Stat 112: Assumptions and prediction intervals

• Administrivia:
  • JMP Homework due next thrusday. You may work in pairs. Try to start on it this weekend so that you can bring in questions/problems on tuesday.

• Course goal: understanding how to use statistics correctly NOT how to compute statistics
• Today: checking our assumptions, "price of admissions" to statistics
• true line --> data --> fit line --> residuals
  • Which fits better orginal line or fit line?
  • Which do we care more about?
• Two questions:
  • Is it a line?
  • Are the residuals well behaved? (sub questions: no patterns, independent, same variance, normally distributed, no outliers)
• Ideal regression model handout
  • Histogram residuals, average squared deviation = variance = SD²
  • Normal quantile plot
• Look for:
  • outliers
  • influential points
  • curvature
  • hetroskadisticty
  • lack of independence (requires thought)
• If it passes tests, then you can use the model for predictions
  • Prediction intervals are +/- 2 * SD
  • The computer does something more accurate--show extrapolation penality picture
• Look at Cellular data (cell phones in the US from 1984 - 1995) Grew from 90k to 35M. (handout)
  • Not linear--use Tukey to suggest something to try
  • Try log(y)
  • discuss percentage changes
  • Make up scale so that 1% increase equals .01 increase
  • Does .10 then equal a 10%? Close enough
  • Does 1.00 then equal a 100% increase? No it 271% increase!
  • a .69 increase equals 100% growth.
  • Try it on the data. Not great. Bummer
  • Generate a forecast anyway for next year

Last modified: Thu Feb 3 12:17:48 2000