I am establishing a prediction model in R (rms, Hmisc) with around 15 predictor variables. One of those has a large proportion of missingness (>80%) and this variable seems to have high prognostic value. After assuming missingness at random, this variable is imputed n=100 using MICE. In the validation and calibration procedure I also calculate and plot the impact of each variable plot(anova(pred_model)) to obtain the chi2 value for each variable (or rather chi2 minus the predictor degrees of freedom). When calculating the chi2 value for each variable taking all imputations into account (fit.mult.impute), the imputed variable (>80% missing) has a high chi2-value but when calculating chi2 value for each of the imputed datasets, the the same missing variable has a very high chi2 (i.e. much higher the calculations that takes all imputations into account). For both analys (taking alla imputations into account and for each imputed dataset), the nomograms looks almost similar.
Question: Is the ”lower” chi2 value we observe (when taking all imputations into account) due to an uncertainty in this variable given the large proportion of missingness?
Thanks in advance!
We require site users to put their real first and last names in the profiles. Please also edit your post to use a proper title, select major and minor topic categories, and add appropriate tags. On more thought this would be better to put as a new thread at the bottom of RMS Discussions
Thank you for reply!
I will move my question to the RMS Discussion. However, I can see that I have the possibility to start a new thread there?
Right it’s just one long thread but still the best place for this. I’ll remove these last 2 posts but you still need your last name specified in your profile.