Introduction to experimental design PART 2
Written by Akshila Anurangi, Machine Learning Engineer at OCTAVE and Dr. Rajitha Navarathna, Principal Data Scientist at OCTAVE
In part I of the article we understood of the Experimental Structure and the Basic Principals of Design. In part II we understand the “Treatment Structure” and a basic understanding about how to design your experiment.
Treatment Structure
Treatment structure is the set of treatment combinations the experimenter has selected for comparison. For example, this could be about different drugs, doses, administration times, and combinations thereof. When we have only one set of treatments, it is called a “one-way treatment structure.” There is no assumed relationship among those.
If we have treatment combination such as drugs* doses, drugs* administration time, drugs*does*administration time it is called a factorial arrangement. It could be two-way, three way or n- way structure.
Design Structure
The design structure is the grouping of the units into homogeneous groups or blocks.
Popular choices of design structures are completely randomised design, randomised complete block design, Latin square design, and incomplete block design.
Completely randomized design (C.D.R)?
In a C.R.D. structure, all experimental units are assumed to have similar characteristics (homogeneous). Treatments are assigned to the experimental units completely at random, generally with equal replication.
Randomized Complete Block Design (R.C.B.D.)
An R.C.B.D. structure is a blocking scheme in which the number of experimental units within a block is a multiple of the number of treatments. The treatments are randomly allocated to the experimental units inside each block, and most importantly, each treatment appears in each block.
Incomplete Block Design
Incomplete block design is very similar to randomized complete block design, but here the number of treatments exceeds the number of experimental units inside the block, hence not all treatments will occur inside each block.
Latin Square Design
In a Latin square design, blocking is done in two directions. Latin square designs are effective in controlling two nuisance factors simultaneously. Each treatment occurs once and only once in each row, as well as in each column, and the treatments are assumed to be independent from both row and column factors.
Experimental design is a combination of both design and treatment structure.
Now that we have a good understanding of the experimental design basics, let’s see what the important steps of experimental design are that we need to think through. To understand the steps easily, let’s walk through an example.
Imagine a scientist wants to compare the effect of popular diets on losing weight. He is about to test out the effects of the paleo diet, the low-crab diet, and the intermittent fasting diet on a group of people who are trying to lose weight.
Before designing the experiment, the first step is to recognize and state the problem. Here the problem statement is which diet is the most effective when it comes to losing weight.
Then comes the choice of factors, levels, and range. Here the factor would be different diet plans, and the factor levels could be the amounts of carbs, proteins, etc. included in each diet.
The next step would be the selection of a response variable or variables. Since this is about losing weight, the response variable can be the number of kilograms lost in a certain period.
Then comes the choice of experimental design. Depending on the desire of the scientist, he can choose any of the given design structures and treatment structures. Let’s assume he decided to go with a randomized complete block design, having four blocks with three individuals each. Since we have three individuals in each block, the scientist can randomly allocate the three different diets for them. Reasonable blocks could be identified based on gender and the starting weight.
Once the experimental design is finalized, the experiment will be performed, and the data will be statistically analyzed. Based on the choice of experiment and experimental design, there are different ways to conduct the statistical analysis.
The concluding step would be to give the conclusion and the recommendation. Based on the results, the scientist can conclude which diet is the most effective and what other recommendations he can give based on the observations.
Other things to be mindful when designing an experiment are,
- Nuisance factors that should be investigated or controlled
- Number of replications of each treatment
- What is the randomization procedure?
- What statistical analysis methods will be used to analyze the results?
- What difference between the results will be considered important?
With this article I hope you got a basic understanding about how you could design your experiment. To conclude what we have reviewed in both parts of this article/blog? are.
- Introduction to experimental design and a few basic definitions
- 3 Basic Principles of Experimental Design
- Treatment Structure and Design Structure
- Steps for planning and executing experiments
