Calculating averages and percentiles

Learn how to model and query percentile-based metrics alongside averages so skewed numeric distributions are represented accurately in Cube.

Use case

We want to understand the distribution of values for a certain numeric property within a dataset. We’re used to average values and intuitively understand how to calculate them. However, we also know that average values can be misleading for skewed distributions which are common in the real world: for example, 2.5 is the average value for both (1, 2, 3, 4) and (0, 0, 0, 10).

So, it’s usually better to use percentiles. Parameterized by a fractional number n = 0..1, where the n-th percentile is equal to a value that exceeds a specified ratio of values in the distribution. The median is a special case: it’s defined as the 50th percentile (n = 0.5), and it can be casually thought of as “the middle” value. 2.5 and 0 are the medians of (1, 2, 3, 4) and (0, 0, 0, 10), respectively.

Data modeling

Let’s explore the data in the users cube that contains various demographic information about users, including their age:

name	age
Abbott, Breanne	52
Abbott, Dallas	43
Abbott, Gia	36
Abbott, Tom	39
Abbott, Ward	67

Calculating the average age is as simple as defining a measure with the built-in avg type.

Calculating the percentiles would require using database-specific functions. However, almost every database has them under names of PERCENTILE_CONT and PERCENTILE_DISC, Postgres and Snowflake included. For BigQuery, you’d need to use the APPROX_QUANTILES function.

YAML

cubes:
  - name: users
    # ...

    measures:
      - name: avg_age
        type: avg
        sql: age

      - name: median_age
        type: number
        sql: PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY age)

      - name: p95_age
        type: number
        sql: PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY age)

Result

Using the measures defined above, you can explore statistics about the age of your users. For a typical dataset, the average age closely matches the median age, and the 95th percentile reveals the upper bound for the vast majority of users — for example, if p95_age returns 82, then 95% of all users are younger than 82 years.