Use design to support a causal claim

Try

Students who attend an optional revision session score higher than those who do not. Is the session responsible? List at least two preexisting differences that could affect attendance and performance. Motivation and prior preparation are plausible candidates. An observed association is a useful clue, but the comparison has not separated the session from those other influences.

Predict

Predict what random assignment can accomplish and what it cannot. It can make treatment groups comparable by a chance mechanism rather than participant choice. It does not guarantee identical groups in one realized assignment, ensure everyone follows instructions or make a volunteer sample representative of all students.

Experiment

Design an original synthetic study with 48 volunteers and two revision methods. Assign 24 to each method using a shuffled list of participant identifiers. Give both groups the same study duration and an assessment scored under a common rule. Before seeing outcomes, specify the response as the assessment score rather than selecting whichever outcome looks most favorable.

Now suppose baseline skill strongly predicts scores. Divide participants into prespecified skill blocks and randomize within each block, balancing allocations as closely as possible. Draw a flowchart: volunteers → baseline blocks → random assignment within blocks → common measurement. Compare this with assigning all high-skill students to one method; that alternative would confound method and baseline skill.

Understand

An observational study records naturally occurring exposures. An experiment assigns treatments. A confounding variable is associated with the explanatory variable and affects the response, making their separate contributions difficult to identify. Random assignment helps break such systematic links in expectation, including links with unmeasured variables. Replication across experimental units provides information about variation; taking many measurements of one treated person is not the same as assigning many people.

A control or comparison group gives a reference against which to interpret change. A placebo can help separate the treatment component from expectations when appropriate. Blinding participants or outcome assessors reduces certain behavioral or measurement effects; it may be infeasible for visible treatments. Describe what is actually blinded rather than assuming every study is double blind.

Blocking groups similar units before assignment. Matched pairs are a special structure: two closely matched units receive contrasting treatments, or one unit receives both. With the same participant using two keyboard layouts, randomize order and provide comparable practice. Carryover and learning can otherwise masquerade as a layout effect. The analysis should preserve the pairs.

For four after-minus-before changes 2, −1, 5 and 2, the mean change is 8/4=2 units. These are four differences, not eight unrelated measurements. However, a before-and-after change without a suitable control does not by itself isolate the treatment from time trends.

Random selection and random assignment answer different questions. Selection can support generalization to a population; assignment can support a causal comparison in the study setting. A well-designed randomized experiment on volunteers still needs care when transporting results beyond those volunteers.

Apply

Repair a flawed study in which one class receives a new method and another class receives the old one. Teacher, class and method may be inseparable. Propose an assignment that respects practical constraints, state its experimental unit, and identify what cannot be blinded. Include a plan for recording noncompliance without silently deleting inconvenient participants.

Continue

Next summarize categorical outcomes using tables. Keep the collection design alongside the table: no arrangement of percentages can retroactively create random assignment.