TROUBLESOME CONCEPTS IN STATISTICS: r2 AND POWER N. Scott Urquhart Director, STARMAP Department of Statistics Colorado State University Fort Collins, CO 80523-1877
STARMAP FUNDING Space-Time Aquatic Resources Modeling and Analysis Program The work reported here today was developed under the STAR Research Assistance Agreement CR-829095 awarded by the U.S. Environmental Protection Agency (EPA) to Colorado State University. This presentation has not been formally reviewed by EPA. The views expressed here are solely those of the presenter and STARMAP, the Program he represents. EPA does not endorse any products or commercial services mentioned in these presentation. This research is funded by U.S.EPA – Science To Achieve Results (STAR) Program Cooperative Agreement # CR - 829095
INTENT FOR TODAY To discuss two topics which have given some of you a bit of confusion r2 in regression Power in the context of tests of hypotheses Thanks for Ann Brock and Harriett Bassett for suggesting these topics Approach: Visually illustrate the idea, Then talk about the concepts illustrated The sequences of graphs are available on the internet right now (address is at the end of this handout) Questions are welcome
r2 IN REGRESSION r2 provides a summary of the strength of a (linear) regression which reflects: The relative size of the residual variability, The slope of the fitted line, and How good the observed values of the predictor variable are for prediction Mainly the range of the Xs Let’s see these features in action, then Look at the formulas
WHAT MAKES r2 TICK? varying one thing, leaving the remaining things fixed r2 increases as residual variation decreases r2 increases as the slope increases r2 increases the range of x increases
WHAT IS r2? r2 provides A measure of the fit of a line to a set of data which incorporates The amount of residual variation, The strength of the line (slope), and How good the set of values of “x” are for estimating the line Some areas of endeavor tend to overuse it!
HOW DOES r2 TELL US ABOUT VARIATION? The following graph illustrates this: The data scatter has r2 = 0.5 (approximately) The red points have the same values, but all concentrated at X = 5. {Strictly speaking the above formulas apply only in the case of bivariate regression.} {Estimation formulas involve factors of n-1 and n-2.}
r2
FORMULAS FOR r2 But these have little intuitive appeal ! We’ll decompose observations into parts: Mean Regression Residual
DECOMPOSING REGRESSION This is really n equations Square each of these equations and add them up across i. The three cross product terms will each add to zero. (Try it!)
DECOMPOSING REGRESSION (continued)
POWER OF A TEST OF HYPOTHESIS Power = Prob(“Being right”) = Prob(Rejecting false hypothesis) Power depends on two main things The difference in the hypothesized and true situations, and The strength of the information for making the test Sample size is very important factor In regression it depends on the same factors as the ones which increase r2. Again, see it, then talk about it Power increases as D = m1 - m2 increases
POWER VARIES WITH DIFFERENCE (D = m1 - m2) and SAMPLE SIZE (n)
ON TESTS OF HYPOTHESES ( ON THE WAY TO POWER) TRUE SITUATION HYPOTHESIS FALSE HYPOTHESIS TRUE ACTION FAIL TO REJECT THE NULL HYPOTHESIS CORRECT ACTION TYPE II ERROR REJECT THE NULL HYPOTHESIS TYPE I ERROR CORRECT ACTION Tests of hypotheses are designed to control a = Prob (Type I Error) While getting Power = 1- Prob (Type II Error) as large as possible
ON TESTS OF HYPOTHESES (AN ASIDE) Which is worse, a type I error, or a type II error? It depends tremendously on perspective Consider the criminal justice system Truth: Accused is innocent (HO) or guilty (HA) Action: Accused is acquitted or convicted Type I error = Convict an innocent person Type II error = Acquit a guilty person Which is worse? Consider the difference in view of the Accused Society – especially if accused is terrorist
COMPUTING THE CRITICAL REGION Consider a simple case X ~ N( m, 1) HO: m = 4 versus HA: m ¹ 4 Critical Region (CR) is X £ l and X ³ u , so 0.025 = P(X £ l ) = P((X-4)/1 £ ( l - 4)/1) = P(Z £ -1.96) l = 2.04, similarly, u = 5.96
COMPUTING POWER Consider a simple case X ~ N( m, 1) HO: m = 4 versus HA: m ¹ 4 Power (at m = 5) = ? = Prob(XA in CR| m = 5) XA ~ N( 5, 1) Prob(XA £ 2.04) + Prob(XA ³ 5.96) = Prob(Z £ -2.96) + Prob(Z ³ 0.96) = 0.0015 + 0.1685 = 0.1700
POWER VARIES WITH DIFFERENCE (D = m1 - m2) and SAMPLE SIZE (n)
COMPUTING POWER USING A MEAN BASED ON n = 2 OBSERVATIONS Consider a simple case: When the mean of two observations follows: HO: m = 4 versus HA: m ¹ 4 Power (at m = 5) = ? Critical Region (CR) is £ l and ³ u , so 0.025 = P( £ l ) = P(( -4)/0.707 £ ( l - 4)/0.707) = P(Z £ -1.96) So l = 4 – (1.96)(0.707) = 2.61, similarly, u = 5.39
COMPUTING POWER USING A MEAN BASED ON n = 2 OBSERVATIONS (continued)
POWER VARIES WITH DIFFERENCE (D = m1 - m2) and SAMPLE SIZE (n)
COMPUTING POWER USING A MEAN BASED ON n = 4 OBSERVATIONS (continued) (This page is not in the handout – so it all would fit on one page)
POWER VARIES WITH DIFFERENCE (D = m1 - m2) and SAMPLE SIZE (n)
POWER VARIES WITH DIFFERENCE (D = m1 - m2) and SAMPLE SIZE (n)
DIRECTIONAL NOTE As the alternative has been two-sided throughout this presentation, the power curves are symmetric about the vertical axis. By examining only the positive side, we can see the curves twice as large.
YOU HAVE ACCESS TO THESE PRESENTATIONS You can find each of the slide shows shown here today at: http://www.stat.colostate.edu/starmap/learning.html Each show begins with authorship & funding slides You are welcome to use them, and adapt them But, please always acknowledge source and funding You are free to reorder the graphs if it makes more sense for r2 to decrease than increase. Urquhart is available to talk to AP Stat classes about statistics as a profession. See content on the web site above.