Handling Large Numbers of Entries

Stratification of materials
group entries on the basis of certain traits - maturity, market class, color, etc. then analyze in separate trials if entries are random effects, you can conduct a number of smaller trials and pool results to get better estimates of variances

Good experimental technique

be careful with land choice, preparation, and husbandry during the experiment choose seeds of uniform viability use consistent methods for data collection

Use of controls
systematically or randomly placed controls can be used to identify site variability and adjust yields of the entries

Incomplete Block Designs

We group into blocks to Increase precision Make comparisons under more uniform conditions But the problem is As blocks get larger, the conditions become more heterogeneous - precision decreases So small blocks are preferred, but in a breeding program the number of new selections may be quite large In other situations, natural groupings of experimental units into blocks may result in fewer units per block than required by the number of treatments (limited number of runs per growth chamber, treatments per animal, etc.)

Incomplete Block Designs

Plots are grouped into blocks that are not large

enough to contain all treatments (selections) Good references:

Kuehl Chapters 9 and 10 Cochran and Cox (1957) Experimental Designs

Types of incomplete block designs

Balanced Incomplete Block Designs
each treatment occurs together in the same block with every other treatment an equal number of times - usually once all pairs are compared with the same precision even though differences between blocks may be large can balance any number of treatments and any size of block but...treatments and block size fix the number of replications required for balance often the minimum number of replications required for balance is too large to be practical

Partially balanced incomplete block designs

different treatment pairs occur in the same blocks an unequal number of times or some treatment pairs never occur together in the same block mean comparisons have differing levels of precision statistical analysis more complex

Types of incomplete block designs

May have a single blocking criterion Randomized incomplete blocks Other designs have two blocking criteria and

are based on Latin Squares

Latin Square is a complete block design that requires N=t2. May be impractical for large numbers of treatments. Row-Column Designs either rows or columns or both are incomplete blocks Youden Squares two or more rows omitted from the Latin Square

Resolvable incomplete block designs

Blocks are grouped so that each group of blocks

constitute one complete replication of the treatment

Trials can be managed in the field on a replication-by-

replication basis
Field operations can be conducted in stages (planting,

weeding, data collection, harvest)

Complete replicates can be lost without losing the whole

experiment
If you have two or more complete replications, you can

analyze as a RBD if the blocking turns out to be ineffective

Lattice designs are a well-known type of resolvable

incomplete block design

Balanced Incomplete Block Designs

t = s*k and b = r*s t + r - 1

t = number of treatments s = number of blocks per replicate k = number of units per block (block size) b = total number of blocks in the experiment r = number of complete replicates

take home message - often the minimum number of

replications required for balance is too large to be practical

Lattice Designs
Square lattice designs number of treatments must be a perfect square (t = k2) blocks per replicate (s) and plots per block (k) are equal (s = k) and are the square root of the number of treatments (t) for complete balance, number of replicates (r) = k+1 Rectangular lattice designs t = s*(s-1) and k = (s-1) example: 4 x 5 lattice has 4 plots per block, 5 blocks per replicate, and 20 treatments Alpha lattices t = s*k more flexibility in choice of s and k

Randomization
Field Arrangement blocks composed of plots that are as homogeneous as possible Randomization Using Basic Plan randomize order of blocks within replications randomize the order of treatments within blocks

The Basic Plan for a Square Lattice

Block 1 2 3 Rep I 123 456 789 Rep II Rep III Rep IV 147 159 168 258 267 249 369 348 357

Balance - each treatment occurs together in the

same block with every other treatment an equal number of times ( = 1)

Example of randomization of a 3 x 3 balanced lattice (t = 9)

1 Assign r random numbers
Random Sequence Rank 372 1 2 217 2 1 963 3 4 404 4 3 2 From Basic Plan Block 1 2 3 Rep I 123 456 789 Rep II Rep III Rep IV 147 159 168 258 267 249 369 348 357

3 Randomize order of replications Block 1 2 3 Rep I 147 258 369 Rep II Rep III Rep IV 123 168 159 456 249 267 789 357 348

4 Randomize blocks within reps Rep I 3 2 1 II 2 1 3 III 3 1 2 IV 1 3 2

5 Resulting new plan

Block 1 2 3 Rep I 369 258 147 Rep II 456 123 789 Rep III Rep IV 357 159 168 348 249 267

Partially Balanced Lattices

Simple Lattices Two replications - use first two from basic plan 3x3 and 4x4 are no more precise than RBD because error df is too small Triple Lattices Three replications - use first three from basic plan Possible for all squares from 3x3 to 13x13 Quadruple Lattices Four replications - use all four Do not exist for 6x6 and 10x10 - can repeat simple lattice, but analysis is different

Analysis
Details are the same for simple, triple and

quadruple
Nomenclature: yij(l) represents the yield of the j-th treatment in the l-th block of the i-th replication Two error terms are computed Eb - Error for block = SSB/r(k-1) Ee - Experimental error = SSE/((k-1)(rk-k-1))

Computing Sums of Squares

SSTot = S yij(l)2 - (G2/rk2)

SSR = (1/k2) S Rj2 - (G2/rk2)

SSB = (1/kr(r-1)) S Cil2 - (1/k2r(r-1)) S Ci2
Cil = sum over all replications of yields of all treatments in the l-th block of the i-th replication minus rBil Bil = sum of yields of the k plots in the l-th block of the i-th replication Ci = sum of Cil

SST = (1/r) S T2 - (G2/rk2) SSE = SSTot - SSR - SSB - SST

Adjustment factor
Compare Eb with Ee - If Eb < Ee
then blocks have no effect analyze as if it were RBD using replications as blocks

If Eb > Ee then compute adjustment factor A

A = (Eb - Ee )/(k(r-1)Eb) used to compute adjusted yields in data table and table of totals

Compute the effective error mean square

Ee = (1+(rkA)/(k+1))Ee except for small designs (k=3,4), used in t tests and interval estimates

ANOVA
Source Total Rep Treatments Block(adj) Intrablock error df rk2-1 r-1 k2-1 r(k-1) (k-1)(rk-k-1) SS SSTOT SSR SST SSB SSE MS

Eb Ee

Testing differences
To test significance among adjusted treatment

means, compute an adjusted mean square

SSBu = (1/k)SBil2 - (G2/rk2) - SSR SSTadj = SST- A k (r-1) [ ((rSSBu)/(r-1)(1+kA))-SSB]

Finally, compute the F statistic for testing the

differences among the adjusted treatment means

F = (SSTadj / (k2-1))/ Ee with k2 - 1 degrees of freedom (k-1)(rk-k-1) from MSE

Standard Errors
SE of adjusted treatment mean
= Ee / r

SE of difference between adjusted means in

same block
= (2Ee/r)(1+(r-1)A)

SE of difference between adjusted means in

different blocks
= = (2Ee/r)(1+rA) 2Ee / r

For larger lattices (k > 4) sufficient to use

Relative Precision
Compute the error mean square of a RBD ERB = ((SSB+SSE)/(k2-1)(r-1)) Then the relative precision of the lattice is RP = ERB/Ee

Numerical Example - Simple Lattice

Rep Blk I 1 (19) 18.2 2 (12) 13.3 3 (1) 15.0 4 (22) 7.0 5 (9) 11.9 (16) 13.0 (13) 11.4 (2) 12.4 (24) 5.9 (7) 15.2 Yield (18) 9.5 (15) 14.2 (3) 17.3 (21) 14.1 (10) 17.2 (17) 6.7 (14) 11.9 (4) 20.5 (25) 19.2 (8) 16.3 (20) 10.1 (11) 13.4 (5) 13.0 (23) 7.8 (6) 16.0 Sum Bil 57.5 64.2 78.2 54.0 Cil 17.6 4.2 5.3 -.5 Adj

1.54 .37 .46 -.04

76.6

3.9

.34
2.67

330.5 30.5

Second Rep
RepBlk II 1 2 (23) 7.7 (5) 15.8 (22) 10.2 (14) 10.9 (6) 20.0 (18) 15.2 (20) 18.0 (12) 11.5 (24) 4.7 (16) 21.1 Yield (3) 19.1 (10) 18.8 (2) 17.0 (9) 10.9 (11) 16.9 (8) 15.5 (15) 14.4 (17) 11.0 (4) 16.6 (21) 10.9 (13) 14.7 (25) 20.0 (7) 15.3 (19) 9.8 (1) 15.0 Sum SUM
Bil Cil

Adj

72.2 87.0 65.0 52.9 83.9 361.0 691.5

-9.9 -13.3 -10.4 15.5 -12.4 -30.5 0.0

-.86 -1.17 -.91 1.35 -1.08 -2.67 0.00

3
4 5

Unadjusted Yield Totals

(1)
(6)

30.0 (2) 29.4

36.0 (7) 30.5

(3)
(8) (13) (18) (23)

36.4
31.8 26.1 24.7 15.5

(4)
(9) (14) (19) (24)

37.1
22.8 22.8 28.0 10.6 SUM

(5) 28.8
(10) 36.0 (15) 28.6 (20) 28.1 (25) 39.2 691.5

(11) 30.3 (12) 24.8 (16) 34.1 (17) 17.7 (21) 25.0 (22) 17.2

ANOVA
Source df SS MS

Total
Replication

49
1

805.42
18.60 621.82

Selection (unadj) 24

Block in rep (adj) 8

Intrablock error 16

77.59
87.41

9.70=Eb
5.46=Ee

Eb is greater than Ee so we compute the adjustment factor, A A = (Eb - Ee )/(k(r-1)Eb ) = (9.70 - 5.46)/((5)(1)(9.70)) = 0.0874 Multiply A by the treatment/block sums (C) to get the adjusted totals

Computing Effective Error

Once you have calculated the Adjustment factor

(A), calculate the effective mean square (Ee )

Ee = (1+(rkA)/(k+1))Ee = (1+(2*5*0.0874)/6)*5.46 = 6.26

To test significance, compute SSBu

1/k S Bil2 - Correction factor - SSR for this example 9834.23-9563.44-18.60=252.18

Then SST(adj)=
SST A*k(r-1){ [ r * SSBu/(r-1)(1+kA) ] - SSB} 621.82-(0.0874)(5)(1) { [ (2 x 252.18)/1 * (1+5*0.0874)]-77.59} = 502.35

Test Statistics
FT to test differences among the adjusted

treatment means:
(SST(adj)/(k2-1))/Ee (502.35/24)/6.26 = 3.34

Standard Error of a selection mean = Ee/r = 6.26/2 = 1.77 LSI can be computed since k > 4 ta 2Ee/r = 1.746 (2x6.26)/2 = 4.37

Relative precision
How does the precision of the Lattice compare

to that of a randomized block design?

First compute MSE for the RBD as: ERB = (SSB+SSE)/(k2 - 1)(r -1) = (77.59 + 87.41)/(24)(1) = 6.88

Then % relative precision = (ERB / Ee )100 = (6.88/6.26)*100 = 110.0%

Report of Statistical Analysis

Because of variation in the experimental site and

because of economic considerations, a 5x5 simple lattice design was used LSI at the 5% level was 4.37 Five new selections outyielded the long term check (12.80kg/plot) One new selection (4) with a yield of 19.46 significantly outyielded the local check (1) None of the new selections outyielded the late release whose mean yield was 19.00 Use of the simple lattice resulted in a 10% increase in precision when compared to a RBD

Cyclic designs
Incomplete Block Designs discussed so far require extensive tables of

design plans. Must be careful not to make mistakes when assigning treatments to experimental units and during field operations
Cyclic designs are a type of incomplete block design that are relatively

easy to construct and implement Block 1 2 3 4 5 Treatment Label 0, 1, 3 1, 2, 4 2, 3, 5 3, 4, 6 4, 5, 7

5, 6, 0

Alpha designs
Patterson and Williams, 1976 Described a way to construct incomplete block designs for

any number of treatments (t) and block size (k), such that t is a multiple of k. Includes a(0,1)-lattice designs.
-designs are available for many (r,k,s) combinations
r is the number of replicates k is the block size s is the number of blocks per replicate number of treatments t = ks

Efficient -designs exist for some combinations for which

conventional lattices do not exist

Can accommodate unequal block sizes

Alpha designs
Design Software
The current version of Gendex can generate designs with up to 10,000 entries

http://www.designcomputing.net/gendex/
Evaluation copy is free

Cost for perpetual/personal license is $199

Others: Alpha+, CycDesigN

Analysis can be done with SAS