SQL PI only samples the top 500 records each second, and submits the top 300 unique combinations (batch,user,database,program etc) analyzed from the sample records each minute.
If there are more than 300 unique combinations each minute, then some loss of data is inevitable in the current design.
The design is meant to gather the top activity most affecting the workload as these are the things that can have the most impact if tuned. It is not aimed to include all information about all workload.
For example when the workload of a specific stored procedure is very small, (e.g. 0.14 seconds/execution) when there are more than 300 batches running each minute, the data of this example stored procedure would probably be missed.