0
$\begingroup$

I have an app in which users go through a form and select options which represent their views about an event they attended. These options are presented like "tags" that the user can select. The user may select zero or more tags per submission. What I'd like to do is analyse what is driving users to select certain tags. Some explanatory variables are the event, day of week, how long they stayed etc. Most of the data is categorical. However my statistical knowledge is very limited.

The data is clearly count data, so I'm considering modelling it using Poisson regression model, or zero-inflated version. Some questions I have:

  • Is this a worthy approach?
  • What assumptions would I be making if I went with this approach. One of the issues is that the counts on some tags are very low.
  • Do I need to break my data down into a daily units which include zero counts for each tag?
  • How would I actually quantify a variable is driving the count?

Some guidance would be greatly appreciated.

UPDATE: I realised I didn't add include any examples of the tags. For what it's worth they are values like "Didn't meet my expectations", "Too long", "Disliked the music", "Not enough options" etc.

$\endgroup$

1 Answer 1

0
$\begingroup$

As a first step to get some understanding of where it's meaningful to start, I would suggest talking to some users. Interviews will only generate qualitative data, but sometimes that's the fastest way to know where to focus quantitative effort.

I don't believe there's a useful answer to your question until you have that context.


If you insist on a quantitative approach before you have enough information to embark on it, a start might be to focus on some of the most common tags and look at them individually against individual predictors. In other words, scatterplot whether a specific tag was used (0/1) against each of the predictors, one by one.

If you find something useful, the next step might be logistic regression. There's a good book from Hosmer and Lemeshow on how to build logistic regression models -- the full method is a bit too much to describe in one answer here.


Why model it as 0/1? If users don't affect each other's tags too much, then once we have the probability that a single user selects a tag, the number of users that select a tag are binomially distributed.

But again, this is assuming users select tags somewhat independently -- talking to the users is a cheap way to avoid assuming!

$\endgroup$
1
  • $\begingroup$ Interesting. What kind of qualitative information would I be seeking from users? As for the quantitative approach I'm struggling to understand how I'd organise the information for the scatter plot. Do you mean adding dummy variables to present the presence of a tag in each submission? $\endgroup$ Commented Jul 19, 2022 at 21:09

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.