I was hoping Joel would have added his own twist to the experiment or at least commented on G0HZU's response about the 8753D upswing that Jeroen observed. For some reason, I envision Joel having a home lab with every VNA HP ever made sitting in it or at least every flavor of the 8753.
I have no home to go to, just always at work.
No, I'm not so much of a hoarder (I was until the big fires in 2017 and lost a lot of gear in my lab to smoke damage). Now I'm down to only 4 PNAs, 2 PXI VNAs, a fieldfox and a streamline. and a quite old 8753D (A.K.A worlds best network analyzer).
On the up-turn of the data at low frequencies in the measurements: I would suppose it might be mismatch due to the source match of the bridge starting to run out of steam. In the 8753, the bias network and blocking cap started to degrade the match pretty seriously at low frequencies, similarly with the PNA. And the impedance of the DUT measured is quite close to the edge of the Smith chart for hi-Z resistors, so source-match error is pretty significant. Oh, and my earlier comment was mistaken; I looked again and I see more on the order of 0.02-.05 pF for these SMT resistors. (0.2 pF was for a through-hole part I wrote about back in the day). BUT, the real thing about SMT parts is how they are positioned on a line. We did a big study pre 2000 for SMT libraries in ADS, and found that series and shunt mounting made a big difference in the model (most of this was for capacitors but holds for resistors as well). So to model your part properly you need to put in into a configuration that matches your use-case. For example for a resistor used in a Pi pad (shunt R, series R, shunt R) you need to build a shunt verison of your resistor model and a series version of your model, and very importantly, consider the pad configuration. For FR4 dual layer board at about 1.5 mm, the 50 ohm line will be wide enough for the resistor pad, but if you are using 4 or higher layer boards with RF lines on the inner layers, then the whole structure of the final-pad layout will have some strong affect on the model, and even more so if you use the resistor shunt to ground through a via. Even the pad around the via can make a difference, odd as it sounds. For example, two 100 ohm resistors in parallel will give a better match than one 50 ohm to ground, but the shunt inductance (which might limit the impedance at high frequency) is not 1/2 for two in parallel, due to coupling between the two resistors. So 2 resistors in a T-shape to ground is different than 2 resistors hung in parallel of the end of the line. It all becomes very 3-D ish.