Transcript Nonparametric Statistics

Nonparametric Statistics
Timothy C. Bates
[email protected]
Parametric Statistics 1
Assume data are drawn from samples
with a certain distribution (usually normal)
 Compute the likelihood that groups are
related/unrelated or same/different given
that underlying model
 t-test, Pearson’s correlation, ANOVA…

Parametric Statistics 2

Assumptions of Parametric statistics
1.
2.
3.
Observations are independent
Your data are normally distributed
Variances are equal across groups
•
Can be modified to cope with unequal ∂2
Non-parametric Statistics?
Non-parametric statistics do not assume
any underlying distribution
 They estimate the distribution AND
compute the probability that your groups
are the related/the same or
unrelated/different

Nonparametric ≠
No parameters
Model structure is not specified a priori
but is instead determined from data.
 The data are parameterised by the
analysis


AKA: “distribution free”
Non-parametric Statistics

Assumptions of non-parametric statistics
1.
Observations are independent
Non-parametric Statistics?
Non-parametric statistics do not assume
any underlying distribution
 Estimating or modeling this distribution
reduces their power to detect effects…


So never use them unless you have to
Why use a Non-parametric
Statistic?

Very small samples (<20 replicates)
High probability of violating the assumption of
normality
 Leads to spurious Type-1 (false alarm) errors

Why use a Non-parametric
Statistic?

Outliers more often lead to spurious Type1 (false alarm) errors in parametric
statistics.

Nonparametric statistics reduce data to
an ordinal rank, which reduces the impact
or leverage of outliers.
Error

Type-I error: False Alarm for a bogus effect


Type-II error: Miss a real effect


reject the null hypothesis when it is really true
fail to reject our null hypothesis when it is really false
Type-III error: :-)

lazy, incompetent, or willful ignorance of the truth
Power

1-alpha
Non-parametric Choices
Data type?
continuous
discrete
χ2
Question?
association
Spearman’s
Rank
Different central
value
Number of
groups?
two-groups
Mann-Whitney U
Wilcoxon’s Rank Sums
Difference in ∂2
BrownForsythe
more than 2
Kruskal-Wallis
test
Non-parametric Choices
Data type?
continuous
discrete
χ2
Question?
Like a
association
Pearson’s R
Spearman’s
Rank
Like
Student’s t
No alternative
Different central
value
Difference in ∂2
Number of
groups?
BrownForsythe
two-groups
Mann-Whitney U
Wilcoxon’s Rank Sums
more than 2
Kruskal-Wallis
test
Like F-test
Like ANOVA
Chi-Squared (Χ2)

χ2 tests the null hypothesis that observed
events occur with an expected frequency


e.g. Ho: “This six-sided dice is fair ”


in large samples frequencies are distributed as Χ2
Expect all 6 outcomes to occur equally often
Assumptions



Observations are independent
Outcomes mutually exclusive
Sample is not small
Small samples require exact test:, i.e., binomial test
Chi-Squared Χ2 formula

Χ2 = the sum of each squared difference
between the observed and expected
frequencies divided its expected
frequency
Χ2 and contingency tables
Χ2 essentially tests if each cell in a
contingency table has its expected value
 In a 2-way table, this expectation will be
the value of an adjacent cell

Example: coin toss
Random sample of 100 coin tosses, of a
coin believed to be fair
 We observed number of 45 heads, and
and 55 tails
 Is the coin fair?

Coin toss

If ho is true, our test statistic is drawn from a Χ2
distribution with df = 1
(45-50)2 + (55-50)2 = 0.5 + 0.5 = 1
50
50
Χ2(1) = 1, p > 0.3
Coin toss Χ2 in R

chisq.test(c(45,55), p=c(.5,.5))
Chi-squared test for given probabilities
 Χ2 = 1, df = 1, p = 0.3173

Spearman Rank test (ρ (rho))


Named after Charles Spearman,
Non-parametric measure of correlation

Assesses how well an arbitrary monotonic
function describes the relationship between
two variables,
Does not require the relationship be linear
 Does not require interval measurement

Spearman Rank test (ρ (rho))

Mathematically, it is simply a Pearson’s r
computed on ranked data



d = difference in rank of a given pair
n = number of pairs
Alternative test = Kendall's Tau (Kendall's τ)
Mann-Whitney U

AKA: “Wilcoxon rank-sum test


Mann & Whitney, 1947; Wilcoxon, 1945
Non-parametric test for difference in the
medians of two independent samples

Assumptions:
• Samples are independent
• Observations can be ranked (ordinal or better)
Mann-Whitney U
U tests the difference in the medians of
two independent samples
 n1 = number of obs in sample 1
 n2 = number of obs in sample 2
 R = sum of ranks of the lower-ranked
sample

Mann-Whitney U or t-test?

Should you use it over the t-test?

Yes if you have a very small sample (<20)
• (central limit assumptions not met)



It is less prone to type-I error


Possibly if your data are inherently ordinal
Otherwise, probably not.
(spurious significance) due to outliers.
But does not in fact handle comparisons of
samples whose variances differ very well

(Use unequal variance t-test with rank data)
Aesop: Mann-Whitney U
Example
Suppose that Aesop is dissatisfied with
his classic experiment in which one
tortoise was found to beat one hare in a
race.
 He decides to carry out a significance test
to discover whether the results could be
extended to tortoises and hares in
general…

Aesop 2: Mann-Whitney U



He collects a sample of 6 tortoises and 6 hares,
and makes them all run his race. The order in
which they reach the finishing post (their rank
order) is as follows:
tort = c(1, 7, 8, 9, 10,11)
hare = c(2, 3, 4, 5, 6, 12)

Original tortoise still goes at warp speed, original
hare is still lazy, but the others run truer to
stereotype.
Aesop 3: Mann-Whitney U
wilcox.test(tort, hare)
 Wilcoxon = W = 25, p-value = 0.31

Tortoises are not faster (but neither are
hares)
 tort = c(1, 7, 8, 9, 10,11) (n2 = 6)
 hare = c(2, 3, 4, 5, 6, 12) (n1 = 6, R1 =32)

Aesop 4: Mann-Whitney U

Wilcoxon = W = 25, p-value = 0.31


Tortoises are not faster (but neither are hares).
Welch Two Sample t-test



t = 1.1355, df = 10, p-value = 0.28
Alternative hypothesis: true difference in means is
not equal to 0
95 percent confidence interval:
-2.25 ~ 6.91

sample estimates:
• mean of x = 7.6 mean of y = 5.3
Power comparison with continuous
normal data
tort = 1 74 79 81 100 121
 hare = 4 9 16 17 18 144
 Wilcoxon

W = 25, p = 0.31

t.test
t.test(tort, hare, var.equal = TRUE)
 t(10) = 1.5, p = 0.16

Wilcoxon signed-rank test
(related samples)
Same idea as MW U, generalized to
matched samples
 Equivalent to non-independent sample ttest

Kruskall-Wallis





Non-parametric one-way analysis of variance
by ranks (named after William Kruskal and W.
Allen Wallis)
tests equality of medians across groups.
It is an extension of the Mann-Whitney U test to
3 or more groups.
Does not assume a normal population,
Assumes population variances among groups
are equal.