I use both, but for different jobs, and the deciding factor is almost always the stub, not the via type on its own. At 28 GHz a through via with even a short unused stub is a resonator in the band, so I never route a 28 GHz signal through a through via and leave a stub hanging below the signal layer. If the signal has to change layers, I use a blind or stacked microvia, and if it has to leave a BGA pad, I use a filled and capped via-in-pad.
On a coupon with the same trace length, my measured comparison was consistent: a filled via-in-pad transition was typically 0.2 to 0.4 dB lower in insertion loss and 5 to 8 dB better in return loss across 24 to 32 GHz than a dogbone pad with a through via, and the dogbone showed a clear resonance dip that the via-in-pad did not. A microvia layer change was close to the via-in-pad result, provided the microvia landed on a solid ground reference and the antipad was not oversized. The difference was not dramatic at a single frequency, but it was repeatable across several coupons, and it grew worse after reflow for the unfilled via-in-pad samples, because solder wicking changed the pad impedance.
Three practical rules have saved me more than any solver setting. First, put ground vias tight around every signal via, within roughly a tenth of a wavelength, and keep the ground return continuous through the transition; a signal via with a distant ground return is an inductor, whatever its type. Second, keep the stub under 0.2 mm, or back-drill it, and model the antipad clearance as actually etched, because a 0.1 mm error in antipad diameter moves the resonance noticeably. Third, qualify the fill: ask the board shop for cross-sections, check for voids, and X-ray a sample after reflow. A voided or unfilled via-in-pad is worse than a careful dogbone.
For measurement, compare the transitions on the same coupon, de-embedded with a through-reflect-line kit, over the full 24 to 32 GHz band rather than at 28 GHz alone, and repeat after reflow and after a thermal cycle. If the return loss degrades after reflow, suspect solder in the via or a pad lifting, not the simulation.