A reference dimension occurs when the key column for the dimension is joined indirectly to the fact table through a key in another dimension table.  This results in a snowflake schema design.

The following figure shows one fact table named InternetSales, and two dimension tables called Customer (regular or intermediate dimension) and Geography (reference dimension), in a snowflake schema:

Note that for performance reasons, it’s better to not use reference dimensions.  Instead, merge the Geography info into the Customer table (see Denormalizing dimension tables).  So the Customer table would add the fields City, StateProvidinceCode, and StateProvinceName populated by the ETL, resulting in one table and a true Star Schema.

More info:

Gotchas With Referenced Dimensions

Stars and Snowflakes and Bears, Oh My!