draw_normal_icc() doesn't allow for full range of ICC [0,1]

[clarabicalho]

Designers in the DesignLibrary relying on draw_normal_icc () should allow for to return values when ICC = 0 or 1 without requiring the designer to condition the distribution function on the icc value, oder? Also odd that it doesn't allow for ICC <= 0.001 or ICC >= 0.999. Could the default behavior be a warning instead?

clusters = rep(1:5, 10)
draw_normal_icc(clusters = clusters, ICC = 0.001)

Error in draw_normal_icc(clusters = clusters, ICC = 0.001) : An ICC of 0 with a finite within-cluster variance implies zero between-cluster variance. You can generate data with zero ICC using R's standard rnorm command to generate normal data independent of the cluster variable.

draw_normal_icc(clusters = clusters, ICC = 1)

Error in draw_normal_icc(clusters = clusters, ICC = 0.999) : An ICC of 1 with a finite within-cluster variance requires division by zero to infer between-cluster variance. Try a lower ICC or specify between- and within-cluster variance (sd_between and sd) to infer ICC.

[macartan]

for the first case just implementing the normal in this case would make sense (eg so that users can compare over a range tha includes 0 without modifying code) for the second I think the error should be " An ICC of 1 is not possible with positive within-cluster variance. " It ought to be possible with within cluster variance of 0?

…

On Mon, Aug 20, 2018 at 4:30 PM Clara Bicalho ***@***.***> wrote: Designers in the DesignLibrary relying on draw_normal_icc () should allow for to return values when ICC = 0 or 1 without requiring the designer to condition the distribution function on the icc value, oder? Also odd that it doesn't allow for ICC <= 0.001 or ICC >= 0.999. Could the default behavior be a warning instead? clusters = rep(1:5, 10) draw_normal_icc(clusters = clusters, ICC = 0.001) Error in draw_normal_icc(clusters = clusters, ICC = 0.001) : An ICC of 0 with a finite within-cluster variance implies zero between-cluster variance. You can generate data with zero ICC using R's standard rnorm command to generate normal data independent of the cluster variable. draw_normal_icc(clusters = clusters, ICC = 1) Error in draw_normal_icc(clusters = clusters, ICC = 0.999) : An ICC of 1 with a finite within-cluster variance requires division by zero to infer between-cluster variance. Try a lower ICC or specify between- and within-cluster variance (sd_between and sd) to infer ICC. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#149>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AMJO_UkDBOJvJnB2jAmvAR74ipnQIKslks5uSsgdgaJpZM4WEHdY> .

[graemeblair]

Hi all -- @aaronrudkin weighed in on this:

The way we set up draw_normal_icc was designed to have the user specify one of sd_between or sd_within and ICC. Mathematically, sd_between^2 = (ICC * sd_within^2) / (1 - ICC). So when ICC = 1, this is a divide by zero. Where ICC is very near 1, sd_between goes to near infinity. And similar problems happen with ICC=0. I believe we preferred errors to producing near-nonsense data or allowing R to NaN and causing less interpretable errors later.

I think we should do special casing in the designer, but that this is the right behavior in fabricatr. Feel free to reopen if you disagree.

[macartan]

Understand why this is happening but not sure this is the best handling of these cases.

• for the first case just implementing the normal in this case would make sense --- why not just do that instead of giving an error
• for the second I think the error should be " An ICC of 1 is not possible with positive within-cluster variance. " not "An ICC of 1 with a finite within-cluster variance requires division by zero to infer between-cluster variance. "
• in the second case also it ought to be possible with within cluster variance of 0 by assigning same values to all units in a cluster

[graemeblair]

ok sounds good. think this should be done with the #133 change and also draw_binary_icc should have the same behaviors so a bit of work. given time constraints, will leave out for this version.