Closed adam2392 closed 8 years ago
Looking at your plot I think that you found a great reason to perhaps not use that test statistic! Try using wilcoxon, or some other statistic? This might be a great opportunity to simply test a few and have that be a thing you learned from this inference - what tests are robust to your type of data as samples tend to infinity. :)
Greg Kiar gkiar07@gmail.com
On Wed, Feb 24, 2016 at 2:28 PM, Adam Li notifications@github.com wrote:
Assigned #6 https://github.com/Upward-Spiral-Science/the-vat/issues/6 to @gkiar https://github.com/gkiar.
— Reply to this email directly or view it on GitHub https://github.com/Upward-Spiral-Science/the-vat/issues/6#event-564201956 .
i'm not quite sure what you mean, but under the null, the power cannot be above alpha, no matter what (assuming the null is that F0=F1, if your null is that they are different one way, and the alternate is that they are differently different, that is different :)
if that's not clear to you, maybe re-read hypothesis testing chapter? happy to chat on thursday.
On Wed, Feb 24, 2016 at 2:44 PM, Greg Kiar notifications@github.com wrote:
Looking at your plot I think that you found a great reason to perhaps not use that test statistic! Try using wilcoxon, or some other statistic? This might be a great opportunity to simply test a few and have that be a thing you learned from this inference - what tests are robust to your type of data as samples tend to infinity. :)
Greg Kiar gkiar07@gmail.com
On Wed, Feb 24, 2016 at 2:28 PM, Adam Li notifications@github.com wrote:
Assigned #6 https://github.com/Upward-Spiral-Science/the-vat/issues/6 to @gkiar https://github.com/gkiar.
— Reply to this email directly or view it on GitHub < https://github.com/Upward-Spiral-Science/the-vat/issues/6#event-564201956> .
— Reply to this email directly or view it on GitHub https://github.com/Upward-Spiral-Science/the-vat/issues/6#issuecomment-188424814 .
the glass is all full: half water, half air. neurodata.io
My null is that D is uniformly distributed and alternative is that D is say normally distributed. If what you said is true, might just be a bug in my implementation?
What hypothesis testing chapter are you referring to? Art of data science pdf, or another book?
i'm not sure what you mean. ks tests test the difference between 2 distributions.
if we are testing whether two groups are different, the null must have a distribution for each, and the alternative has a distribution for each.
sounds like you have only specified an alternative...
On Wed, Feb 24, 2016 at 5:15 PM, Adam Li notifications@github.com wrote:
My null is that D is uniformly distributed and alternative is that D is say normally distributed. If what you said is true, might just be a bug in my implementation?
What hypothesis testing chapter are you referring to? Art of data science pdf, or another book?
— Reply to this email directly or view it on GitHub https://github.com/Upward-Spiral-Science/the-vat/issues/6#issuecomment-188477619 .
the glass is all full: half water, half air. neurodata.io
https://en.wikipedia.org/wiki/Kolmogorov%E2%80%93Smirnov_test
I referred to that to get a basic understanding of the test. For a one sample test it tests the null hypothesis that my data comes from a reference distribution (in this case uniform)? Then the alternative is that it doesn't?
swing by tomorrow 1-3, i'l clarify
On Wed, Feb 24, 2016 at 7:32 PM, Adam Li notifications@github.com wrote:
https://en.wikipedia.org/wiki/Kolmogorov%E2%80%93Smirnov_test
I referred to that to get a basic understanding of the test. For a one sample test it tests the null hypothesis that my data comes from a reference distribution (in this case uniform)? Then the alternative is that it doesn't?
— Reply to this email directly or view it on GitHub https://github.com/Upward-Spiral-Science/the-vat/issues/6#issuecomment-188534149 .
the glass is all full: half water, half air. neurodata.io
Sure I'll try to come by your office after lunch.
In conclusion, learned that Kolmogorov and Wilcoxon return statistic and p-value in different order.
Better print statements needed for next time :)
I fixed my simulation for the Kolmogorov in the inferential_simulation notebook by AL. However, it seems that the null also converges to 1 also.
I was just wondering if I'm doing the simulation correctly, or if this is perhaps a weakness in the test statistic? I am testing if a distribution comes from a uniform of [0,1] vs. say... normal (0,1).
Just wanted a quick 3rd person check to see if I should be looking at another test statistic.
Thanks!