Dealing with the Normality requirement

Generally the 'Roll's Royce' inferential statistics - the most powerful and sophisticated
ones - require DV's in the form of interval numerical scores and the assumption that
the population sampled has a particular distribution, the so-called 'normal' distribution
(and often other prerequisites as well). I.e. the mathematics of these tests assumes
that the population sampled has the scores of all the individual cases spread round the
average in all a particular bell-shaped pattern that you see illustrated in all the books.
Given that distribution of scores in the population, the statistician then has formulae
that predict how often you would get samples of a given size with a particular average
if you picked them randomly from that population over and over again. This is what is
used to decide if your sample(s) is unusual or not, and hence if your sample suggests a
significant difference or whatever. Statistical tests etc. with this assumption are often
referred to as 'parametric' statistics; they generally come on the left of my flow charts.
Mainly they are the t tests and ANOVA/GLM.
What do we mean by 'distribution'? We mean the shape of the histogram when you
plot the possible scores on the bottom scale and the number of cases scoring each
score on the side scale. The shape of the distribution of a set of scores is not captured
either by the mean or the SD of the set: in practice you have to look at the histogram.
What does 'normal' mean. It just refers to a particular shape, also called the Gaussian
curve, with a known formula. It does not mean you 'normally' find this shape.
Why are such tests called 'parametric'? A parameter is a limit within which something
can operate. E.g. the police operate within parameters defined by the law - so for
example they are not relevant in disputes which involve civil rather than criminal law.
Similarly parametric tests work within limits defined by the distribution shape of the
populations. In principle they are not relevant where the populations sampled do not
have this distribution. But non-parametric tests work on populations with any
distribution shape (it is just a bit unfortunate that there do not exist non-parametric
tests that are as versatile and powerful as the parametric ones, otherwise it would be
simpler to use those all the time!).

How do we know if the distribution has the normal shape?

The problem of course is how to know what the distribution of scores in a population
is like, since usually nobody has measured the entire population and drawn a
histogram of scores for the whole population, and seen what shape it is. The normal
distribution is known to arise commonly in the natural world, but certainly can't be
assumed to arise everywhere.
Two common ways of checking are:

Make histograms of the scores for each group on each variable. Get SPSS to
insert the normal curve shape (ElementsShow distribution curve). Decide
by inspection if the histograms fit the normal shape. BUT remember it is the
population that has to have this shape, not the samples and common sense
says we can't take it that the distribution in the population is exactly that in the
sample. The usual common sense about sampling tells us that one might get
samples whose distribution looks quite different from that of the population
they are from.

Do a side investigation to answer the question 'If the population is normally

distributed, how often would you get the distribution in my sample coming up
1

in random samples taken from it?'. If the answer is "often", we'll take it that
the population is normally distributed. If the answer is "rarely" (taking perhaps
5% or less of the time as = "rarely"), then we will say not. Put more formally,
you test the hypothesis that there is no difference between the 'observed'
distribution in the sample and an 'expected' normal distribution". There are
several inferential and other techniques that can be used to tackle this
question. The best and simplest I find is the one sample Kolmogorov-Smirnov
(K-S) test, using the SPSS program (look in AnalyzeNonparametric). Use
it on each separate subgroup of subjects if you have several. You end up with a
probability (between 0 and 1) that a sample is from a normally distributed
population. In interpreting this probability, unlike in the use of many
inferential tests, you are usually "wanting" a p greater than .05 (or 5%) rather
than less. It is usually helpful for you if the null hypothesis is confirmed, so
you conclude that your sample's distribution is no different from the normal
distribution, but for departures just due to sampling. Then you can feel
justified in using parametric inferential stats to compare groups etc..

What to do if the shape is not normal

In practice, researchers take a variety of lines. Choosing parametric tests or not is
often a matter for judgment, not maths.
* Some say, "If previous people working in the same area have assumed the normal
distribution, I will too without further ado".
* Others say "It may be a false assumption, but I'll assume it anyway since in practical
use of statistics one is always violating some assumptions anyway, and there is some
research that shows that you don't go far wrong if you use parametric stats where the
population distribution is not exactly normal".
* Others (e.g. Langley) again say, "Let's be cautious and never assume it without good
evidence, and there is research that shows that you don't lose much in power by using
non-parametric stats instead". If in doubt choose 'No'! The problem is, what if there is
no nonparametric test to use instead?
One of the above is the easiest choice for the beginner. The next requires work!
* Sometimes the distribution of the population is known from a lot of past work to be
of some specific shape that is definitely not the 'normal' one. Or you feel able to
assume from the distribution of your sample(s) that your population has a non-normal
shape for some reason. Often known shapes that are not normal can be "made into"
the normal shape by transforming the scores of each case. Some researchers do such
mathematical conversions of all the scores in their sample(s) to achieve this and then
use parametric stats on the transformed scores. The commonest conversions of data to
make them more normal are:
Percentage scores (e.g. subjects each get a % correct score for use of
the -s third singular), and other scores on scales with two fixed ends
(like 5 point rating scales): use the arcsine transformation.
Response time scores, which are of course on a scale fixed at one end
(0 milliseconds) but extending indefinitely at the other end: use the
logarithmic (log) transformation.
What is a transformation? A 'transformation' mathematically is any conversion of
scores to a related set of scores, following some mathematical rule. So for instance if

you make a new set of scores that is the original ones multiplied by two, or with one
subtracted from all of them, those are simple transformations - but not one that
changes the distribution shape. Re-expressing a set of scores as percent out of some
total is another. Arcsine and logarithmic transformations simply involve more esoteric
mathematical functions which have special effects on distribution shape.
In SPSS you can create new columns of figures from existing ones using
Transform...Compute then enter a name for the new target variable and in numeric
expression put a suitable formula including the variable to be transformed.
Scores on a scale fixed at one end: the log transformation
Sometimes where a histogram shape of a sample (i.e. the 'distribution' of scores) is not
very obviously 'normal', it may have a recognised non-normal shape which can be
transformed into 'normal'. This then makes it more appropriate then to use the stats
that assume the normal distribution shape (so-called 'parametric' tests like t tests and
ANOVA).
For example, response times in psycholinguistic research often are 'skewed' to the left
(positively). They are usually recorded in milliseconds, and also incidentally often
found referred to as reaction times, or 'latencies'. For example in resptim.sav there are
a whole lot of response times gained from a single subject called Ettore in a
psycholinguistic experiment by a student working on the processing of regular and
irregular verb forms in Italian. Subjects had to say as fast as possible whether the
forms existed or not as they appeared on computer screen.
Load the file and see the shape of the histogram.
Why do you think it is that the main 'heap' is not centred on the scale
with symmetrical tails on both sides? Think of the nature of response
time data.
A heap on the left of the distribution of scores is also called 'positive
skew', a heap on the right 'negative skew'. Which have we here?
Check with the one sample K-S test to see if the distribution could be
from a population with the so-called 'normal' shape.
Now try using the log transformation:
Log transformed score = logarithm of raw score
Obtain the logs of the original scores in a new column. From the top menu use
Transform.... Compute... and choose either the LN(numexpr) or LG10(numexpr)
function from the list. Click it into the numeric expression box and replace the word
numexpr in the function with the resptime variable. In the target variable box enter
a name for the new column (e.g. logrt). Then OK.
Look at the histogram of the transformed set of scores. Is it more
normal looking?
Try the K-S test. Does the new distribution pass the test for being a
likely one to be obtained from a population with a normal distribution?
LN stands for 'natural logarithms' and LG10 for logarithms to base ten.
Try both and see if it makes any difference to the histogram or K-S test
result.

For those who don't know what logarithms are..... they convert numbers in the
following way. Logs to base ten (LG10 or log10) re-express a number as the power to
which ten would have to be raised to be the same value. So LG10 of 100 is 2 (because
102 = 100); LG10 of 1000 is 3 (because 103 = 1000). Other logs are intervening
figures: e.g. LG10 of 200 = 2.301.
On that system, LG10 of 10 would be.... what? LG10 of 1 would be...?
Can you see how this conversion makes the differences between larger
figures on the original scale smaller on the new scale, so it contracts
the right hand side of a histogram and moves a heap on the left hand
side of a distribution more towards the middle?
Natural (or Napierian) logarithms (LN) work exactly the same except that the base
figure is not 10, but a figure that has special mathematical properties, usually referred
to as e, which = approximately 2.7183. So the original numbers get re-expressed as
the power to which 2.7183 would have to be raised to be equivalent. E.g. LN of 100 =
4.605 because 2.71834.605 = 100. (This peculiar figure e is as important in maths as
another peculiar figure you possibly know better, pi. While pi has all sorts of uses
when calculating things to do with circles, e, the exponential constant, has great
importance in areas of maths where one calculates growth. E.g. it helps you calculate
the compound interest on an investment. But that is irrelevant here). In principle one
can make logs to any base you like, but these two types are the commonest.
Percent scores: the arcsine transformation
The arcsine transformation is a trigonometric function concerned with the
relationships between the lengths of sides of right-angled triangles and the angles
between them. If you have never done trigonometry, then skip to the next paragraph.
If you have, then you will have come across sines. The sine of an angle in a right
angled triangle is the length of the side opposite the angle divided by the long side of
the triangle (hypotenuse). Maybe you remembered that with the mnemonic OHMS
(opposite over hypotenuse means sine). A little thought shows that with an angle of 0
degrees (or 0 radians) the sine will be 0. With an angle of approximately 90 degrees
(or /2 radians = 3.1416/2 = 1.57 radians) then the sine of the angle will be 1. And
with a negative angle of 90 degrees (or 1.57 radians) then the sine will be 1. The
arcsine is the reverse function of the sine. So since sine 0 = 0, then arcsine 0 = 0; as
sine 1.57 radians = 1, then arcsine 1 = 1.57; and so on.
However, for our purposes, all we need to know is that this transformation stretches
the ends of a fixed end scale but leaves the middle relatively untouched. So it tends to
get rid of ceiling or floor effect i.e. where your distribution of scores is not normal,
because it is bunched in a skewed way at the top or bottom of the scale.
Things are complicated by the fact that the arcsine transformation requires input in the
form of a scale running within the limits -1 to +1, so any fixed-ended scale that you
want transformed has to be converted first to that. A commonly used form of the
arcsine transformation that gets over such difficulties (Rietveld and van Hout 1993
p126) is as follows:
Arcsine transformed score =
2 x arcsine of (square root of (raw score/max possible score))
In SPSS you get this via TransformCompute. You will find ARSIN(numexpr)
and SQRT(numexpr) available in a parallel way to LN(numexpr) described above.
4

If your scores to be transformed are in a column called rscore and all out of 15, then
you create a formula of the type: 2*ARSIN(SQRT(rscore/15)). If the scores are
percentage scores, then the formula is 2*ARSIN(SQRT(rscore/100)), and so on.
The new scale always runs between 0 and 3.14. To see the effect, I here show what
some percent scores come out as:
Original % score

Transformed to
an arcsine score

rscore
arcscor
0
0.0
10
0.64
20
0.93
30
1.16
40
1.37
50
1.57
60
1.77
70
1.98
80
2.21
90
2.5
100
3.14
See how the original scale is stretched at the ends and so, if scores were bunched at
one end (floor or ceiling effect), the shape of the distribution would get stretched on
the side nearest the end of the scale and so end up more normal-like.
For data where scores at the exact ends of the scale are not observed, the log
probability ratio transformation has a similar effect. For %correct scores it would take
the form:
Transformed score = Log10 (%correct/%incorrect)
It stretches the ends rather more than the arcsine transformation, but cannot cope with
scores of 100% or 0%.
Original % score

Transformed to a
log probability ratio

rscore
.00001
10
20
30
40
50
60
70
80
90
99.9999

lgpscor
-7
-.95
-.6
-.37
-.18
0
.18
.37
.6
.95
7

The different effect of the two transformations can be seen below. Clearly the logbased transformation has a much more drastic stretching effect near the ends of the
scale than the arcsine transformation does.