I was hoping someone could help me fit an unusual data set which includes a large amount of 0's in my dependent variable in R.


My goal is to find which variables (tyre brand (qualitative), tyre variant (qualitative), price (quantitative), etc.) that are most important for our sales compared to visitors (called conversion). I am not looking into any interactions between my variables or anything else fancy. In the future I might be looking to use these variables to predict the conversion of a new tyre.


The frequency numbers of the conversion (in %) is:

Grouped Value   Frequency  
0.0 - 0.2       6592  
0.2 - 0.4        522  
0.4 - etc.       395  
0.6              288  
0.8              189  
1.0              133  
1.2               90  
1.4               66  
1.6               54  
1.8               27  
2.0               16  

Most of the 0.0 - 0.2 are 0's, which means no sales.


I'm an not very strong in statistics, but I thought I could try a Weibull or Exponential distribution. Not sure how to do this in R though. I have already tried a normal and quasipoisson but there are some definite trends in fitted vs the real values.

I am happy to include anything else one might need to help me out.

I would appreciate any help, also simple reminders saying remember to check for [something].


2 Answers 2


Try seeing if your data is zero inflated. For an ideal example, it may be counted data (in tens of successes), for example. In which case, zero inflated Poisson fits.

Weibull and other EVT is for exceedences, so you will need to know what you are exceeding, some number of inches past your maximum tolerance, for example.

  • 3
    $\begingroup$ Hurdle models might also be appropriate depending on how the data was collected and/or the process that generated the data. $\endgroup$ Mar 19, 2015 at 19:51
  • $\begingroup$ Thank you both for the suggestions. I think 1 of the two suggestions might to be spot on, but not sure I fully understand either. I have added a bit info in the my original question to make it more clear what data we are dealing with. I think we can rule out my suggestion of Weibull as I am not trying to exceed anything. Thank you both for being so fast to respond. Greatly appreciated. $\endgroup$
    – Alexander
    Mar 20, 2015 at 11:07

It surely looks like your data are zero inflated. If you convert from % to actual number of conversions you could use a zero inflated count model such as zero inflate Poisson or zero inflated negative binomial regression. This thread discusses how to fit a ZINB model in R (it has also been discussed in other threads). You might have to then include whatever the denominator of the % of conversions was to the model.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.