Model building

• Administrivia:
  • Courtney Bennett will say a few words about being subjects in her study.
  • Stay after class to collect some data if you want

Principled model building: Example

• Suppose Y = blood presure
• Xs in include: Age, time-of-day, at-work, exercising, sex, race, drugs-taken, billing-address....
• Maybe some interactions also
• Naive rule: Keep all X variables which have t's of 2 or more
• For example, which do you think is better to include:
  • X10 = Age with a t = 2.2
  • X25 = street-address with a t = 2.7
  • If you think of them as X10 and X20, the question is easy--X25 is better
  • If you think of htem as science--X10 is better, X25 is pure noise!

• Principled approach = use your brain
• unprincipled approach = be a statistician and don't know the science

Unprincipled approach = stepwise regression

• Story of GI's during Viet Nam war

  • American GI's were easy for the communist to "brain wash"
  • Americans tried a program of "America is wonderful."
  • led to No effect
  • Americans tried a program of "what is it really like in China."
  • Big effect--much harder to "brain wash" GIs
  • This is my reason for telling you a full lecture about stepwise regression

• run stepwise regression on Car89 data (using quadradic surface: control-click to highlight continuous variables--use "surface" to add them)
• Run forward, change p-value to .001, run mixed
• Lots of variables
• Clear variables--rerun. generates a "natural" answer

Last modified: Thu Apr 6 10:47:48 2000