Type A Evaluation of Uncertainty | Devin Alex
Devin Alex
← Back to Uncertainty

Type A Evaluation

A Type A evaluation of standard uncertainty is a method of evaluation of uncertainty by the statistical analysis of a series of observations. In metrology and calibration, this method is fundamentally used to quantify the dispersion or scatter of repeated measurement results.

Unlike Type B evaluations, which rely on scientific judgment and external information, a Type A evaluation is strictly empirical. It is derived directly from the data collected during the measurement process itself. The core assumption is that the repeated observations are independent and randomly distributed around the true value of the measurand (or the best estimate thereof).

The Arithmetic Mean

When a quantity $q$ is measured $n$ times independently under the same conditions of measurement, the resulting values $q_1, q_2, ..., q_n$ will generally differ. The best estimate of the expected value of $q$ is the arithmetic mean, or average, of the $n$ observations:

$\bar{q} = \frac{1}{n} \sum_{k=1}^{n} q_k$

The arithmetic mean $\bar{q}$ serves as the measurement result. However, to complete the measurement statement, we must evaluate the uncertainty associated with this estimate.

Experimental Standard Deviation

The dispersion of the individual observations around the mean is characterized by the experimental standard deviation of the sample, denoted as $s(q_k)$:

$s(q_k) = \sqrt{\frac{1}{n-1} \sum_{k=1}^{n} (q_k - \bar{q})^2}$

This value, $s(q_k)$, represents the standard deviation of a single measurement. It provides an estimate of the variability of the measurement process itself. If we were to take one more measurement, $s(q_k)$ describes the expected deviation of that single new measurement from the mean.

Standard Deviation of the Mean

However, our reported measurement result is not a single observation; it is the arithmetic mean of all $n$ observations. The mean is always a more reliable estimate than any single measurement. Therefore, the uncertainty of the mean is smaller than the uncertainty of a single measurement.

The standard uncertainty of the measurement result (the mean) is calculated as the experimental standard deviation of the mean, denoted as $s(\bar{q})$ or simply $u_A$:

$u_A = s(\bar{q}) = \frac{s(q_k)}{\sqrt{n}}$

This equation demonstrates a fundamental principle in metrology: taking more measurements (increasing $n$) reduces the Type A uncertainty of the resulting mean.

Degrees of Freedom ($\nu$)

The degrees of freedom, denoted by the Greek letter nu ($\nu$), is a critical parameter associated with a Type A evaluation. It reflects the reliability of the standard uncertainty estimate. A standard uncertainty based on only 2 or 3 measurements is much less reliable than one based on 30 measurements.

For a simple Type A evaluation consisting of $n$ independent observations used to calculate a single mean, the degrees of freedom is:

$\nu = n - 1$

The degrees of freedom parameter becomes essential later in the uncertainty analysis process when combining uncertainties and determining the appropriate coverage factor ($k$) using the Welch-Satterthwaite equation to calculate expanded uncertainty.

Practical Considerations

  • Minimum Sample Size: While calculating a mean and standard deviation is mathematically possible with just $n = 2$ measurements, it is statistically poor practice. A small sample size leads to very few degrees of freedom, which ultimately forces the use of a much larger coverage factor $k$ to achieve a 95% confidence interval, drastically inflating the expanded uncertainty. In precision metrology, a sample size of at least $n = 10$ is often recommended when characterizing a process.
  • Pooled Standard Deviation: If a measurement process is well-characterized over time across many different calibrations, it is often more accurate to use a "pooled" standard deviation derived from historical control charts rather than relying solely on the $n$ measurements taken during a single specific calibration event. This pooled value will have significantly higher degrees of freedom.
  • Normality Assumption: Type A evaluations fundamentally assume the underlying distribution of the repeated measurements is approximately normal (Gaussian). If a measurement process is heavily skewed or exhibits a non-normal distribution (such as a uniform distribution caused by the finite resolution of a digital display dominating the variance), different statistical treatments may be necessary.