Checking what attributes of a population made objects more likely to do certain action

Question

I'm pretty lost at how to approach this, and particularly with regards to the terminology; so please feel free to point to resources, fix my question name, etc. Thanks.

I used R back in college, I've got Orange installed. I've heard of MiniTab. I've got Excel installed. I've got access to SQL Server 2008 w/ Analysis Services. That's the breath of my toolkit. [I'm a developer, but that's all I know of for statistical analysis.]

I'm marketing a product, and I've got a few datasets.

Customers:

• Customer_id
• Campaign they clicked on, and the date.
• Campaign they converted on, and the date.
• The amount of biz they did with us in Q1 2010
• The amount of biz they did with us in Q2 2010
• The amount of biz they did with us in Q3 2010 ... etc.

Campaigns:

• Day of the week
• Time of Day
• Specific pitch

Results:

• Number of purchases per customer
• Value of purchases per customer

I could easily go hogwild, guessing at patterns and testing them to see if there's any accuracy; but it's not a best use of time. What I'd like to know are a few things...

• A ranking of the various pitches in their effectiveness, and an indication of how much the pitch affects the result.
• Same for the day of the week and time.
• And, trickier -- How much of a correlation does their previous business have to do with who converts now? Is it the previous big spenders that convert? Are they the ones buying the most product?
• And finally, of all the attributes above, how do they rank in their importance to conversions?

I assume these would all come with something like an R-squared indicating their 'accuracy.'

How would I go about starting? What reading would be essential to understanding this? Is there an approach/tool I can use that'll allow me to do a basic version of these sorts of analyses [and what would they be called] rapidly? I'm just looking for trends and hints of where it might be best to focus our energy; it can be rough.

Finally, does anyone know of a resource for basic data mining? analysis? for advertising in particular?

Much obliged,

Answer 1

I will just try to share my experience as I did some marketing research. Of course everything depends on the data itself, type of business, frequency constraints, etc... So I want you to understand that there is no exact answer for your question.

A ranking of the various pitches in their effectiveness, and an indication of how much the pitch affects the result.

I would just compare the cumulative results of the customers who were involved into pitch with the cumulative results for all the rest. Then you can see relative success of each pitch and rank them afterwards.

Same for the day of the week and time.

This depends a lot of the nature of your pitches. You can either do the same procedure as described above for every day of the week, or for a pair "pitch+day".

And, trickier -- How much of a correlation does their previous business have to do with who converts now? Is it the previous big spenders that convert? Are they the ones buying the most product?

You are right, that's a little trickier. I would first segment your customers into groups (as you said "previous big spenders" can be one of the groups). And after that I will have a look at the impact of the pitch for every group separately. You will see that some groups are converted more (the relative success of the pitch is higher).

And finally, of all the attributes above, how do they rank in their importance to conversions?

That can vary a lot from one industry to another. So I cannot answer.

Hope some of my suggestions would help!

Answer 2

In general, what you want is a covariance matrix. In MatLab, this is relatively simple to implement. The function corrcoef (correlation coefficient) will tell you the correlation between every pair of variables.

You will want an nx11 matrix, where n is the number of instances (one per customer) and 11 is your number of variables (6 customer variables, 3 campaign variables, and 2 results)

Corrcoef will multiply (and normalize) M'M to get you an 11x11 result. The diagonal of this matrix will just be how the i'th variable correlates with itself, namely, a diagonal of 1's. Any i,j entry in the correlation matrix will represent how variable i correlates to variable j. Answers close to 1 mean very high correlation. Answers close to -1 mean anti-correlation. Values near 0 mean no correlation.