Correlation and causation is touchy subject for mathematicians and marketers alike. Despite what every person passionate about anything would like you to think, just because two sets of data are correlated it does not mean one caused the other. There may be a link, or they may be just co-incidental. As a marketer, you do not want to be making decisions unless you know there is a causal relationship.
Whenever we send out a holiday email, our web visits drop off!
For the last 4 years, we have sent out a holiday-themed email toward the beginning of the month of December. It is the only time we sent out holiday themed emails, and every time the number of web visits we receive drops after we do this.
Should we conclude that because these two data points are correlated, we should never send out holiday emails? Of course not; in fact the holiday emails are are a response to the fact that (I’m convinced) marketers take off the entire month of December, and we are doing our best to stay engaged with them anyway.
Whenever we send more MQLs to sales, they generate more SQLs
In the above MQL and SQL data, there is a clear correlation between MQL and SQL generation numbers. In fact, the correlation is even more noticeable if you see the 30-day shift; i.e. if you shift the SQL line one month to the left, the two graphs line up even more closely.
Because MQLs and SQLs should be related, seeing correlation is expected. In addition to that, the fact that a 30-day shift makes the correlation even stronger, it indicates a lead-lifecycle time of somewhere around 30 days to make the journey from MQL to SQL. Is it guaranteed that MQL and SQL are causally related? No, but based on the data and a relationship and MQL has to an SQL it would be pretty safe to say they are.
We decreased our PR spend and at the same time, we were seeing a decrease in form conversion
The above graph shows that form-submissions and PR spend are slightly related (the behavior in Q2’13 and Q3’13 point strongly to that) there are clearly a number of other factors affecting form submission that are not account for simply by PR spend.
Not a lot could be concluded by looking at data correlation in this case. There was some correlation indicating the two things were possibly related, but not a strong enough correlation to attribute causality to the PR spend affecting form submissions.
What to do?
Scientists will conduct double-blind tests to avoid the errors of assuming correlation implies causation. By randomly splitting test subjects into two groups and applying a change to one group and leaving the other group unchanged, the true result of the change is much more easily (and confidently) observed. This concept is identical to A/B testing; e.g. randomly splitting a group into a test group and a control group to observe behavior differences.
Don’t be caught saying, “Lets just do this and see what happens” forcing you to sift through numerous groups of data looking for correlation. Be a good marketer and do a proper A/B test which will get you the answers you are looking for quickly and reliably.