5 ms·
It's pretty heteroskedastic, which practically makes the errors not tooo bad. For example, say my wife is low (< 70), in that case the error rate will be small.
by RA_Fisher 11y ago
It's pretty heteroskedastic, which practically makes the errors not tooo bad. For example, say my wife is low (< 70), in that case the error rate will be small. That's where precision matters. Now suppose she's 200+, in that case the errors will be large but it doesn't really impact how to take care of it. In terms of error, What worries me much more is a server at a restaurant giving my wife a regular coke rather than a diet coke or vice versa.
Where smaller blood is really beneficial is that it requires a less painful finger prick.
- Fede_V 11y agoIf the error is proportional to the measurement - can't you just take the log to make it additive again?
- RA_Fisher 11y agoHmm, it's not clear to me what the benefit of that would be in terms of information (a transformation). It's univariate data, it's been a bit since I looked at it but a histogram showed readings were lognormal / weibull / gamma distributed (point is asymmetric, right-skewed), I believe. In terms of inference, it's kind of mind-bending to think of a "population" of blood and each person having a specific variation.