You are on page 1of 23

6.3.

2 Analysis of Factorial Designs for a Randomized Block Design (RBD)

 In this section, we will be considering the analysis of a two-factor factorial design (call them factor A and factor B)
in a randomized block design.
 There are levels of factor A and levels of factor B.
 Furthermore, suppose we have blocks.
 In this design, observations are collected for all the factor level combinations/treatments within each of the blocks.
 The layout of the data is similar to that of a CRD factorial design. However, it is repeated for each block.
 We will only be considering this design with 1 observation per treatment per block.
 is the response to the level of factor A and the level of factor B within the block.
 and
Block Block Block
Factor B Factor B Factor B
… … …

… … …

… … …
Factor A
… … …

… … …

 This layout of the data can sometimes require a lot of space, and thus it is usual to represent the data above in a more
condensed form…
Block
Treatment


Some more notation…
𝑏 𝑟
𝑌 =
o 𝑖.. ∑ ∑ 𝑌 𝑖𝑗𝑘is the total of the responses for the level of factor A.
𝑗=1 𝑘=1

𝑎 𝑟

o 𝑌 . 𝑗. = ∑ ∑ 𝑌 𝑖𝑗𝑘is the total of the responses for the level of factor B.


𝑖=1 𝑘 =1

𝑎 𝑏

o 𝑌 .. 𝑘= ∑ ∑ 𝑌 𝑖𝑗𝑘is the total of the responses for the block.


𝑖=1 𝑗=1

o𝑌 𝑖𝑗. = ∑ 𝑌 𝑖𝑗𝑘is the total of the observations for the treatment corresponding to level of factor A
𝑘 =1 and level of factor B.

𝑎 𝑏 𝑟

o𝑌 .. . = ∑ ∑ ∑ 𝑌 𝑖𝑗𝑘is the grand total of all the observations.


𝑖=1 𝑗=1 𝑘=1
Block Treatment
Treatment
… Total

Block Total … Grand Total


Example:
An experiment was set up to compare the effect of soil ph (factor A with levels 4, 5, 6, 7} and calcium additives (factor
B with levels 100, 200, 300) on the increase in diameters for orange trees. All factor level combinations of these two
factors were used at each of Limpopo, Mpumalanga and KZN Research Stations. Thus, the experiment was a 4 x 3
factorial design in a randomized block design – with research stations as the blocks.
At the end of a two-year period, the diameter of a randomly chosen tree at each station and factor level combination
was measured.

 Since location is a blocking factor, we will assume that there are no interactions between the blocks and ph level, and
between the blocks and calcium levels.

 There are treatments in total.


Block
Treatment
Treatment (A,B)
Mpumalanga Limpopo KZN Total

(4, 100) 5.2 5.9 6.3 17.4


(4, 200) 7.4 7.0 7.6 22.0
(4, 300) 6.3 6.7 6.1 19.1
(5, 100) 7.1 7.4 7.5 22.0
(5, 200) 7.4 7.3 7.1 21.8
(5, 300) 7.3 7.5 7.2 22.0
(6, 100) 7.6 7.2 7.4 22.2
(6, 200) 7.6 7.5 7.8 22.9
(6, 300) 7.2 7.3 7.0 21.5
(7, 100) 7.2 7.5 7.2 21.9
(7, 200) 7.4 7.0 6.9 21.3
(7, 300) 6.8 6.6 6.4 19.8

Block Total 84.5 84.9 84.5 253.9


The Statistical Model
The statistical model for a two-factor factorial design in a RBD (a fixed effects model) has the following form:

𝑌 𝑖𝑗𝑘=¿𝜇 + 𝜏 ¿ 𝑖 +¿𝛽 𝑗 + ¿𝜌 𝑘 +¿𝜏𝛽𝑖𝑗 +¿𝜖 𝑖𝑗𝑘


¿ 𝜇 𝑖𝑗 . + 𝜌 𝑘 +𝜖 𝑖𝑗𝑘
Where:
 is the usual overall population mean;

 is the effect of the level of factor A (also known as the main effect of factor A);
 is the effect of the level of factor B (also known as the main effect of factor B);

 is the effect of the block;

 is the interaction effect of the level of factor A and level of factor B;

 is the mean of the combination of the level of factor A and the level of factor B, i.e., the treatment mean;

 The are the random error terms which are assumed to be independently and normally distributed with a mean of 0 and
constant variance .
Since this is a fixed effects model, the usual assumptions apply:

 The parameters , and are fixed constants;


𝑎 𝑏 𝑟 𝑎 𝑏
 ∑ 𝜏 𝑖 =∑ 𝛽 𝑗 =∑ 𝜌 𝑘=∑ ( 𝜏𝛽 )𝑖𝑗= ∑ ( 𝜏𝛽 )𝑖𝑗=0
𝑖=1 𝑗 =1 𝑘=1 𝑖=1 𝑗=1

 The same hypothesis tests are carried out for this RBD factorial design as those for a CRD factorial design in the
previous section.
 However, we can now carry out an additional hypothesis test concerning blocking:

: Blocking was not effective

: Blocking was effective

or
ANOVA
The total sum of squares is calculated as follows:
𝑎 𝑏 𝑟
1 2
SSTO¿ ∑ ∑ ∑ 𝑌 − 𝑎𝑏𝑟 𝑌 . ..
2
𝑖𝑗𝑘
df
𝑖=1 𝑗=1 𝑘=1

SSTO is decomposed into 3 sources of variation now:

𝑎 𝑏
1 2 1 2
 ¿ ∑ ∑ 𝑌 𝑖 𝑗. −
SST 𝑌... df
𝑖=1 𝑗=1 𝑟 𝑎𝑏𝑟
𝑟
1 2 1
¿
 SSBlock ∑ 𝑎𝑏
𝑌 . .𝑘 −
𝑎𝑏𝑟
𝑌 2. .. df
𝑘=1

 SSE SSTO – SST – SSBlock


df
Just as the case of a factorial design in a CRD, the treatment sum of squares (SST) is decomposed into:
𝑎
1 2 1
SSA∑
2
 ¿ 𝑌 𝑖 .. − 𝑌 .. . df
𝑖=1 𝑏 𝑟 𝑎𝑏𝑟

𝑏
1 2 1 2
¿
 SSB ∑ 𝑎𝑟 𝑌 . 𝑗 . − 𝑎𝑏𝑟 𝑌 . .. df
𝑗=1

 SSAB SST – SSA – SSB df


The ANOVA table for a factorial design in a RBD:

Sum of Mean sum of


Source of Variation df F-statistic
Squares squares

F=
Blocks SSBlock 𝑟 −1 MSBlock

Treatment SST 𝑎𝑏 −1 MST F=

F=
Factor A SSA 𝑎− 1 MSA

F=
Factor B SSB 𝑏 −1 MSB

F=
Interaction AB SSAB (𝑎 −1)(𝑏 −1) MSAB

Error SSE (𝑎 𝑏 −1)(𝑟 −1) MSE

Total SSTO 𝑎𝑏𝑟 −1


Let us go back to the previous example:
An experiment was set up to compare the effect of soil ph (factor A with levels 4, 5, 6, 7} and calcium additives (factor B
with levels 100, 200, 300) on the increase in diameters for orange trees. All factor level combinations of these two factors
were used at each of Limpopo, Mpumalanga and KZN Research Stations.

 Using the data from this example, let us obtain the ANOVA table in the form of that on the previous slide.
 Recall for this example:
Block
Treatment (A,B)
Mpumalanga Limpopo KZN First, calculate SSTO:
(4, 100) 5.2 5.9 6.3
(4, 200) 7.4 7.0 7.6 𝑎 𝑏 𝑟
SSTO ¿ ∑ ∑ ∑ 𝑌 −
1 22
(4, 300) 6.3 6.7 6.1 𝑌 . ..
𝑖𝑗𝑘
𝑖=1 𝑗=1 𝑘=1 𝑎𝑏𝑟
(5, 100) 7.1 7.4 7.5
2 2 2 ( 253.9 )2
(5, 200) 7.4 7.3 7.1 ¿ ( 5.2 ) + ( 5.9 ) + …+ ( 6.4 ) −
36
(5, 300) 7.3 7.5 7.2
(6, 100) 7.6 7.2 7.4
¿ 10.8098
(6, 200) 7.6 7.5 7.8
(6, 300) 7.2 7.3 7.0
(7, 100) 7.2 7.5 7.2
(7, 200) 7.4 7.0 6.9
(7, 300) 6.8 6.6 6.4
Treatment
Total 𝑎 𝑏
1 1
17.4 ¿ ∑ ∑ 𝑌 𝑖 𝑗. −
SST
2 2
𝑌...
𝑖=1 𝑗=1 𝑟 𝑎𝑏𝑟
22.0
19.1
( 17.4 )2 ( 22.0 )2 ( 19.8 )2 ( 253.9 )2
22.0 ¿ + +…+ −
3 3 3 36
21.8
22.0 ¿ 9.18311
22.2
22.9
21.5
21.9
21.3
19.8
Block Total 84.5 84.9 84.5

𝑟
1 2 1
SSBlock SSStation¿ ∑ 𝑎𝑏
𝑌 . .𝑘 −
𝑎𝑏𝑟
𝑌 2. ..
𝑘=1

( 84.5 ) 2 ( 84.9 )2 ( 84.5 )2 ( 253.9 )2


¿ + + −
12 12 12 36

¿ 0.00891

SSE SSTO – SST – SS Station ¿ 10.8098 − 9.18311−0.00891¿ 1.6178

Next we need to calculate the decomposed parts of SST…


Block
Treatment Totals for soil ph factor
Treatment (A,B)
Mpumalanga Limpopo KZN Total levels (factor A):
(4, 100) 5.2 5.9 6.3 17.4
(4, 200) 7.4 7.0 7.6 22.0 58.5
(4, 300) 6.3 6.7 6.1 19.1
(5, 100) 7.1 7.4 7.5 22.0
(5, 200) 7.4 7.3 7.1 21.8 65.8
(5, 300) 7.3 7.5 7.2 22.0
(6, 100) 7.6 7.2 7.4 22.2
(6, 200) 7.6 7.5 7.8 22.9 66.6
(6, 300) 7.2 7.3 7.0 21.5
(7, 100) 7.2 7.5 7.2 21.9
(7, 200) 7.4 7.0 6.9 21.3 63.0
(7, 300) 6.8 6.6 6.4 19.8
Soil ph level
4 5 6 7
58.5 65.8 66.6 63.0

𝑎
1 2 1 2( 58.5 )
2
( 65.8 )2 ( 66.6 )2 ( 63.0 )2 ( 253.9 )2
¿∑
SSA = SSPh 𝑌 𝑖 .. − 𝑌¿ .. . + + + −
𝑖=1 𝑏 𝑟 𝑎𝑏𝑟 9 9 9 9 36
¿ 4.4608
Similarly, calculating the total of the responses for the different levels of calcium additives (factor B):
Calcium additive level

100 200 300

83.5 88.0 82.4


𝑏
1 1 (283.5 ) 2 ( 88.0 )2 ( 8 2 .4 )2 ( 253.9 )2
¿∑
SSB = SSCalcium
2
𝑌. 𝑗 .− ¿𝑌 . .. + + −
𝑗=1 𝑎𝑟 𝑎𝑏𝑟 12 12 12 36
¿ 1.4672
Lastly, we can find the sum of squares for the interaction between soil ph level and calcium additive level:
SSAB SSPh*Calcium SST – SSA – SSB

9.18311 – 4.4608 – 1.4672

3.2551

The degrees of freedom for each sum of squares:

df(SSTO) ¿ 36 − 1=35
df(SST) ¿ 12 −1=11
These 3 sum to df(SSTO) df(SSBlock) ¿ 3 − 1=2
df(SSE) ¿ 22
df(SSPh) ¿ 4 −1=3
df(SSCalcium) ¿ 3 − 1=2 These 3 sum to df(SST)

df(SSPh*Calcium) ¿6
The ANOVA table for this factorial design example of a RBD:

Sum of Mean sum of


Source of Variation df F-statistic
Squares squares

0.00891 F
Station 0.00891 2 =0.004
2
9.18311 F
Treatment 9.18311 11 11
=0.835

4.4608
Ph 4.4608 3 3
=1.487 F

1.4672
Calcium 1.4672 2 2
=0.734 F

3.2551 F
Ph*Calcium 3.2551 6 6
=0 .543

1.6178
1.6178
Error 22 22
=0.074

Total 10.8088 35
Now, we start the ANOVA test by performing a hypothesis test concerning the treatments using a level of
significance of 1%:

Treatment Effects:

The treatments (factor level combinations) have no effect on the response

The treatments (factor level combinations) have a significant effect on the response
or

All the are equal versus At least two differ

The critical value for this test: (from the statistical F-tables).

Degrees of freedom of Degrees of freedom of


MST MSE

Therefore, (from the ANOVA table) since is rejected a 1% l.o.s., therefore we can conclude that the treatments have a
significant effect on the response (or that at least two treatment means differ).
Since on the previous slide was rejected, we can then perform a hypothesis test for the presence of an interaction effect.

Interaction Effects:

All the equal zero (there are no interaction effects)

Not all the equal zero (there are interaction effects)

The critical value for this test: (from the statistical F-tables).

Degrees of freedom of Degrees of freedom of


MSPh*Calcium MSE

Therefore, (from the ANOVA table) since is rejected a 1% l.o.s., therefore we can conclude that there are interaction effects
present between soil ph level and calcium additives.

 Since there ARE interaction effects present between soil ph and calcium, there is now no reason to perform hypothesis
tests concerning their individual main effects.
Was blocking effective?:

: Blocking was not effective

: Blocking was effective

or

The critical value for this test: (from the statistical F-tables).

Degrees of freedom of Degrees of freedom of


MSBlock MSE

Therefore, (from the ANOVA table) since is not rejected a 1% l.o.s., therefore we can conclude the that blocking for the
location of the research station was not effective, and thus it was unnecessary.

You might also like