TS11 Clinical Trial Reporting with SAS ODS Graphics

Transcription

1 TS11 Clinical Trial Reporting with SAS ODS Graphics Hans-Rainer Pauli, SAS Inst. AG, Brüttisellen, Schweiz David J Garbutt, BIOP AG, Centralbahnstrasse 9, 4051, Basel, Schweiz 20 October 2009 PhUSE Clinical trials generate large amounts of data and a clear understanding of their results by all stakeholders is essential. The information contained in the trial data must therefore be analysed and presented in the clearest way possible. Increasingly it is being accepted throughout the industry that visual presentations of data must become central to standard reporting. We show, using examples of clinical data, how ODS graphics and the new graphics procedures in SAS 9.2 have an important role to play. The new SGPLOT and SGPANEL procedures emphasize high-level components rather than the low level data points of annotate data sets. We will show how to convert data into information using a gradual approach to building the desired graphic. The datasets we use include Liver Function Tests (LFTs), Haematology, and a meta-analysis for which we build up a Forest Plot. We also exemplify the Graphics Template Language (GTL) and illustrate its potential for building standardised but flexible graphics in colour and monochrome. INTRODUCTION Using graphics (and pictures) is a powerful way to tell a story [Cleveland, 1984], [Roam, 2008], [McCloud, 1993] and although we like to think our work is more serious than telling a story 1 we ignore the rules of perception and neuronal processing [Medina, 2008], [Pinker, 1997] at our peril. Clinical trials generate a large amount of information and much of it is currently analysed and presented in a sub-optimal way [Harrell, 2005, Garbutt, 2007]. But, while tables can be obtained almost free, the area of graphics is renowned for being hard and time consuming. Why is that? I believe it is the extra degrees of freedom that come from drawing (in a picture) rather than typing (with 26 letters) in a table. And in computer land parameters mean more typing in order to be specific enough about what you mean. Compounded with this is the fact that we all have ideas and preferences about aesthetics which surface when we assess a graphic rather than a table. So the challenge in the area of graphics is to combine flexibility and ease of use with high quality aesthetics. We can compare classic SAS/GRAPH to fitting Generalised Linear Models Previous address: Business & Decision, Life Science Division, Schweiz 1 Actually, if you find stories not so serious go to a performance of King Lear 1

2 using only data steps, PROC REG and SAS MACRO. It does work, you can do anything, but PROC GENMOD is much, much easier to write, read and validate. Of course that it is possible relies on the development of a model for the tasks needed and development of a syntax and a grammar to drive the design of commands to be used. Similarly ideas and usages of Statistical Graphics have moved a long way since SAS/GRAPH was released around 1982 and now, with the ODS and Statistical graphics in Version 9.2, so has SAS. This paper has two aims. First to exemplify how the new SAS Statistical Graphics procedures (called SG PROCS informally) can be used to build up useful and clear graphical representations of data. To do this we apply SG PROCS to some typical examples of Clinical Graphics that were not previously easy to construct, and we build some of the complex graphics layer-by-layer illustrating how each program statement affects the output. Second we demonstrate with two examples of how the new Graphical Template Language can be exploited to create new graphics and make them easily accessible to all your users. We will bear in mind along the way that an important aspect of graphics creation is how the data need to be structured to be usable and we will demonstrate the power of the layering philosophy used by the SG PROCS which allow complex graphics to built up piece-by-piece. We briefly cover set up tips for those unfamiliar with ODS and the conventions used for ODS graphical output. While there are now more and more papers and code samples covering SG PROCS (see Further Reading) we will however not avoid repeating material from elsewhere because our aim is to provide a guide for SAS Clinical Graphics that can make a good first stop on the way to solving your problems with these exciting new tools. We will use the following two main examples in detail but here are others covering scatter, series, vector, bar plots, and all the possible panel and lattice layouts. There will be examples of custom axes, labelling, transparency in histograms and other techniques. The two examples we dissect in detail illustrate exactly how the graph is built layer by layer and how the appearance changes as features are added. These two examples are 1. Creation of a dose response plot by overlaying multiple plots in one graph allowing change of treatment for some patients during the study (Figure 2). 2. Creation of a Butterfly plot (Figure 16) in which a back-to-back histogram used for comparing the counts split into two groups, in this case males and females. We also discuss the construction of a Forest plot for a meta analysis illustrating the Graphical Template Language (GTL), Figure 20. When describing graphics and what can, or should, be produced we need to distinguish three kinds of variables. There is general confusion about this issue in both texts and the documentation of various software systems and programming traditions. We want to avoid this debate but in order to be clear what is meant we will define the variable types here: variable meaning a continuous variable of SAS numeric type (of any length). This is regardless of how many distinct values it, in fact, has. If a format is applied then it is a numeric one and not a date/time format. ordered factor. This can be a SAS character variable or a numeric variable. If a numeric variable then it might either have a few distinct values or have a character format in effect. If so, then the ordering is by value or data. 2

3 1 ODS SG PROCS Example Program Listings A variable with date or datetime values and with an appropriate format is automatically treated this way. factor A character variable, or a formatted numeric variable, with ordering set to formatted. Factors are effectively used as classifiers for grouping, or facetting, a variable. Example Program Listings 1 Suggested ODS settings for SG PROCS Example 1 Clinical response by visit per treatment Example 2 Grouped dose-response graph with labelling Example 3 Plasma Concentration Example 4 Required sample sizes under differing assumptions Example 5a Code for Box plots of Age distribution by treatment in lattice layout Example 5b Demography data in a Panel layout in one row Example 6 P-values in a row lattice Example 7 US Cancer data HBAR lattice by Gender Example 7b Butterfly plot of USA Cancer data Example 8a LFT safety panel Example 8b LFT panel per patient Example 9 Combining a Histogram, box plot and fringeplot with GTL Example 10 Forest plot extract of template code Example 10 Calling the Forest plot GTL template ODS Statistical Graphics These graphics are produced optionally when running SAS/STAT procedures when using the Output Delivery System (ODS). They use the same underlying systems as the SG PROCS and some examples are shown in Figure 1. There are two types single and multi-cell graphics and both are shown here. The top row and the last figure of the second row are single cell graphics and the first two plots of the second row are multi-cell graphics. Multi-cell plots could on first thought be created using PROC GPLOT followed by a call of PROC GREPLAY 2 ; multi-cell graphics are different because they are produced together as a unit. This means that complicated issues involved with the previous methods, for example for using common axis scaling and labeling, or the need for a different template depending on how many treatment groups there are, can all be forgotten. 1.1 The new ODS Statistical Graphics procedures The SG PROCS use the graphics core developed for ODS Statistical Graphics and are designed to produce publication quality graphics directly. They work differently from the standard SAS/GRAPH procedures and do not use PROC GREPLAY or annotate datasets, they produce image files, or documents with embedded images, directly without using graphics catalogs. A SAS/GRAPH like syntax is used for analytical graphs and is designed for programmers. (So therefore: concise vs. verbose; flexible vs. limited in options; combinable pieces vs. monolithic list of everything possible; extensible (access to primitives via Graphics Template Language (GTL)) vs. fixed layouts). L A T E X, HTML, RTF 2 If you want to follow this up see[holland, 2005, Spruck, 2008] for workable approaches 3

4 1.1 ODS SG PROCS 1 ODS SG PROCS Figure 1: Sample Single- and multi-cell ODS Statistical Graphics There are fundamentally two types of layouts available and provided by different procedures: Single-cell layouts Multi-cell layouts PROC SGPLOT PROC SGSCATTER PROC SGPANEL PROC SGSCATTER Each cell in a layout can include many (overlaid, in order) plot types, but has one axis frame. Multi-cell plots have axis frames in each cell. A plot type takes paired data and places points on the plane; the exact representation used varies. For example, a SERIES has a joined line (with or without markers), a SCATTER plot has just points and a BAND plot shades areas. In SAS/Graph these different presentations are combinations of plot commands and symbol statements. Plot types also include types that make calculations and plot the results of those calculations rather than the points themselves, examples would be REG, LOESS, and BPSPLINES. There are restrictions on which plot types can be overlaid; for example, it does not make sense to overlay a scatter plot with a histogram. All the available plot types are listed separately in the SG PROCS manual under each procedure, which is a little confusing because they are almost all available in each SG PROC. How are graphics produced in ODS and SG procs? There are three ways to use the new Statistical Graphics infrastructure in SAS 9.2. These are: using plain ODS GRAPHICS, using the new SG procedures, and rolling your own template. It is not obvious but the underlying process is the same for all these routes: The data will be passed with a data= option but can also be generated internally, for example the plot might use residuals from a model fit. data + plot instructions + template graphic output file Depending on which method you choose, different parts of the process are automated for you. There is a progression from fully automated (ODS Statistical graphics) to flexible (build your own template) and the SG PROCS fall in between. 4

5 1 ODS SG PROCS 1.1 ODS SG PROCS In the next figures we see the process defined above and the grey-shaded areas that represent the operations the programmer does not see it is a black box, but here made grey to show what is inside. In ODS GRAPHICS the programmer just calls a SAS/STAT procedure and all components are assembled by the procedure. In the centre plot the programmer calls an SG proc and supplies some plot instructions (such as VBOX), so in this case only the template generation is hidden. In the third figure a template is created in one call of PROC TEMPLATE and this template is then called (probably later and perhaps from a different program). The Appendix table explains the how each method implements the process outline and the differences are illustrated below. ODS Statistical Graphics are designed for many SAS/STAT procedures and are builtin. Data to be used are passed to the STAT procedure in the normal way and may be calculated by the procedure itself, for example fitted values or boxplot limits. Some plots may be omitted but no new standard plots defined. The SAS/STAT procedure calls routines that render the output automatically. Note that these facilities are available in more than 50 Base SAS, SAS/STAT SAS/ETS and SAS/QC procedures as defined in the SAS user manuals 3. Note: this is equivalent to calling PROC SGRENDER. The SG PROCS allow more control by providing a custom SAS/GRAPH-like language for defining the plot. The procedure generates a GTL template on the fly and uses it automatically. It is gives the same result that a call of PROC SGRENDER would have. Two of the SG PROCS can save the dynamically generated GTL template to form the basis of a new customised plot. 3 List of SAS procedures supporting ODS Graphics for an up to date list 5

6 1.1 ODS SG PROCS 1 ODS SG PROCS The Graphics Template Language is used with a call of PROC TEMPLATE to make a plot available. To use it call PROC SGRENDER to create a graphic output file, possibly passing parameters (DYNAMIC elements). The GTL has elements not used in the standard templates (e.g. a FRINGEPLOT see Figure 19) and also the possibility to make calculations within the tempate. New in SAS 9.2 is the ODS Graphical Designer (in pre-production state at the moment) which is an interactive interface to PROC TEMPLATE, see the documentation mentioned in Further Reading and [HOLLAND, 2009]. How do groups, sub groups and BY variables map to plot cells? The single-cell layouts support BY group processing so they can actually automatically produce separate single-cell graphics. It is also important to note that all the multi-cell procedures use layout algorithms that split graphics over images as needed. So, for example, two factors both with 6 levels in a PROC SGPANEL does not produce a 6 x 6 lattice in one image by default. The example of a series plot is interesting. When plotting a series there is always the question of whether to use a factor to distinguish sub-groups, or to put the groups in different variables. The SG PROCS allow three approaches: Groups in observations PROC SGscatter ; plot x= y= ; by factor ; /* or */ PROC SGpanel ; panelby factor / layout= lattice_etc ; Groups in observations PROC SGscatter ; plot x= y= /group=factor ; run; Groups in variables PROC SGscatter ; series x=onex y=grp1 ; series x=onex y=grp2 ;... run; Plot types and how to combine them Different plot types can be combined and used in both single and multi-cell layouts. Not all combinations make sense however, and there are four groups of plot types that can be used together: Basic plots (Two-variable plots and models (2 variables, optional factor)) scatter, series, band, needle (~ dot plot),vector [plots] 6

7 1 ODS SG PROCS 1.1 ODS SG PROCS loess, regression, penalized b-splines, ellipse [models] Distributions (one variable, optional factor) histograms, normal curves, kernel density Box plots (one variable, optional factor) Categorization (one variable, factor) dot plots, bar charts, line charts Plot types in the same item can be contained in the one graphic cell together. Scatter plots never join up their points, although they can add ranges. To get joined up points use a SERIES plot. A third type is the JOIN plot, this type of plot does not require sorted data, points are sorted by X axis value internally. SAS 9.2 Phase 2 A new feature is that each plot command (i.e. layer) can be named. Some plots types can take a name and then will follow the named plot. This is useful for band plots, for example, where the bands will then follow a line defined elsewhere, perhaps a smoothed line created with PBSPLINE. In this sense some plot types are essentially annotations but without the complexity of an annotate dataset and macros. This feature is also useful to ensure that elements look uniform without repeating appearance code in each layer. Box plots are available in two orientations and are a distribution plot but in SG PROCS they are not combinable with any other plot type and are therefore separated here. It is also worth mentioning that there is one factor only in the But they can also GROUP= option allowed with box plots. This might seem an arbitrary limitation be put in a separate until it is realised that box plots can be also be used in the multi-cell procedures frame in the and therefore put in to layouts with one, two, three or more additional factors, see, same graphic. See for example, Figure 13. In addition to the plot types above there are also statements for controlling legends, labeling and scaling axes, axis transformations (such as log) and adding reference lines and labels. The use of GROUP= options creates automatic colouring for separate lines and by default this will cycle around fixed style names. These same style names can be used directly in other plot components, for example to apply colours. The cycling can be controlled and can also be set to the same as another named plot. Styles can be called in the ODS options (Listing Algorithm 1) and will set many characteristics of the plot elements, such as line colours, thicknesses, or dotting style. This makes for a flexible and configurable system. Appearance can be changed by setting the style (e.g. switch from colour for PDF to monochrome for publication), and new house styles can be created. And the style name does not need to be mentioned in the plot (or its template) itself. Figure 19 What s the big deal with layering? Layering means the components are added sequentially in the order defined by your program. The last component will appear to be on top of the others. This is very useful for many clinical plots, where for example we can add placebo observations last and therefore more easily see where treatment observations differ. Also, as you add more components SAS will resize and move components to keep the plot looking good. If you are using an output type that supports transparency then that can now be set in the plotting command, which is excellent news for those of us with lots of Lab data and over-plotting problems. 7

8 1.2 ODS set-up 1 ODS SG PROCS 1.2 ODS graphics set-up tips Using ODS as a way to generate output has many advantages, notably the ability to create more than one output format at a time and the ability to generate separate image files along with PDF or RTF. There are no intermediate graphics catalogs to manage and portable formats are available. Some defaults although sensible just cry out to be overridden (Listing Algorithm 1). Especially if you are developing a graphic interactively in SAS. In this case the RESET INDEX option is essential. The GPATH is likewise important because the graphic output files are otherwise put in the SAS 9.2 installation directory. Specifying the IMAGENAME will make distinguishing different sections of a program and different outputs possible. Listing 1 Suggested ODS settings for SG PROCS options nobyline ; ods HTML body= <progname>.html path= C:\Data\Docs\PhUSE2009 (URL=none) gpath="c:\data\docs\phuse2009\images" (URL=none) image_dpi=600 /* print quality */ style = statistical ; /* journal is monochrome, statistical otherwise preferred*/ ods graphics on / reset=index imagename="p-g-eg1_final" imagefmt=png width=21.0cm /* PA5 paper shape, landscape half A4/Letter*/ /* height=14.0cm ; /* 14.8 for A5 */ /* width=10.5cm height=14.0cm ;* PA6 paper shape, quarter A4/Letter ; /* 14.8 for A6 */ *================= part 1 individual box plots =======; ods rtf file="c:\data\docs\phuse2009\p-g-eg1_final.rtf"; ods pdf file="c:\data\docs\phuse2009\p-g-eg1_final.pdf"; proc sgplot data=phuse des= TS11 Example 1 ; title font=tahoma "TS11 Example 1"; title2 font=tahoma Height=0.72 #BYVAL1 ;... ods graphics off ; ods pdf close; ods rtf close ; /* This program creates png images, an HTML file, a PDF and an RTF vers in one pass. The HTML is a great way to browse all your output in the order of cr */ Notes on ODS settings 1. Always set the path option of the ODS destination. If you do not do this (or change your default setting using the SAS registry editor) your graphics files will otherwise disappear, possibly into the SAS installation directories. 8

9 2 EXAMPLES OF SINGLE CELL GRAPHICS 2. The aspect ratio is preserved by default. The default aspect ratio is 4:3, it can be changed in the SAS registry or here if both height and width are given. For A size papers a ratio of 1: 2 makes sense. The PA4 size of paper is the largest size that will fit both A4 and letter size. 3. The default resolution is chosen according to ODS destination (defaults are 100 DPI for HTML, 200 for listing) but we prefer to increase it. 4. The fonts used by SG PROCS are system fonts, so we can set our favourites in titles and footnotes 5. Setting descriptions for each plot is also vital for managing your results. Notes on examples numbering and figures These examples are based on programs in the SAS sample library and are self contained because they include either datasets in the SAS help library or code to simulate the data used. The example numbering is based on the dataset used; different plots of the same data are distinguished by letters. In some cases where there are several figures used to demonstrate components of the full plot they are also labeled by layer L1, L2,......,. Therefore Example 7c L3 is the picture/code/discussion of the data of Example 7 (US Cancer data) plot c, layer three. Code snippets are presented with their example plot (i.e. in the text) and for each example the whole code is presented separately to facilitate code reuse. 2 Examples of Single Cell Graphics We will show four examples of single cell graphs: 1. Clinical Response By Visit (Efficacy simple dose response) 2. Percent Change from Baseline: More complex dose / response graph that will illustrate how to layer 5 plots into one graph with complex annotations. 3. Plasma Concentration Profile highlighting an area under the curve (AUC) and drawn with a band plot. 4. A Graph from Trial Design to show how to use CURVELABELS, UNICODE FONTS to display uppercase and lowercase Greek letters, and how to overide automatic attribute cycling so all curves are drawn in the same way. In addition the example used to demonstrate GTL is a single-cell graphic (Figure 19). 2.1 Example 1 Clinical Response by visit for two treatment groups Our aim is simple: plot the response against time expressed as visit number, and separately for each treatment group. We would like to join the responses separately in each treatment group so we need a SERIES plot. The data for the series are X and Y values for each point which are joined up without point markers by default. If a GROUP= option is given then (as here with TRT) the values for each group identified by a common value of the factor TRT are joined. This is a Grouped Series Plot. 9

10 2.1 Grouped Series plot 2 EXAMPLES OF SINGLE CELL GRAPHICS Listing 2 Example 1 Clinical response by visit per treatment ods style=listing; proc sgplot data=response; format visit visitfmt. trt trtfmt.; series x=visit y=response / group=trt markers; refline 1 / axis=x; xaxis grid; yaxis grid; run; Figure 2: Example 1 Grouped Series plot of Clinical Response Slide 10 The data used for this plot would be typical for a condition such as Hypertension where the purpose of the drug under development is to reduce a measure such as Systolic blood pressure. In other words, the bigger the reduction the better. In PROC SGPLOT (and other SG PROCS) we have the SERIES statement which plots the data from two variables and optionally, as in this case, with a group factor trt. The lines for each group are distinguished automatically and because point markers are off by default we turn them on with the MARKERS option. It makes sense with these data to flag the start of treatment and so we add a reference line at day 1. By default a reference line is set on the Y axis so we need to add the AXIS=X option. The legend is generated automatically and drawn outside the plot by default. We also add grid lines on both X and Y axes. Note that the X axis values are automatically thinned to keep the plot uncluttered. Thinning means that not every day value supplied in the format is used. In this case the axis is thinned to even numbered days. This is another reason to mark day 1 when treatment starts. In the data we have two variables and the treatments are given by a factor. If we had the two treatment values in two separate variables we could use two series plots instead: series x=visit y=resp_super / markers ; series x=visit y=resp_old / markers ; It is not obvious from this plot but there is no need to sort the data or adjust the formats to fewer points: meaning SAS is using defaults much closer to what is most often needed. 10

11 2 EXAMPLES OF SINGLE CELL GRAPHICS 2.2 Dose Response plot 2.2 Example 2 Dose response graph with three treatment groups and washout This example is similar to Example 1 except there are three treatment groups and there is a more complicated treatment design. At week 12 half of the patients in each treatment group cross-over to a placebo. Weeks 12 to 15 are the washout period for these patients. We would expect the lines for each group to split at week 12 to be able to show this. Obviously we would like these lines and patients to get a different visual treatment to distinguish them from the patients in those groups remaining on drug. We need a grouped series plot - but there is an issue because a series plot cannot add error bars to the points. The error bar information is in separate variables (low and high) and therefore we use a SCATTER plot to get the points and intervals, and then add only the lines with a SERIES plot. We would like to highlight the washout period with a shaded area and suitable label on the X axis. Figure 3 shows the final result. Figure 3: Example 2 Final plot - Grouped dose-response graph This graph is built in six layers. These follow section by section, the full code is in Listing Algorithm 3. Slide 11 Layer 1 The first step is to set up the axes this is done with XAXIS and YAXIS statements. The second step is to draw the shaded band for the washout period. It is next because it should be behind all other lines or points. To draw a band we need two X values for each Y value. These four points define a rectangle, essentially what could be done with an annotate macro when using SAS/GRAPH. Two observations have values of YMAX and the values are are : 0 and 15. The remaining obs have ymax=missing values so they are ignored by the band. So the band draws from Y=0 X=12.1 to 15 (bottom edge), and upper edge is defined by the observation with Y=15 and X=12.1 to 15. Note that only the Y values come from the dataset. The X range (which depends on the treatment design) is embedded in the program. 11

12 2.2 Dose Response plot 2 EXAMPLES OF SINGLE CELL GRAPHICS Listing 3 Example 2 Grouped dose-response graph with labelling proc sgplot data=response; title1 Mean Percent Change from Baseline band y=ymax lower=12.1 upper=15 / transparency=.8 fillattrs=graphdata1; xaxis integer values=(0 to 15 by 3) label="weeks in Treatment"; yaxis label="percent Change"; scatter x=week y=pct / group=trt yerrorlower=low yerrorupper=hi markerattrs=(symbol=circlefilled) name="scat"; keylegend "scat" / title="" noborder; series x=week y=pct / group=trt lineattrs=(pattern=solid); series x=week y=end / group=trt lineattrs=(pattern=shortdash) markers markerattrs=(symbol=circle); refline 0; refline 13.5 / axis=x lineattrs=(thickness=0px) label=" -- Washout -- " labelloc=outside labelpos=min; run; We paint the band with the transparency set to 0.8 which is very light but we do this so the band does not overwhelm the graph. Transparency goes from 0 (completely opaque) to 1 (completely transparent). We do not specify the fill colour directly but by naming a style component because doing it this way ensures ease of maintenance and compatibilty with current and future styles. band y=ymax lower=12.1 upper=15 / transparency=.8 fillattrs=graphdata1; xaxis integer values=(0 to 15 by 3) label="weeks in Treatment"; yaxis label="percent Change"; Layer 2 We make a scatter plot of the means, minimums, and maximums for each treatment/time. It has to be done this way because the scatter plot allows error bars but the series plot does not. The low and hi variables are calculated in the dataset. scatter x=week y=pct / group=trt yerrorlower=low yerrorupper=hi markerattrs=(symbol=circlefilled) name="scat"; keylegend "scat" / title="" noborder; In the LISTING style (used here) GRAPHDATA1 has an open circle for its marker symbol, GRAPHDATA2 has a plus sign, and GRAPHDATA3 has an X cross. By setting the marker attributes to the filled-circle we prevent (only) that attribute changing with the treatment. But the colour still cycles with the group. 12

13 2 EXAMPLES OF SINGLE CELL GRAPHICS 2.2 Dose Response plot Figure 4: Layer 1 Figure 5: Layer 2 Figure 6: Layer 3 (code on on the next page) Figure 7: Layer 4 (code on on the preceding page) Figure 8: Layer 5 (code on the next page) Figure 9: Layer 6 (code page on page 15) Figure 10: Example 2 Dose Response graph with three treatment groups 13

14 2.2 Dose Response plot 2 EXAMPLES OF SINGLE CELL GRAPHICS There are 3 group values: placebo, 10mg and 20 mg. The appearance attributes are assigned from the first 3 group elements in the style: GRAPHDATA1, GRAPH- DATA2 and GRAPHDATA3. So the placebo scatter plot get its marker colour/symbol/size and errorbar line colour/pattern/thickness as set in the element GRAPHDATA1. The 10mg group gets its attributes from GRAPHDATA2, etc. On the previous slide the band plot fill was set to GRAPHDATA1 with a.8 transparency. GRAPHDATA1 is the same blue in the band plot and the first group value (the placebo group). The scatter plot is named (scat) - this allows us to select the plot which gets the legend generated for it with the KEYLEGEND statement. The legend is to be drawn with a blank title and no border. It is placed at the bottom by default so no position is specified. Layer 3 The next step is to add the lines joining the points for each treatment. A grouped series plot with the same X= Y= and GROUP= as the scatter plot will do this. The series joins the markers from the scatter plot. series x=week y=pct / group=trt lineattrs=(pattern=solid); Each grouped plot starts from element GraphData1. So the colours of the grouped series plot match the colours in the grouped scatter plot. Instead of taking the line patterns from the style, this code is setting it manually to a solid line. As with the marker in the scatter plot, the override is applied to all 3 groups. Layer 4 A second grouped series plot to draw the week 12 to week 15 results of the patients who were crossed over to placebo in the washout period. /* placebo in washout */ series x=week y=end / group=trt lineattrs=(pattern=shortdash) markers markerattrs=(symbol=circle); The same X (WEEK) and group (TRT) variables are used as in the scatter and first series plots. The Y variable (END) is valued with the week 15 responses for the cross-over patients only, it is set to missing values elsewhere. Again, by default all grouped plots derive their appearance attributes from the GRAPHDATA1 to GRAPHDATAN style elements. So the colours of the second grouped series plot match the colours in the grouped scatter and series plot. We are overriding two attributes in this series plot: the LINEPATTERN is set to a dashed line (for all groups) and the marker symbol is set to an open circle (for all groups). As before, the overrides are applied automatically to all 3 groups. Layer 5 A reference line is drawn at Y=0. Slide 16 refline 0; 14

15 2 EXAMPLES OF SINGLE CELL GRAPHICS 2.3 Plasma Concentration Layer 6 A reference line is specified at X=13.5. But the line thickness (attrs=) is set to 0 so the line does not actually draw. But we can still set a label on the reference line. The label is the text --Washout--, it is drawn outside (the plot area) and positioned at the minimum point ( = 0) of the reference line. Slide 17 refline 13.5 / axis=x lineattrs=(thickness=0px) label=" -- Washout -- " labelloc=outside labelpos=min; Notice there is nothing in Listing Algorithm 3 requiring a macro variable to deliver the plot. There is nowhere a dependency on the number of treatments or their labels for example. The number of weeks treatment, position and length of the washout are the only fixed assumptions. Of course the data set has needed some work to implement the treatment cross-over but given knowledge of the cross-over week that could potentially be automated too. Techniques similar to this plot could be adapted for wider use. 2.3 Example 3 Plasma Concentration Profile and labeled AUC Slide 18 Listing 4 Example 3 Plasma Concentration proc sgplot data=plasma noautolegend; band x=time lower=0 upper=conc; series x=time y=conc / markers markerattrs=graphdata1(symbol=circlefilled); vector xorigin=2 yorigin=1.5 / /* AUC */ x=vx y=vy lineattrs=(color=black) datalabel=vlabel arrowshape=filled; vector xorigin=0 yorigin=6.76 / /* Cmax */ x=vx2 y=vy2 lineattrs=(pattern=mediumdash color=black) datalabel=vlabel2 noarrowheads; vector xorigin=.75 yorigin=0 / /* Tmax */ x=vx3 y=vy3 lineattrs=(pattern=shortdash color=black) datalabel=vlabel3 noarrowheads; xaxis offsetmin=0 label= Time relative to dose (hours) ; yaxis offsetmin=0 max=8.0 label= Plasma concentration [ng/ml] ; run; Here we have a series plot of the concentration values but we wish to visually illustrate the area under the curve (AUC) by shading the space between the lines and X axis. We start with a horizontal band plot - it has two Y values for each X. The band FILLCOLOR is from GRAPHCONFIDENCE 1. SERIESPLOT with x=time (same) and y=conc (same as upper point in band), markers set to GRAPHDATA1 2. Vector plot for the AUC starts at (2,1.5) to (3.5,3) and 15

16 2.4 Sample Size 2 EXAMPLES OF SINGLE CELL GRAPHICS a) overrides the default line colour to black b) overrides the default marker symbol (open circle) to closed circle c) adds a label on the shaded area (AUC) as a vector for Cmax, with LINEPAT- TERN = MEDIUMDASH and without arrowheads d) adds a vector labelling Tmax, LINEPATTERN = SHORTDASH, also with no arrowheads e) minimizes white space around the axes by adding OFFSETS = 0 on the axes and customised labels Figure 11: Example 3 Plasma Concentration profile with Labelling (Listing Algorithm 4) 2.4 Example 4 Sample Size Requirements This example plots fitted values as multiple lines and them all the same colour and line style. It also uses Greek letters in the axis labels and demonstrates how to make an text insertion to the graph. The data have been calculated in a previous datastep and then stored as variable pairs for each value of n (25, 50, 100, 500). For theoretical reasons it makes most sense to smooth the calculated points with a 2nd degree polynomial so a FIT plot type is used. Rather than using a legend we would like to label the curves directly because this often gives an easier to understand graphic. NOCYCLEATTRS take styles for all lines from the same style element. This graph has 4 regression statements in it. The default behavior is to cycle through GRAPHDATA1 to GRAPHDATAN when there are multiple occurrences of the same plot statement. NOCYCLEATTRS stops the cycling so each regression is drawn with the same attributes, from the GRAPHFIT element. The first Regression Fit Line with CURVELABEL text set to n=500 and with CURVELABEL position set to max it puts text at the right-hand end of fitted line. Note it is not curved label, but a label for a curve. Markers are suppressed and we just display the line. We repeat this code for each of the four sample sizes. 16

17 2 EXAMPLES OF SINGLE CELL GRAPHICS 2.4 Sample Size Figure 12: Example 4 Sample Size Requirements (Listing Algorithm 6) Listing 5 Example 4 Required sample sizes under differing assumptions ods escapechar="~"; /*used below */ proc SGplot data=txsize nocycleattrsnoautolegend; reg x=c500 y=t500 / curvelabel="n=500" curvelabelpos=max degree=2 nomarkers; reg x=c100 y=t100 / curvelabel="n=100" curvelabelpos=max degree=2 nomarkers; reg x=c50 y=t50 / curvelabel="n=50" curvelabelpos=max degree=2 nomarkers; reg x=c25 y=t25 / curvelabel="n=25" curvelabelpos=max degree=2 nomarkers; refline.5 / axis=x; xaxis values=(0.0 to 1.0 by.1) grid label="response rate in the control group" ; yaxis min=0.0 max=.5 grid label="anticipated smallest treatment effect /*line-break*/ (~{unicode delta_u} vs control)" ; inset ( "~{unicode alpha} =" = "5%" "~{unicode beta} =" = "20%" ) / position=topright textattrs=graphtitletext noborder; run; 17

18 3 MULTI-CELL GRAPHICS In the next step a reference line is added at the 50% response rate. The ODS ESCAPECHAR signals unicode specification following between curly brackets {unicode delta_u}. This will display as a in the Y axis label. The INSET statement allows arbitrary labelling to be added and is drawn in the upper right, and consists of two value pairs, α = 5% and β = 80%. TEXTATTRS assigns a title (bolder) font and font size to the inset characters ensuring they match the style of the graph title. 3 Multi-cell graphics As a general rule, multi-cell graphs are a series of single cells generated with a list of variables and/or factors. Plots are created automatically depending on the number of factor levels. A lattice contains 1 column factor and 1 row factor. A panel contains 1 or more factors, nested in a series, from major to minor. A column lattice contains 1 factor which forms the columns, all cells are displayed in 1 row. A row lattice is exactly the opposite, with 1 classifier and all cells are displayed in 1 column. A scatter plot matrix displays a matrix of scatter plots, that is, given a list of variables it generates a square plot with each variable plotted against the others. The diagonal can be filled in various ways, e.g. with a histogram. Independent scatter plots or restricted matrix plots can also be created: given two lists of variables, all in the first list are plotted against all in the second list. There is no diagonal in this plot type and the resulting panel is not usually a square. 3.1 Example 5 Lattice, panel, and row or column lattice? Demography Age distribution Box plot by Treatment in Lattice layout This example shows how to split a box plot against a factor by two more factors using a lattice or panel layout. This is defined in the PANELBY statement of PROC SGPANEL. A panel has a rectangular layout fitting into the defined plotting space, it allows one, two or more factors and allocates the factor combinations in order into the cells of the plot. A Lattice also has a rectangular layout but accepts exactly two factors. The dimensions then depend on the number of levels of each factor. It is akin to a table or crosstabulation. Row lattice and COLUMNLATTICEs accept only one factor and then layout the cells in one dimension. The terminology might be confusing: Remember a ROW- LATTICE has lots of rows, and a COLUMNLATTICE has lots of columns. 18

19 3 MULTI-CELL GRAPHICS 3.1 Box Plots in Boxes Figure 13: Example 5a Box plots of Age distribution by treatment in lattice Figure 14: Example 5b Box plots of Age distribution by treatment in column lattice Layout of Demography Data 19

20 3.2 Column Lattice 3 MULTI-CELL GRAPHICS Listing 6 Example 5a Code for Box plots of Age distribution by treatment in lattice layout proc sgpanel data=demog; panelby gender race / layout=lattice; hbox age / category=trt datalabel; rowaxis display=(nolabel); run; Example 5a Demographic data in a lattice layout The example shown in Figure 14 uses the PROC SGPANEL procedure and demonstrates two possible layouts and how to achieve them. A lattice contains a column factor and a row factor and in this case we get a square panel because both factors have two levels. We plot an horizontal box plot of AGE, displayed by the factor TRT. The factors are set on PANELBY statement: the first GENDER is the column and the second RACE is the row. The layout type is LATTICE. In this plot Outliers values are labeled (DATALABEL). Note that the SGPANEL procedure uses ROWAXIS and COLUMNAXIS instead of XAXIS and YAXIS used in the single-cell plots. The Row Axis label is suppressed it would otherwise be the variable name. Slide 22 Example 5b Demographic data in a panel layout A panel contains 1 or more factors, nested in a series, from major to minor (cycling fastest). This is the same data as the previous LATTICE example. Only the layout is different and is in a PANEL with the number of columns set to four. This makes it equivalent to a COLUMNLATTICE in shape, but here we can use two factors, unlike a COLUMNLAT- TICE layout. Listing 7 Example 5b Demography data in a Panel layout in one row proc sgpanel data=demog; panelby gender race / layout=panel columns=4; hbox age / category=trt; rowaxis display=(nolabel); colaxis values=( ); run; The PANELBY has the same classifiers. GENDER is first and RACE is second. The number of columns is set explicitly to 4. As before, the horizontal box plot is of AGE, displayed by the category TRT however in this plot the outliers values are not labeled. The Row Axis label is suppressed and the column axis values are set explicitly. Slide Example 6 Primary and Secondary endpoints This is an example taken from analysis of a clinical trial that has primary and secondary endpoints. The primary endpoints are a gatekeeper to the secondary endpoints. H11 and H12 are the primary endpoints, H21 and H22 are the secondary endpoints. Primary are 28-day mortality and number of days off a ventilator during the 28 day trial. Either are enough for a regulatory claim. The secondary endpoints are number of days out of the ICU and quality of life -these can be additional regulatory claims if significant. 20

21 3 MULTI-CELL GRAPHICS 3.3 Row Lattice cancer data Figure 15: Example 6 A COLUMNLATTICE display of test results A COLUMNLATTICE contains one and only one factor (here it is TRT), with all cells displayed in one row. The panelby specifies NOVARNAME, which suppresses the TRT= in the cell header. There are 2 SERIES plot so the procedure automatically switches to using the GRAPHDATA1 and GRAPHDATA2 elements and the two series are blue and red, just as they were in previous grouped plots that used GRAPHDATA1 and GRAPHDATA2. In the first series the only customization is the size of the marker. The open circle is the default marker. In the second series both the marker size and marker symbol are set. ROWAXIS and COLUMNAXIS labels are set and a dashed reference line is drawn at Y=0.05. Note that the legend is drawn by default, and identifies the 2 series plots. There is no default legend title in the non-grouped case and of course it takes no space in that case. Listing 8 Example 6 P-values in a row lattice proc sgpanel data=hyp; panelby trt / novarname layout=columnlattice; series x=hyp y=rawp / markersmarkerattrs=(size=6pt); series x=hyp y=adjp / markersmarkerattrs=(size=6pt symbol=square); rowaxis label= P-value ; colaxis label= Hypothesis ; refline 0.05 / lineattrs=(pattern=shortdash); run; 3.3 Example 7a US Cancer data lattice by Gender In this example we have the number of case and number of deaths from different types of Cancer in the USA from 2007 (American Cancer Society). Here we want 21

22 3.3 Row Lattice cancer data 3 MULTI-CELL GRAPHICS to show bar charts but split by gender and adjacent to make comparison easier. Although the profile of common cancers in the two sexes is very different PROC SGPANEL will let us do that. We use a side by layout with a row lattice. Slide 24 A ROWLATTICE has one factor like a COLUMNLATTICE but the orientation is the opposite: all cells are displayed in 1 row. We want the category values to display in response order not category order. The procedure will force a sort if it detects that the input data set is not sorted in order of the PANELBY variables. So we sort by our one factor (sex) and then by descending order of the response (number of cases in the variable cases). SAS 9.2 Phase 2 ONEPANEL forces all rows to appear in the same graph (regardless of how small each row gets). NOVARNAME suppresses the SEX= in the cell header. UNISCALE=COLUMN tells the procedure not to display category values if they do not have any data for the classifier. HBAR creates a horizontal bar chart. Listing 9 Example 7 US Cancer data HBAR lattice by Gender proc sort data=cancer ; by sex descending count; run; proc sgpanel data=cancer ; panelby sex / layout=rowlattice novarname uniscale=column; hbar cause / response=cases stat=sum; rowaxis display=(nolabel) discreteorder=data ; colaxis display=(nolabel); run; The row axis DISCRETEORDER=DATA prevents reordering of the categories and preserves the presorted order of the descending responses. So we have a ROWLATTICE that displays the types of cancers diagnosed for each sex in order. Lung Cancer is the second most diagnosed cancer for both men and women, with number one for women being breast and prostate for men. Note what UNISCALE=COLUMN does here. It tells the procedure not to display category values if they do not have any data for the classifier. So the Female row does not contain any PROSTATE, and the Male row does not contain any OVARIAN. By default UNISCALE=ALL. This plot while useful does not make it easy to compare the distribution of cancers between the two sexes. For one, there are different rows in each so the profile is not visible. We will therefore return to these data below with the view to making a better comparison using a butterfly plot. We will also add a way of illustrating the difference between deaths and number of cases reported. 22

23 3 MULTI-CELL GRAPHICS 3.3 Row Lattice cancer data Example 7b USA Cancer Data Butterfly Plot A butterfly plot is used to compare two halves of a whole by drawing bars reflected along a vertical axis of symmetry. The trick is to make negative values for the bars on the left and retain positive values to make bars on right X axis. To conceal this trick we use a Picture Format for the X axis to hide the minus signs. Output format - We would also like to show the comparison between number of deaths and number of cases. Using transparency this can easily be done. So we will use overlaid bar charts with different bar widths and transparencies. There is not a standard way to do this. No PROC BUTTERFLY exists. And neither is it needed because with the SG PROCS we can break the picture into four HBAR charts and then build it layer by layer. The data are arranged with four variables for each type of cancer, these being male cases & deaths, and Female cases and deaths. This example with code and data are available at: Layer 1 First assign the user-defined format to the male and female cases and deaths variables. Then we draw bars for the male cases. Note that lung cancer mcases = -114,760. proc sgplot data=cancer; format mcases mdeaths fcases fdeaths positive. ; hbar cause / response=mcases fillattrs=graphdata1 transparency=.65 legendlabel="new Cases (Male)" name="mcases" ; run; The bars use the fill colour from GRAPHDATA1 with a lighter transparency applied. Notice that in this layer we have not specified the order and therefore the Y axis is sorted aphabetically. There is only one plot and so no legend is created. Also the X axis shows just the data range (all negative so zero is on the right). this Butterfly plot and its layers were created as postscript files (using the L A T E X destination), most of the other plots were created as PNG files (a bit mapped format). 23

24 3.3 Row Lattice cancer data 3 MULTI-CELL GRAPHICS (a) Layer 1 (b) Layer 2 (c) Layer 3 (d) Layer 4 (e) Example 7b Complete - Layer 5 Figure 16: Example 7b Butterfly plot of USA Cancer Data Layer by Layer 24

25 3 MULTI-CELL GRAPHICS 3.3 Row Lattice cancer data Layer 2 cases. Then the bars for the male deaths are overlaid onto the bars for male hbar cause / response=mdeaths barwidth=.5 fillattrs=graphdata1 transparency=.25 legendlabel="deaths (Male)" name="mdeaths" ; These bars also use the fill colour from GRAPHDATA1 with more saturation transparency is only.25 instead of.65. The bars are the same colour but you can distinguish between them. Note that for lung cancer mdeaths= Also note that a legend has appeared, the consequent shrinking of the plotting space has automatically thinned the Y axis labels. Layer 3 Then the bars for the female cases are drawn. These bars use the fill colour from GRAPHDATA2 instead of GRAPHDATA1. They have the same saturation as the male cases transparency=.65. hbar cause / response=fcases fillattrs=graphdata2 transparency=.65 legendlabel="new Cases (Female)" name="fcases" ; Note that female cases are drawn on the positive side of the axis, lung cancer fdeaths= Layer 4 Next the bars for the female deaths over overlaid onto the bars for female cases. The same approach for fill colour and transparency is used on the female side. hbar cause / response=fdeaths barwidth=.5 fillattrs=graphdata2 transparency=.25 legendlabel="deaths (Female)" name="fdeaths"; Note that lung cancer fdeaths= Layer 5 Then we finish the graph by moving the legend to better match the panel layout, removing the default axis labelling by variable name and ensuring the Y axis is ordered according to the total number of deaths: Slide 38 keylegend "mcases" "fcases" "mdeaths" "fdeaths" / across=2; yaxis label=" " discreteorder=data; xaxis label=" " grid; The KEYLEGEND draws the keys in the order we define using the names we set on each HBAR statement; the left column is male cases and deaths, the right column is female cases and deaths. The Y axis is told to retain the categories in their presorted order, and for both axes the default labels are removed by setting them to. 25

26 3.3 Row Lattice cancer data 3 MULTI-CELL GRAPHICS Listing 10 Example 7b Butterfly plot of USA Cancer data proc sgplot data=cancer; format mcases mdeaths fcases fdeaths positive.; hbar cause / response=mcases fillattrs=graphdata1 transparency=.65 legendlabel="new Cases (Male)" name="mcases" ; hbar cause / response=mdeaths barwidth=.5 fillattrs=graphdata1 transparency=.25 legendlabel="deaths (Male)" name="mdeaths" ; hbar cause / response=fcases fillattrs=graphdata2 transparency=.65 legendlabel="new Cases (Female)" name="fcases"; hbar cause / response=fdeaths barwidth=.5 fillattrs=graphdata2 transparency=.25 legendlabel="deaths (Female)" name="fdeaths"; keylegend "mcases" "fcases" "mdeaths" "fdeaths" / across=2; yaxis label=" " discreteorder=data; xaxis label=" " grid; run; 26

27 3 MULTI-CELL GRAPHICS 3.4 LFT Panel plot So this graph shows the breakdown of cancer diagnoses and cancer deaths by sex. Because of the way we presorted the data it also shows that lung cancer has the highest number of deaths, and colorectal cancer for the 2nd highest number of deaths. Note that breast cancer has the 2nd highest number of deaths for Females, but colorectal cancer has the highest number of total deaths. Example 8a LFT safety panel study vs baseline Now let us construct some panels of laboratory values using the techniques we have just seen. Slide 27 This firrst example constructs a panel of Liver Function Tests and for each parameter plotting Baseline values vs all those on study. This is a ungrouped, graphical version of a shift table. Listing 11 Example 8a LFT safety panel proc SGpanel data=labs; format visitnum wk. labtest lab. drug $trt.; panelby labtest visiting / novarname layout=lattice onepanel; scatter x=pre y=result / group=drug; refline / axis=x lineattrs=(pattern=dash); refline / axis=y lineattrs=(pattern=dash); rowaxis integer min=0 max=4 label= Study (/ULN) ; colaxis integer min=0 max=4 label= Baseline (/ULN) * ; keylegend / title=" " noborder; run; Lines for ULN Thresholds Customized Row and Column Axis Values Slide 29 Figure 17: Example 8a LFT baseline vs On trt 3.4 Example 8b LFT panel per patient Having perhaps identified patients of interest from the previous, or other plots we often wish to examine certain patients in more detail. Here we have selected a 27

28 4 GTL Slide 30 subset of patients and want to show the time course of their readings by visit. To do this we set up a series plot for each LFT variable and simply panel by patient. It is important in this case to show the treatment duration, which can of course be different for each patient. We use a shaded area to accomplish this. Figure 18: Example 8b labs LFT data per patient We use PANELBY which defaults to Layout=Panel because we want a general plot without a dependency on the number of patients in the selected subset. Data Here we are expecting the LFT data in a separate variable for each LFT parameter, hence there are no GROUP= options used. 4 The Graph Template Language (GTL) A template is a program that describes individual graphs, written using the Graph Template Language, and stored in a SAS itemstore (SAS7BITM) file. The Graph Template Language is an extension to the TEMPLATE procedure that combines data representations, layouts, and plots in flexible ways. For example it supports statistical computations and plot types and, for speed, ODS Styles for a variety of usages. GTL is a powerful tool for creation of complex analytical graphs and as already discussed GTL is the same tool that is used to create the automatic graphics in 28