Every well-designed experiment has a logical architecture. It isn't just a matter of applying a treatment and seeing what happens. This architecture is composed of two primary parts: the treatment structure and the design structure.

The treatment structure is all about the what. It consists of the set of treatments, or treatment combinations, the researcher has chosen to study. It's the collection of factors, their levels, and their relationships that you want to compare.

The design structure is the how. It describes how the treatments are assigned to the experimental units. This involves the grouping of experimental units and the method of randomization. The design structure dictates the logic of the experiment and how we control for sources of variation.

A common point of confusion in research is the distinction between what is being treated and what is being measured. Getting this right is critical for a valid analysis.

This is the entity that is randomly assigned to a treatment group. It could be a person, an animal, a plot of land, or even a petri dish.

Sometimes the experimental and observational units are the same. For example, if you assign a diet to a person and then measure their weight, the person is both. But often they are different. Imagine a study testing new teaching methods. An entire classroom (the experimental unit) might be assigned to a method, but the test scores of individual students (the observational units) are what's measured.

Randomization is the cornerstone of a valid experiment. The physical act of randomly assigning treatments to experimental units is what validates the statistical assumptions we make later. Specifically, it ensures that the errors in our model are independent and identically distributed. Without proper randomization, our statistical tests and confidence intervals are not trustworthy.

Replication, making multiple observations for a given treatment, is equally fundamental. You cannot estimate experimental error with only one observation per treatment. Replication allows us to measure the natural, random variation that occurs among experimental units that receive the same treatment. This estimate of random error becomes the benchmark for determining if the differences we see between treatments are real or just due to chance.

The design structure, dictated by the randomization process, tells us exactly how to model this error. For instance, if experimental units are grouped into blocks (a common technique to control for variation), the randomization is restricted within each block. This physical constraint must be reflected in the statistical model by including a term for the block effect. The analysis must always follow the design.

These structural components all come together in a linear statistical model. This model is a mathematical equation that represents our understanding of how the data were generated. It provides a formal framework for linking the observed response to the different components of the experiment.

In this simple model, we assume the response is a sum of an overall average, a specific treatment effect, and a random error component. By understanding the treatment and design structures, we can formulate the correct model, estimate the treatment effects ($\tau_i$), and properly test our hypotheses using the variation estimated in the error term ($\epsilon_{ij}$). Every decision in the physical conduct of the experiment has a direct consequence on the form and validity of this model.

Building a solid experiment starts with understanding its fundamental structure. These principles ensure that the bridge between your physical research and your mathematical analysis is sound.