Because the error will distribute around the true value, thus even with sampling errors you can still extract a signal, even a very weak one, if you have many samples.
But surely we have some which survived, or can manufacture using methods of the time to check.
Also, physics was in a pretty good state at the time, at least regarding temperature/length measurement. Surely the scientists of the time measured the bias.
> One possible reason for the lower temperature estimates today than in the past is the difference in thermometers or methods of obtaining temperature. To minimize these biases, we examined changes in body temperature by birth decade within each cohort under the assumption that the method of thermometry would not be biased on birth year.
I think that the assumption that the method of thermometry is not biased based upon birth year is probably correct.