Growth Hacker

Internet marketing is my area of expertise. Past consulting clients include eBay, CafePress, DocuSign, myfab, and DeepDyve. Visit my consulting site at Market Extend.
← Back to blog
Randomized Controlled Experiments: A Data Analysis Excercise To Raise Tea Sales
November 10, 2009 Jamie Shiller Data Analysis


Let’s pretend we run a tea company called Mr. Tea. We have locations across the country. The country is in the midst of a recession, and our overall tea sales are down. From an in-store survey of customers we find that customers in some locations do not believe they are receiving a good value for their money. Let’s be honest, our tea is expensive!

How do we increase tea sales?


The value perception seems to be our real problem.

Since some customers have indicated the tea is too expensive, we could try lowering the price.

Alternatively, we could try to convince them that they are getting BIG VALUE! The idea here is that if we tell them our tea is a good value, they just might believe us. We’ll make some fancy signs to tell them that our tea contains organic* ingredients (very few really, ha ha!).

So how do we structure a test to see if lowering the price or messaging value to customers will increase sales?

We could lower the price this month and compare to last month or even the same month last year. Such a historical comparison is an option, but it’s not optimal. We should be testing our variables against a control at the same time.

If we keep the original price on tea at one location, but lower the price at a location around the corner, customers will undoubtedly flock to the location with the lower price. Stores located in the same general area should be part of the same test group to avoid this type of problem.

Going a step further, let’s take the country and divide it into regions (e.g. North West or even Northern California) and micro regions (e.g. San Francisco). For example, all the tea stores in San Francisco will now list the lower price.

Each micro region will fall into one of three buckets: the control group (nothing changes), the lower price group, or the showing value group.

But how do we figure out which bucket each micro region falls into?

To structure a good controlled experiment we need to apply randomness.

Let’s head over to MS Excel. Looking at one region, let’s say it’s the North West region, we see the micro regions within the North West.

Create a new column called “Random” next to the micro regions column and apply randomness to the micro regions by using the =RAND ()  formula in the random column (copy the formula down the column). This will create a column of random numbers. Now sort the data by the random column.

Here’s where we’ll break up the micro regions into test buckets. Let’s say we have 60 micro regions. Create a new column called “Test”. For the first 20 results enter “Control” into the Test column as these micro regions will belong to the control group. For this group the price of tea will remain $4.00 a cup. Now take the next 20 results and enter “Lower Price” in the Test column. In these micro regions we will lower the price of tea to $3.75 per cup. Finally, take the last 20 results and enter “Show Value” in the Test column. Here tea will still be $4.00 a cup, but we’ll have our fancy signs strategically positioned to convince customers that $4.00 for flavored water is a steal!

After running our experiment for a month it’s time to view the results (see chart below). Comparing the Control group to the Lower Price group, we see that lowering prices didn’t make much of a difference in terms of motivating sales. On the other hand, the Show Value group did result in an increase in sales, so that strategy paid off. It’s clear we should investigate other ways to show value to customers!

Source: rehashed chapter 2 of Head First Data Analysis

View comments →