Why does JMP not support Simulate or Bootstrap when using a BY variable?

[#1]

Hi All,

  I came across something interesting, that I wasn't really expecting with JMP Pro. Normally, in several platforms, JMP Pro offers the option to run Simulate or Bootstrap to get better estimates and statistics on metrics of interest. However, this appears to not be available when using a BY variable in a supported Platform.

  For example, with the Big Class.jmp file, if you do a bivariate fit of :weight by :height and fit a line to it, you get the option to bootstrap or simulate the fit estimates. However, if you do the same bivariate, but include :sex as a BY variable, this option goes away. Is there a reason for this?

  Just curious, as I find it odd that this should limit the ability to use either of those techniques. There is a workaround where you can do the normal bivariate fit and then use a local data filter and and select either F or M from :sex.

  The first image shows the first example of the regular bivariate fit. The second, when using a BY variable (and that the simulate and bootstrap options are gone when right clicking a statistic). The third shows the regular bivariate fit with a local data filter and the ability to still perform the simulate or bootstrap.

  If anyone knows the answer, I'm curious as to why the option goes away with a BY variable.

Thanks!,

DS

[#2]

If you try to "force" bootstrap by scripting you will get message to log

Bootstrap not supported for By groups

and same can be found from Basic Analysis > Bootstrapping > Bootstrapping Window Options (similar text can be found for Simulate Basic Analysis > Simulate > Launch the Simulate Feature ). No specific reason though.

[#3]

Hi @jthi ,

  I had skimmed over that note in the help for both methods, and didn't really make that connection until I tried to do it myself and found out. But, yes, there's no explanation that I can find why it's not available. Just curious why not.

Thanks!,

DS

[#4]

I'm not sure why it is not supported as a BY variable except to say that a BY variable is not needed. Using Big Class like you did, build the model with the effects: Height, Sex, Height*Sex. This is a single model that is equivalent to the two "subset" models. This then has the ability to do the Bootstrap and Simulate also.

[#5]

Hi @Dan_Obermiller ,

  Yes, in this particular case, that is possible to run that additional cross term. I have another data table from work that is setup differently and there I am using a BY variable to do a certain kind of fit, but for whatever reason those methods are not allowed when using the BY variable. It doesn't seem like that should be so problematic, but for whatever reason, it's not allowed. It's not a big deal to work around that issue, but I just find it interested and I'm curious why.

Thanks!,

DS