Skip to content

Experimental Design

Introduction to Experiments

Experiments

Components:

  • Experimental unit
  • Treatment

Variables:

  • Explanatory variable
  • Responsive variable

Observational Study

Do not influence the individual

Types:

  • Retrospective
    • Data is collected then analyzed
  • Prospective
    • A sample is observed for a period of time into the future

Well-Designed Experiments

Confounding Variables

graph LR;

A[Explanatory] --> C[Responsive]
B[Confounding] --> C
A --> B
  • In an experiment, confounding values need to be controlled.
  • In an observational study, confounding values should not and cannot be controlled

Well-Designed Experiments

  • \(\geq 2\) treatment groups (a control group counts)
  • Randomization: Treatment groups are formed by randomly assigning treatments
  • Replication: \(\geq 1\) experimental unit per treatment group
  • Confounding variables are controlled

Control Groups, Placebos & Blind Experiments

Control Groups

A treatment group in an experiment that are purposefully given either

  • an inactive form of the treatment
  • or the pre-existing treatment

Placebo

A dummy treatment that the experimental units believe is an active treatment

  • Control groups can use placebos as inactive forms of treatment

The placebo effect is the response made by experimental units to a placebo

Blind Experiments

A single-blind experiment is the one in which the participants do not know which treatment they are receiving

  • Helps reduce the placebo effect

A double-blind experiment is the one in which participates and researchers do not know which groups are assigned which treatments

  • Helps reduce the placebo effect
  • Reduce any bias

Completely Randomized Design

All experiment units are randomly assigned

  • Creates treatment groups that are as similar as possible
  • Helps balance out any effects of confounding variables across the groups

Randomized Block & Matched Pairs Design

Block

A group of experimental units who have something in common that may affect how they respond to a treatment

  • It should only be done if the researcher believes the blocking variable could affect the results
  • Blocking allows natural variations in response to treatments to be distinguished from those variations that were due to the blocking variable

Randomized Block Design

  1. Experimental units are separated out into blocks
  2. Experimental units are randomly assigned the different treatments within each block
  3. Each block needs to be randomly assigned all of the treatments
  • A randomized block design is generally better than a completely randomized design
  • If blocking variables are unknown or the sample size is very large, completely randomized designs should be used
    • Larger samples tend to introduce more blocking variables
    • So each block has smaller sample size

Matched Pairs Design

  • Each block has only two experiment units which are matched either naturally or manually based on some common factor
  • Two treatments
  • A matched pairs design is better than completely random design
    • Easier to distinguish the effectiveness of the treatment