As part of a broader study I am analysing 30 websites that fall into 3 categories:

  • Consumer (10 sites)
  • Commercial (10 sites)
  • Health (10 sites)

The approach I used was a 'tick and flick' spreadsheet with 24 dichotomous variables that represent features of the website that are either absent or not (i.e. they receive a tick if they exhibit that particular feature).

Here is an example of the data.

The numbers represent how many websites from each category contain each particular feature (variable).

I want to know which kind of statistical test would be used to find if there are any systematic patterns about which 'Category' of website tends to correlate with particular variables. For example, which websites tend to share power with users to edit/contribute web content (measured by variables 2,3,4,5,6,7,8,13,16,19,23,24)?

I would rather use a more robust/rigorous statistical approach than simply counting up totals, or 'eye-balling' patterns in the data.

Thank you in advance.

  • $\begingroup$ the link is not working. also have you considered machine learning procedures ? Another approach is using entropy although i am not sure how would you do testing of traditional hypothesis testing nature with these. $\endgroup$ – htrahdis Oct 29 '13 at 14:29
  • $\begingroup$ Hi @htrahdis, the link seems to be working for me? Sorry about that. Here is a simplified version of the data example: Category Consumer Commercial Health var1 4 1 0 var2 17 5 6 var3 5 7 9 var4 1 8 2 var5 7 11 4 var...n $\endgroup$ – timothyjgraham Oct 29 '13 at 22:09
  • $\begingroup$ the data works now. but nowhere in the inputs is it given as to how to decide if a category has a particular feature which is a mixture of multiple features. you need to decide how much weightage to give to each variable when deciding a feature which is not present in the given features. $\endgroup$ – htrahdis Oct 30 '13 at 14:52
  • $\begingroup$ @htrahdis Ok thanks. What if each variable simply has equal weight? $\endgroup$ – timothyjgraham Oct 30 '13 at 20:25

It seems you don't really know what you want except some sort of pattern. So why not making a Principal Component Analysis to reduce the complexity and get the direction of greatest variabilities.

| cite | improve this answer | |
  • $\begingroup$ Thanks, I will try this approach and post back the results. $\endgroup$ – timothyjgraham Oct 29 '13 at 22:13
  • $\begingroup$ Principal Component Analysis worked well, but I am finding it difficult to interpret the biplot of the results. Some of it makes a lot of sense, other aspects I'm just scratching my head about. $\endgroup$ – timothyjgraham Oct 31 '13 at 21:40

In order to check if a variable is significant for a category, do a hypothesis testing for a binomial variable assuming that the probability of getting a 1 is 0.5. In order to answer these other questions, do the same for each of the constituent variables. if the p value in the two cases is sufficient to reject the null hypothesis, then you can claim whether that variable or the set of variables is characteristic or not for the category. if you cannot reject it, then the variables behave randomly for that category. Check the link Binomial test.

| cite | improve this answer | |
  • $\begingroup$ Thanks, is there any literature you could point me towards for how to do a hypothesis testing of a binomial variable? For example, how to calculate the statistics. $\endgroup$ – timothyjgraham Oct 31 '13 at 21:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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