You are on page 1of 18

Subject: Statistics

Paper: Design of Experiments and Sample Survey


Module: Block Design-II
Module No: DOE-7

DOE and SS -7 1 /18


Development Team

Principal Dr. Bhaswati Ganguli,


investigator: Professor, Department of Statistics, University of Calcutta

Paper Dr. Bikas Kumar Sinha,


co-ordinator: Retired Professor, Indian Statistical Institute, Kolkata

Content writer: Dr. Santu Ghosh,


Lecturer, Department of Environmental Health Engineering,
Sri Ramachandra University, Chennai

Content reviewer: Dr. Sugata SenRoy,


Professor, Department of Statistics, University of Calcutta

DOE and SS -7 2 /18


Disclaimer

I take this opportunity to thank Dr Bhaswati Ganguly of the Department of Statistics, Calcutta University for
approaching me with a specific purpose. In this UGC initiated epathsala programme, for the subject of
’Statistics’, in her capacity as the Principal Investigator, she wanted me to act as a coordinator for the topic :
Design of Experiments [DoE] and Sample Survey [SS]. I gladly accepted her proposal and volunteered to
prepare all the 40 Modules as asked for. I have followed a distinctive style while preparing the modules viz., as
that of a dialogue between an Instructor [Professor Bikram Kanti Sahay(BKS)] and his two students [Ms.
Sagarika Ghosh(SG) and Mr. Subhra Sankar Gupta(SSG)]. I fondly hope this instructional discourse and my
efforts on two of my favorite topics in Statistics will be appreciated and found useful.
In the video recordings, I will impersonate as BKS. Mr. Samopriya Basu [MSc (Statistics), Calcutta
University] and Ms. Moumita Chatterjee [University of Calcutta, Kolkata] will impersonate as the students
[SSG and SG] respectively.

Professor Bikas k Sinha


Retired Professor of Statistics
Indian Statistical Institute
Kolkata
July 10, 2015

DOE and SS -7 3 /18


Analysis of Block Design

BKS
Let’s discuss about data analysis now in the framework of a general block
design......
Under normality of errors, standard least squares theory applies for
derivation of the MLEs of the model parameters.....

SG
Sir, I would like to try to deduce these routine results. Once I tried but gave up.
It involves some algebra and this time, I must be slow and steady !

DOE and SS -7 4 /18


Normal Equation

The Least Squares Estimates of the parameters are derived from the NE’s. We
obtained
X X
µ = y.. − kj βj /n − ri τi /n (1)
j i
X
(µ + βj ) = y.j − nij τi /kj , 1 ≤ j ≤ b & (2)
i
X
Cτ = Q, C = ((Cii 0 )), Cii 0 = ri δii 0 − nij ni 0 j /kj (3)
j
X
Q = (Q1 , Q2 , · · · , Qv )0 , Qi = Ti − nij Bj /kj (4)
j

Ti (Bj ) being the total yield under the i th treatment(resp. the j th block).

DOE and SS -7 5 /18


C-Matrix

BKS
Very good..... you have come to a stage wherefrom I can take up.....

The C-matrix in (3) is central to the analysis of data as also to our


understanding about the efficiency of the underlying design.
In a very general framework of a block design, we have already introduced
the ’incidence matrix’ N of order v × b where nij is the number of
appearances of the i th treatment in the j th block.
Note that nij ≥ 0 subject to the column totals being the block sizes
k1 , k2 , ..., kb - usually specified from the context.

DOE and SS -7 6 /18


C-Matrix

The row totals of N are the replication numbers r1 , r2 , ..., rv and these may
be subject to choice by the experimenter.
P P
However, i ri = j kj = n = total number of observations in the
experiment.
It is known that
1 C = r d − Nk −d N 0 , k −d = (k d )−1 (r d is a diagonal matrix);
2 C is v × v and with C 1 = 0 ( i.e., its row and column sums are 0’s ) ;
3 Rank (C ) = v − 1 iff the design is connected ;

DOE and SS -7 7 /18


C-Matrix

4 b + Rank(C ) = v + Rank(D) , D = k d − N 0 r −d N, r −d = (r d )−1 ;


X
5 For a connected design, C = λi ξi ξi0 , ξi0 1 = 0 , 1 ≤ i ≤ v − 1,
1≤i≤v −1
λi > 0, ξi0 ξj = δij , 1 ≤ i, j ≤ v − 1.
X
6 C + = Moore-Penrose g-inverse of C = λ−1 0
i ξi ξi
1≤i≤v −1

DOE and SS -7 8 /18


Q-vector

The vector Q is called the vector of ”Adjusted Treatment Totals”. We have


Q = T − Nk −d B
I T = vector of treatment totals
I B = vector of block totals

It is readily verified that


1 Q 01 = 0
2 E (Q) = C τ, D(Q) = σ 2 C

DOE and SS -7 9 /18


Reduced Normal Equations

The system of equations in (3) :


’C τ = Q’ is called the ”reduced normal equations”
One solution is given by :
τ̂ = C + Q.
From (2) ,
X
(µ\+ βj ) = y.j − nij τ̂i /kj , 1 ≤ j ≤ b.
i
It is known that (1) follows from (2) by summing over all j.
By convention, we take µ̂ = y...

DOE and SS -7 10 /18


Analysis of Block Design

SG
Sir, may I interrupt for a moment ? We took a course on linear models but that
went half hearted. I believe you are referring to some results without explicit
mention.....
Is it not true that the solution to τ̂ mentioned above is not unique?
However, so far as treatment differences are concerned, τi − τi 0 are all
estimated uniquely in the sense of blue.
I mean, all pairwise differences possess unique solution and that’s all we are
concerned with in this context.

DOE and SS -7 11 /18


Analysis of Block Design

SG
Moreover, the same holds for block effects differences ! I can also see a notable
difference in the formulae for derivation of expressions for τ̂ and β̂.... whereas
τ̂ is derived from C τ = Q,
β̂ follows readily from (2).
My questions are:
What about a parallel development wherein β̂ needs to be derived from a
similar matrix equation and ..... are the results different?
If so, how should we be guided in practice ?

DOE and SS -7 12 /18


Analysis of Block Design

SG
This question makes sense to me since.... after all ..... we are dealing with v × v
or b × b matrices.... so v << b or b << v may make a difference in
computations ..... of course · · · now-a-days..... we have computer power..... it
does not really matter.... still ..... from a theoretical perspective....

BKS
SG, I very much appreciate your mature thoughts and queries in this context. Let
me complete two more computations before I get back to you. Right ?

DOE and SS -7 13 /18


Sum of Squares

The Sum of Squares are computed as usual . We have


X
1 SS due to blocks ( unadj. )= Bj2 /kj − CF
j
0 0
2 SS due to treatments ( adj. )= τ Q = Q C + Q
3 SS due to Error = Total SS - SS due to Blocks - SS due to Treatments

DOE and SS -7 14 /18


ANOVA Table

We assume the design to be connected so that all treatment contrasts (as also all
block contrasts) are estimable. Here is the ANOVA Table :

Table: ANOVA Table


Source df SS MS F-ratio
Bj2 /kj − CF
P
Block ( unadj. ) b−1 j - -
Treatment (adj) v −1 Q0C +Q SS / df Valid for Testing Equality
Error n−b−v +1 by subtraction SS / df of Treatement Effects
P 2
Total n−1 yijh − CF -

DOE and SS -7 15 /18


Analysis if Block Design

SG
Sir, thanks for the computations above. I can possibly see answers to my queries!
First of all.... you wrote ’Blocks.... Unadjusted’ and ’Treatments-Adjusted’.
I believe, you are referring to the way treatment effects contrasts/difference
parameters have been estimated · · · by eliminating the block effects and
that is also reflected in the expression for Q-vector....

DOE and SS -7 16 /18


Analysis if Block Design

SG
On the other hand, when you say.... Blocks-Unadjusted..... you are referring
to Blocks in the absence of treatments so that .... as if.... you are in a CRD
set-up...
Hence Block SS [unadjusted] assumes a very simple expression.... derived
from a CRD model....
I now recall.... there are two ways of viewing this decomposition.... the one you
wrote, Sir, has a counter-part.... I mean....(to be continued)

DOE and SS -7 17 /18


Thank You

DOE and SS -7 18 /18

You might also like