The completely randomized design, or CRD, is the most straightforward experimental layout. Its strength lies in its simplicity. We take a group of experimental units, like plots of land or patients in a study, and randomly assign them to different treatments. The goal is to see if the treatments have different effects on some outcome we care about.

The statistical heart of a CRD is its linear model. This model assumes that any observation we make is a combination of three distinct parts: a baseline average, the specific effect of the treatment applied, and some unavoidable random noise.

Imagine testing three new fertilizers (treatments) on wheat yield. The overall mean, $\mu$, would be the average yield across all plots. The treatment effect, $\tau_i$, would be how much a specific fertilizer increases or decreases the yield compared to that average. The error term, $\epsilon_{ij}$, accounts for all the other little differences between plots, like slight variations in soil quality or sunlight that randomization helps to average out.

To determine if the treatments actually have different effects, we need to analyze the variation in our data. The core idea behind the analysis of variance (ANOVA) for a CRD is to partition the total variation into distinct sources.

We start with the Total Sum of Squares (SST), which measures the overall variability of all observations from the grand mean. We then break this down into two pieces: the variation between the treatment groups (SSTr) and the random variation within each treatment group (SSE).

This partitioning is powerful. If the variation between treatment groups (SSTr) is large compared to the random variation within the groups (SSE), it's strong evidence that the treatments are causing different outcomes.

To formally test the null hypothesis that all treatment means are equal ($H_0: \mu_1 = \mu_2 = ... = \mu_t$), we use the F-test. This test compares the two sources of variation we just calculated.

First, we convert the sums of squares into mean squares by dividing by their respective degrees of freedom (df). The degrees of freedom represent the number of independent pieces of information used to calculate the sum of squares.

• The df for treatments is $t - 1$, where $t$ is the number of treatments.
• The df for error is $N - t$, where $N$ is the total number of observations.

The ratio of these two mean squares gives us our F-statistic.

To properly interpret the F-test, we also consider the Expected Mean Squares (EMS). These are the long-run average values of the mean squares if we were to repeat the experiment many times. For a CRD, they are:

• $E(MSE) = \sigma^2$
• $E(MSTr) = \sigma^2 + \frac{\sum n_i \tau_i^2}{t-1}$

Notice that the expected value of MSE is just the error variance, $\sigma^2$. The expected value of MSTr is the error variance plus a term that reflects the variation due to the treatment effects. If the null hypothesis is true (all $\tau_i = 0$), then the second term vanishes, and both mean squares estimate the same error variance. This is the theoretical justification for our F-test.

The validity of the F-test in ANOVA hinges on a few key assumptions about the model's error terms ($\\{\epsilon_{ij}\\}$). Before you can trust your conclusions, you must perform diagnostic checks.

The two most critical assumptions are:

1. Homogeneity of Variance: The variance of the errors is constant across all treatments. In other words, the spread of the data within each group should be roughly the same.
2. Normality of Residuals: The errors are normally distributed.

We check these assumptions by analyzing the residuals, which are the differences between the observed values and the values predicted by our model: $e_{ij} = Y_{ij} - \bar{Y}_{i.}$. A residual is our best estimate of the unobservable random error term, $\epsilon_{ij}$.

To check for normality, you can create a normal probability plot (also known as a Q-Q plot) of the residuals. If the residuals are normally distributed, the points on this plot will fall approximately along a straight line. Significant deviations from this line suggest that the normality assumption may not hold.

If these diagnostic checks reveal serious violations, the results of your ANOVA may be unreliable. In such cases, you might need to transform your data (e.g., using a log or square root transformation) or use a different type of statistical test that doesn't rely on these assumptions.

The Completely Randomized Design provides a foundational model for comparing treatments. By partitioning variance and checking assumptions, we can confidently determine whether observed differences are real or just due to chance.