# Which test should I use to determine a relationship between Social media activity and calls booked from activity?

We run a business and I am trying to calculate the worth of spending time on our social media groups to book calls with potential clients.

I have a list of social media activity for the last 60 days (Posts made, reactions, comments, etc) and all calls booked per day for the last 60 days. If I wanted to find out if a relationship exists between "post reactions" and "calls booked" would calculating the correlation (in google sheets I'm using =correl) be the right way to do this? Or is there a test I should use for testing all of the data at once?

I took stats in college, but that was over a decade ago and never thought I'd have to use it again. But here I am wishing I didn't toss all of my notes.

EDIT

Here is a plot the OP posted in a comment below:

• What exactly are you interested in, specifically? (Pearson) Correlation tests for linear relationships. What are your variable types (numeric, categorical, ...)? Oct 19, 2018 at 7:28
• before doing any actual looking at data, it's best to come up with a clear "theory"/hypothesis - outlining what you expect to be the case - what causes what - are there any exceptions... are there confounding effects or 'reverse causation' ...en.wikipedia.org/wiki/Correlation_does_not_imply_causation. Oct 29, 2018 at 16:57
• Well, my hypothesis is that social media engagement in our group helps build goodwill and excitement about our business, which would lead to more calls about purchasing our product. Especially since every third or fourth post by us is a direct call to action to book a call. So far when I ran a pearson correlation there was a very near zero correlation between any group metric and calls booked. I want to make sure that I'm measuring the data correctly before we dump social media efforts. Oct 31, 2018 at 16:01
• Also, my data is absolute numbers. Posts made, comments made, number of reactions, and calls booked. And yes I do know that correlation does not imply causation, but if I remember my stats correctly you can't have causation without correlation. Oct 31, 2018 at 16:02
• Its not correct necessarily that causation always can be seen as correlation. For instance, if you make three social media posts each day, which unfalteringly causes five calls each, the correlation would be zero---learning is not possible without variation! Nov 1, 2018 at 9:03

So you have some daily data for 60 days. That is a kind of time series, but yours is probably not a standard time series problem. First, just computing some correlation will not solve your problem, even after defining what to correlate (which do not seem obvious). One reason is that for a time series problem you must look at what time it takes from some event (your posting something in some social media) until some effect occurs (calls booked). Do you really expect that the effects all be contemporaneous, that is, the social media posting only has an effect that same day, and then forgotten completely?

Therefore, I advice you start analysis by some plotting, a timeline of events, and in parallel, a timeline of calls booked. If you could post/link to your data we could propose something more concrete. Your problem seems interesting, and could probably also help others.

EDIT

Thank for posting the plot. But it is not easily interpreted ... You say the correlations are "close to zero". I guess that is the contemporaneous correlations. But this is time-series data, so we need to look at the auto-correlation functions, and especially the cross-correlation. Can you calculate those and post them?

In a comment you say: if I remember my stats correctly you can't have causation without correlation and if that is true, no correlation would imply no causation. But, that isn't true! As a though-experiment, if every social media post by you generated five calls, and you posted 10 times a day, constantly, then ultimately the daily number of calls and posts both would be constant, so no correlation. So you need at least to give the correlation (if it exists) a chance to reveal itself. You give it that chance by introducing variation into the experiment. That is, make a plan of social media posting, decide in advance of the number of posts to make a day, with variation, and follow that plan. That is called experimental design and there must be written about experimental design in marketing. Some googling gives this and there is much more. But you could maybe try a simple design with a few days with zero postings, then some days with 20 postings, some with zero, ... or something similar.

A little anecdote. Many years ago I watched a TV program about some (cruel) research with kittens. Some researchers (I will not disclose where, to not inspire future ecoterrorists---we don't need another unabomber). The researchers had built some very nice quarters for kittens, they had everything they needed. Except all was painted in white and black horizontal stripes. When the kittens had grown up, they were translated to identical quarters, except everything was painted in white and black vertical stripes. They were as blind.

Statisticians are like those kittens---they need variation to be able to thrive and learn!

• I would love to, but the data is so riddled with customer information there is no way I could post it publicly. But I do have a screenshot of what the scatter plot looks like between how many posts were made that day and how many calls were booked. Graph Also I when I ran the correlation function in sheets for each category of social media activity they all returned a near zero number. Oct 31, 2018 at 15:55
• That plot is difficult to interpret ... but you have count data, maybe try poisson regression, or use the variance-stabilizing transformation $\sqrt{\text{count}}$ in linear model. Then also compute the autocorrelation functios, and the cross-correlation function. What do you see? Nov 1, 2018 at 8:54
• For the future you could also try a designed experiment: Plan in advance about how many social media posts to make, and lay in some intervals without posts --- then you will see how the calls react (or fails to react) to that. Nov 1, 2018 at 8:56