Correlated errors between variables [between / within subjects, within variables, everything]!

Question

I am trying to figure out good model to analyze social dynamics using a model. I have commenting data from a website. Such as 500 articles, 10000 users, and their 75000 comments. Every user does not comment on every article.

The opinions posed by each user is linked to the opinion of the previous users, as well as the opinion of the article itself. The opinions of previous users are also linked to the previous users and the article.

I want to estimate the impact of opinion of previous user, average user opinion, article opinion and variation of opinions on the current opinion. Can you suggest a methodology for such a problem? Everything is correlated.

I also would like to ask, what software and code would you use to model a problem like this? Codes are very very appreciated.

Answer 1

I don't think you can statistically analyse all the responses to the same thread (article?) because:

1. for parametric statistics, you're going to have a lack of independence in later answers to the questions (i.e. they are going to be influenced by earlier answers), which violates assumptions for a number of methods, and
2. for non-parametric statistics (e.g. decision tree), you're possibly going to produce a model with inflated estimated accuracy due to the correlation between responses in the same thread - and longer threads may have proportionately more comments that are similar (thinking of Reddit here as an example).

Can you take a random sample of answers across all the threads, only one answer per thread (just randomise which answer is selected, e.g. the 3rd, 15th, etc) and then model using the predictors you suggest in your third paragraph, and include the answer number as another predictor? You may wish to think about how you wish to specify the current opinion, e.g. binary positive/negative, scale of positiveness, etc, as that will suggest which technique(s) are open for you to use.