Reliability is a more subjective term, referring usually to interpretations but sometimes to
measurements. Reliability is affected by both precision and accuracy, but it also depends on the validity of any assumptions that we have made in our measurements and calculations. Dubious assumptions, regardless of measurement precision and accuracy, make interpretations unreliable.
Random and Systematic Errors
Random errors are produced by multiple uncontrolled and usually unknown variables, each of which has some influence on the measurement results. If these errors are both negative and positive perturbations from the true value, and if they have an average of zero, then they are said to affect the precision of replicate measurements but they do not bias the average measurement value.
If the errors average to a nonzero value, then they are called systematic errors. A constant systematic error affects the accuracy but not the precision of measurements; a variable systematic error affects both accuracy and precision. Systematic errors cause a shift of individual measurements, and thus also of the average measured value, away from the true value. Equipment calibration errors are a frequent source of systematic errors. Inaccurate calibration can cause all values to be too high (or low) by a similar percentage, a similar offset, or both. An example of a systematic percentage bias is plastic rulers, which commonly are stretched or compressed by about 1%. An example of an offset bias is using a balance without zeroing it first. Occasionally, systematic errors may be more complicated. For example, a portable alarm clock may be set at a slightly incorrect time, run too fast at first, and run too slowly when it is about ready for rewinding.
Both random and systematic errors are ubiquitous. In general, ‘random errors’ only appear to be random because we have no ability to predict them. If either random or systematic errors can be linked to a causal variable, however, it is often possible to remove their adverse effects on both precision and accuracy.
One person’s signal is another person’s noise, I realized when I was analyzing data from the Magsat satellite. Magsat had continuously measured the earth’s magnetic field while orbiting the earth. I was studying magnetism of the earth’s crust, so I had to average out atmospheric magnetic effects within the Magsat data. In contrast, other investigators were interested primarily in these atmospheric effects and were busily averaging out crustal ‘contamination.’
Random errors can be averaged by making many replicate, or repeat, measurements. Replicate measurements allow one to estimate and minimize the influence of random errors. Increasing the number of replicate measurements allows us to predict the true value with greater confidence, decreasing the confidence limits or range of values within which the true value lies. Increasing the number of measurements does not rectify the problem of systematic errors, however; experimental design must anticipate such errors and attenuate them.
Representative Sampling
Most experiments tackle two scientific issues -- reducing errors and extrapolating from a sample to an entire population -- with the same technique: representative sampling. A representative sample is a small subset of the overall population, exhibiting the same characteristics as that population. It is also a prerequisite to valid statistical induction, or quantitative generalization. Nonrepresentative sampling is a frequent pitfall that is usually avoidable. Often, we seek patterns applicable to a broad population of events, yet we must base this pattern recognition on a small subset of the
