Transcript The Assumptions of ANOVA - ROHAN Academic Computing

The Assumptions of
ANOVA
Dennis Monday
Gary Klein
Sunmi Lee
May 10, 2005
Major Assumptions of Analysis of
Variance
• The Assumptions
– Independence
– Normally distributed
– Homogeneity of variances
• Our Purpose
– Examine these assumptions
– Provide various tests for these assumptions
• Theory
• Sample SAS code (SAS, Version 8.2)
– Consequences when these assumptions are not met
– Remedial measures
Normality
• Why normal?
– ANOVA is an Analysis of Variance
– Analysis of two variances, more specifically, the ratio of
two variances
– Statistical inference is based on the F distribution
which is given by the ratio of two chi-squared
distributions
– No surprise that each variance in the ANOVA ratio come
from a parent normal distribution
• Calculations can always be derived no matter
what the distribution is. Calculations are
algebraic properties separating sums of squares.
Normality is only needed for statistical inference.
Normality
Tests
• Wide variety of tests we can perform to test if the
data follows a normal distribution.
• Mardia (1980) provides an extensive list for both
the univariate and multivariate cases,
categorizing them into two types
– Properties of normal distribution, more specifically, the
first four moments of the normal distribution
• Shapiro-Wilk’s W (compares the ratio of the standard
deviation to the variance multiplied by a constant to one)
– Goodness-of-fit tests,
• Kolmogorov-Smirnov D
• Cramer-von Mises W2
• Anderson-Darling A2
Normality
Tests
proc univariate data=temp normal plot;
var expvar;
run;
proc univariate data=temp normal plot;
var normvar;
run;
Tests for Normality
Tests for Normality
Test
--Statistic---
-----p Value------
Test
--Statistic---
-----p Value------
Shapiro-Wilk
Kolmogorov-Smirnov
Cramer-von Mises
Anderson-Darling
W
D
W-Sq
A-Sq
Pr
Pr
Pr
Pr
Shapiro-Wilk
Kolmogorov-Smirnov
Cramer-von Mises
Anderson-Darling
W
D
W-Sq
A-Sq
Pr
Pr
Pr
Pr
0.731203
0.206069
1.391667
7.797847
<
>
>
>
W
D
W-Sq
A-Sq
<0.0001
<0.0100
<0.0050
<0.0050
0.989846
0.057951
0.03225
0.224264
<
>
>
>
W
D
W-Sq
A-Sq
0.6521
>0.1500
>0.2500
>0.2500
Normal Probability Plot
8.25+
|
*
|
|
|
*
|
|
*
|
+
4.25+
**
|
++++
** +++
|
*+++
|
+++*
|
++****
|
++++ **
|
++++*****
|
++******
0.25+*
* ******************
+----+----+----+----+----+----+----+----+----+----+
Stem
8
7
7
6
6
5
5
4
4
3
3
2
2
1
1
0
0
Leaf
0
#
1
Boxplot
*
1
1
*
2
5
4
588
3
59
00112234
56688
00011122223444
55555566667777778999999
000011111111111112222222233333334444444
----+----+----+----+----+----+----+----
1
1
1
3
1
2
8
5
14
23
39
*
0
0
0
0
|
|
|
+--+--+
*-----*
+-----+
Normal Probability Plot
2.3+
++ *
|
++*
|
+**
|
+**
|
****
|
***
|
**+
|
**
|
***
|
**+
|
***
0.1+
***
|
**
|
***
|
***
|
**
|
+***
|
+**
|
+**
|
****
|
++
|
+*
-2.1+*++
+----+----+----+----+----+----+----+----+----+----+
-2
-1
0
+1
+2
Stem
22
20
18
16
14
12
10
8
6
4
2
0
-0
-2
-4
-6
-8
-10
-12
-14
-16
-18
-20
Leaf
1
7
90
047
6779
469002
2368
005546
228880077
5233446
3458447
366904459
52871
884318651
98619
60
98557220
963
584
853
0
4
8
----+----+----+----+
Multiply Stem.Leaf by 10**-1
#
1
1
2
3
4
6
4
6
9
7
7
9
5
9
5
2
8
3
3
3
1
1
1
Boxplot
|
|
|
|
|
|
|
+-----+
|
|
|
|
*-----*
| + |
|
|
|
|
+-----+
|
|
|
|
|
|
|
|
Consequences of Non-Normality
• F-test is very robust against non-normal data,
especially in a fixed-effects model
• Large sample size will approximate normality by
Central Limit Theorem (recommended sample
size > 50)
• Simulations have shown unequal sample sizes
between treatment groups magnify any departure
from normality
• A large deviation from normality leads to
hypothesis test conclusions that are too liberal
and a decrease in power and efficiency
Remedial Measures for NonNormality
• Data transformation
• Be aware - transformations may lead to a
fundamental change in the relationship between
the dependent and the independent variable and
is not always recommended.
• Don’t use the standard F-test.
– Modified F-tests
• Adjust the degrees of freedom
• Rank F-test (capitalizes the F-tests robustness)
– Randomization test on the F-ratio
– Other non-parametric test if distribution is unknown
– Make up our own test using a likelihood ratio if
distribution is known
Independence
• Independent observations
– No correlation between error terms
– No correlation between independent variables and error
• Positively correlated data inflates
standard error
– The estimation of the treatment means are more
accurate than the standard error shows.
Independence Tests
•
If we have some notion of how the data was
collected, we can check if there exists any
autocorrelation.
• The Durbin-Watson statistic looks at the
correlation of each value and the value before it
– Data must be sorted in correct order for meaningful
results
– For example, samples collected at the same time would
be ordered by time if we suspect results could depend
on time
Independence Tests
proc glm data=temp;
class trt;
model y = trt / p;
output out=out_ds r=resid_var;
run;
quit;
proc glm data=temp;
class trt;
model y = trt / p;
output out=out_ds r=resid_var;
run;
quit;
data out_ds;
set out_ds;
time = _n_;
run;
proc gplot data=out_ds;
plot resid_var * time;
run;
quit;
data out_ds;
set out_ds;
time = _n_;
run;
proc gplot data=out_ds;
plot resid_var * time;
run;
quit;
First Order Autocorrelation
0.90931
Durbin-Watson D
0.12405
First Order Autocorrelation
0.00479029
Durbin-Watson D
1.96904290
Remedial Measures for Dependent
Data
• First defense against dependent data is proper
study design and randomization
– Designs could be implemented that takes correlation
into account, e.g., crossover design
• Look for environmental factors unaccounted for
– Add covariates to the model if they are causing
correlation, e.g., quantified learning curves
• If no underlying factors can be found attributed to
the autocorrelation
– Use a different model, e.g., random effects model
– Transform the independent variables using the
correlation coefficient
Homogeneity of Variances
• Eisenhart (1947) describes the problem of
unequal variances as follows
– the ANOVA model is based on the proportion of the
mean squares of the factors and the residual mean
squares
– The residual mean square is the unbiased estimator of
2, the variance of a single observation
– The between treatment mean squares takes into account
not only the differences between observations, 2, just
like the residual mean squares, but also the variance
between treatments
– If there was non-constant variance among treatments,
we can replace the residual mean square with some
overall variance,  a2, and a treatment variance,  t2,
which is some weighted version of  a2
– The “neatness” of ANOVA is lost
Homogeneity of Variances
• The omnibus (overall) F-test is very robust
against heterogeneity of variances,
especially with fixed effects and equal
sample sizes.
• Tests for treatment differences like t-tests
and contrasts are severely affected,
resulting in inferences that may be too
liberal or conservative.
Tests for Homogeneity of Variances
– Levene’s Test
• computes a one-way-anova on the absolute value (or
sometimes the square) of the residuals, |yij – ŷi| with t-1, N –
t degrees of freedom
• Considered robust to departures of normality, but too
conservative
– Brown-Forsythe Test
• a slight modification of Levene’s test, where the median is
substituted for the mean (Kuehl (2000) refers to it as the
Levene (med) Test)
– The Fmax Test
• Proportion of the largest variance of the treatment groups
to the smallest and compares it to a critical value table
• Tabachnik and Fidell (2001) use the Fmax ratio more as a
rule of thumb rather than using a table of critical values.
– Fmax ratio is no greater than 10
– Sample sizes of groups are approximately equal (ratio of
smallest to largest is no greater than 4)
• No matter how the Fmax test is used, normality must be
assumed.
Tests for Homogeneity of Variances
proc glm
class
model
means
run;
quit;
data=temp;
trt;
y = trt;
trt / hovtest=levene hovtest=bf;
Homogeneous Variances
The GLM Procedure
proc glm
class
model
means
run;
quit;
data=temp;
trt;
y = trt;
trt / hovtest=levene hovtest=bf;
Heterogenous Variances
The GLM Procedure
Levene's Test for Homogeneity of Y Variance
ANOVA of Squared Deviations from Group Means
Source
DF
Sum of
Squares
Mean
Square
TRT
Error
1
98
10.2533
1663.5
10.2533
16.9747
Levene's Test for Homogeneity of y Variance
ANOVA of Squared Deviations from Group Means
F Value
Pr > F
Source
DF
Sum of
Squares
Mean
Square
0.60
0.4389
trt
Error
1
98
10459.1
27921.5
10459.1
284.9
Brown and Forsythe's Test for Homogeneity of Y Variance
ANOVA of Absolute Deviations from Group Medians
Source
DF
Sum of
Squares
Mean
Square
TRT
Error
1
98
0.7087
124.6
0.7087
1.2710
F Value
Pr > F
36.71
<.0001
Brown and Forsythe's Test for Homogeneity of y Variance
ANOVA of Absolute Deviations from Group Medians
F Value
Pr > F
Source
DF
Sum of
Squares
Mean
Square
0.56
0.4570
trt
Error
1
98
318.3
333.8
318.3
3.4065
F Value
Pr > F
93.45
<.0001
Tests for Homogeneity of Variances
• SAS (as far as I know) does not have a procedure
to obtain Fmax (but easy to calculate)
• More importantly:
VARIANCE TESTS ARE ONLY FOR ONE-WAY
ANOVA
WARNING: Homogeneity of variance testing and Welch's
ANOVA are only available for unweighted one-way
models.
Tests for Homogeneity of Variances
(Randomized Complete Block Design and/or
Factorial Design)
• In a CRD, the variance of each treatment
group is checked for homogeneity
• In factorial/RCBD, each cell’s variance
should be checked
H0: σij2 = σi’j’2, For all i,j where i ≠ i’, j ≠ j’
Tests for Homogeneity of Variances
(Randomized Complete Block Design and/or
Factorial Design)
•
•
Approach 1
–
–
Code each row/column to its own
group
Run HOVTESTS as before
Approach 2
–
–
–
data newgroup;
set oldgroup;
if block = 1 and treat = 1 then newgroup
if block = 1 and treat = 2 then newgroup
if block = 2 and treat = 1 then newgroup
if block = 2 and treat = 2 then newgroup
if block = 3 and treat = 1 then newgroup
if block = 3 and treat = 2 then newgroup
Recall Levene’s Test and BrownForsythe Test are ANOVAs based on
residuals
Find residual for each observation
Run ANOVA
proc sort data=oldgroup; by treat block; run;
= 1;
= 2;
= 3;
= 4;
= 5;
= 6;
run;
proc glm data=newgroup;
class newgroup;
model y = newgroup;
means newgroup / hovtest=levene hovtest=bf;
run;
quit;
proc means data=oldgroup noprint; by treat block;
var y;
output out=stats mean=mean median=median;
run;
data newgroup;
merge oldgroup stats;
by treat block;
resid = abs(mean - y);
if block = 1 and treat = 1 then newgroup = 1;
………
run;
proc glm data=newgroup;
class newgroup;
model resid = newgroup;
run; quit;
Tests for Homogeneity of Variances
(Repeated-Measures Design)
• Recall the repeated-measures set-up:
Treatment
a1
a2
a3
s1
s1
s1
s2
s2
s2
s3
s3
s3
s4
s4
s4
Tests for Homogeneity of Variances
(Repeated-Measures Design)
• As there is only one score per cell, the variance
of each cell cannot be computed. Instead, four
assumptions need to be tested/satisfied
– Compound Symmetry
• Homogeneity of variance in each column
– σa12 = σa22 = σa32
• Homogeneity of covariance between columns
– σa1a2 = σa2a3 = σa3a1
– No A x S Interaction (Additivity)
– Sphericity
• Variance of difference scores between pairs are equal
– σYa1-Ya2 = σYa1-Ya3 = σYa2-Ya3
Tests for Homogeneity of Variances
(Repeated-Measures Design)
• Usually, testing sphericity will suffice
• Sphericity can be tested using the Mauchly test in
SAS
proc glm data=temp;
class sub;
model a1 a2 a3 = sub / nouni;
repeated as 3 (1 2 3) polynomial / summary printe;
run; quit;
Sphericity Tests
Variables
Transformed Variates
Orthogonal Components
DF
2
2
Mauchly's
Criterion
Det = 0
Det = 0
Chi-Square
6.01
6.03
Pr > ChiSq
.056
.062
Tests for Homogeneity of Variances
(Latin-Squares/Split-Plot Design)
• If there is only one score per cell, homogeneity of
variances needs to be shown for the marginals of
each column and each row
– Each factor for a latin-square
– Whole plots and subplots for split-plot
• If there are repititions, homogeneity is to be
shown within each cell like RCBD
• If there are repeated-measures, follow guidelines
for sphericity, compound symmetry and additivity
as well
Remedial Measures for
Heterogeneous Variances
• Studies that do not involve repeated measures
– If normality is not violated, a weighted ANOVA is suggested
(e.g., Welch’s ANOVA)
– If normality is violated, the data transformation necessary to
normalize data will usually stabilize variances as well
– If variances are still not homogeneous, non-ANOVA tests
might be your option
•
Studies with repeated measures
– For violations of sphericity
• modify the degrees of freedom have been suggested.
– Greenhouse-Geisser
– Huynh and Feldt
• Only do specific comparisons (sphericity does not apply since
only two groups – sphericity implies more than two)
• MANOVA
• Use an MLE procedure to specify variance-covariance matrix
Other Concerns
• Outliers and influential points
– Data should always be checked for influential
points that might bias statistical inference
• Use scatterplots of residuals
• Statistical tests using regression to detect outliers
– DFBETAS
– Cook’s D
References
•
Casella, G. and Berger, R. (2002). Statistical Inference. United States: Duxbury.
•
Cochran, W. G. (1947). Some Consequences When the Assumptions for the Analysis of
Variances are not Satisfied. Biometrics. Vol. 3, 22-38.
•
Eisenhart, C. (1947). The Assumptions Underlying the Analysis of Variance. Biometrics.
Vol. 3, 1-21.
•
Ito, P. K. (1980). Robustness of ANOVA and MANOVA Test Procedures. Handbook of
Statistics 1: Analysis of Variance (P. R. Krishnaiah, ed.), 199-236. Amsterdam: NorthHolland.
•
Kaskey, G., et al. (1980). Transformations to Normality. Handbook of Statistics 1: Analysis
of Variance (P. R. Krishnaiah, ed.), 321-341. Amsterdam: North-Holland.
•
Kuehl, R. (2000). Design of Experiments: Statistical Principles of Research Design and
Analysis, 2nd edition. United States: Duxbury.
•
Kutner, M. H., et al. (2005). Applied Linear Statistical Models, 5th edition. New York:
McGraw-Hill.
•
Mardia, K. V. (1980). Tests of Univariate and Multivariate Normality. Handbook of Statistics
1: Analysis of Variance (P. R. Krishnaiah, ed.), 279-320. Amsterdam: North-Holland.
•
Tabachnik, B. and Fidell, L. (2001). Computer-Assisted Research Design and Analysis.
Boston: Allyn & Bacon.