TRB PG Assistant - Design of experiments - principles of experimentation, analysis of variance, completely randomized design, randomized block design, Latin square design E-Book Study Material

Design of Experiments

In statistics, the design of experiments (DOE) is a crucial branch that deals with planning, conducting, and analyzing controlled tests to understand the relationship between factors and the response variable. It's about making efficient use of resources to gain the maximum amount of information. This involves carefully considering how to vary the input factors and observe the output response. The goal is to isolate the effect of specific variables on an outcome while minimizing the influence of confounding factors.

Principles of Experimentation

At the heart of any well-designed experiment lie fundamental principles that ensure the validity and reliability of the results. These principles are not mere guidelines but the bedrock upon which sound conclusions are built.

1. Randomization

Randomization is the process of assigning experimental units to different treatment groups using a random mechanism. This ensures that each unit has an equal chance of being assigned to any treatment. The primary benefit of randomization is that it helps to eliminate systematic bias. By distributing the effects of lurking variables (unforeseen factors that could influence the outcome) randomly across all treatment groups, it prevents them from systematically favoring one group over another. This allows us to attribute any observed differences in the response variable primarily to the treatments themselves.

For example, if we are testing the effect of different fertilizers on crop yield, we would randomly assign plots of land to receive fertilizer A, fertilizer B, or no fertilizer. This prevents any inherent differences in soil fertility or sunlight exposure in certain plots from unfairly influencing the results of one fertilizer over another.

2. Replication

Replication is the repetition of an experiment or a treatment on multiple experimental units. It is essential for several reasons. Firstly, it increases the precision of the estimates of treatment effects. By averaging the results from multiple units receiving the same treatment, we reduce the impact of random error associated with individual units. Secondly, replication allows us to estimate the experimental error. Without repetition, we cannot distinguish between the variability due to the treatment and the inherent variability in the experimental units. Thirdly, replication increases the power of the statistical tests used to detect differences between treatments. A larger sample size, achieved through replication, makes it easier to find statistically significant differences if they truly exist.

Consider a clinical trial testing a new drug. Replication means administering the drug to several patients and comparing their outcomes to several patients receiving a placebo. This repetition helps confirm that the observed effect is due to the drug and not just a random variation in how a few individuals respond.

3. Blocking

Blocking is a technique used to reduce variability within the experimental units by grouping them into homogeneous blocks. Experimental units within a block are more similar to each other than units in different blocks. The treatments are then applied randomly within each block. The purpose of blocking is to control for known sources of variation that might otherwise obscure the treatment effects. By isolating these sources of variation, blocking increases the efficiency of the experiment, making it easier to detect treatment differences.

Imagine an experiment to test the effectiveness of different teaching methods. If students are taught in different classrooms, and there are known differences in classroom environments (e.g., lighting, noise levels), then classrooms can be used as blocks. Within each classroom, the teaching methods would be randomly assigned to student groups. This ensures that differences in teaching methods are compared within similar classroom environments, reducing the impact of classroom variations on the results.

Analysis of Variance (ANOVA)

Analysis of Variance, commonly known as ANOVA, is a powerful statistical technique used to analyze differences between the means of two or more groups. It was developed by Ronald Fisher. ANOVA works by partitioning the total variation observed in the data into different sources of variation. Specifically, it compares the variance between the group means to the variance within the groups. If the variance between groups is significantly larger than the variance within groups, it suggests that the group means are indeed different.

The Core Idea of ANOVA

ANOVA tests the null hypothesis (H₀) that all group means are equal against the alternative hypothesis (H₁) that at least one group mean is different. It does this by calculating an F-statistic. The F-statistic is the ratio of the variance between the groups (explained variation) to the variance within the groups (unexplained variation or random error).

A large F-statistic indicates that the variation between groups is much larger than the variation within groups, leading us to reject the null hypothesis and conclude that there is a significant difference among the group means.

ANOVA Table

The results of an ANOVA are typically summarized in an ANOVA table, which provides a structured overview of the sources of variation, degrees of freedom, sum of squares, mean squares, and the F-statistic.

Key components of an ANOVA table:

Source of Variation: This typically includes 'Between Groups' (or 'Treatment') and 'Within Groups' (or 'Error').
Degrees of Freedom (df): For 'Between Groups', df = k - 1, where k is the number of groups. For 'Within Groups', df = N - k, where N is the total number of observations.
Sum of Squares (SS): Measures the total variation for each source. SS_Total = SS_Between + SS_Within.
Mean Squares (MS): Calculated by dividing the Sum of Squares by its corresponding degrees of freedom. MS_Between = SS_Between / (k - 1) and MS_Within = SS_Within / (N - k).
F-statistic: The ratio of the mean squares: F = MS_Between / MS_Within.
P-value: The probability of observing an F-statistic as extreme as, or more extreme than, the one calculated, assuming the null hypothesis is true. A small p-value (typically < 0.05) leads to rejection of the null hypothesis.

ANOVA Shortcut: Think of ANOVA as dividing the total 'story' (variation) in your data into chapters: the chapter about the treatments (Between Groups) and the chapter about random noise (Within Groups). If the treatment chapter is much more exciting (larger variation) than the noise chapter, then treatments matter!

Completely Randomized Design (CRD)

The Completely Randomized Design (CRD) is the simplest form of experimental design. In a CRD, experimental units are randomly assigned to the different treatment groups. There are no restrictions on how the units are assigned, other than that the assignment must be random. This design is most effective when the experimental units are homogeneous (very similar to each other) and when the number of experimental units is large enough to allow for effective randomization.

Assumptions of CRD

The experimental units are independent.
The errors (residuals) are normally distributed with a mean of zero and constant variance (homoscedasticity).
The observations within each treatment group are independent.

When to Use CRD

When experimental units are homogeneous.
When the number of experimental units is large.
When there are no obvious sources of variability that need to be controlled through blocking.
In laboratory settings where conditions can be tightly controlled.

Example of CRD

Suppose a company wants to test four different formulations of a new detergent. They have 20 identical washing machines available for testing. In a CRD, the 20 washing machines would be randomly assigned to the four detergent formulations, with 5 machines assigned to each formulation. The random assignment ensures that any differences in washing performance are not systematically due to a particular machine's characteristics.

ANOVA for CRD

For a CRD with 'k' treatments and 'n' observations per treatment (total N = nk observations), the ANOVA is set up as follows:

Null Hypothesis (H₀): μ₁ = μ₂ = ... = μ_k (All treatment means are equal)

Alternative Hypothesis (H₁): At least one μ_i is different.

The ANOVA table for CRD is as follows:

Source of Variation	Degrees of Freedom (df)	Sum of Squares (SS)	Mean Squares (MS)	F-statistic
Between Treatments (Trt)	k - 1	SS_Trt = Σ n_i(ȳ_i - ȳ_..)²	MS_Trt = SS_Trt / (k - 1)	F = MS_Trt / MS_Error
Within Treatments (Error)	N - k	SS_Error = Σ Σ (y_ij - ȳ_i)²	MS_Error = SS_Error / (N - k)	F = MS_Trt / MS_Error
Total	N - 1	SS_Total = Σ Σ (y_ij - ȳ_..)²

Where:

k = number of treatments
n_i = number of observations in treatment i
N = total number of observations (N = Σ n_i)
ȳ_i = mean of treatment i
ȳ_.. = grand mean (mean of all observations)
y_ij = j-th observation in the i-th treatment

Randomized Block Design (RBD)

The Randomized Block Design (RBD) is an extension of the CRD that incorporates the principle of blocking. It is used when there is a known source of variability among the experimental units that can be grouped into blocks. The key idea is to make comparisons between treatments within relatively homogeneous blocks. This helps to remove the variability due to the blocking factor from the experimental error, thereby increasing the precision of the experiment.

How RBD Works

1. Identify Blocking Factor: Determine a characteristic of the experimental units that is expected to influence the response variable (e.g., different batches of raw material, different operators, different days of production). 2. Form Blocks: Group the experimental units into blocks based on this characteristic. Units within a block should be as similar as possible. The number of units in each block should be equal to the number of treatments (or a multiple of it). 3. Randomize within Blocks: Randomly assign the treatments to the experimental units within each block. This ensures that each treatment is applied once within each block, and the assignment is random within that block.

Assumptions of RBD

The effects of treatments and blocks are additive.
Errors are independent, normally distributed with mean zero and constant variance.
There is no interaction between treatments and blocks (i.e., the effect of a treatment is the same across all blocks). If there is a significant treatment-block interaction, RBD may not be appropriate, and a more complex design might be needed.

When to Use RBD

When there is a known source of variation that can be controlled by blocking.
When the experimental units are heterogeneous.
To increase the efficiency and precision of the experiment compared to CRD.

Example of RBD

A food scientist wants to compare the taste of five different brands of cookies. They recruit 10 participants for the taste test. Since participants might have different taste sensitivities, the participants themselves can be used as blocks. Each participant (block) will taste all five brands of cookies. The order in which each participant tastes the cookies would be randomized. This design controls for individual differences in taste perception.

ANOVA for RBD

For an RBD with 'k' treatments and 'b' blocks, the ANOVA partitions the total variation into three sources: Treatments, Blocks, and Error.

Null Hypothesis (H₀): μ₁ = μ₂ = ... = μ_k (All treatment means are equal)

The ANOVA table for RBD:

Source of Variation	Degrees of Freedom (df)	Sum of Squares (SS)	Mean Squares (MS)	F-statistic
Treatments (Trt)	k - 1	SS_Trt	MS_Trt = SS_Trt / (k - 1)	F = MS_Trt / MS_Error
Blocks (Blk)	b - 1	SS_Blk	MS_Blk = SS_Blk / (b - 1)
Error	(k - 1)(b - 1)	SS_Error	MS_Error = SS_Error / ((k - 1)(b - 1))
Total	N - 1 (where N = bk)	SS_Total

Where:

k = number of treatments
b = number of blocks
N = bk = total number of experimental units
SS_Trt, SS_Blk, SS_Error, SS_Total are the sums of squares for treatments, blocks, error, and total, respectively.

RBD Shortcut: RBD is like having multiple mini-experiments (blocks) instead of one big one. You test all treatments within each mini-experiment. This helps you see the treatment differences clearly because you've already accounted for the differences between the mini-experiments (blocks).

Latin Square Design (LSD)

The Latin Square Design (LSD) is a more complex experimental design used when we want to control for two sources of variation simultaneously. It is particularly useful when there are two blocking factors that are believed to influence the experimental outcome. The design ensures that each treatment appears exactly once in each row and exactly once in each column. This structure allows for the independent estimation of treatment effects, row effects, and column effects.

How LSD Works

1. Identify Two Blocking Factors: Choose two factors that are likely to cause variation in the response, and these factors must have the same number of levels as the number of treatments. For example, if you have 'k' treatments, you need two blocking factors, each with 'k' levels. 2. Arrange Treatments in a Square: The experimental units are arranged in a k x k square. The rows represent one blocking factor, and the columns represent the other blocking factor. The 'k' treatments are then assigned to the k² experimental units such that each treatment appears exactly once in each row and exactly once in each column.

Assumptions of LSD

The effects of treatments, rows, and columns are additive.
Errors are independent, normally distributed with mean zero and constant variance.
There are no interactions between treatments and rows, treatments and columns, or rows and columns. This is a strong assumption, and if interactions are present, LSD may not be appropriate.

When to Use LSD

When there are two sources of variation (blocking factors) that need to be controlled.
When the number of levels for both blocking factors is equal to the number of treatments.
When the experimental units can be arranged in a square grid.

Example of LSD

Suppose a chemist wants to test four different catalysts (treatments) on the rate of a chemical reaction. The reaction rate might be affected by the temperature at which the experiment is conducted (a row factor) and the specific operator performing the experiment (a column factor). If there are four temperature levels and four operators, and four catalysts, an LSD can be used. The experimental units are arranged in a 4x4 square. The rows represent temperatures, columns represent operators, and the cells contain the catalysts. Each catalyst is used once at each temperature and once by each operator.

A typical 4x4 Latin Square might look like this:

	Operator 1	Operator 2	Operator 3	Operator 4
Temp 1	Catalyst A	Catalyst B	Catalyst C	Catalyst D
Temp 2	Catalyst B	Catalyst A	Catalyst D	Catalyst C
Temp 3	Catalyst C	Catalyst D	Catalyst A	Catalyst B
Temp 4	Catalyst D	Catalyst C	Catalyst B	Catalyst A

ANOVA for LSD

For an LSD with 'k' treatments, 'k' rows, and 'k' columns, the ANOVA partitions the total variation into four sources: Treatments, Rows, Columns, and Error.

Null Hypothesis (H₀): μ₁ = μ₂ = ... = μ_k (All treatment means are equal)

The ANOVA table for LSD:

Source of Variation	Degrees of Freedom (df)	Sum of Squares (SS)	Mean Squares (MS)	F-statistic
Treatments (Trt)	k - 1	SS_Trt	MS_Trt = SS_Trt / (k - 1)	F = MS_Trt / MS_Error
Rows (Row)	k - 1	SS_Row	MS_Row = SS_Row / (k - 1)
Columns (Col)	k - 1	SS_Col	MS_Col = SS_Col / (k - 1)
Error	(k - 1)(k - 2)	SS_Error	MS_Error = SS_Error / ((k - 1)(k - 2))
Total	k² - 1	SS_Total

Where:

k = number of treatments (also number of rows and columns)
k² = total number of experimental units
SS_Trt, SS_Row, SS_Col, SS_Error, SS_Total are the sums of squares for treatments, rows, columns, error, and total, respectively.

LSD Shortcut: Think of LSD as a tic-tac-toe board for experiments. You want to test different 'X's (treatments). You arrange them so each 'X' appears once in every row and once in every column. This helps you figure out if the 'X's are different, without being confused by differences between rows or columns.