Professional Documents
Culture Documents
An Introduction to Non-parametric
Statistics for Health Scientists
Christy C Tomkins MSc, CK
Faculty of Rehabilitation Medicine
what constitutes a small sample, Pett has non-normally distributed populations. of stringent assumptions associated
suggested that the choice of parametric However, it seems that most researchers with non-parametric tests implies that
or non-parametric tests just ‘depends’ using parametric statistics often just there is little probability of violating
(1). It depends on sample size, level ‘assume’ normality. Micceri et al. (11) state assumptions, which implies robustness.
of measurement, the researcher’s that the naïve assumption of normality The application of non-parametric tests
knowledge of the variables’ distribution appears to characterize research in many in health science research is wide, given
in the population, and the shape of the fields. However, empirical studies have that they can be applied to constructs
distribution of the variable of interest. If documented non-normal distributions for which it is impossible to obtain
in doubt, try using both parametric and in literature from a variety of fields quantitative measures (descriptive
non-parametric techniques. (11). Micceri et al. (11) investigated studies), as well as to small sample sizes.
the distribution in 440 large sample
Normality achievement and psychometric measures. Misconceptions about
It was found that all of the samples were non-parametric statistics
According to Pett, in choosing a test significantly non-normal (p<0.01). It was
we must consider the shape of the concluded that the underlying tenets of The lack of use of non-parametric
distribution of the variable of interest. normal-assuming statistics appeared to techniques is owing to a series of
In order to use a parametric test, we be fallacious for the commonly used data common misconceptions about this
must assume a normal distribution of in these samples (11). It is likely that if branch of statistics. Non-parametric
the dependent variable (1). However, a similar study, investigating the nature statistics have long taken the back seat
in real research situations things do not of the distributions of data were to be to parametric statistics, often being
come packaged with labels detailing conducted with some of the measures portrayed as inferior in practice and
the characteristics of the population commonly used in health science research, teaching. It has been suggested that
of origin (8). Sometimes it is feasible a similar result would ensue, given that researchers are hesitant to use these
to base assumptions of population not all variables are normally distributed. techniques, due to fears that peer
distributions on empirical evidence, reviewers may not be completely familiar
or past experience. However, often with these statistics, and therefore unable
When and why to use
sample sizes are too small, or experience to properly interpret, and review the
non-parametric techniques
too limited to make any reasonable results (12). This opinion could be owing
assumptions about the population It is apparent that there are a number to a widespread case of limited exposure
parameters (1). Generally in practice, one of factors involved in choosing whether of research clinicians to this type of
is only able to say that a sample appears or not to use a non-parametric test, statistics. Non-parametric techniques are
to come from say, a skewed, very peaked, including level of measurement, sample often left out of basic statistics courses,
or very flat population (9). Even when size and sample distribution. Table 6 and relegated to the last chapter of texts
one has a precise measurement (ratio summarizes a number of situations (12), making them seem less important,
scale), it may be irrational to assume that would implicate the use of a non- while reinforcing the focus on parametric
a normal distribution, because this parametric technique. When the choice of statistics.
implies a certain degree of symmetry statistical technique for a set of data is not
Another common misconception
and spread (9). Non-parametric statistics clear, there is no harm in analyzing the
concerning non-parametric statistics
were designed to be used when we know data both ways: using both parametric
is that they are restricted in their
nothing about the distribution of the and non-parametric tests. For each of the
application (12). It is thought that there
variable of interest. Thus, we can apply main parametric techniques there is a
are only a limited number of simple
non-parametric techniques to data from non-parametric counterpart; Experiment
designs that can be analyzed using
which the variable of interest does not with the data to determine which test
these techniques. However, there are
belong to any specified distribution (i.e. provides the best power, and the greatest
non-parametric techniques which
normal distribution). level of significance.
span from simple 2-group analysis, to
Although there are many variables in There are a number of advantages to complex structural equation modelling
existence that are normally distributed, using non-parametric techniques in (12). Basically, for any parametric test,
such as weight, height and strength, health science research (Table 7). The there is a non-parametric equivalent that
this is not true of all variables in health most important of these advantages would be equally, or in some cases, more
science (10). The incidence of rare disease are the generality and wide scope of appropriate for use.
and low prevalence conditions are both non-parametric techniques. The lack
Adapted from: Pett MA. Nonparametric Statistics for Health Care Research. London, Thousand Oaks, New Delhi: Sage
Publications, 1997.
best suited for a substantial proportion of application of statistics. New York: 14. Blair RC, Higgins JJ. Comparison of teh
the current research in this field. McGraw-Hill, 1990. Power of the Paired Samples t Test to That
of Wilcoxon’s Signed-Ranks Test Under
8. Gibbons JD. Nonparametric Statistical
Various Population Shapes. Psychological
References Inference. New York: Marcel Dekker Inc.,
Bulletin 1985;119-28.
1971.
1. Pett MA. Nonparametric Statistics for
15. Pierce A. Fundamentals of Nonparametric
Health Care Research. London, Thousand 9. Sprent P, Smeeton NC. Applied
Statistics. Belmont, California: Dickenson
Oaks, New Delhi: Sage Publications, 1997. Nonparametric Statistical Methods. New
Publishing Company, 1970.
York: Champan and Hall/CRC, 2001.
2. Fairbank JC, Pynsent PB. The Oswestry
16. Blair RC, Higgins JJ. A comparison of the
Disability Index. Spine 2000;2940-52. 10. Berg KE, Latin RW. Essentials of Research
power of Wilcoxon’s rank-sum statistic to
Methods in Health, Physical Education,
3. Cole JC, Rabin A.S., Smith T.L. that of Student’s t statistic under various
Exercise Science, and Recreation.
Development and Validation of a Rasch- non-normal distributions. Journal of
Philadelphia, Baltimore, New York,
Derived CES-D Short Form. Psychological Educational Statistics 1980;309-35.
London, Buenos Aires, Hong Kong,
Assessment 2004;360-72. Sydney, Tokyo: Lippincott Williams and 17. Conover WJ, Iman RL. Rank
4. Cohen J. Statistical Power Analysis for Wilkins, 2004. transformations as a bridge between
the Behavioural Sciences. Hillsdale, NJ: parametric and nonparametric statistics.
11. Micceri T. The Unicorn, The Normal
Lawrence Erlbaum, 1988. The American Statistician 1981;124-8.
Curve, and Other Improbable Creatures.
5. Kraemer HC, Thiemann S. How Many Psychological Bulletin 1989;156-66. 18. Lehmann EL. Nonparametrics: Statistical
Subjects? Newbury Park, CA: Sage, 1987. Methods Based on Ranks. San Francisco,
12. Hunter MA, May RB. Some myths
New York: Holden-Day Inc. and McGraw-
6. Siegel S, Castellan NJ. Nonparametric concerning parametric and nonparametric
Hill International Book Company, 1975.
Statistics for the Behavioural Sciences. tests. Canadian Psychology 1993;384.
New York: McGraw-Hill Book Company, 19. Conover WJ. Practical Nonparametric
13. Kerlinger FN. Foundations of Behavioural
1988. Statistics. New York, Chichester, Brisbane,
Research. New York: Holt, Rinehart and
Toronto: John Wiley and Sons, 1980. Q
7. Wampold BE, Drew CJ. Theory and Winston, 1964.