Sie sind auf Seite 1von 41

Two-way ANOVA

Introduction to Factorial
Designs
and their Analysis

Two-Way ANOVA Data Layout


Factor
A
1
2
:
a

1
X111
X11n
X211
X21n
:
Xa11
Xa1n

Factor B
2
...
X121
...
X12n
...
X221
...
X22n
...
:
:
Xa21 ...
Xa2n ...

There are

b
X1b1
X1bn
X2b1
X2bn
:
Xab1
Xabn

a X b treatment

Observation k
in each cell

Xijk
Level i
Factor
A

Level j
Factor
B

i = 1,,a
j = 1,
,b

Motivating Example:
Capsule Dissolve Time
Suppose are looking at two capsule types
(C or V) & two digestive fluids (Gastric
or Duodenal)
Randomly assign 5
capsules of each
type to each of
type of digestive
juice and observe
dissolve time.
Xijk = measured
dissolve time for
capsule k in
digestive juice i
and capsule type j.
i = 1 or 2 (i.e. G
and D)
j = 1 or 2 (i.e. C

Questions of Interest
What effect does capsule type
have on the time to dissolve?
What effect does fluid type
have on the time to dissolve?
Do both capsule types dissolve
in the same manner in the
two different fluid types?

Time Until Bubbles


(seconds)

Plotting the Results ~ Capsule Effect

Capsule
Type

X 1 mean time for type C capsules to dissolve 43.05 seconds


X 2 mean time for type V capsules to dissolve 42.85 seconds

There appears to be very little difference


between the capsule types in terms of the time

Plotting the Results ~ Fluid Effect

Fluid Type

X 1 mean time for capsules to dissolve in gastric juice 45.7 seconds


X 2 mean time for capsules to dissolve in duodenal juice 40.2 seconds

Capsules take 5.5 seconds longer on average to


dissolve in gastric juice compared to duodenal
juice.

Preliminary Conclusion
There is very little difference between the
capsule types in terms of the length it time it
takes them to dissolve.
Capsules take about 5 seconds longer on
average to dissolve in gastric juice than in
duodenal juice.
THESE CONCLUSIONS ARE
COMPLETELY WRONG!!
WHY ?!?

Plotting the Results ~ Capsule Effect Separately


Type C capsules dissolve faster in
duodenal juice than do type V
capsules where for gastric juice
the opposite is true.

X 11 mean time for C capsules to dissolve in gastric juice 49.8 seconds


X 12 mean time for V capsules to dissolve in gastric juice 41.6 seconds
X 21 mean time for C capsules to dissolve in duodenal juice 36.3 seconds
X 22 mean time for V capsules to dissolve in duodenal juice 44.1 seconds

Clearly the time to dissolve depends on what


capsule is being used and which juice it is

Interactions
The capsule study is an example
of
Type C capsules
dissolve
faster than
situation where there is an
interaction
V studied
in duodenal
between the two factors Type
being
in
juice, where
terms of their effect on the
numeric
opposite
is true
response.
when gastric juice
is used to dissolve
the capsules.

An interaction occurs when the effect of


one factor depends on the level of
another factor. Here the effect of
capsule depends on the type of digestive
juice used to dissolve it and vise versa.

Interactions can mask


main effects

We say the interaction


masks the main effect
of capsule type.

The apparent
lack of a capsule
effect is caused
by the
interaction of
capsule type
and fluid type.

Types of Interactions and


Interpreting Interaction Plots

Here the mean


response is the
same for both
levels of both
factors.

Here both effects are


masked by the
interaction. This type
of interaction is called
a difference in
direction of the

Types of Interactions and


Interpreting Interaction Plots

Here the mean


response differs
depending on the
level of B but not
A.

Here the A main effect is


masked by the
interaction. The B main
effect is significant,
although cannot be
talked about

Types of Interactions and


Interpreting Interaction Plots

Here the effect of A is


the same for both
levels of B. There is
minimal separation
between the two
profiles for the levels
of B, thus B is not

Here the A main effect is


differs depending on the
level of B. The B main
effect is masked by the
interaction as the means
for B1 and B2 are the

Types of Interactions and


Interpreting Interaction Plots
This type of interaction is
a difference in
magnitude the effect.
The direction of A main
effect is the same for
both levels of B, however
the A effect is larger
Here the effect of A
is
Here the A main
st
when
B isfor
at both
the 1levels
level. effect is differs
the same
of B and vise versa. The
response differs across
the level of both factors
and both differences
suggest significant A &

depending on the
level of B. Neither
the A or B main
effects are masked by
the interaction.

Types of Interactions
In summary there are types of
interactions:
Differences in Direction
Differences
in Magnitude

Always construct an interaction plot to


visualize the interaction or lack thereof !

Questions of Interest
Generally, the questions of interest here
(i.e. hypotheses to be tested) concern
three questions regarding the potential
effects of the factors on the response
variable.
Question 1: Do the effects that factors A
and B have on the response variable
interact, i.e. is there a significant
interaction between factors A and B ?

Questions of Interest
If we conclude there is a significant
interaction then we conclude the
effects of both factors
A and B are significant!
When we have an interaction we cannot
consider the effect of either factor
independently of the other, therefore
both factors matter.

Questions of Interest
If there is not a significant
interaction effect then we can
consider the main effects
separately, i.e. we ask the
following:
Question 2: Does factor A alone have
a significant effect?
Question 3: Does factor B alone have
a significant effect?

Tests of Hypotheses
Just as we had Sums of Squares and
Mean Squares in One-way ANOVA, we
have the same in Two-way ANOVA:
Recall, Mean Squares are measures of
variability across the levels of the relevant
factor of interest.
In balanced Two-way ANOVA, we measure
the overall variability in the data by:
a

SS T ( X ijk X ) 2
i 1 j 1 k 1

df N 1

Tests of Hypotheses
Sum of Squares for factor A
a

SS A ( X i X ) bn ( X i X ) 2
2

i 1 j 1 k 1

df a 1

i 1

Measures variation in the response due


to the fact that different levels of factor
ASum
were
ofused.
Squares for factor B
a

SS B ( X j X ) an ( X j X ) 2 df b 1
i 1 j 1 k 1

j 1

Measures variation in the response due


to the fact that different levels of factor
B were used.

Test of Hypotheses
Interaction Sum of Squares
a

SS AB ( X ij X i X j X ) 2

df (a 1)(b 1)

i 1 j 1 k 1

Measures the variation in the response due


to the interaction between factors A and B.
If the interaction plot is perfectly parallel
this will be 0!
Error or Residual Sum of Squares
a

SS E ( X ijk X ij ) 2

df ab(n 1)

i 1 j 1 k 1

Measures the variation in the response


within the a x b factor combinations.

Tests of Hypotheses
So the Two-way ANOVA Identity is:

SS T SS A SS B SS AB SS E
This partitions the Total Sum of
Squares into four pieces of
interest for our hypotheses to be
tested.

Tests of Hypotheses
As in One-way ANOVA, we obtain mean
squares for the different effects by dividing
the sums of squares by their respective
degrees of freedom
SS effect
i.e. MS effect
df effect
These are our measures of variance for the
analysis.
If an effect is not
we expect
MSsignificant
effect MS E
and if it is we expect

MS effect MS E

Test of Hypotheses
F-Statistic for Testing an Effect
Fo

MS effect
MS E

~ F distribution

Numerator df =
dfeffect

Denominator df =
If the F-statistic is large we reject that the df
effect
is zero in
error
favor of the alternative that the effect of the factor is non-zero.

Two-way ANOVA Table


Source of Degrees of
Variation Freedom

Sum of
Squares

Mean
Square F-ratio

P-value

Factor A

a1

SSA

MSA

FA = MSA / MSE

Tail area

Factor B

b1

SSB

MSB

FB = MSB / MSE

Tail area

Interaction

(a 1)(b 1)

SSAB

MSAB

FAB = MSAB / MSE

Tail area

Error

ab(n 1)

SSE

MSE

Total

abn 1

SST

This is our initial


focus which is the pvalue for Question
1: Is there an
interaction effect?

Tests of Hypotheses
If the interaction is not statistically significant
(i.e. p-value > 0.05) then we conclude the
main effects (if present) are independent of
one another.
We can then test for significance of the main
effects separately, again using an F-test.
If a main effect is significant we can then use
multiple comparison procedures as usual to
compare the mean response for different
levels of the factor while holding the other
factor fixed.

Tests of Hypotheses
If an interaction is significant (p-value < .05) we
conclude the main effects are not independent of
one another and that both effects are important!
In this case (i.e. the interaction is significant) the tests
for main effects in the Two-way ANOVA table are
MEANINGLESS!

We must compare levels of factor A


within each level of factor B (and
vise versa).

Example 1: Capsule
Dissolve Time

Enter the n = 5
replicates for each
treatment combination:
Gastric, C
Gastric, V
Duodenal, C
Duodenal, V

Example 1: Capsule
Dissolve Time

1st Highlight
both factors
in this list.
Next
highlight Full
Factorial from
the Macros
pull-down
menu.

Then click Run


Model leaving
everything else
unchanged.

Example 1: Capsule
Dissolve Time

Lots of extra CRAP we dont


need. Turn off the plots as they
are unnecessary when
considering two-way ANOVA.
Also we really only need to
consider the Effect Tests portion
of the numeric output initially.

Example 1: Capsule
Dissolve Time

Interaction effect is statistically significant


(p = .0049). Therefore we conclude that
both the capsule and fluid type effects are
significant, however we cannot talk about
their effects in term of the mean time until
bubbles are observed independently. We can
compare capsule effects for the same fluid

Capsule x Fluid Type


Interaction
Because the interaction is
statistically significant we are
interested in comparing fluid types
Treatment
for a given capsule type or
comparing capsules for a combination
given fluid
means.
type.
The table on the right contains the
results of all pair-wise treatment
mean comparisons, however we are
only interested in those as
described above.
Interaction
Here we find that there is a
Plot
significant difference in the fluid
types for the type C capsules
however there are no significant
differences between the capsules
themselves for given fluid type, nor
is there a fluid effect when
dissolving type V capsules.

Checking Assumptions
To check the assumptions of normality
of the response and equality of
variance for the difference treatment
combinations we can examine the
residuals. For a two-way ANOVA
the residuals are the deviations of
the observations from their
respective etreatment
combination

x
ijk
ijk
ij
sample means, i.e.

Checking Assumptions
To check the assumption of normality,
xijkresiduals
xij
assess the normalityeof
the
ijk
The residuals
from the capsule
experiment look
approximately
normal with the
exception of two
outliers, but
neither are
extreme enough
to warrant any

Checking Assumptions
To check the equality of variance for
the difference treatment
combinations we can examine the
xijk xij
residuals plotted vs. theeijk
different
treatment combination
means
There appears to be more
variation for the dissolve times
for type C capsules being
dissolved in gastric fluid. These
combination produced the two
mild outliers seen in the normal
quantile plot. Generally we
worry when the variation
increases with the treatment
combination mean.

Example 2: Comparing the Effectiveness of


Three Forms of
Psychotherapy for Alleviating Depression

Suppose that a clinical psychologist is interested


in comparing the relative effectiveness of three
forms of psychotherapy for alleviating depression.
Fifteen individuals are randomly assigned to each
of three treatment groups: cognitive-behavioral,
Rogerian, and assertiveness training. The
Depression Scale of MMPI serves as the response.
The psychologist also wished to incorporate
information about the patients severity of
depression, so all subjects in the study were
classified as having mild, moderate, or severe
depression. Thus we have two factor of interest in
this study: the treatment they received and the
initial severity of their depression.

Example 2: Comparing the Effectiveness of


Three Forms of
Psychotherapy for Alleviating Depression
Interaction Plot

Therapy Effect Plot

Degree of Severity
Effect Plot

Example 2: Comparing the Effectiveness of


Three Forms of
Psychotherapy for Alleviating Depression

Is there a significant interaction effect ? NO, p = .9717


Is there a significant therapy effect ? YES, p = .0356
Is there a significant degree of severity of effect ? YES, p < .0001

Now we can conduct multiple comparisons on each


factor separately.

Example 2: Comparing the Effectiveness of


Three Forms of
Psychotherapy for Alleviating Depression

We see that Rogerian


therapy differs
significantly from
Cognitive-Behavioral
therapy, with
Rogerian having
larger mean by
between .8 and 9.87
units.
We see that the initial
severity of
depression levels significantly differ
from each other in terms of mean
depression score. In particular we
see that those with a severe
classification have a mean
depression score exceeding that for
those with mild depression by
between 8 and 18 points and those
with moderate depression by

Example 2: Comparing the Effectiveness of


Three Forms of
Psychotherapy for Alleviating Depression

Residuals look
approximately
normal.

Residuals
indicate constant
variation within
each treatment
combination.

Summary
These ideas can be extended to
more than
two factors.
When interactions exist, the main
effects are involved are important,
but cannot discussed separately.
Multiple comparisons can still be
conducted

Das könnte Ihnen auch gefallen