MultipleLinearRegression - test for minimum number of samples

• Key: MATH-279
• Type: Bug
• Status: Closed
• Resolution: Fixed
• Priority: Minor
• Assignee: Luc Maisonobe
• Reporter: Michael Bjorkegren
• Affects Version/s: 2.0
• Fix Version/s: 2.0

Description

It's currently possible to pass in so few rows (samples) that there isn't a row for each column (predictor). Does this look like the right thing to do?

Index: AbstractMultipleLinearRegression.java
===================================================================
--- AbstractMultipleLinearRegression.java       (revision 786758)
+++ AbstractMultipleLinearRegression.java       (working copy)
@@ -91,6 +91,9 @@
                   "dimension mismatch {0} != {1}",
                   (x == null) ? 0 : x.length,
                   (y == null) ? 0 : y.length);
+        } else if (x[0].length > x.length){
+            throw MathRuntimeException.createIllegalArgumentException(
+                    "not enough data (" + x.length + " rows) for this many predictors (" + x[0].length + " predictors)");
         }
     }

Comments

[Luc Maisonobe]

fixed in subversion repository as of r791244
patch was only slightly modified to avoid a NPE
thanks for the patch

[Luc Maisonobe]

closing resolved issue for 2.0 release