Correlation, Causation, and Confounding
Separate association from causal evidence by identifying temporal order, confounders, mediators, selection effects, and alternative explanations.
Separate association from causal evidence by identifying temporal order, confounders, mediators, selection effects, and alternative explanations.
Core science
Correlation describes variables changing together; it does not identify why. Temperature and irrigation volume may rise together because hot weather causes both. Diseased plants may receive more treatment because symptoms prompted intervention, creating an association between treatment and poor outcome even when the treatment did not cause the disease. Reverse causation and management response are common in observational cultivation records.
A confounder influences both the proposed cause and outcome. Cultivar, plant size, room position, root-zone volume, crop stage, worker, season, and prior health can confound comparisons. A mediator lies on the causal pathway and should not always be “controlled away.” Conditioning on a collider—an outcome influenced by two variables—can create a false association. Causal diagrams help state assumptions before analysis.
Strong causal evidence combines a defined intervention, suitable control, randomization or credible adjustment, temporal order, mechanism, dose or gradient where relevant, replication, and consistency with alternative evidence. Observational data remain valuable for surveillance and hypothesis generation. Models can adjust only for measured, correctly modeled variables; a high correlation coefficient or predictive accuracy does not prove the intervention will work.
Why this matters in cultivation
- Before stating “X caused Y,” draw the plausible pathways and list competing explanations. Phrase conclusions at the level supported: observed association, treatment effect under the experiment, mechanism, or transfer hypothesis.
Measure and record
Record 1
Define exposure, outcome, population, time order, and the causal question explicitly. Draw or document plausible confounders, mediators, colliders, selection mechanisms, and alternative explanations before choosing adjustment variables.
Record 2
Record allocation or intervention status, controls, repeated measures, missingness, model specification, effect estimates, uncertainty, sensitivity analyses, and whether key variables were measured before or after the exposure.
Record 3
Match the final wording to the design: use association language for observational evidence unless stronger causal assumptions are justified, and state which assumptions remain untested.
Common misconceptions
Evidence limits and uncertainty
Causal conclusions depend on design plus assumptions about confounding, selection, measurement, interference, and model structure; some assumptions cannot be verified from the observed data alone.
Sensitivity analysis can show how conclusions change under alternative assumptions but cannot prove that unmeasured bias is absent.
Check your reasoning
- For "Correlation, Causation, and Confounding", explain the mechanism behind this objective: Separate association from causal evidence by identifying temporal order, confounders, mediators, selection effects, and alternative explanations. Which observation or measurement would best test whether that mechanism is operating in the real crop?
- A learner claims, "A strong correlation proves a strong causal effect." Use the lesson’s science and evidence limits to explain why that claim is unreliable, then name one observation or measurement that could separate the competing explanations.
- Applied case — Before stating “X caused Y,” draw the plausible pathways and list competing explanations. Phrase conclusions at the level supported: observed association, treatment effect under the experiment, mechanism, or transfer hypothesis. Build a verification plan using the lesson’s record set (Exposure and outcome definitions; time order; population; candidate confounders/mediators/colliders; allocation; control; repeated measures; missingness; model; effect estimate/uncertainty; sensitivity checks; alternative explanations; causal wording decision.). What would you compare before and after the action, and what result would make you revise the original interpretation?
Require lesson-specific evidence, not memorized universal targets. Open the rationales after you have written or discussed your own answer.
Answer rationale 1: Mechanism / workflow rationale
- A strong answer should connect the response to the lesson objective: Separate association from causal evidence by identifying temporal order, confounders, mediators, selection effects, and alternative explanations.
- Correlation describes variables changing together; it does not identify why. Temperature and irrigation volume may rise together because hot weather causes both. Diseased plants may receive more treatment because symptoms prompted intervention, creating an association between treatment and poor outcome even when the treatment did not cause the disease. Reverse causation and management response are common in observational cultivation records.
- A confounder influences both the proposed cause and outcome. Cultivar, plant size, room position, root-zone volume, crop stage, worker, season, and prior health can confound comparisons. A mediator lies on the causal pathway and should not always be “controlled away.” Conditioning on a collider—an outcome influenced by two variables—can create a false association. Causal diagrams help state assumptions before analysis.
- The most useful verification evidence includes Define exposure, outcome, population, time order, and the causal question explicitly. Draw or document plausible confounders, mediators, colliders, selection mechanisms, and alternative explanations before choosing adjustment variables..
- Keep this limit explicit: Causal conclusions depend on design plus assumptions about confounding, selection, measurement, interference, and model structure; some assumptions cannot be verified from the observed data alone.
Answer rationale 2: Misconception rationale
- The shortcut is unreliable because the lesson explicitly teaches a more conditional explanation.
- Representative misconception: A strong correlation proves a strong causal effect. Correlation can be produced by confounding, reverse causation, selection, shared trends, or measurement structure.
- Correlation describes variables changing together; it does not identify why. Temperature and irrigation volume may rise together because hot weather causes both. Diseased plants may receive more treatment because symptoms prompted intervention, creating an association between treatment and poor outcome even when the treatment did not cause the disease. Reverse causation and management response are common in observational cultivation records.
- A useful discriminator is Record allocation or intervention status, controls, repeated measures, missingness, model specification, effect estimates, uncertainty, sensitivity analyses, and whether key variables were measured before or after the exposure..
- Do not overextend the conclusion beyond this limit: Causal conclusions depend on design plus assumptions about confounding, selection, measurement, interference, and model structure; some assumptions cannot be verified from the observed data alone.
Answer rationale 3: Applied verification rationale
- In practice: Before stating “X caused Y,” draw the plausible pathways and list competing explanations. Phrase conclusions at the level supported: observed association, treatment effect under the experiment, mechanism, or transfer hypothesis.
- Record before action: Define exposure, outcome, population, time order, and the causal question explicitly. Draw or document plausible confounders, mediators, colliders, selection mechanisms, and alternative explanations before choosing adjustment variables..
- Also record: Record allocation or intervention status, controls, repeated measures, missingness, model specification, effect estimates, uncertainty, sensitivity analyses, and whether key variables were measured before or after the exposure..
- After the action, repeat the same measurement or observation so the comparison is valid.
- Revise the interpretation if the result conflicts with the lesson limit or the expected response: Causal conclusions depend on design plus assumptions about confounding, selection, measurement, interference, and model structure; some assumptions cannot be verified from the observed data alone.
Worked example: Higher light is correlated with higher yield
Scenario: Historical production data show that crops with higher measured light also had higher yield, and the team concludes that increasing light caused the yield increase.
Reasoning path
- Confirm temporal order and how both light and yield were measured.
- List variables that changed with light, including cultivar, plant density, CO2, nutrition, crop age, room, season, and management.
- Distinguish confounders from mediators and from consequences of the treatment.
- Use a controlled or quasi-experimental comparison when a causal claim matters.
- Phrase the historical result as an association unless the design rules out credible alternative explanations.
Evidence to collect
- raw light measurements and locations
- yield definition and sampling
- cultivar and crop-stage data
- room/season identifiers
- CO2/nutrition/density records
- management changes
- controlled comparison data if available
Common weak answers
- Correlation proves the higher light caused the yield change.
- A statistically strong relationship cannot be confounded.
- Adding more variables after seeing the result automatically establishes causality.
Verification: A causal interpretation becomes more credible when treatment precedes outcome, assignment or design reduces confounding, measurements are comparable, and alternative explanations fail to account for the observed effect.
Applicability boundary: Observational cultivation data are valuable for hypothesis generation and forecasting but often cannot support causal claims without additional design or analysis.
Related lessons
Sources and evidence
- NIST/SEMATECH e-Handbook of Statistical MethodsV21-SRC-009
Primary reference for experimental design, measurement process characterization, control charts, uncertainty, regression, and statistical graphics.
- Montgomery — Design and Analysis of ExperimentsV21-SRC-030
General experimental-design reference for randomization, replication, blocking, factorials, interactions, and model checking.
- Gelman and Hill — Data Analysis Using Regression and Multilevel/Hierarchical ModelsV21-SRC-032
Confounding, partial pooling, repeated measures, multilevel data, and model-based interpretation.
- THC Master Content Compilation and Merge Register v1.0V21-SRC-034
Permanent IDs, completion definitions, review gates, source hierarchy, visual requirements, page package, and release-state separation.
Internal controlled file
Downloads
No lesson-specific download is approved for this release. Use browser print/save-to-PDF when you need an offline reading copy.