Thanks to Kit Baum there is a new package named -ros- at the SSC.
The command ros is for estimating upper reference bounds for a dataset with possibly non-detectable/censored values and possibly contaminated in the upper end.
The upper reference bounds are from the mean and standard deviation from regressing the observed values on the empirical (normal) z-values. The mean and the standard deviation estimated from ros are the intercept and slope of the regression.
Box-Cox optimization is done by choosing a transformation with a high R square from the regressions.
A dataset can have non-detectable/censored values to the left and can also be contaminated to right by non-healthy participants, so it might be only a small linear part of the qnorm scatter plot that is acceptable for estimation.
This means that some of the estimated upper bounds are extrapolations from the acceptable part. The estimated, not extrapolated, upper bounds, are similar to the empirical percentiles.
The Ideas behind the command ros were demonstrated at the 2022 Northern European Stata Conference on 12 October in Oslo. The ros is inspired by the work in the references.
However, more tools for model control, the rsquare for the goodness of fit, the rsquare theta plot, and the qnorm scatter plot are available.
The only modification from the presentation is that the Box-Cox transformation here is defined as: bct(x; theta) = x^theta / abs(theta) if theta != 0 and bct(x; 0) = log(x).
This modification means that negative thetas can be included in the Box-Cox transformations.
Also in the package is the command boxcoxsim, which simulates data from a BoxBox-Cox normal distributed data with a possible percentage degree of left truncation and in a mixture of possible percentage degrees of extreme values.
It is a simulation tool for ros.
For now, it works for Stata 15 and up.
An update is on the way such that it works from Stata 12.
Enjoy
The command ros is for estimating upper reference bounds for a dataset with possibly non-detectable/censored values and possibly contaminated in the upper end.
The upper reference bounds are from the mean and standard deviation from regressing the observed values on the empirical (normal) z-values. The mean and the standard deviation estimated from ros are the intercept and slope of the regression.
Box-Cox optimization is done by choosing a transformation with a high R square from the regressions.
A dataset can have non-detectable/censored values to the left and can also be contaminated to right by non-healthy participants, so it might be only a small linear part of the qnorm scatter plot that is acceptable for estimation.
This means that some of the estimated upper bounds are extrapolations from the acceptable part. The estimated, not extrapolated, upper bounds, are similar to the empirical percentiles.
The Ideas behind the command ros were demonstrated at the 2022 Northern European Stata Conference on 12 October in Oslo. The ros is inspired by the work in the references.
However, more tools for model control, the rsquare for the goodness of fit, the rsquare theta plot, and the qnorm scatter plot are available.
The only modification from the presentation is that the Box-Cox transformation here is defined as: bct(x; theta) = x^theta / abs(theta) if theta != 0 and bct(x; 0) = log(x).
This modification means that negative thetas can be included in the Box-Cox transformations.
Also in the package is the command boxcoxsim, which simulates data from a BoxBox-Cox normal distributed data with a possible percentage degree of left truncation and in a mixture of possible percentage degrees of extreme values.
It is a simulation tool for ros.
For now, it works for Stata 15 and up.
An update is on the way such that it works from Stata 12.
Enjoy
Comment