Inference App

[Helge Gabert]

Please take a look at the inference app, method: regression, type: mean response vs. type: prediction interval.

For a given data set (Xlist, Ylist), the output, as calculated for a X-value, and confidence level, is identical for both types (mean response and prediction interval). The only apparent difference is the labeling on the output table for y hat.

[mcjtom]

I think there may be a display bug in all(?) hypothesis testing, when Alt. hypothesis is set to mju <> mju null (double tail). Try this:

Use the default example data for Z-test: 1 mju (reset Inference App. if needed), but change the Alt Hyp. to not equal. Plot screen looks like this:

   

Now, in Num screen, change Sample Mean to 0.539... (or anything greater than null Hyp mean). The screen should look about the same, but instead looks like this:

   

The p-value seems to be reported correctly - it's just the shaded area that is not. It seems to be the case with all test types with double tail Alt Hyp.

When I'm at it:

- the notation (red arrow in previous screen shot): it should probably read mju null not Sample Mean. Similarly in proportions there is: p, p-hat, and Pi null. Wouldn't two of them suffice?

- one improvement suggestion: in Plot screen the X-axis has dual scale (Z-score and actual mean values). If alpha button is checked, Critical Z range is reported on top of the screen but the actual critical values for the mean are not. They are available in the Results screen as Lower and Upper. I think being able to see them in Plot as well would be v. helpful in interpreting the results.

[mcjtom]

Would anyone know how the confidence interval/critical values on proportions (i.e. test type: 1 p and p1-p2) is calculated (i.e. what's the distribution and formulas being used)?

It looks to me that the sample proportion's bionomial distribution is simply approximated as normal, regardless of the sample size or proportion value being close to 0 or 1. If that's the case, it may at least deserve some warning on screen or in Help (e.g. http://en.wikipedia.org/wiki/Binomial_pr...e_interval).

But is it true?

[mcjtom]

I just looked at this again with the new firmware (7820), and the histograms have not changed. Am I right about what I described above to begin with?

Cheers!

[Wes Loewer]

(08-12-2014 09:03 AM)mcjtom Wrote:  Would anyone know how the confidence interval/critical values on proportions (i.e. test type: 1 p and p1-p2) is calculated (i.e. what's the distribution and formulas being used)?

It looks to me that the sample proportion's bionomial distribution is simply approximated as normal, regardless of the sample size or proportion value being close to 0 or 1. If that's the case, it may at least deserve some warning on screen or in Help (e.g. http://en.wikipedia.org/wiki/Binomial_pr...e_interval).

But is it true?

Yes, it is using the normal distribution.