error bars. I would
> like to
> make them as meaningful as possible.
(1) The errbar() function in the Hmisc package will allow you to set any
``spread'' that you wish on your error bars.

(2) In respect of maximal meaningfulness:  The naive viewer tends to
interpret error bars by concluding that if the ranges of two pairs of
error bars do not overlap then the two quantities being estimated are
``significantly different''.  Hence it strikes me that you might want to
imitate what is done for the notches in boxplots, which are designed
to make such an interpretation roughly correct.

 From the help on boxplot.stats():
> The notches (if requested) extend to |+/-1.58 IQR/sqrt(n)|. This seems
> to be based on the same calculations as the formula with 1.57 in
> Chambers /et al./ (1983, p. 62), given in McGill /et al./ (1978, p.
> 16). They are based on asymptotic normality of the median and roughly
> equal sample sizes for the two medians being compared, and are said to
> be rather insensitive to the underlying distributions of the samples.
> The idea appears to be to give roughly a 95% confidence interval for
> the difference in two medians.
     cheers,

         Rolf Turner


	[[alternative HTML version deleted]]