Title of entry
Random assignment
Synonyms
Definition
Random assignment defines the assignment of participants of a study to their
respective group strictly by chance.
Introduction
Statistical inference is based on the theory of probability and effects investigated in
psychological studies are defined by measures that are treated as random variables.
The inference about the probability of a given result with regard to an assumed
population and the popular term “significance” are only meaningful and without bias
if the measure of interest is really a random variable. To achieve the creation of a
random variable in form of a measure derived from a sample of participants, these
participants have to be randomly drawn. In an experimental study involving different
groups of participants, these participants have to additionally be randomly assigned to
one of the groups.
Main Text (headings chosen by authors)
Why is random assignment crucial for statistical inference?
Many psychological investigations, such as clinical treatment studies or
neuropsychological training studies, comprise different participant groups. The goal of
most studies is to generalize results obtained for the sample to a larger population of
people. These generalizations are based on the theory that all observations made in the
experiment are randomly drawn from a population with a given random distribution.
The observations are usually called xi, referring to the concrete realization of a random
variable Xi, that stands for all possible values the observation of person i may take on.
If the distribution of the population, this observation was randomly drawn from, is
adequately described by a normal distribution with parameters and 2, this random
variable is described by:
Xi ~ N(; 2).
Moreover, the statistical measures computed from the random observations are also
random variables (because they are derived from random variables). Often, means of
different groups are compared, such as in neuropsychological training studies. For
example, a plausible question would be, if the mean reaction time of a training group
and a control group ( and ) differ after the reaction time training. If the
random variables of the observations of the groups follow a normal distribution with
variance ², the two means computed from the random variables in the two groups of
equal size n are random variables defined by:
and .
To see if the training of the two groups differs, the difference between these two
means is computed, which, in turn, is another random variable:
Thus, it becomes clear that the theoretical ground for statistical inference from
experimental results can only hold if the observations in the groups are randomly
drawn from a population. However, in a study comprising more than one group of
participants, randomness of observations is not guaranteed by randomly drawing a
sample from the population in order to conduct an experiment: all groups have to
comprise random samples of observations. Notably, random variables are based on
drawing observations with replacement, meaning that a single observation may be
drawn several times. This is not feasible in psychological research, yet it constitutes a
negligible bias if the sample constitutes less than five percent of the total population.
The importance of random assignment
The previous section explains why statistical inference and unbiased significance
testing cannot be achieved without drawing cases for observations at random from a
population to create a random sample. However, if more than one group is needed in a
psychological investigation, the random process is not finished with drawing a random
sample from the population. The assignment of the randomly drawn participants to the
groups has to be at random as well in order to ensure any measures computed from
cases within a group are still random variables.
Leaving aside the fundamentals of statistical inference and random variables, it is easy
to imagine cases in which non-random distribution of the subjects to the groups may
lead to biases or incorrect results. If, for some reason, one of two groups is assigned
the subjects that are free to be tested in the evening and the other is assigned the
subjects that are free in the morning, the first group may only comprise employed
participants and the other group solely unemployed ones, possibly providing an
experimental confound. The same assigning the participants to different groups by
age, sex, or other attributes. If this attribute is not the independent variable of choice,
group effects cannot be clearly traced to the independent variable. Random assignment
therefore, also provides a “safeguard” for the researcher, protecting him from letting
his wished or opinions (unconsciously) influence the outcome of a study by
systematically assigning different types of participants to particular groups
(Gigerenzer et al., 1989).
The only way to ensure that the random variable assumption for the measure of
interest still holds is to distribute all subjects strictly at random over the groups.
Random distribution can be achieved by an automatized process, like a random
number generator, pre-existing lists for assigning the subjects to the groups in a certain
order, or even a coin flip.
Importantly, random assignment cannot ensure that the distributions of gender, age,
and other potential confounds are the same across all groups. It can only make it
possible that the deviation is at random, which is what the distribution of the random
variables account for and what is ultimately the basis for significance testing and
confidence intervals.
Conclusion
Conclusion
Measures computed from observations in samples are random variables. These
random variables are defined by their random distribution, which, in turn, is defined
by its shape derived from random selection or random sampling. When multiple
groups are involved in an investigation, random sampling does not stop after sampling
participants at random from the population but continues with randomly distributing
them over the different groups to enable unbiased estimates within groups. This
random distribution is called random sampling and can be accomplished by any
random process, such as a random number generator.
Cross-References (related entries selected from the table of contents)
Random selection; confound (experimental)
References (up to 10 for short entries; up to 30 for medium and long entries)
Gigerenzer, G., Swijtink, Z., Porter, T., Daston, L., Beatty, J., & Kruger, L. (1989).
The empire of chance: How probability changed science and everyday-life. New York:
Cambridge.