Any measure of some aspect of human behavior, referred to as an observed score, consists of a true score plus some random error. If there were no measurement errors, then the observed and true scores would be the same, something that never occurs with living organisms. Random error arises when the measurement is affected by chance factors. An example would be fluctuations in behavioural states during an experiment or period of observation. Such error does not affect the mean performance of a group, but adds variability to the data (hence random error is sometimes referred to as ‘noise’). Moreover, random errors typically assume a Gaussian normal distribution. In contrast to random error there is also systematic error, error due to factors that systematically affect the measurement of a variable. For example, if group testing was carried out when then was persistent noise outside the testing location, then all participants are likely to be affected in a systematic manner, with errors likely to have a negative influence on performance. Systematic error can also affect performance in a more positive direction. Thus, such error can be labelled as a ‘bias’ in measurement. It occurs when there is some bias in the process of measurement, and not chance as is the case with random error. Systematic error that changes during an experiment (referred to as ‘drift’) are easier to detect, but cannot be detected statistically. Thus, drift may occur, for example, as a result of error fluctuations if a measuring instrument changes due changes in the ambient temperature. Random errors can be reduced by averaging multiple measurements. In fact, there are a number of ways for reducing measurement errors. One is to conduct pilot studies to gain information on the easiness or difficulty in carrying out performance measures. Another is the thorough training of interviewers. observers, and experimenters so they less likely to introduce errors. Then, there is always need to double-check data as thoroughly as possible. Finally, there are a range of statistical procedures to adjust for measurement (ransom) error, other than averaging over multiple measurements. They can range from the relatively simple to rather more complex modeling. Bringing together random and systematic errors results in the following additive model underlying the nature of observed scores:
Observed score = true score + random error + systematic error
See Behavioral state, Measurement error, Measurement theory, Reliability, True score, Variable