This shows the importance of understanding how capacitors behave, especially ones that employ high dielectric constant ceramic materials like XR7 and such! Your data shows a decrease in effective capacitance value of 5X over the 90V DC bias span! Also this points to the well known behavior amongst older "seasoned" folks, if you need a minimum value capacitance under DC bias, then either use a much greater capacitance value, or much higher breakdown voltage, or a much less voltage dependent dielectric like COG or NPO ceramic, or all of the above, or a film type dielectric!!
Edit: One of the interesting stories behind these capacitors dates back into the 70s, we had put in quite an effort to understand the capacitor materials and capacitors themselves and had the resources on-site to do so. This was to support the requirements for nuclear survivability in which one common requirement was to survive the EMP, the Radiation impulse and long term effects. Some required work-through, others required surviving allowing a equipment "reset" . We had developed considerable knowledge into electrolytic, tantalum, film and ceramic materials/capacitors which was held as proprietary, and knew the amount of energy required from different capacitors types and materials that would destroy a chip under radiation impulse which causes all the PN junctions to become highly conductive.
One program called XM21 (Remote Sensing Chemical Agent Detector) required surviving a point-blank Nuclear Detonation!! This was a battlefield instrument purposed to protect solders by early warning of a chemical attack utilizing a remote very controversial (another story, I'll discuss if folks want) technique. We unsuccessfully argued that nuclear battlefield survivability made no sense, as anyone still around surely wouldn't be concerned about a chemical attacks after a nuclear attack!! So we had to design for survivability and the simultaneous EMP and Radiation pulse from a point blank detonation!! This imposed a limit on the decoupling capacitance and type we could use, and required that all the supply inputs and outputs were swiftly crow-bared not to mention all the other things required to pass, which we did but this drove the instrument cost up by ~4X!!
Best,