Hi,
Is there a way to use the descriptive statistics feature to calculate the 90% Confidence Interval for the Median.
Thanks
Hey there,
there are different approaches to calculate CI for the median.
Bootstrap or binomial quantiles are not implemented in PHX.
By the way some simpler way is possible, see here:
http://exploringdatablog.blogspot.ru/2012/04/david-olives-median-confidence-interval.html
I attached implementation in Phoenix. Note that you need PHX7 to run it. Winnonlin6.4 does not have inverse t distribution function.
BR,
Mittyright
PS: I suppose Simon won’t like it due to some poorly documented ‘tricks’, but it works!
CIMed.phxproj (679 KB)
Forgot to say that here I’m calculating 95%CI
To change the alpha please see the second DataWizard Object (alpha assignment)
Mittyright
Wow, in your project the transformations you do seem quite interesting. I can barely understand what you are doing, but it does not seem like an easy way to work with large simulated datasets. I run descriptive stats on the observed pharmacology for several timepoints throughout a treatment (could be up to a year) and would like to plot the median +/- 90% CI. Is there a reason why Phoenix does not offer an easier option, would it be added in the future? Are most PD modelers using mean or median?
By the way thank you for your project, it is very helpful!
Hi csheme,
I can barely understand what you are doing, but it does not seem like an easy way to work with large simulated datasets. I run descriptive stats on the observed pharmacology for several timepoints throughout a treatment (could be up to a year) and would like to plot the median +/- 90% CI.
Phoenix NLME is another story
If you have simulated datasets, I assume there are some replicates. So you can get the median for each timepoint in each replicate and then get descriptive statistics for the medians. I can show you with the data you have if you can share something.
Is there a reason why Phoenix does not offer an easier option, would it be added in the future?
Good question, next question! 
As I said above, for relicates in simulated datasets it is not so difficult,
For non-modelling purposes I think it is not so interesting. By the way even if you ask here, some big bosses can take your opinion into account 
Are most PD modelers using mean or median?
It depends on the goals. For VPC plots majority of packages are using 50% percentile (aka median)
BR,
Mittyright
Hi mittyright,
I love that you used the “Old Faithful” data set. I’m skeptic whether it makes sense to use this method (link). The median is nonparametric. I don’t see any reason to introduce a CI based on parametric statistics.
@csheme: I suggest to use Phoenix’ Descriptive Statistics and activate in the Options the box “Include percentiles”. For the “Old Faithful” waiting we get a median of 76 and 5%/95% of 47.65 and 89. Yes, this nonparametric interval is by nature asymmetric. Have a look at the data:
The distribution is bimodal… Hence, it is even questionable whether the median makes much sense. IMHO, this “90% CI” is close to nonsense.