| Basic Commands | Probability Puzzles | Hypothesis Test, Count Data | Hypothesis Test, Measured Data | Confidence Interval, Count Data | Confidence Interval, Measured Data | Association / Correlation | Regression | Other Examples |

George Bush Poll Results

What is the sampling error of a poll of 1500 people that yields 56% in favor of Bush? Use the "bootstrap" procedure. If we could, we would take additional samples from the universe of voters. That is too time-consuming and too expensive.

Instead, we let our 56%/44% sample serve as a proxy universe. We can imagine replicating each of the sample observations millions of times to create our "bootstrap" universe. Even this is time-consuming, so we achieve the same effect by putting each observation back in after we sample it -- sampling with replacement. We draw samples of size 1500 because this is the size sample whose behavior we wish to observe.

MAXSIZE b 1500

technical command to create enough room in vector a (which would otherwise be limited to 1000 elements).

URN 56#1 44#2 a

the universe suggested by the sample

REPEAT 1000
  SAMPLE 1500 a b

sample 1500 with replacement

  COUNT b =1 c

how many "bush" voters?

  DIVIDE c 1500 d

express as proportion

  SCORE d z

keep score

END
HISTOGRAM z
PERCENTILE z (2.5 97.5) k
PRINT k