The classical analysis of variance (ANOVA) technique that is based on least squares assumes that the underlying experimental errors are normally distributed. However, data often contain outliers as a result of recording or other errors. In other cases, extreme responses occur when control variables in the experiments are set to extremes. It is important to distinguish among these extreme points and determine whether they are outliers or important extreme cases. You can use the ROBUSTREG procedure for robust analysis of variance based on M estimation. Usually there are no high-leverage points in a well-designed experiment, so M estimation is appropriate.

This example shows how to use the ROBUSTREG procedure for robust ANOVA.

An experiment studied the effects of two successive treatments (`T1`

, `T2`

) on the recovery time of mice that had certain diseases. Sixteen mice were randomly assigned to four groups for the four
different combinations of the treatments. The recovery times (`time`

) were recorded (in hours) as shown in the following data set:

data recover; input T1 $ T2 $ time @@; datalines; 0 0 20.2 0 0 23.9 0 0 21.9 0 0 42.4 1 0 27.2 1 0 34.0 1 0 27.4 1 0 28.5 0 1 25.9 0 1 34.5 0 1 25.1 0 1 34.2 1 1 35.0 1 1 33.9 1 1 38.3 1 1 39.9 ;

The following statements invoke the GLM procedure (see Chapter 45: The GLM Procedure) for a standard ANOVA:

proc glm data=recover; class T1 T2; model time = T1 T2 T1*T2; run;

Output 86.2.1 indicates that the overall model effect is not significant at the 10% level, and Output 86.2.2 indicates that neither treatment is significant at the 10% level.

The following statements invoke the ROBUSTREG procedure and use the same model:

proc robustreg data=recover; class T1 T2; model time = T1 T2 T1*T2 / diagnostics; T1_T2: test T1*T2; output out=robout r=resid sr=stdres; run;

Output 86.2.3 shows some basic information about the model and the response variable `time`

.

The "Parameter Estimates" table in Output 86.2.4 indicates that the main effects of both treatments are significant at the 5% level.

Output 86.2.4: Model Parameter Estimates

Parameter Estimates | |||||||||
---|---|---|---|---|---|---|---|---|---|

Parameter | DF | Estimate | Standard Error |
95% Confidence Limits | Chi-Square | Pr > ChiSq | |||

Intercept | 1 | 36.7655 | 2.0489 | 32.7497 | 40.7814 | 321.98 | <.0001 | ||

T1 | 0 | 1 | -6.8307 | 2.8976 | -12.5100 | -1.1514 | 5.56 | 0.0184 | |

T1 | 1 | 0 | 0.0000 | . | . | . | . | . | |

T2 | 0 | 1 | -7.6755 | 2.8976 | -13.3548 | -1.9962 | 7.02 | 0.0081 | |

T2 | 1 | 0 | 0.0000 | . | . | . | . | . | |

T1*T2 | 0 | 0 | 1 | -0.2619 | 4.0979 | -8.2936 | 7.7698 | 0.00 | 0.9490 |

T1*T2 | 0 | 1 | 0 | 0.0000 | . | . | . | . | . |

T1*T2 | 1 | 0 | 0 | 0.0000 | . | . | . | . | . |

T1*T2 | 1 | 1 | 0 | 0.0000 | . | . | . | . | . |

Scale | 1 | 3.5346 |

The reason for the difference between the traditional ANOVA and the robust ANOVA is explained by Output 86.2.5, which shows that the fourth observation is an outlier. Further investigation shows that the original value of 24.4 for the fourth observation was recorded incorrectly.

Output 86.2.6 displays the robust test results. The interaction between the two treatments is not significant.

Output 86.2.7 displays the robust residuals and standardized robust residuals.

Output 86.2.7: PROC ROBUSTREG Output

Obs | T1 | T2 | time | resid | stdres |
---|---|---|---|---|---|

1 | 0 | 0 | 20.2 | -1.7974 | -0.50851 |

2 | 0 | 0 | 23.9 | 1.9026 | 0.53827 |

3 | 0 | 0 | 21.9 | -0.0974 | -0.02756 |

4 | 0 | 0 | 42.4 | 20.4026 | 5.77222 |

5 | 1 | 0 | 27.2 | -1.8900 | -0.53472 |

6 | 1 | 0 | 34.0 | 4.9100 | 1.38911 |

7 | 1 | 0 | 27.4 | -1.6900 | -0.47813 |

8 | 1 | 0 | 28.5 | -0.5900 | -0.16693 |

9 | 0 | 1 | 25.9 | -4.0348 | -1.14152 |

10 | 0 | 1 | 34.5 | 4.5652 | 1.29156 |

11 | 0 | 1 | 25.1 | -4.8348 | -1.36785 |

12 | 0 | 1 | 34.2 | 4.2652 | 1.20668 |

13 | 1 | 1 | 35.0 | -1.7655 | -0.49950 |

14 | 1 | 1 | 33.9 | -2.8655 | -0.81070 |

15 | 1 | 1 | 38.3 | 1.5345 | 0.43413 |

16 | 1 | 1 | 39.9 | 3.1345 | 0.88679 |