Use the impedance curves, when you can find them, or estimate them from known parameters.
A filter is an impedance network, doesn't really matter if the elements are strictly capacitive or inductive. No true capacitance exists in real life, after all. An inductor divider for instance is still a divider.
Electrolytics are typically modestly capacitive at low frequencies, <Hz to ~kHz. Midband, ESR dominates. At high frequencies, ESL dominates (>100s kHz), and there's probably some higher level structure up there too (electrolyte physics? body stub resonance in the 100s MHz?).
The capacitance is given. ESR is usually given, or impedance at 100kHz or something like that. Possibly it can be reconstructed from ripple current ratings, too (but I would suggest, just keep shopping to find one that is rated for impedance/ESR).
ESL isn't usually given, but is typical for package size -- lead length dominates. 10s of nH.
Knowing this, we can already approach some misconceptions --
To answer the original question then, I would expect a 1500µF capacitor to be inductive at hundreds of kHz and up, so the 68µF capacitors are likely to be better. The more diigent manufacturers have charts of impedance vs. frequency in their datasheets, so you can see where they start to become ineffective.
This may be true, but impedance magnitude is more important than phase. The 1500uF cap will have much lower ESR, and not much more ESL. If ESL is 10nH and you add a 1uH filter choke between a voltage source and the capacitor, the asymptotic attenuation at high frequencies will be around 0.01 / 1 or ~40dB. That's nothing to sneeze at.
Even better, use multiple in parallel, with the traces routed such that signal enters from one end, goes through the capacitors as a ladder network of their own, and out the other side. This puts some trace inductance between capacitors, further dividing the high frequency response even though the trace inductance may be tiny (5 or 10nH) and puts their ESLs effectively in parallel. Well, better than parallel, because it makes a cascaded impedance divider (ladder network). Most times, this is a tiny or negligible impact (because ESR dominates, or ESL is a few times larger than the trace inductance), but it can be better at basically no cost, even if the advantage might be fractional dB.
Where I need wideband filtering, I use a cascade of lowpass filters with increasing cutoffs. For example, 1uH into 100uF (electrolytic), then 0.22uH into 10uF (ceramic), then FB (might be ballpark 0.3uH || 30R near cutoff) into 0.1uF. This improves the shortcomings of each proceeding stage with the next, so the attenuation at very high frequencies can be essentially total.
If the offender is at low frequencies, such lengths will not be necessary, and merely using appropriate capacitors will suffice. That may require larger or smaller values. Whatever works, shop around, do simulations.
Note that we're mostly discussing differential filters. This hasn't been explicitly mentioned very often in this thread. Such a high attenuation filter isn't very useful, as common mode noise will quickly dominate, or radiation directly off the PCB. The solution is to address common mode in the same way you would address differential or normal mode noise, taking a ground reference plane (preferably a metallic enclosure, but a ground plane PCB will work in a pinch) and filtering with respect to that, using paired connections where necessary (e.g., data pairs where the differential bandwidth can't afford to be filtered so heavily, or beefy power supply lines for which a current compensated choke is cheaper).
Tim