Saturday, April 2, 2011

SG – Topic 2 – Common Designs


 

Completely Randomised Designs

-          randomization and replication

-          treatments allocated at random to experimental units

-          each treatment repeated equal number of times

-          allocation

o   by lot

-          Analysis

o   provides one-way classified data according to levels of single factor (treatment)

o   Anova

§  how much variation is due to differences between treatments

§  how much variation is due to differences within each set of observations for same treatment

o   linear statistical model

§     y  =  u  +  α  + ε

·         y  -- response

·         u  -- general mean

·         α  --  effect of treatment

·         ε --  error

o   error components assumed to be independently and normally distributed.

·         total variation partitioned

o   variation among treatments (treatment)

o   variation among units within treatment units

-          Advantages

o   flexible design

o   simple statistical analysis if some missing data / loss of info small in comparison to other designs

o   design provides maximum degrees of freedom for estimate of error variance à which increases precision for small experiments.

o   maximum use of experimental units à all experimental material can be used.

-          Disadvantages

o   if experimental material is not uniform, precision is low.

§  whole variation among experimental units is included in residual variance

-          Applications

o   physical science laboratory experiments

o   greenhouse studies, where experimental material is uniform.

o   recommended for situations where large fraction of units likely to be destroyed or fail to respond.

o   where total number of units is limited.

 

 

Randomised Block Designs

 

-          experimental units sorted into subgroups  à where it is believed subgroups will respond differently to treatment

-          Analysis

o   two way ANOVA

§  blocks , treatment , error

o   response = general mean effect (overall mean) + treatment effect + block effect + error

 

 

Source of variation

Degrees of freedom (df)

Sum of Squares

Mean Square

F Statsitic

Treatment

p-1

SST

MST = SST / (p-1)

F = MST / MSE

Block

r – 1

SSB

MSB = SSB / (r-1)

 

Error

(p-1) ( r-1)

SSE

MSE = SSE / (p-1))r-1)

 

Total

rp - 1

SStotal

 

 

 

p treatments

r blocks

 

 

 

-          Advantages

o   design more efficient than CRD  à improves precision by removing one source of variation from experimental error

o   no restrictions on number of treatments or replicates

o   any number of blocks can be used so long as each treatment replicated same number of times in each block.

o   stat analysis simple and rapid

-          Disadvantages

o   not suitable for large number of treatments

o   efficiency decreases as number of treatments and hence block size increases

o   in analysis à missing data can cause some difficulty.

-          applications

o   provides unbiased estimates of means of blocking categories à additional information obtained

o   remove one source of variation

 

 

Latin Square Design

 

-          control 2 sources of variation

-          rows and columns

-          number of treatments = number of replications   [?]

-          p treatments require p * p experimental units

-          [is fig 2.4 on page 14 correct?]

-          Analysis

o   three way anova

o   response = general mean effect (overall mean) + row effect + column effect + treatment effect + error.

 

Source of variation

Df

SS

MS

F Statistic

Rows

p-1

SSR

MSR = SSR / (p-1)

Fr = MSR / MSE

Columns

p-1

SSC

MSC = SSC / (r-1)

Fc = MSC / MSE

Treatment

p-1

SST

MST = SST / (p-1)

Ft = MST / MSE

Error

(p-1) (p-2)

SSE

MSE = SSE / (p-1)(p-1)

 

Total

p^2 -1

SSTotal

 

 

 

 

-          Advantages

o   control two sources of variation

o   more than one factor can be investigated simultaneously and  with fewer trials

o   even with missing data, analysis relatively simple.

-          disadvantages

o   number of treatments restricted to number of replications  [?]

o   fundamental assumption à there is no interaction between different factors à may not be true

o   missing plots – when several units are missing à complex analysis

-          Applications

o   agriculture / food manufacturing / engineering / dairy farm

 

 

Factorial Designs

 

-          CRD / RBD / LSD à primarily concerned with comparison and estimation of effects of a single set of treatments

-          factorial à treatments are all the combinations of different factors

o   effects of each factor

o   interaction effects à variation in one factor as result of different levels of other factors

 

 

Nested Designs

 

-          nested or hierarchical design

-          advantages

o   study effects of sources of variability that manifest themselves over time.

 

 

Repeated Measures

 

-          all subjects participate under all levels of the independent variable

-          controls for subject heterogeneity

o   problem

§  practice effects

No comments: