MGMT7250 Chap.6 From Association to Causal Claims
From Association to Causal Claims
Causal reasoning compares the observed outcome with an unobserved counterfactual for the same unit. Because both potential outcomes cannot be seen, design must create a credible comparison. This chapter separates association, temporal precedence and control of alternative explanations, then uses causal diagrams to expose confounding.
It compares random assignment, matching, stratification and weighting, with attention to the assumptions and residual alternatives each method leaves open.
Define the treatment contrast before analysing outcomes. Assignment, uptake, dose and the alternative condition can differ, and the resulting estimate belongs to that particular comparison.
Time order is necessary but not sufficient: common causes can affect both exposure and outcome, while the outcome may influence exposure through reverse causation.
A causal diagram records assumptions before coefficients influence the story. Identify plausible common causes, mediators and colliders, then decide which variables should be measured and when.
Random assignment seeks exchangeability by breaking systematic links to prior causes on average. Matching, stratification and weighting pursue comparability on measured characteristics, so their credibility depends on adequate measurement and overlap.
A model cannot create a comparison where treated and untreated units occupy entirely different regions.
Report the population, contrast, effect magnitude, uncertainty and specific remaining vulnerability. Statistical significance cannot establish managerial importance or transfer to another context. Distinguish whether the design supports the effect internally from whether mechanisms and implementation conditions travel.
A sensitivity thought experiment asks how strong an unmeasured factor would need to be to reverse the result. End with the next discriminating observation, such as a randomised offer, improved baseline measure or replication in a different setting. This makes uncertainty an acquisition plan rather than a generic disclaimer.
What this chapter covers
- 01
Potential outcomes and counterfactuals
- 02
Association and temporal precedence
- 03
Common causes and confounding
- 04
Exchangeable comparison groups
- 05
Random assignment and observational design
- 06
Residual alternative explanations
Worked example · free
Explain a reversal after stratification
- 2Identify case complexity as a common cause of coaching and resolution.
- 2Compare coached and uncoached agents within complexity bands.
- 1Report the adjusted direction without treating stratification as randomisation.
- 1Name motivation or supervisor selection as residual alternatives.
Key terms
- Counterfactual
- A Counterfactual is the unobserved outcome a unit would have experienced under the alternative exposure or action.
- Confounding
- Confounding is distortion created when a common cause affects both exposure and outcome.
- Exchangeability
- Exchangeability means comparison groups can stand in for each other's potential outcomes apart from treatment.
From Association to Causal Claims FAQ
Why is association insufficient for causation?
The exposure may follow the outcome, or another variable may cause both. A causal claim needs the cause to precede the effect and a design or assumptions that make the comparison approximate the missing counterfactual.
What does random assignment contribute?
It breaks systematic links between treatment and prior characteristics on average, supporting exchangeability. Noncompliance, attrition and chance imbalance can still threaten the estimate, so implementation and analysis remain important.
Can matching remove all confounding?
Matching can improve comparability on measured variables with adequate overlap. It cannot balance unmeasured causes automatically, and poor measurement or model choices can leave residual differences that should be reported explicitly.
Assessment move
For each causal claim, write the observed exposure, outcome and time order. Draw plausible common causes, then decide how the comparison was created. Explain what randomisation, matching, stratification or weighting would buy and what could still break. Close with one sentence that states the estimated direction and the specific alternative explanation left open. Rehearse with one randomised and one observational example.
In each, define treatment versions, identify noncompliance or selection and inspect whether groups overlap. Practise deciding why a mediator or collider should not automatically be controlled. Then write two conclusions: an internal-validity claim for the studied population and a transfer claim for a new context. Add a sensitivity sentence naming an unmeasured factor and the strength it would need to reverse the result.
This sequence keeps design reasoning ahead of model output. Redraw one example after changing the timing of a variable and explain how its adjustment role changes. Inspect the assignment mechanism before reading any outcome difference, then predict the direction of confounding. Construct a case with poor overlap and state the narrower population for which the comparison remains supportable.
Translate a coefficient into a bounded causal sentence containing contrast, period, population and vulnerability. Finish with the single new observation most capable of separating the causal account from its strongest rival.
Working through From Association to Causal Claims in MGMT7250? Sia is AskSia’s AI Management tutor — ask any MGMT7250 From Association to Causal Claims question and get a clear, step-by-step explanation grounded in how MGMT7250 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.