Check for outliers, equal variance, normality, and model validity
You just ran your SAS analysis; you are almost done.
Now you have to examine your dataset for three potential problems in your data. If one or more of these occur, you will have to make corrections and then rerun your SAS analysis. It is uncommon for datasets to have any of these problems, but if they occur, your conclusions can be altered. So you have to check.
Go to Step 10 of your analysis and click on your example in the gray Example output table to display the output in a popup window. It will be helpful to have the analysis example ouput for your particular design available, to refer to as you move through this diagnostics module. In the example output, yellow, clickable arrows identify comments for each of the diagnostics.
The ANOVA tests depend on data that are normally distributed with equal variance. You use probability values to make decisions about treatment differences. These probabilities are incorrect if the original data are not normal. Equal variance is needed because variability within each treatment is pooled to create an error term. Clearly, if variances are not equal, this one pooled error term will be too large for some treatments and too small for others. Again, this produces incorrect probabilities. Generally, these two problems can be corrected by transforming your data. And finally, outliers are a primary cause of unequal variance and non-normality. Since by definition outliers are bad data points, these should be removed before making final conclusions.
If your analysis includes block or covariate effects (if you chose something besides CRD in column 1 and/or chose Covariate in column 3 of Choose Design), there is an additional check that needs to be done for each. These will be covered in Step D. These steps will not be used if you determine your model does not include these terms.
Therefore, this module will lead you through identifying and correcting these problems. Examples will be shown to illustrate each of the problems. You can also click the Examples tab to view real datasets having these problems. If you are running your own dataset, compare its diagnostics to the illustrations.
You check for outliers first, in Step A, because they affect all other diagnostics. Then you continue with Steps B and C, to determine if you need to transform your data to correct unequal variance and/or non-normality (and possibly also apparent outliers). The module will show you how to transform your data. You will finish with Step E, which will summarize what you should do next. You are sent to Step F only if necessary.
The diagnostics checking process that you will learn here is complicated by each diagnostic potentially affecting others. In this module, unlike other modules in which we can lead you more clearly through sequential steps, you will bounce back and forth among the various steps. For example, outliers may cause non-normality, but also, non-normality may cause apparent outliers. So you will first check all diagnostics before fixing any one problem.
Any corrections of diagnostic problems will require re-analysing the dataset (re-doing the ANOVA module that led you here), either because outliers were deleted or your model was transformed. After this re-analysis, you will need to go through the complete diagnostics checking steps again. This is due to the possibility that deleting some outliers will cause other outliers to be identified. Also, the chosen transformation may not completely correct normality or unequal variance.
Click next >> at the bottom of each page to move through this module.
|H I N T S :|
This ANOVA Diagnostics module serves all experimental and treatment design combinations. After finishing it,
you may need to return to your analysis module and re-run SAS.
|For a more detailed explanation of ANOVA output, click the Examples tab.|