Today I stumbled across the help file for "inlet~ / outlet~" which contains some interesting info about how PD performs upsampling.

Basically you have a choice of sample & hold, zero-padding, and linear interpolation.

Can anyone comment on the pros and cons of each option?

Cool! I did not know about this! Linear interpolation is equivalent to a 1 point avrg filter eg: [rzero_rev~ 1] -> [*~0.5] and zero padding it's opposite: [rzero_rev~ -1] -> [*~ 0.5]. So basically they are linear low and highpass filters. Try downsampling something to 22.05 inside a subpatch, then play with the outlet's method and listen to the differences/inspect the spectrum.

Thanks for the reply. I was hoping for thoughts about what's cleanest to save me doing the analysis myself
I am doing 8x oversampling to try and alleviate aliasing and wondering the pros and cons of each one in terms of both sound and time/CPU efficiency.

Non of the options will impact your final downsampled sounds much (downsampling is done simply by dropping the extra samples.), but interpolation is probably the way to go as it dampens high frequency content. I believe sample and hold is the default?

How low is your tolerance for aliasing? I get the spikes down below -78db and call it a day with polybleps, 2x oversampling and a biquad low pass...

I am fairly relaxed about aliasing but I am using a lot of non-linear stuff like FM, wavefolding etc and it still seems like the most efficient thing to do is upsample/filter/downsample.

I am currently into the same, trying to figure.
It depends on the usecase.
Linear interpolation, while being cheap, introduces artefacts, like ringing, and [tabsend~] with [cyclone/wave~] has more options, such as cúbic, spline, Hermite and Lagrange as [tabread4~], which all introduce different artefacts but might be worth a try, too.

And am very interested in a PD-patch that does Mip-Mapping, @seb-harmonik.ar mentioned here:
https://forum.pdpatchrepo.info/topic/13247/few-questions-about-oversampling/2