I often do a first pass filtering the raw data. Might for example read times, add and shift right 3 steps. The need for this depends on if the hw contains any electrical noise on the input data or the voltage reference etc.

Next step is that I may add a low-pass filter on the resulting data. This also depends a bit on what I am measuring. If level of a fluid in a container of a moving vehicle, then the surface level will obviously jump up and down with the movement of the car. In this case, it is irrelevant to filter for electrical noise affecting voltage reference or input signal.

Another issue is what speed you need to sample. If sampling 50kHz, then you don't want to perform 50000 conversions from raw data to a presentable unit. Much better to perform the filtering on raw data.

If sampling once/second, then it's natural that every sample will also be displayed, obviating the need for a conversion. Any filtering may be to decide if you should generate an alarm - maybe after having had 5 measurements in a row (5 seconds) outside acceptable level. So the filtering would be made on the already converted values.