search

carlganz / svrepmisc

6 stars 2 forks source link

svymultinom and negative df.residual #3

Closed larmarange closed 6 years ago

larmarange commented 6 years ago

Dear @carlganz,

When you have a svrepdesign with not enough replicates, a call to svymultinom or similar function will result with a negative number of df.residual resulting with an impossibility to compute confidance intervals using t distribution.

Would consider to add a specific warning in wR to inform user?

carlganz commented 6 years ago

Hello,

I am pretty surprised (in a good way) to discover that another human is using this package. I don't believe I spent sufficient time ensuring that this package generates statistically accurate results, so I urge you to please use this package with caution.

I am not actively supporting this package, if you would like to make a pull request to add a warning I will gladly merge it.

Kind Regards

larmarange commented 6 years ago

Dear @carlganz ,

thank you very much for your feedback. I'm facing the need for multinomial and ordinal (clm) models with complex survey data and to be honest there are very few explanations online.

The only detailed other example I found online was https://rpubs.com/corey_sparks/249896 but it seems that there is a mistake as the author is using the mean attribute of the var attribute instead of the result of withReplicates.

By the way, this author is using normal distribution instead of t distribution to compute confidence intervals. My feeling is that t distribution is probably better.

I will prepare a pull request to add this Warning message.

In addition, I will prepare two additional PR if it suits you: (i) a svyclm function (using wR it's really easy) and (ii) a tidy method for svrepstatmisc class (I'm using broom package quite a lot).

larmarange commented 6 years ago

integrated in https://github.com/carlganz/svrepmisc/pull/7/