Back to the on-screen lesson ·
More figures of the same kind settle nothing; a comparison built so the explanations disagree settles it.
Paper packet. Every task here also exists on screen, where it is checked automatically; answers written on paper are not assessed by Nydus. When you are back at a device, enter your answers there.
You will inspect how a comparison was made, identify a rival explanation it does or does not address, calculate simple observed group differences, and state a causal conclusion with the design's assumptions and limits visible. You will distinguish a volunteer comparison, a confounded treatment-location comparison, and a randomized comparison within blocks.
You can write out the four things that could be true when two things move together. This lesson is what to do next: each of the four predicts something different, and a comparison built on purpose is what makes them disagree out loud.
| Term | What it means |
|---|---|
| Comparison group | A group used to estimate what might happen without the particular intervention. |
| Random assignment | Allocation by chance that avoids deliberately linking treatment to initial characteristics. |
| Confounding | A mixing of influences that prevents a comparison from isolating the proposed cause. |
| Blocking | Making treatment comparisons within groups sharing an important known background feature. |
| Observed effect estimate | The measured difference used as evidence about an intervention, subject to the design's assumptions. |
An association starts a causal question rather than ending it. To investigate whether changing A affects B, identify an important alternative explanation and look for evidence that distinguishes it from the proposed cause. Timing, comparable groups, controlled interventions, and repeated observations can all help, but none is a magic guarantee when its conditions are ignored.
Checking that a proposed cause came first rules out a simple account in which that later event produced the earlier one. It does not remove every common cause. Comparing cases with the same measured background factor can address that factor, while unmeasured differences may remain. Random assignment makes treatment assignment less dependent on existing characteristics and balances groups in expectation; it does not make small groups identical in every respect.
The central question is what the comparison lets you infer under its stated assumptions. A well-designed experiment can provide strong causal evidence. A before-and-after count or a volunteer comparison can still provide useful descriptive information, but additional explanations must be considered before attributing its difference to the proposed intervention alone.
Another way: picture
Draw two groups with treatment assigned by a random process. Keep care, measurement time, and measurement method the same. Mark any departures, missing outcomes, or additional changes as questions that could affect the comparison.
Another way: steps
Specify the causal question and outcome. Name a plausible rival explanation. Design or inspect a comparison that addresses that rival. Calculate the observed difference. Report what the design supports and what uncertainty remains.
Suppose a town installs new lamps in streets with high incident reports. A table of current streets may show more lamps where more incidents are reported. Dates can reveal that reports rose before the lamps were installed. That finding undermines a simple claim that those newly installed lamps caused the earlier increase. It does not tell us the later effect of the lamps. The causal question must specify which outcome period is being explained.
For a before-and-after comparison, record what else changed between the periods. A reading club may start in September just as students return from summer break and receive a new schedule. A rise in scores after the club begins is compatible with an effect of the club, ordinary school instruction, changes in the tested material, or several influences together. A larger number of students can make the average change more precisely estimated without automatically separating those explanations.
A comparison group helps by showing what happened over the same period without the particular intervention. The relevant assumption is that, apart from the intervention, the comparison gives a suitable guide to what would have happened to the treated group. This is not established merely by calling one group a control. Inspect starting conditions, schedules, measurement methods, and selection. A morning class and an evening class may differ in ways that matter to the outcome.
Volunteers create a particular concern. People choosing an extra reading program may differ in motivation, available time, prior achievement, or family support. Those differences can influence later scores. A volunteer comparison can accurately report that participants improved more while still leaving the causal contribution of the program uncertain. Do not describe the report as worthless; describe the distinction between the observed group difference and the causal attribution being proposed.
Matching groups on measured characteristics can reduce some differences. If every treated student is compared with a student of similar starting score and age, those recorded features become less plausible explanations for a remaining difference. However, matching does not automatically address motivation or another unmeasured factor. It can also fail when the measurements poorly capture the relevant characteristic. State which alternatives the comparison actually addresses rather than claiming that everything else has been held fixed.
The best next observation depends on the rival being tested. Dates address simple reverse timing. Measurements of enrolment address a size-related explanation for school resource counts. A within-slope comparison addresses the fact that all fertilized trees happened to be on the sunny slope. These are targeted improvements to an inquiry. They work by making the proposed explanations less interchangeable, not by invoking the word comparison as a guarantee.
For a low-stakes plant experiment, start with similar pots and seedlings and assign a treatment using a random process. Random assignment means the plants do not receive treatment because they already look stronger or occupy a sunnier position. Across repeated assignments, it avoids a systematic connection between those initial characteristics and the treatment label. In one small experiment, chance imbalances can still occur. Record the starting measurements so an obvious imbalance is visible.
Keep the other planned care the same: pot size, water schedule, measurement interval, and the way growth is measured. If the treated pots also receive more water, the outcome difference does not isolate the treatment named in the report. The experiment would be comparing packages of changes. That can be a legitimate question if stated honestly, but it is a different question from the effect of the added feed alone.
A blocking design can address a known environmental difference. Suppose plants occupy a north bench and a south bench. Randomly assign some treated and some comparison plants within each bench. This ensures that treatment is not completely confounded with bench location. Compare treated and untreated outcomes within each bench before combining the results. The method is stronger than treating every south-bench plant and leaving every north-bench plant untreated, where treatment and location cannot be disentangled.
Use a clear outcome measured in the same way for every plant. 'Looks healthier' can be affected by expectations. Height gain over a stated number of days, measured from a fixed reference point, is easier to compare. Where practical, the person measuring can be unaware of which label represents treatment. This reduces one possible measurement bias. It does not fix every design problem, but it addresses a specific route through which expectations could affect the recorded result.
Track every assigned plant and report missing outcomes. If weak plants disappear from only one group's final record, the reported average may describe a selected subset rather than the assigned group. Similarly, contamination can reduce the intended contrast if untreated pots accidentally receive the feed. Record what actually happened instead of treating the plan as proof that it was followed. The strength of an experiment depends on its conduct as well as its initial design.
After collecting the data, calculate an observed difference and keep it distinct from a universal law. A two-centimeter larger average gain in this run is a result for this run under these conditions. Random variation, measurement error, and the limited range of plants still matter. Repeating a well-designed comparison can strengthen evidence about reliability. Replication reduces concern about a one-off chance pattern; it does not logically prove that chance or every hidden problem is impossible.
A useful causal report has four parts: the question, the comparison, the observed result, and the limits. For example: 'We tested whether this feed changes two-week height gain in these seedlings. We randomly assigned treatment within each bench and kept other planned care the same. The treated mean exceeded the comparison mean by two centimeters. The result is evidence under these conditions; further runs would help assess its stability.' Each sentence contributes something different.
Do not hide an uncertain inference behind the word proves. A result can provide meaningful evidence without absolute certainty. Conversely, listing every imaginable doubt can make a well-controlled comparison sound indistinguishable from an unsupported guess. State material limitations tied to the actual study: a small sample, a known measurement issue, one site, or incomplete follow-up. Explain how those limitations bear on the conclusion instead of adding a generic disclaimer that anything might be wrong.
Transporting a result to a new setting is another inference. A feed tested on one plant variety under one watering schedule may behave differently elsewhere. The original experiment can be well designed while the wider generalization remains uncertain. Ask which features are likely to matter to the mechanism and whether the new setting preserves them. Further testing in different conditions addresses that extension more directly than simply repeating the original conclusion more confidently.
Finally, some causal questions cannot be answered by casually assigning people to conditions. In ordinary reasoning, you may need to evaluate existing records, naturally occurring changes, or carefully designed studies rather than conduct an intervention yourself. The same intellectual task remains: identify the comparison, state its assumptions, examine rival explanations, and match the conclusion to the evidence. The classroom plant example makes these principles visible without implying that every important question has an easy experiment available.
A fictional gardening club compares seedling growth over two weeks. In its first attempt, all fed plants sit on the sunny south bench and all unfed plants on the north bench. The fed group grows 8 centimeters on average and the unfed group 4 centimeters. The observed four-centimeter difference is real in the records, but treatment and bench location changed together. The data do not separate those two candidate explanations.
For a better comparison, the club assigns equal numbers of fed and unfed plants randomly within each bench and keeps other planned care the same. On the north bench, fed plants gain 6 centimeters on average and unfed plants 4, a difference of 2. On the south bench, fed plants gain 8 and unfed plants 6, also a difference of 2. With equal numbers in the two blocks, the average within-bench difference is 2 centimeters.
The result is consistent with the feed contributing to growth under these conditions, and it is harder to explain solely by putting every fed plant on a sunnier bench. It does not guarantee an identical gain for every plant or prove that all possible sources of error vanished. Starting sizes, missing plants, measurement consistency, and random variation still belong in the report.
Notice the improvement in the question answered. The first comparison contrasts a feed-and-location package. The second creates treatment comparisons within each location. A planning team can now discuss the feed's observed contribution under the tested care schedule rather than mistakenly treating the entire original four-centimeter gap as its effect. The design has made an important rival explanation less able to account for the result.
Thinking random assignment makes groups identical. It balances in expectation; chance differences can remain in a particular small sample.
Thinking more data never help. More suitable observations can reduce uncertainty, but merely enlarging a confounded comparison does not remove its design problem.
Thinking replication eliminates chance with certainty. Repeated findings can strengthen the evidence without providing a logical guarantee.
Calling a volunteer comparison meaningless. It describes a real difference while leaving selection as a possible explanation of that difference.
State the observation.
Scores rose after the reading club began.
This records a temporal association.
Name the proposed cause.
The reading club caused the rise.
That claim is stronger than saying the club came first.
Identify a concurrent change.
The school term and ordinary lessons resumed at the same time.
Another process could contribute to the later scores.
Specify useful comparison information.
Scores for a suitable group over the same period without the club.
A concurrent comparison can help assess change that might have occurred anyway.
Limit the current conclusion.
A rise is observed; attribution to the club alone remains uncertain.
Temporal order does not by itself remove the competing explanation.
Record the group difference.
Volunteers improved 9 points; nonparticipants improved 3.
The recorded contrast is 6 points.
Inspect how groups formed.
Students chose whether to join.
Treatment assignment may track pre-existing motivation or available time.
Name a relevant rival explanation.
Motivation may influence both joining and later improvement.
A common influence can contribute to the observed difference.
Describe a design improvement.
Use an appropriate random allocation to an offered program where the study permits it.
Random allocation reduces systematic selection into the compared conditions.
Preserve the original finding's proper scope.
Participants improved more; the entire 6-mark gap is not thereby established as the program's effect.
A descriptive result and a causal attribution are distinct claims.
Identify the first comparison's confounding.
All fed plants south; all unfed plants north.
Feed and bench location vary together, so their effects are mixed.
Create both conditions within each bench.
Randomly assign fed and unfed plants on north and south benches.
Each location now contains a treatment comparison.
Calculate the north-bench difference.
6 − 4 = 2 cm.
This compares plants sharing the north-bench setting.
Calculate the south-bench difference.
8 − 6 = 2 cm.
This compares plants sharing the south-bench setting.
Combine equal-sized blocks.
(2 + 2) ÷ 2 = 2 cm average difference.
Equal numbers justify giving the two block differences equal weight.
Report the supported interpretation.
Evidence for a 2-cm mean difference under this design, with uncertainty still to assess.
The comparison addresses bench location without guaranteeing identical effects or eliminating every error.
Record the nominal assignment.
Pots were assigned randomly to feed or no feed.
The assignment method addresses systematic selection at the start.
Check a later departure.
Fed pots also received twice as much water.
Two planned causes changed together during the experiment.
State the comparison actually achieved.
Volunteers in a reading program improve 9 points while nonparticipants improve 3. Which issue prevents attributing the whole difference to the program from this report alone?
Equal-sized bench blocks compare fed with unfed plants. North means are 7 and 4 cm; south means are 10 and 5 cm. Complete the within-bench differences and their equal-weight average.
Subtract within the north bench.
North fed mean minus north unfed mean = north cm.
Both groups in this contrast share the north-bench setting.
Subtract within the south bench.
South fed mean minus south unfed mean = south cm.
Both groups in this contrast share the south-bench setting.
Average the equal-sized block contrasts.
Sum of the two block differences divided by two = average cm.
The block sizes are stated to be equal, so equal weighting is appropriate.
In a recorded plant comparison, treated seedlings gain 7 cm on average and comparison seedlings gain 5 cm. What is the observed treated-minus-comparison difference in centimeters?
Answer:
A plant study randomizes feed within each bench. Care records confirm the same water schedule in both conditions. The measurer does not know which label is feed. Connect each design feature to the particular concern it addresses, without claiming every possible error disappears.
This task has no paper form; do it on a device.
Equal-sized bench blocks are used. On the north bench, treated and comparison mean gains are 6 and 4 cm. On the south bench, they are 9 and 5 cm. What is the equally weighted average of the two within-bench differences, in centimeters?
Answer:
A gardening report puts every fed tree on the south slope and every unfed tree on the north slope. The south-fed group averages 8 units of fruit and the north-unfed group 4. Construct the supported links. No design information separates feed from slope.
This task has no paper form; do it on a device.
Lesson test: one question per skill, one attempt each, no hints. Your answers are checked when you submit.
An experiment randomly allocates otherwise consistently cared-for seedlings within each bench. Complete records give a treated-minus-comparison mean difference of 2 cm. A later audit finds no planned care difference except the feed. Construct the supported links while preserving uncertainty and the tested setting.
This task has no paper form; do it on a device.
You can explain why a group difference alone does not identify its cause, calculate a within-block difference, and say how random assignment and consistent measurement strengthen an inference without guaranteeing identical groups or universal effects.
13. Inspect an experiment before accepting its label, step 3
The result compares feed-plus-extra-water with neither change.
Random assignment alone does not isolate feed when subsequent care differs too.