2
$\begingroup$

My name is Abhi and I am fairly new to statistics. I found some sample exercises online & I am trying to solve them to get a better understanding of model development.

Problem Statement

Assume a forum like stack exchange or cross validated. Using parameters like number of comments, number of upvotes, reputation points of the user, etc (there are about 10 more fields) can you predict the likelihood of the question getting answered. There about 4000 records to consider

Current Attempt

It seems to me that number of up votes and comments should be strong indicators. However when I graph them (and do the chi square significance test) I get very poor results enter image description here enter image description here

number of up votes - p value much less than 0.05
number of comments - p value much less than 0.05

My Question

What should be my next step from here? Is there something obvious that I am missing? Are there any transformations that I should consider. Any help would be much appreciated

Update #1

I graphed the distribution for number of comments & number of likes. They are not normally distributed

Update #2

Based on the suggestions below, I calculated the means and standard deviation of the 2 groups - number of comments for the people whose question was answered (m-2.094,sd-4.008) & number of comments for people whose question was unanswered (m-5.22,sd-5.688). Both of them are within 1 standard deviation of each other. I ran the t test and the difference between the 2 means is 1.91 with a p value <<< 0.0001 Does this mean this feature is useless or do I need to transform this feature

Improve this question
$\endgroup$
9
  • 2
    $\begingroup$ Forget regression; just do a t-test for a difference in means. I'm tempted to say that there's no relationship and so should you. Your eyes are smarter than your model. However you should always jitter your points (just add uniform-distributed noise to your y variable) when you're making this kind of plot, because overplotting makes it impossible to get a sense of relative densities. $\endgroup$
    – shadowtalker
    Jul 1, 2014 at 6:38
  • $\begingroup$ I would advise never fitting a model until you've taken a peek at the means. On average, do the questions that are answered have more upvotes than those which aren't answered? Do they have more comments? Your model can tell you if your results are statistically significant (as can a t-test, as @ssdecontrol points out), but it's those averages tell you what's actually going on. $\endgroup$
    – Eoin
    Jul 1, 2014 at 10:56
  • $\begingroup$ Can you clarify what the problem is? Is it just that the graphs look surprising given the numerical results? $\endgroup$
    – gung - Reinstate Monica
    Jul 2, 2014 at 3:11
  • $\begingroup$ @gung Assume 2 data sets which have means within 1 standard deviation. Further assume that the data is not normally distributed. Should I (a) Hunt for some transformation to normalize the data & then feed it into logistic regression (b) Switch from logistic regression to some other modelling technique (c) Is there a 3rd option? $\endgroup$
    – Abhi
    Jul 2, 2014 at 4:21
  • 1
    $\begingroup$ Knowing nothing beyond what's posted here, I don't see that you need to do anything else necessarily. I disagree w/ the previous advice that you should do t-tests instead, that answers a different question & you should run the analysis that addresses the question you have. (a) You should not try to transform your predictor data to normalize them, LR makes no assumptions about the distribution of X, (b) no, as just discussed, (c) option for what, exactly? $\endgroup$
    – gung - Reinstate Monica
    Jul 2, 2014 at 14:12

1 Answer 1

Reset to default
0
$\begingroup$

First: I am not sure that the number of comments is a good indicator for the likelihood of an answer. Maybe it's a nonlinear relationship? E.g.: Zero comments are a bad sign, while a few comments are a good indicator and dozens of comments are indicating a wild discussion but no obvious solution? (this hypothesis is based on my feeling of stackoverflow).

Second: Do your data capture the time of the upvote or comment? Many comments and upvotes are occuring when the answer was given, so they have not influence whatsoever. Try to capture the number of comments / upvotes immediately after the answer.

Third: I would second @ssdecontrol: Use a simple t-test to compare group means. Or is logistic regression a requirement?

Last: which fields are also available?

Improve this answer
$\endgroup$
1
  • $\begingroup$ Logistic Regression is not a requirement. What do you recommend? Other attributes are (a) number of questions posted on the forum (b) number of tags attached to the question (c) number of days on the forum (d) number of days since first post on the forum (e) number of times the OP has commented on his post $\endgroup$
    – Abhi
    Jul 1, 2014 at 21:07

Your Answer

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

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