Graphics in EG and R HRP223 2009 November 16th, 2009
Copyright Leland Stanford Junior University.All rights reserved. Warning:This presentation is protected by copyright law and international treaties.Unauthorized reproduction of this presentation, or any portion of it, may result in severe civil and criminal penalties and will be prosecuted to maximum extent possible under the law. Robbins Creating More Effective Graphics by Naomi Robbins is a wonderful book showing the right and wrong ways to visualize scientific data.Read it when you have an afternoon off.It is an ideal read on a transcontinental flight. Why Do Data Visualization?
Well designed pictures will show you the details and the whole pattern in your data. Numeric descriptions can easily hide important patterns. Some patterns are hard to detect in tables. Whenever data is reported over time or locations, you need art. YOU CAN LEARN A LOT BY JUST LOOKING. -Yogi Berra Fishers Plot Data Reported in Cleveland
Year 1 Year 2 Based on code written by Robert Allison at SAS Institute Scatter Plot for Correlations
Anscombe 1973, Graphs in Statistical Analysis All have r2 = .67 Bad Things First, I want to talk about bad graphics that I frequently see. 3d Pie Donuts Stacked graphics General 3D graphics Dont, Dont, Dont
While the SAS implementation of 3D graphics is relatively good, dont use 3D effects, unless you are measuring something in 3D.Even then, dont. Tufte is a God to many. The empiricist in me is very nervous about the amount of pontificating in his books I want to have evidence-based advice. His best advice is to put no extra ink on the page. Think about the ink-to-information ratio. Remove all chart junk. Note: the irony of the chart junk on this slide. You can remove ink rather than adding .
Example Bar Chart Serum Samples in Each Trimester You can remove ink rather than adding . Ink-to-Information Ratio
How much ink for seven numbers? Based on Soukup & Davidson, 2002 Visual Data Mining Cleveland If you want to know how to do scientific visualization, you must read William Clevelands work. He attempted to quantify what makes a good graphic good. His early work on graphics is one of the reasons why R/S-plus is taking over the statistical world. Pie is bad. Work by Cleveland (and experimental psychologists) suggests that: people are bad at judging the relative magnitude of angles if you twist the rotation of the pie you can cause people to systematically misjudge the size of the angles a 3rd dimension makes judgment worse If you get a glossy handout with a 3D pie, assume someone is lying to you. Dont use them. Dont Explode! This exploded 3D pie (brought to you by Excel) is nearly useless for judging amounts. Forbidden Donut. Donut plots have the same problems as pies (if not worse) . Stacking is Bad Cleveland also quantified the fact that people are bad at judging the relative height of stacked data. Wow, a cinnamon roll plot!
Good luck making rapid judgments using this stacked 3D pie. What is a good graphic? Dont make your audience think unnecessarily!
Minimize the amount of ink on the page. This needs to be studied. Show the central tendency and the variability. Plot the quantity (inference) that you want people to notice. Be sure colorblind people can understand it. Use a black and white photocopier and make sure you can distinguish all groups. Avoid Thinking But labels on the graphic directly instead of using a key. If you want people to compare the difference between two lines, plot the difference, not the two lines. Bivariate Comparisons with Lines
People are extremely bad at judging the distance between two curves.Never ask people to judge up and down (vertical) distances between curves. The distance between the two curves is the same at all points. Based on: Robbins Creating More Effective Graphs, 2005 Plot Types Univariate (one variable) Categorical variables
Bar charts Dot plots Waffle plots Continuous variables Histogram Box plot Violin plots Bar Charts The ink-to-information ratio is lousy.
A one dimensional quantity is being expanded into two dimensions. Doubling of the amount corresponds to how much of an increase in area? SAS Bar Charts SAS makes the reader do extra work by rotating the axis labels in ActiveX images. They pointlessly include variable labels by default. Notice you can Edit the data and apply filters.
How to do it? Notice you can Edit the data and apply filters. You can right click on variables and apply user-defined formats off the Properties dialog. First create the format.
In the Data windowpane of the Bar Chart GUI, right click on the variable and change the format to the User Defined format you had created. The GUI is Solid My only complaints are that the rotate grouping values text does not work (position in this example) and the summary statistics do not show up when you request ActiveX images. Saving the Graphic for Publication
The easiest way to get publication quality graphics is to set the output type to be RTF. .PNG format ActiveX image format Default Output and Graphics
The default graphic format in EG is ActiveX.These images can be edited (even on the web) but they only display with Internet Explorer.I have set my graphics to display as ActiveX images.Tweak this with Tools> Options > Graph. Types of Images The default formats of the images are determined by the ODS destinations you are using: LISTING: pgn visible in the Windows Image Fax Viewer HTML: png, gif, jpg contained in web pages and visible in Internet Explorer, Firefox or Opera LATEX: PostScrpt, epsi, gif, jpeg, pgn are visible in GhostView PCL or PS: contained in Postscript file are visible in GhostView PDF: contained in pdf, which is visible with Adobe Reader RTF: visible in MS Word I Typically Use HTML This is the appearance template.For optimal results use: Analysis: color Default : overdistinguishes symbols for color or B&W Journal or journal2, etc: black and white Statistical or statistical2, etc: color Include image_dpi = 200 to set the resolution to be higher than the default 100 dots per inch.Try 200 for final images pasting into MS Office. This says the images should show tooltips with extra statistical details when you hover the mouse over parts of the graphic.(I cant image these.) Useful ods graphics Options
After the ods graphics on statement, type a / then: imagename = fileName reset resets the counter of images back to 0. imagefmt = jpg width = 4.5 in height = 4.5 in If you set only width or height, it will use a 4:3 aspect ratio. Different appearance templates
What is ODS? The Output Delivery System (ODS) controls the type and appearance of SAS output. Different appearance templates Different outputdestinations/types. You can browse the ODS appearance templates from the Style Manager on the Tools menu. ODS Graphics Compared to the competition, for the last 10 years SAS graphics have been between poor and pathetic. Graphics procedures that rendered okay quality, at best . No what you see is what you get editing. Many plots were nearly impossible to render. Custom graphics required extensive programming. SAS 9.x has attempted to solve this problem. Old vs. New Procedures The old (commonly used) graphics procedures were gchart, gplot. Now most analysis procedures have built in high quality graphics that can be invoked with an ODS graphics on statement. Early on in the class I told you to tweak the EG options to include ODS graphics on with every run. There are also new easy to use statistical graphics (sg) procedures. New Graphics Statistical Graphics Procs
proc sgPlot general plotting procedure that replaces gplot proc sgScatter lots of tools for scatterplots and scatter matrices proc sgPanel quick and easy trellis/lattice/matrix/panel of plots Proc sgRender used with proc template to make totally custom plots It replaces proc greplay Plot Types Univariate (one variable) Categorical variables
Bar charts Dot plots Waffle plots Continuous variables Histogram Box plot Violin plots Quantile and QQ plots You can get an okay looking graphic using sgpanel. I was able to get exactly the graphic I wanted using R. If you want to use R Download R for Mac or PC cran.cnr.berkeley.edu/bin/macosx/ cran.cnr.berkeley.edu/bin/windows/base If you use a PC, also get PERL and Tinn-R
PERL is a text manipulation language that is used by a couple of key R packages.It ships with Mac OS X.PC users can get ActivePerl (what I use) or Strawberry Perl for Windows. Tinn-R is a text editor that knows the R language. sourceforge.net/projects/tinn-r/ R Help R help files are user hostile.To learn about the options for dotchart type: ?dotchart Use: rseek.org Browse To see why people use R for graphics look here:
addictedtor.free.fr/graphiques/thumbs.php Additional Libraries If you see sample code that includes require() or library(), you will need to do a onetime download of the additional package.If you are using Vista, run R as the administrator (by right clicking on the R icon instead of just double clicking ) to install and update packages. Waffle Plots I have not found software to do them. I need to find their real name Image from: Visual language for Designers by Connie Malamed 2009. Continuous Outcomes The Distribution Analysis menu option can do basic plots. The resolution of the histogram is okay but the others are unacceptable. Use sgplot for high resolution plots. Violin A violin plot mirrors the shape of the histogram (density). They can be done in R. Grouped Categorical Data
To graph categorical data in SAS you need to get Michael Friendlys Visualizing Categorical Data.Unfortunately, his macros are copyrighted with the bookSo I will show you the R versions. Fourfold plots Mosaic plots Association plots Fourfold Plots They draw 4 slices of pie with the area corresponding to the number of people in each cell of a 2x2 table and they have confidence bands such that if the confidence bounds overlap on adjacent pie pieces, they are not statistically significantly different. More males were admitted than females.
There is clear evidence of sexist policies in admissions! Department A admitted more females than males and every other department had no bias!
The joy of Simpsons paradox. Mosaic Plots So you have an contingency table and you want to know if there is as an association.You do a chi-square test and it says there are associations between the rows and columns.What next? Some basic voodoo in R shows which combinations are over (in blue) or under represented (in red).
values = c(5, 29, 14, 16, 15, 54, 14, 10, 20, 84, 17, 94, 68, 119, 26, 7); values = matrix(values, nrow = 4, byrow=TRUE) rownames(values) =c("Green", "Hazel", "Blue", "Brown") colnames(values) =c("Black", "Brown", "Red", "Blond") mosaicplot(values, shade = TRUE) I prefer the simpler association plots.
values = c(5, 29, 14, 16, 15, 54, 14, 10, 20, 84, 17, 94, 68, 119, 26, 7); values = matrix(values, nrow = 4, byrow=TRUE) rownames(values) =c("Green", "Hazel", "Blue", "Brown") colnames(values) =c("Black", "Brown", "Red", "Blond") marg Options>Enhanced Editor then click User Defined Keywords to add the coloring. I want to add in a reference line showing what is normal and put the categories in order. Grids You can produce lattices full of graphics with proc gpanel. Spaghetti Plots Data from Singer and Willett: