Log Student-t comparing differences in scale

I am doing a log student t distribution. All nu values are above 2 so both mean and variance exist.
I want to compare the log sigmas between groups, so taking the difference in the posteriors of log sigma_group1 - log sigma_group2.

If this was a gaussian I would take the CV to control for different means or equally for lognormal model I would first convert the sigma to the arithmetic CV sqrt(exp(log_var) - 1), and for gamma use 1/shape .

Although I don’t believe I can apply this to a student T. So would the perferable thing to do is to take the delta of the log sd? But just not interpret it as the arithmetic CV? Cause I want to make sure that the difference I observe in the residuals isnt due to the difference in the means.

brm.formula <- bf(
  log(y) ~ 0 + group +
    (1 | gr(ID, by = group)),

  sigma ~ 0 + group
)

fit <- brm(
  formula = brm.formula,
  data = dat,
  family = student()
)


Thanks,
Hunter

Hi, @Hunter24. Sorry nobody has responded in a week. I’m afraid I don’t really know brms, but I thought I might be able to help clarify the question.

There’s no problem comparing log difference as you’re doing. It’s just comparing the group scales.

But I don’t know what “CV” means in this context or why sqrt(exp(log_var) - 1) comes up.

I’m not sure what you mean by the residuals resulting from means. By residuals, do you mean what contributes to the underling sigma values? Why would it matter how the residual arose?

Hi Hunter, a fee point/questions for clarification.

The CV (coefficient of variation) of a log-t-distribution with a finite number of degrees of freedom (nu) is undefined as the mean and variance of the distribution are undefined due to its heavy right tail. If I remember correctly.

On the logarithmic scale perhaps, but not after exponentiation back to the natural scale.

When you say “log sigma”, do you mean the logarithm of the standard deviation or the standard deviation on the logarithmic scale?

I mean the scale parameter (\(\sigma\)) estimated on the logarithmic scale, not the logarithm of the standard deviation.

Because I know that this can be approximate to the CV (Julious & Debarnot, 2000). So assuming all nu values are above 2. Shouldn’t this be a good approximation of the CV?

Julious, S. A., & Debarnot, C. A. (2000). Why are pharmacokinetic data summarized by arithmetic means?. Journal of biopharmaceutical statistics, 10(1), 55-71.

The log-t-distribution does not have a defined coefficient of variation, so an approximation doesn’t exist either. You may perhaps be referring to Julious’ & Debarnot’s work on log-normal distributions (which you cite), where for small standard deviations (approx <0.3 on the logarithmic scale), the CV of a log-normal distribution (on the natural scale) is approximately similar to its SD (on the logarithmic scale). This does not extend to the log-t-distribution because it has no defined mean and variance on the natural scale.

Ok so the best is to just report the group scales, but for the reason you explained, its incorrect to call it an approximation of the CV.

@Bob_Carpenter Thanks Bob for the input it cleared things up.

Hunter

sqrt(exp(log_var) - 1) is the equation for a lognormal model to convert the standard deviation on the logarithmic scale to the CV on the arithmetic scale.