Then we can search in the transaction log, using the fn_dblog function, for LOP_BEGIN_XACT log records from before the crash point that have a matching LOP_ABORT_XACT log record after the crash point, and with the same transaction ID. This is easy because for LOP_BEGIN_XACT log records, there’s a Begin Time column, and for LOP_ABORT_XACT log records (and, incidentally, for LOP_COMMIT_XACT log records), there’s an End Time column in the TVF output.
And there’s a trick you need to use: to get the fn_dblog function to read log records from before the log clears (by the checkpoints that crash recovery does, in the simple recovery model, or by log backups, in other recovery models), you need to enable trace flag 2537. Now, if do all this too long after crash recovery runs, the log may have overwritten itself and so you won’t be able to get the info you need, but if you’re taking log backups, you could restore a copy of the database to the point just after crash recovery has finished, and then do the investigation.
Read on for the code, as well as a test.