## Applied Statistics Lesson of the Day – Fractional Factorial Design and the Sparsity-of-Effects Principle

Consider again an experiment that seeks to determine the causal relationships between $G$ factors and the response, where $G > 1$.  Ideally, the sample size is large enough for a full factorial design to be used.  However, if the sample size is small and the number of possible treatments is large, then a fractional factorial design can be used instead.  Such a design assigns the experimental units to a select fraction of the treatments; these treatments are chosen carefully to investigate the most significant causal relationships, while leaving aside the insignificant ones.

When, then, are the significant causal relationships?  According to the sparsity-of-effects principle, it is unlikely that complex, higher-order effects exist, and that the most important effects are the lower-order effects.  Thus, assign the experimental units so that main (1st-order) effects and the 2nd-order interaction effects can be investigated.  This may neglect the discovery of a few significant higher-order effects, but that is the compromise that a fractional factorial design makes when the sample size available is low and the number of possible treatments is high.

## Applied Statistics Lesson of the Day – Additive Models vs. Interaction Models in 2-Factor Experimental Designs

In a recent “Machine Learning Lesson of the Day“, I discussed the difference between a supervised learning model in machine learning and a regression model in statistics.  In that lesson, I mentioned that a statistical regression model usually consists of a systematic component and a random component.  Today’s lesson strictly concerns the systematic component.

An additive model is a statistical regression model in which the systematic component is the arithmetic sum of the individual effects of the predictors.  Consider the simple case of an experiment with 2 factors.  If $Y$ is the response and $X_1$ and $X_2$ are the 2 predictors, then an additive linear model for the relationship between the response and the predictors is

$Y = \beta_0 + \beta_1 X_1 + \beta_2 X_2 + \varepsilon$

In other words, the effect of $X_1$ on $Y$ does not depend on the value of $X_2$, and the effect of $X_2$ on $Y$ does not depend on the value of $X_1$.

In contrast, an interaction model is a statistical regression model in which the systematic component is not the arithmetic sum of the individual effects of the predictors.  In other words, the effect of $X_1$ on $Y$ depends on the value of $X_2$, or the effect of $X_2$ on $Y$ depends on the value of $X_1$.  Thus, such a regression model would have 3 effects on the response:

1. $X_1$
2. $X_2$
3. the interaction effect of $X_1$ and $X_2$

full factorial design with 2 factors uses the 2-factor ANOVA model, which is an example of an interaction model.  It assumes a linear relationship between the response and the above 3 effects.

$Y = \beta_0 + \beta_1 X_1 + \beta_2 X_2 + \beta_3 X_1 X_2 + \varepsilon$

Note that additive models and interaction models are not confined to experimental design; I have merely used experimental design to provide examples for these 2 types of models.

## Applied Statistics Lesson of the Day – The Full Factorial Design

An experimenter may seek to determine the causal relationships between $G$ factors and the response, where $G > 1$.  On first instinct, you may be tempted to conduct $G$ separate experiments, each using the completely randomized design with 1 factor.  Often, however, it is possible to conduct 1 experiment with $G$ factors at the same time.  This is better than the first approach because

• it is faster
• it uses less resources to answer the same questions
• the interactions between the $G$ factors can be examined

Such an experiment requires the full factorial design; in this design, the treatments are all possible combinations of all levels of all factors.  After controlling for confounding variables and choosing the appropriate range and number of levels of the factor, the different treatments are applied to the different groups, and data on the resulting responses are collected.

The simplest full factorial experiment consists of 2 factors, each with 2 levels.  Such an experiment would result in $2 \times 2 = 4$ treatments, each being a combination of 1 level from the first factor and 1 level from the second factor.  Since this is a full factorial design, experimental units are independently assigned to all treatments.  The 2-factor ANOVA model is commonly used to analyze data from such designs.

In later lessons, I will discuss interactions and 2-factor ANOVA in more detail.

## Applied Statistics Lesson of the Day – The Completely Randomized Design with 1 Factor

The simplest experimental design is the completely randomized design with 1 factor.  In this design, each experimental unit is randomly assigned to a factor level.  This design is most useful for a homogeneous population (one that does not have major differences between any sub-populations).  It is appealing because of its simplicity and flexibility – it can be used for a factor with any number of levels, and different treatments can have different sample sizes.  After controlling for confounding variables and choosing the appropriate range and number of levels of the factor, the different treatments are applied to the different groups, and data on the resulting responses are collected.  The means of the response variable in the different groups are compared; if there are significant differences, then there is evidence to suggest that the factor and the response have a causal relationship.  The single-factor analysis of variance (ANOVA) model is most commonly used to analyze the data in such an experiment, but it does assume that the data in each group have a normal distribution, and that all groups have equal variance.  The Kruskal-Wallis test is a non-parametric alternative to ANOVA in analyzing data from single-factor completely randomized experiments.

If the factor has 2 levels, you may think that an independent 2-sample t-test with equal variance can also be used to analyze the data.  This is true, but the square of the t-test statistic in this case is just the F-test statistic in a single-factor ANOVA with 2 groups.  Thus, the results of these 2 tests are the same.  ANOVA generalizes the independent 2-sample t-test with equal variance to more than 2 groups.

Some textbooks state that “random assignment” means random assignment of experimental units to treatments, whereas other textbooks state that it means random assignment of treatments to experimental units.  I don’t think that there is any difference between these 2 definitions, but I welcome your thoughts in the comments.

## Applied Statistics Lesson of the Day – Positive Control in Experimental Design

In my recent lesson on controlling for confounders in experimental design, the control group was described as one that received a neutral or standard treatment, and the standard treatment may simply be nothing.  This is a negative control group.  Not all experiments require a negative control group; some experiments instead have positive control group.

A positive control group is a group of experimental units that receive a treatment that is known to cause an effect on the response.  Such a causal relationship would have been previously established, and its inclusion in the experiment allows a new treatment to be compared to this existing treatment.  Again, both the positive control group and the experimental group experience the same experimental procedures and conditions except for the treatment.  The existing treatment with the known effect on the response is applied to the positive control group, and the new treatment with the unknown effect on the response is applied to the experimental group.  If the new treatment has a causal relationship with the response, both the positive control group and the experimental group should have the same responses.  (This assumes, of course, that the response can only be changed in 1 direction.  If the response can increase or decrease in value (or, more generally, change in more than 1 way), then it is possible for the positive control group and the experimental group to have the different responses.

In short, in an experiment with a positive control group, an existing treatment is known to “work”, and the new treatment is being tested to see if it can “work” just as well or even better.  Experiments to test for the effectiveness of a new medical therapies or a disease detector often have positive controls; there are existing therapies or detectors that work well, and the new therapy or detector is being evaluated for its effectiveness.

Experiments with positive controls are useful for ensuring that the experimental procedures and conditions proceed as planned.  If the positive control does not show the expected response, then something is wrong with the experimental procedures or conditions, and any “good” result from the new treatment should be considered with skepticism.