Reducing bit depth without dithering causes unwanted artefacts
For AU 2019/2020, reducing 24-bit audio to 16-bit audio with no dithering will cause unwanted artefacts in the noise floor.
The last known good version to me is AU 2015.
Attached are screenshots of the mixdown of original file inverted with the file converted in Audition, and comparing the same file converted in iZotope RX with no dithering.

-
Ho Yin Yan commented
https://drive.google.com/open?id=1yhu2JGhuOXas6LPFtlOU9y0cQUdzwrHS
Please see the link above for the files.
The original 24/48 file was recorded directly in Au, saved as _2448.wav.
Then I converted to 16bit inside Au to produce the file with the suffix au1648.
2448.wav imported to RX 7, exported as 16bit (truncated), named with suffix rx1648.The other two files are inverting one of the tracks and null tested with the _2448.wav
Result of the Au null test shows there are differences even before gaining up the audio. Therefore it is not truncating properly.
-
C Van Winkle commented
I did a test with Au CC (6.0) and Au 2020 (13.0) and only saw at most a one sample level difference in the truncation. I believe the explanation for that change is changing the particular rounding method when truncating in order to be consistent across all platforms. Are you able to post the original file as well as your truncated result somewhere?