Class 4. Stat701 Fall 1997

Residuals, leverage and outliers.

Today's material

*
Leverage
*
Residuals
*
Influence
*
Jackknife
*
Masking


Residuals

Residuals are vital to regression because they establish the credibility of the analysis. Never accept a regression analysis without having checked the residual plots.

Residuals come in many flavors:

*
Vanilla residual: tex2html_wrap_inline83 .
*
Standardized residual:

displaymath85

*
Studentized residual: tex2html_wrap_inline87 .

The plain residual tex2html_wrap_inline89 and its plot is useful for checking how well the regression line fits the data, and in particular if there is any systematic lack of fit, for example curvature.

But, what value should be considered as a big residual?

*
Problem: tex2html_wrap_inline89 retains the scale of the response variable (Y).
*
Answer: standardize by an estimate of the variance of the residual.
*
Know, tex2html_wrap_inline93 estimated by tex2html_wrap_inline95 .
*
But, tex2html_wrap_inline83 , which is more than just tex2html_wrap_inline99 .
*
Turns out, tex2html_wrap_inline101 .
*
Use standardized residual, tex2html_wrap_inline103 .
*
The quantity, tex2html_wrap_inline105 is fundamental to regression.
*
A heuristic explanation of tex2html_wrap_inline105 (visually we are dragging a single point upward and measuring how the regression line follows):
*
Think about tex2html_wrap_inline99 the observed value, and tex2html_wrap_inline111 the estimated value (ie the point on the regression line).
*
For a fixed tex2html_wrap_inline113 perturb tex2html_wrap_inline99 a little bit, how much do you expect tex2html_wrap_inline111 to move?
*
If tex2html_wrap_inline111 moves as much as tex2html_wrap_inline99 then clearly tex2html_wrap_inline99 has the potential to drive the regression - so tex2html_wrap_inline99 is leveraged.
*
If tex2html_wrap_inline111 hardly moves at all then clearly tex2html_wrap_inline99 has no chance of driving the regression.
*
In other words tex2html_wrap_inline105 is the measure of ``leverage''.
*
More precisely

displaymath133

and it depends only on the x-values.

*
Understanding leverage is essential in regression because leverage exposes the potential role of individual data points. Do you want your decision to be based on a single observation?

Standardized residuals

Standardized residuals allow the residuals to be compared on the ``standard scale''. Plus/Minus 2 indicates something unusual, Plus/Minus 3 indicates something really out of the ordinary and Plus/Minus 4 is something from outer space (it just shouldn't happen).


Subtle point

Problem. The standardized residuals still start off with tex2html_wrap_inline135 and the problem is that if tex2html_wrap_inline99 is really leveraged then it will drag the regression line toward it, influencing the estimate of the residual itself.

Solution. Fit the regression line excluding tex2html_wrap_inline99 and base the residual on tex2html_wrap_inline141 , where tex2html_wrap_inline143 denotes the fit based on a regression line estimated excluding tex2html_wrap_inline99 .

Notes.

*
This is the leave one out idea and is the basis for much computationally intensive modern statistics. The leave one out idea is often called ``jackknifing''.
*
This ``leave one out'' residual can be used as a basis for judging the predictive ability of a model. Clearly the lower the residual the better, and the sum of the squares of the jackknifed residuals is called the PRESS statistics, or Predicted Sum of Squares.
*
The studentized residual, tex2html_wrap_inline87 , is just a standardized jackknifed residual. This is an extremely good way of judging how much of an outlier in the y-direction a point is.
*
from now on we will use the studentized residual plot to judge outliers in the y-direction.
*
A new plot. Leverage vs. studentized residual. Points that drive the regression have big leverage and extreme studentized residuals.
*
The delete one idea works pretty well, except when there is a second data point lying close by. In this case the second point can drive the regression line, masking the effect of the first point. This leads to the idea of ``delete two'' etc.



Richard Waterman
Mon Sep 15 22:27:08 EDT 1997