[devellis]

I think the reason the dots seem more closely clustered in the inaccurate, imprecise case than in the accurate imprecise case is that the bulls eye was kind of small. So, to move the loose cluster off center but still have the dots on the bull's eye, they needed to be grouped a bit closer together.

Accuracy in this context really means unbiased. Error is random rather than systematic. Bias in this contexts isn't a prejudicial term. It simply refers to the relationship between the central tendency of the observations and whether or not it corresponds to the true score. The dispersion is centered around the true score in an unbiased (accurate) measurement. In a biased instance, the dispersion is not centered around the true score. That is, even if you control for dispersion through sampling, the result would be biased (i.e., wouldn't converge on the true score).

Precision is just another term for low dispersion or lack of random error. Precision doesn't address whether you're actually measuring what you want to measure but simply that you're measuring something with consistency. If I used a really good light meter to measure pitch, I could get good precision (consistency) but the variable I was consistently measuring would be the wrong one (reflected light, not vibrational frequency of the strings).

Calibration, strictly speaking, is a different parameter. In formal measurement theory, accuracy is measured by the correlation between obtained scores and true scores. If the correlation is very high, the measurement is accurate (or valid). But a high correlation doesn't mean the obtained score and the true score agree. It means they're distributionally similar (i.e., the lowest value for the observed score is the same distance from the average of the observed scores as the lowest value of the true score is from the average of the true scores, and thus for every pair of observed and true scores). This is an error not of precision but of calibration. So, for example, if you assess temperature in Fahrenheit but your criterion reference that provides the true scores is in Celsius, the temperature readings may seldom agree but be perfectly correlated. The Fahrenheit values would be accurate but there would be a calibration discrepancy. A linear transformation of the values form one scale would reproduce the values for the other. In situations where there really is only one measurement scale in play, calibration may not be an issue.