MATH 251, Probability and Statistics I, Fall 2007, Wed. Nov. 7, Day 32AfterclassYES

Questions on Testing HW? Day 31
Takehome exam will be available Friday.

Quiz returned.  Many did  pretty well,  except too many people blanked on the binomial formula and especially the proportions.  (Almost) Everyone remembered to sum the variances for X-Y! Makeup! Open book: redo the problems you missed, finding the appropriate formulas from text or webpages, hand in Friday by 2:30, with your original quiz. If you didn't get your quiz back today, pick it up from outside my door. If you're lost on what I was asking for, contact me for hints.

Significance tests use an elaborate vocabulary, but the basic idea is simple: a result that would "rarely" happen if a claim were true--is good evidence that the claim is NOT true.   Notes Day 30
Look at the results from the shoeboxes:   Notice  that there is always a chance of getting a "somewhat unusual" result from a population where the null hypothesis is true.  And if the actual mean is not extremely different from the null, a result may not be detectably different from the null-hypothesis results.

A "Significance level" alpha is a probability level we decide on  in advance as being the "rarely" amount that will push us over into believing (well, sort of) that the H₀ claim  is not true.
Simple benchmarks: .10 (1 in 10), .05 (1 in 20), .01 (1 in 100).
When the P-value is less  than (or equal to) a particular significance level alpha (say .05), we say,
    "The results are significant at the alpha = .05 level," or "The results are significant (P < .05)" , or "Reject the null hypothesis at level alpha = .05"  Details Day 31
Try Applet: Statistical Significance  : Same setup as "P-value" applet only instead of giving P-value it shows whether the xbar is in the "significant" "rejection" region.

What if you don't have the Z-table but only have the t-table (Table D)?
What if you have a demanded level of significance, alpha?
    Table D: a limited list of probabilities  across the top row:
            = Right tail values for the bell curve distribution.
        The value in the bottom (z*) row under p is the corresponding standard normal value.
        "z* is the upper p critical value of the standard normal distribution."
  Do this: Find your z from the data. Make a sketch of the normal curve and mark z on it.  Mark the direction(s) of H_a.
    (If your z is in the direction of H_a , continue.  Otherwise the results are hopelessly not significant: you can quit.)
Find the two z*'s in Table D that bracket your z (ignore minus sign, using symmetry of Normal).  Find the corresponding p's.
    e.g. z =2.111
p      .02     .01
z*   2.054 \/ 2.326
       z = 2.111  So the P-value for your z is: between those 2 p's (one sided test)
                                           between double those 2 p's (two sided test)
(Some versions of the table add another top line, for two-sided tests: just Double the one-sided values)
    Test is significant at the bigger bracketing probability; not sig. at the smaller.
One sided: P-value is less than .02 and greater than .01
       Significant at the .02 level,not at the .01 level
Two sided: P-value is less than .04 and greater than .02
       Significant at the .04 level,not at the .02 level
If you have a specific demanded significance level, compare it with these levels.
            If  a test is significant at level b, then it is significant at every level bigger than b.
            If a test is Not significant at level d, then it is Not significant at every level smaller than d.
    "Significant at a":  probability of getting my results (again) by chance (if H₀ is true) is less than (or =) a.
Results  Significant at    Not significant at
p bigger   .10      .05      .01      .005     .001 smaller
                         /\
                        P-value
                        z-value (one-sided)
z* smaller 1.282   1.645  | 2.326    2.576    3.091 bigger
  You can compare z directly to z* for your desired alpha. The 2-sided is a bit tricky.
          (2-sided: Split the alpha in 2, then find the z*.  Don't halve or double z's--it doesn't work!)

CI's and Two-sided tests (pp. 413-14):
        Your 95% CI doesn't include  µ_o  <==> Reject H_o = µ_o  at the alpha = .05 level  (Seems like common sense.)

Start here Friday, Yes Sec. 6.3
>>Don't do inference on data that doesn't look like probability-model data (All that bias, design flaws stuff was for this!) and check the data for weirdness (Ch. 1)
>>(Not in text any more?) How small a P is "convincing evidence" against H₀?
     In practice...beyond the formal testing.
        How plausible is H₀?  H_a?  Strong evidence needed to reject "conventional wisdom."
        How expensive (mentally, economically) will abandoning H₀ be?
      (May need more than one set of data; replicate, recast, refine.)
>> In reality, no sharp border between "significance" and "not significant"

>>"Statistically Significant" doesn't always mean "Important." (e.g. medicine: "Clinically significant.") Big enough sample sizes will allow you to distinguish even small differences.
>> Lack of significance--doesn't prove H₀ true.  Best we can do: "data are consistent with (not inconsistent with) H₀ "

>>You cannot legitimately test a hypothesis on the same data that first suggested that hypothesis. Every data set will turn up with some unusual pattern if you examine it hard enough.
       (If you must explore and confirm with the same data set, one way is to (randomly) take half the data set, explore and generate hypotheses; then use the other half for confirmatory tests.  You can use P-value to describe unusualness, but be wary of making decisions with it if you didn't expect that particular unusualness.)

>>Multiple Tests: beware!
    If you do 100 tests and use the alpha = .05 significance level for each, then the structure of testing requires this:
    When all 100 null hypotheses H₀ are true, out of your 100, about 5 of the 100 (.05) will give "significant" results by chance alone (falsely indicating the alternative hypothesis is to be preferred.)
Try Applet: Statistical Significance  with some randomly generated "shoebox" data, true mean 20, alpha =.10, and see that you get "significant" results about 1 in 10 times.
    Moral: if you use the testing mechanism as a screening instrument for many questions, a proportion will give falsely significant results.  You can't accept the results from such multiple tests as good evidence, only as indicating questions requiring further, more specific study. The game gives you one shot, not a hundred shots.  (Dr. Pericak-Vance, Sci. Colloq. Sept. 28--screening for disease-causing genes)
- - - - - - - - - - -
Statistical inference in a nutshell:
Am I surprised (If H_ois true)? (Do I reject null?)
    How surprised? (give P-value)
  What would not surprise me?  (confidence interval--estimate the actual value)
(IPS:  Testing is over-used, Confidence interval estimation under-used)

Next: Brief look at issues of 6.4; then Ch. 7