I am starting to think that maybe the impedance does not change (much) up to a certain point in the given graphs, that is why it is not shown.
In fact, impedance does not change [from proportional], when Q factor is very large.
There is an underlying fundamental here: the Kremers-Kronig relations. Indeed, to the extent we can model/measure a component as a one-port (that is, it has two terminals, and behaves with no reference to any other connection ground or otherwise, current into one pin is perfectly balanced by current out the other, there are no common mode effects), the fact that it is a real component necessitates causality/analyticity and therefore the real, imaginary, or magnitude aspects are all related uniquely, i.e. you only need to measure one.
Granted, since derivatives are involved, this relation is highly susceptible to measurement error, so you can't expect an accurate Q or ESR plot from fitting the above inductance curves.
A good way to play around with this, may be the conversion/synthesis tool I wrote some time ago [links to my website]:
Coilcraft SPICE Model Converter | Calculators | Seven Transistor Labs, LLC
Along with similar discussion at the bottom, you can adjust parameters in the interactive calculator (also click-and-drag on values for fast adjustment!) and see the effect on all measurement aspects: |Z|, R, X, Q, etc.
A consequence of the K-K is, a lossy reactive component has an impedance slightly less than proportional (or inversely) with frequency. Anywhere you see an inductor with some Q factor; or for a capacitor with ESR and Z plotted and the two track as parallel asymptotes, the ratio of those two curves (reactance X / resistance R) gives the Q factor; then the slope of impedance |Z| on a log-log plot will lie at a flatter angle, determined by that Q factor. In other words, real lossy components are not \$Z \sim F\$, but \$Z \sim F^{1-\epsilon}\$, where \$\epsilon\$ is a small value when Q is modest to large.
I haven't actually worked out the exact relation, as it's not important to me, other than to note that the important element in the above calculator has equal portions (i.e. Q = 1) and therefore a slope of 1/2 (i.e., \$R = X \sim \sqrt{F})\$. Or, let's say... applying K-K to the sub-proportional \$F^\alpha\$ (\$-1 < \alpha < 1\$) case is an exercise for the student. :)
Finally, the application to modeling, is that we can approximate such an element over a finite frequency range, with finitely many dissipative elements: a geometric series of RC or RL networks. Which, intuitively, has the effect of connecting more capacitance or inductance in parallel, reducing the apparent value of L, or increasing C, as frequency goes up. This can only occur over a finite range, because the impulse response of such an (ideal lossy) element settles very slowly over time, i.e. the difference between the element and a lossless or perfectly lossy element (we only have R, L and C to build lumped-equivalent networks from) diverges at both short and long time scales (high and low frequencies).
Fortunately, real components are finite as well, as inductor DCR dominates at LF, and EPR and C at HF; or capacitor leakage at LF, and ESR and ESL at HF.
Regarding the data in qrk's answer (original research, splendid!), we might possibly see confirmation of the very fitting errors I identified on the above page -- the published flatter curve may be a curve-fitted result, perhaps using the very same parameters used in their SPICE models (the older ones); the fact that it's flatter (decreasing less as frequency rises), implies higher Q than measured -- indeed higher Q than is physically possible given other constraints (namely, losses at high and low frequencies; the fact that the curve ticks up towards 10MHz at a certain rate, also dictates how much it rises, or how quickly it can decline, in the 0.1-1MHz range!).
In other words, what you will find (or, what I've found from testing, anyway), is using their model parameters as a starting point (which, remember, use nonphysical elements: pure resistances and inductances that vary with frequency), a reasonable fit can be had, but it will always be an approximation of the given model, whether by compromising more on L(F) droop, or overestimating Q at LF and HF while underestimating Q at MF (middle frequencies).
For a more dramatic illustration of X and R interdependence, you might look at some ferrite bead curves, perhaps even curve-fit some yourself (set up an RLC network and tweak values and topology until the curves match; this is a good exercise for getting used to whatever simulator you're working in, as well). I've published a few model created in this way myself, e.g. HI0603P600R-10.ckt (fitted using above calculator tool), or MI1206L391R.ckt (fitted by hand, arbitrary network) [links to my website; plain text files].
