Factorial experiment

A factorial experiment is a statistical method used to study how multiple factors influence a specific outcome. This outcome is known as the response variable. In many scientific studies, researchers want to know how different settings change a result. Instead of changing just one thing at a time, a factorial experiment tests many factors simultaneously. Each factor is tested at several distinct values called levels. By testing every possible combination of these levels, researchers gain a complete view of the system. This method is essential for identifying interactions, which occur when the effect of one factor depends on the level of another.

The mechanism of a full factorial design relies on testing every possible combination of factor levels. These combinations are often called runs, points, or cells. For example, in a 2x2 factorial design, there are two factors and each has two levels. This results in four unique combinations to test. If a researcher adds a third factor with two levels, the number of runs grows to eight. This growth is exponential, meaning the number of required tests increases very quickly as more factors are added. If the number of combinations becomes too large to manage, researchers use a fractional factorial design. This approach strategically omits some combinations to make the experiment more efficient while still providing useful data.

There are different ways to categorize and denote these experiments. A design is called symmetric or fixed-level if every factor has the same number of levels, denoted as $s^k$, where $k$ is the number of factors. For instance, a $2^5$ experiment involves five factors with two levels each. If the factors have different numbers of levels, the design is called mixed-level or asymmetric. Researchers use different notation systems to represent these levels. Some use the values 1 and 0, while others use 1 and -1, often abbreviated as + and -. In some mathematical contexts, levels are represented by integers modulo $s$ to allow for advanced algebraic analysis.

The history of these designs is rooted in agricultural and statistical science. In the 19th century, John Bennet Lawes and Joseph Henry Gilbert utilized these designs at the Rothamsted Experimental Station. However, the modern understanding of the method was shaped by Ronald Fisher. In 1926, Fisher argued that factorial designs were much more efficient than the traditional one-factor-at-a-time (OFAT) approach. He noted that a factorial design could determine the effects of several factors and their interactions using the same number of trials required to study just one factor. Fisher officially used the term "factorial" in his 1935 book, *The Design of Experiments*. Additionally, Frank Yates made significant contributions to the field through his work on the analysis of these designs.

Factorial experiments offer significant advantages over OFAT experiments. They are more efficient and provide more information at a similar or even lower cost. Most importantly, they are required to detect interaction effects. If a researcher uses an OFAT design when interactions are present, they may seriously misunderstand how the response changes. A notable example involves engineers at the bearing manufacturer SKF. They wanted to test a new cage design to see its effect on bearing lifespan. By using a 2x2x2 factorial design, they tested three factors: heat treatment, outer ring osculation, and cage design. This allowed them to discover that combining specific levels of heat treatment and osculation could increase bearing life fivefold.

Understanding the mathematical components of these experiments requires looking at cell means and contrasts. The expected response for a specific combination of levels is called the cell mean, often denoted by the Greek letter $\mu$. A contrast is a linear combination of these cell means where the coefficients sum to zero. Contrasts are the building blocks used to define main effects and interactions. A main effect is present if the contrast between different levels of a factor is not zero. Interaction is defined as a lack of additivity between factors. If the interaction is absent, the factors are said to exhibit additivity, which can be visualized as a kind of parallelism in the data.

Factorial designs connect deeply to various scientific and engineering fields. They allow for conclusions that are valid over a wide range of experimental conditions rather than just a single point. In engineering, they help find optimal conditions for products much faster than traditional methods. In complex systems, they help researchers map the connections between different variables. Whether studying motor power at different speeds or chemical reactions at various temperatures, factorial experiments turn isolated observations into a comprehensive map of how the world works.