Enhancement to permtest

[cmbarton]

Enrico and Andy

A colleague and I have run into an interesting issue.

We have a set of dates from several kinds of sites. We'd like to use permtest to compare the SPD of the dates from site type 1 with a model generated from all the other kinds of sites. Permtest compares the SPD of site type 1 with a model generated from all the dates, including the dates from site type 1.

Is there any way to compare site type 1 with a model from just the other dates?

[ahb108]

Hi Michael
It has been awhile since I have looked at this code and I confess that I mostly tended to use this test in a two-sample case: either (1) spd of dates with attribute A vs one with attribute B, or (2) spd of dates with attribute A vs all others simplified as a hypothetical B. But am I right in thinking that you wish to retain the >2 attributes throughout, not shuffle site type 1, but shuffle site types 2, 3, ...n?

[cmbarton]

Hi Michael It has been awhile since I have looked at this code and I confess that I mostly tended to use this test in a two-sample case: either (1) spd of dates with attribute A vs one with attribute B, or (2) spd of dates with attribute A vs all others simplified as a hypothetical B. But am I right in thinking that you wish to retain the >2 attributes throughout, not shuffle site type 1, but shuffle site types 2, 3, ...n?

Hi Andy,
Actually, it seems you can't do what you describe. PermTest will only compare A vs A+B. That is, the model is generated by random pulls from both A and B. We'd like to do what you describe and compare A vs B

[ercrema]

Hi Mike,

I don't a permutation based approach can do that as it's fundamentally just testing for random labelling. If you want to compare A Vs B you would need a proper inferential model. For example, if you consider an exponential growth, you could parametrise this like a regression to see the 'effect' of being B. This can be done in nimbleCarbon but you'll need to write the model in nimble. Also it works only with parameteric models. If you want something non parametric with similar feel (e.g.local differences) I think you would need to fit two non parametric models separately and compare the posterior ranges year by year graphically.