0

I have a dataset of 30 social variables such as Facebook Likes, Posts, Comments, etc. I would like to see if these variables predict Website Views.

MY problem is I only have 3 months of data- or 3 data points. Thus I have a 3 by 30 dataset. My question is, how do I model with such few data points but lots of variables?

In addition, I would love a resource that walks through regression in R. My data does not even remotely look linear, so I don't know what to do next. Suggestions?

Tim
  • 108,699
  • 20
  • 212
  • 390
lmcshane
  • 187
  • 1
  • 7
  • Welcome here. Please split your questions into separate threads – sheß Apr 21 '16 at 22:24
  • What exactly do you mean by "y data does not even remotely look linear". Would LASSO regression or sth be an option? Is your ultimate goal to predict as accurate as possible, to summarise correlations, or something else? – sheß Apr 21 '16 at 22:29
  • 8
    3 observations, 30 variables implies that you need many more observations. Even 30 observations and 3 variables would be a stretch. – Nick Cox Apr 21 '16 at 22:37
  • Do you expect to collect more data points in the future? – dv_bn Apr 22 '16 at 01:59
  • 3
    This can not be done. Your linear regression is what we call *underdetermined*. Because you have way more variables than you do observations, linear regression won't be able to pick out coefficients that minimize the sum of squared errors. There are, literally, an infinite number of coefficients that would do so, which doesn't sound very useful. You are going to need to wait for more data. – Demetri Pananos Jul 07 '19 at 15:51
  • 2
    You need more data. People are suggesting using regularization or local methods, but both will give highly variable results. Frankly, your inference will be questionable even if you take a simple average. – Josh Jul 07 '19 at 16:22
  • 1
    I may be missing something, but why do 3 months of visits result in 3 data points? Why not summarize by day? That would give you 30x more observations. – burger Jul 08 '19 at 19:28

2 Answers2

0

Yes, as Nick mentions, it is definitely a stretch for using models with assumptions. If you just want to predict, I would suggest that you use nearest neighbor method as a non-parametric method for prediction.

If you have little more points, you can use partial least squares (PLS) if more variables than data points.

As pointed out by others, three data points is too few for developing a regression model. I would suggest to collect at least 20 data points to run a regression model.

KarthikS
  • 1,066
  • 1
  • 9
  • 18
  • 4
    This really isn’t appropriate given the size of the data set. Op should collect more data before considering any modeling. – Demetri Pananos Jul 07 '19 at 16:11
  • @DemetriPananos, I agree that it would be ideal to collect more data. But in case, only data point is available, that is the best guess. In case, only two data points are available, the outcome of the closest one to the test data is the best guess, so on and so forth. These are be written into the code until sufficient data has been collected to do any proper modeling. – KarthikS Jul 08 '19 at 01:57
  • 1
    I'm not trying to answer if OP can make predictions, I'm answering *should* OP make predictions. No worthwhile predictions can be made from 3 data points housing 30 measurements. Telling OP otherwise is at worst negligent and at best shows a gross misunderstanding for the modelling processes. – Demetri Pananos Jul 08 '19 at 02:28
  • You are right. I will update my answer to suggest that nearest neighbor is not really a model and for 3 data points, it really works only as a heuristic. – KarthikS Jul 09 '19 at 05:41
0

You have insufficient data. Start with the Rules of thumb for minimum sample size for multiple regression thread, with thirty variables you would need at least a few hundred samples. With three samples, you could try fitting a model with one parameter, intercept-only model, i.e. calculate the mean of the dependent variable, but still, three samples would be very little for this as well.

Tim
  • 108,699
  • 20
  • 212
  • 390