Experimental Design

Quantitative planning of experiments to isolate variables, estimate effects, and evaluate uncertainty.

Conceptual context

A useful experiment is not merely a large collection of measurements. It is a planned comparison in which one quantity changes, relevant conditions remain controlled, and the observed response can be separated from random variation. In this leaf, the $x$ is deliberately varied and the $y$ is measured under comparable conditions. Replication, execution order, and range determine whether the result can support a physical conclusion. The purpose is not to obtain an attractive line, but to estimate the $\beta_{1}$ with a controlled $u_ybar$ and to decide whether the trend is actually resolvable.

🟢 Essential level

Experimental design organizes measurements so that the intended effect can be separated from accidental variation.

Several levels of $x$ are chosen, the $y$ is measured at each level, and the procedure is repeated with a suitable $n$. Replication provides a $y_bar$ and reveals fluctuation through $s_y$. If $y_bar$ changes systematically as $x$ changes, an effect may exist. That change is meaningful only when other relevant conditions remain controlled. The slope represented by $\beta_{1}$ summarizes how strongly the system responds per unit change. The $\epsilon$ reveals what the model does not explain. Experimental design therefore does not remove uncertainty. It makes uncertainty visible, quantifiable, and useful for deciding whether a physical dependency has actually been detected. It also exposes when the available range or replication is insufficient. Randomizing the order in which the different levels are executed prevents a temporal drift of the system, such as equipment warming up, from being systematically confused with the effect that is actually being measured.

🔵 Formal level

Basic formulas

The linear model separates baseline, effect, and unexplained variation:

Linear response model

The $y$ is interpreted as the combination of $\beta_{0}$, the product of $\beta_{1}$ and $x$, and $\epsilon$. The $\beta_{1}$ has response units per input unit. Its sign indicates whether the response rises or falls as the treatment increases.

Replicates at one condition are summarized through:

Mean of replicates

Real student questions

Do more data always improve an experiment? Only when they provide independent information and cover the relevant range.

Does a positive slope prove causation? No. $\beta_{1}$ supports an association under the model, while causation also requires control of confounding variables.

Why is one measurement per level insufficient? It cannot separate random variability from the underlying trend.

What does overlap between error bars mean? It does not settle the question by itself. The full model and uncertainty of $\beta_{1}$ must be considered.

May I force $\beta_{0}$ to zero? Only with a strong physical and metrological justification.

Final synthesis

Experimental design builds a comparison able to separate effect, noise, and bias. The $x$ must cover an informative range, the $y$ must be measured consistently, and $\beta_{1}$ must be reported with uncertainty, units, and validity limits. Graphs and $\epsilon$ are part of the physical argument.