# [Analytic Block] AB Testing with Statistical Significance

Looker has created this block to make analyzing data easier and more efficient. This Data Block is made available on an “as is” basis, meaning there will not be updates moving forward.

AB testing lets you compare how a group (often of users) acts under two sets of conditions, and can be integral for making scientifically informed decisions about your business. For example, a company might be considering redesigning their web site to increase the time visitors spend on the site. The company can show two different versions of the site, option A and option B, to two different groups of visitors. By doing this, the company will have a set of data to compare and can more easily understand how best to meet their goals.

You can also use your existing data to create groups of users to compare. Adding statistical significance to this test ensures you are adding the proper rigor to your analysis, and helps avoid erroneous conclusions. This block enables you to quickly perform this analysis inside Looker.

Note: Please be sure to use this block with the guidance of a trained statistician.

# Ideal Data Types

For a proper AB test to work with statistical significance, we need the following:

• A large sample of users/events/observations from which to draw, so we can get statistically significant results.
• Some numerical or binary `(yes/no)` dimension of those observations in which we are interested. These dimensions could be anything from bounce rates on a website, to purchases made by shoppers, to the number of defects in a product, and so on.

An easy way to begin our analysis is to split the observations into two groups. This could be by ad campaigns, versions of a website, method of manufacturing, and so on.

Once we have sufficient data, all we need to do is choose our variable of interest and our splitter variable, and we are ready to go.

## Expected Output

Once we have chosen our variables, we can dive right into our AB analysis.

In this simplified example, we are going to see how a user’s gender affects the number of orders they will make in their lifetime as a customer. In the following Look, we have split this up by age group (any arbitrary dimension will work) and can see the significance at each age, which is broken into `tiers`.

## Try It Yourself!

To create the AB testing template, we just need to add some filters, dimensions, and several measures to any existing view file. Note that you can have any number of split variables as long as they are reflected both as filters and as parts of the SQL in the `a_b` dimension.

Note: There are many ways to modify the mathematics specific to your use case, so be sure to consult a statistician to ensure the math is correct for you.

We will be using several filtered and otherwise complex measures to do statistical calculations for us. At the root, this pattern relies on a filter the user will enter in the UI to determine what variable will split our two groups for the purposes of testing.

Below is example code — with the necessary parts blocked out — in which you can place your own variables.

Starting in Looker 7.4, the filters parameter syntax has changed. See the `filters` parameter documentation page to view the new syntax.

``````view: ab_test {

# Add any variables here for splitting your data into A and B groups (in this case gender and age)
filter: a_b_gender {}
filter: a_b_age {}

dimension: a_b {
type: yesno
sql:
{%condition a_b_gender %} \${gender} {% endcondition %}
AND {%condition a_b_age %} \${age} {% endcondition %}
-- AND ...
-- AND ...
-- AND ...
-- Add more variables here for splitting
;;
}

measure: count_a {
type:count
filters: {
field: a_b
value: "yes"
}
}

measure: count_b {
type: count
filters: {
field: a_b
value: "no"
}
}

# Put in the measures for your variable of interest (in this case, lifetime orders)
type: average
filters: {
field: a_b
value: "yes"
}
}

type: average
filters: {
field: a_b
value: "no"
}
}

type: number
sql: 1.0 * STDDEV(all CASE WHEN \${a_b} THEN \${user_order_facts.lifetime_orders} ELSE NULL END);;
value_format_name: decimal_2
}

type: number
sql: 1.0 * STDDEV(all CASE WHEN NOT \${a_b} THEN \${user_order_facts.lifetime_orders} ELSE NULL END ;;
value_format_name: decimal_2
}

measure: t_score {
type: number
sql:
SQRT(
) ;;
value_format_name: decimal_2
}

measure: significance {
sql:
CASE
WHEN (ABS(\${t_score}) > 3.291) THEN '(7) .0005 sig. level'
WHEN (ABS(\${t_score}) > 3.091) THEN '(6) .001 sig. level'
WHEN (ABS(\${t_score}) > 2.576) THEN '(5) .005 sig. level'
WHEN (ABS(\${t_score}) > 2.326) THEN '(4) .01 sig. level'
WHEN (ABS(\${t_score}) > 1.960) THEN '(3) .025 sig. level'
WHEN (ABS(\${t_score}) > 1.645) THEN '(2) .05 sig. level'
WHEN (ABS(\${t_score}) > 1.282) THEN '(1) .1 sig. level'
ELSE '(0) Insignificant'
END ;;
}
``````

# Further Analysis

Note that the preceding code only works for continuous scalar variables as the target variable of interest. For a binary variable, we will need to change our formula for standard deviation. Also remember that the significance level is calculated here only for sufficiently large data sets.

To create variations in this analysis and for proper use of these methods, consult with a statistician.

Additionally, there are a lot of other statistical methods out there for other types of analysis. For relatively simple ones like this, it can most likely be coded directly into LookML.

5 Likes

converted to new LookML