Skip to content

Dissimilarity measure for rank data

3 messages · Tania Bird, Sarah Goslee, Hanna Tuomisto

Original
Tania Bird
# 
I have species data that I would like to use for ordination. With
regular abundance data I would apply a Hellinger Transformation and
then use the Bray-Curtis distance.

Since the data are ranked (0 to 5) I will not transform it. But what
dissimilarity measure should I use instead of Bray-Curtis?

Thanks


Tania Bird MSc
"There is a sufficiency in the world for man's need but not for man's
greed" ~ Mahatma Gandhi
Sarah Goslee
# 
Hi Tania,

That's not really an R question, and there's no one perfect answer.
Googling "distance metric for ordinal data" turns up some discussions of
the pros and cons of the various options.

You need to choose the one best able to address your hypothesis. You might
get better ideas on a statistics forum, rather than an R-specific list.

Sarah
On Thu, Jun 8, 2017 at 6:54 AM Tania Bird <taniabird at gmail.com> wrote:

            

      
      
      
    

        

      
      
    

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Hanna Tuomisto
    

    

      
      #
    

  


  

      
Hi Tania,

Even though there is no perfect answer: don't use the combination Hellinger transformation + Bray-Curtis distance. The appropriate combination is Hellinger transformation + Euclidean distance, which gives you Hellinger distances (this is an asymmetric dissimilarity measure, which does not suffer from the double-zero problem that plagues the Euclidean distance). Bray-Curtis distance is in itself asymmetric, and it can be used directly both on regular abundance data and on ranked data.

Cheers,
Hanna Tuomisto

        
On 8 Jun 2017, at 14:45, Sarah Goslee wrote: