I've just been reading
"A versatile thermoelectric temeprature controller with 10mK reproducibility and 100mK absolute accuracy" by K.G.Libbrecth and A.W. Libbrecht, Rev. Sci. Instrum. 80 page 126107 (December 2009).
The authors don't mention self-heating in their thermistor, and dissipate 2.5mW in the device. It isn't obvious what the self-heating coefficient of their thermistor probe might be, but since it seems likely to be a packaged YSI 46000, which does 10mW/C in a well-stirred oil-bath, they are probably seeing 250mK of self-heating, whcih makes nonsense of their claimed 100mK absolute accuracy.
Since I've seen a bunch of smaller glass-encapsulated thermistors provoked into showing unstable resistances in 2002 by dissipating 1mW in them, 2.5mW sounds much too high for comfort.
The 2002 event involved getting calibration data by measuring the resistances of a bunch of thermistors in a well-stirred water-bath over a range of tmperatures. The resistances were being measured by a big Thurlby-Thandar multimeter, set to autoranging, and the resistances read went unstable - in the last few digits - at highish temperatures when the multimeter had moved down to the 1k or less range.
By disabling autoranging and moving up to the 10k or less range, we went back to seeing stable resistances.
I've poked around since then looking for discussions of "hot channel" formation in NTC thermistors, but what analysis I've seen of self- heating and its consequences seems to assume that if 10mW/C is the self-heating at zero power dissipation, it is the right figure at any dissipation.
In fact, even if NTC thermistors were homogenous objects (which they aren't) self-heating is going to move the current paths in towards the axis of the thermistors, so that the heat is going to have to diffuse through more thermistor to get out, and the heat flux per unit area around the core is going to rise as the current gets moved in towards the axis, so the self-heating per watt is going to rise with increasing dissipation.
In reality, thermistors are inhomogenous lumps of incompletely sintered metal oxides, and the conduction paths are going to be a chaotic network; as the dissipation in the thermistor increases, the shorter inner paths are going to get to carry more of the current and the self-heating per watt is going to rise, but it is going to be a somewhat random process, which could explain the unstable resistance readings that I saw back in 2002.
There's probably a paper in this if anybody has access to a useful batch of themistors, a well-stirred water-bath and a six-digit multimeter or resistance bridge.
-- Bill Sloman, Nijmegen