Dealting with outliers in Prism

Identifying and excluding outliers when fitting curves with nonlinear regression

While Prism does not identify outliers automatically, it does let you manually exclude values you consider to be outliers. From a data table, select the outlier(s), drop the Edit menu, and click the Exclude button

Prism will display excluded values in blue italics and append an asterisk. Such values will be ignored by all analyses and graphs. We suggest that you document your reasons for excluding values in a floating note.

Note that sometimes outliers provide very useful information. Think about the source of variation before excluding an outlier.

When you use Prism to fit curves with nonlinear regression, you can choose automatic outlier rejection using a method we developed (reference below). This is a choice on the first tab of the nonlinear regression dialog. Because this method combines robust regression and outlier detection, we call it the ROUT method.

You can also use this method to detect one or more outliers in a stack of data in a Column table. To do this you have to convert your table to be an XY table, and then use nonlinear regression to fit a straight line model, constraining the slope to equal 0.0. In this case, Prism fits only one parameter which is labeled intercept, but is really the mean. Follow these steps:

1.	Go to the Format Data Table dialog and change the table so it is an XY table.

2.	In that dialog, choose to make X a series so you don't have to bother entering data. It doesn't matter what starting value and increment you use, as the X values will be ignored.

3.	Click Analyze and choose Nonlinear regression from the list of XY analyses.

4.	On the fit tab, choose the classic model "Polynomial: First order (straight line) and choose the fitting method to be "automatic outlier elimination"

5.	Go to the Constrain tab, and constrain slope to equal 0.

6.	Go to the Weights tab and check the value of the ROUT coefficient. We recommend setting it to 1%. Increasing the value tells Prism to be more aggressive when defining a point to be an outlier, and decreasing the value makes Prism be more conservative about defining points to be outliers.

8.	At the bottom of the tabular results, Prism will report how many outliers it excluded. You can view a table of these values as a separate results page.

9.	If you want to exclude these values from graphs or t tests or anovas, go to the data table, select the questionable values, right click and choose Exclude. The values will appear in blue italics, and will be ignored by all graphs and analyses.

Motulsky HM and Brown RE, Detecting outliers when fitting data with nonlinear regression – a new method based on robust nonlinear regression and the false discovery rate, BMC Bioinformatics 2006, 7:123. Download from http://www.biomedcentral.com/1471-2105/7/123.