COMM5000 Chap.8 Causality, Privacy and Data Ethics
Causality, Privacy and Data Ethics
Distinguish association from intervention effects, evaluate identification assumptions, and connect privacy, governance, fairness and monitoring to responsible business action. A causal question compares potential outcomes under different interventions for the same target units. Because both outcomes are not observed simultaneously, design and assumptions determine whether another group supplies a credible comparison.
Random assignment can balance confounders on average, while observational studies require an explicit argument about selection, timing, measurement and common causes. Regression adjustment alone does not guarantee that every relevant pathway has been addressed. Responsible analysis begins before modelling.
Define the legitimate purpose, collect only necessary data, control access, set retention and consider whether people understand or can challenge the use. Evaluate performance and error distribution across groups affected by the decision. A model can be accurate overall while imposing concentrated costs.
Communication should distinguish description, prediction and causal evidence, document uncertainty and prevent use outside the validated population or context. Deployment needs an accountable owner, monitoring metrics, review triggers and a retirement condition. Ethical reasoning is operational when it changes collection, modelling, communication or action.
A generic paragraph cannot repair an analysis whose purpose, data or decision process created avoidable harm.
What this chapter covers
- 01
Counterfactual questions
- 02
Randomisation and confounding
- 03
Privacy and data minimisation
- 04
Fairness and accountability
- 05
Monitoring and review
Worked example · free
Audit a feature-effect claim
- +1The four marks used here are not an official university marking scheme. Identify self-selection between users and non-users.
- +1List plausible confounders such as motivation and prior savings.
- +1Rewrite the result as a 46-dollar observed mean association.
- +1Propose an appropriate stronger design and harm monitoring.
Key terms
- Counterfactual
- The potential outcome for the same unit under an alternative intervention or exposure.
- Data minimisation
- Collecting and retaining only data necessary for a defined and legitimate purpose.
- Model monitoring
- Ongoing review of performance, drift, errors and harms after deployment.
Causality, Privacy and Data Ethics FAQ
Can regression prove causality?
Regression can adjust for included variables, but causal interpretation requires a credible design and assumptions that address confounding, selection, timing, measurement and the target intervention.
When should ethics be reviewed?
Review begins when the question and collection plan are designed, continues through modelling and communication, and remains active after deployment through monitoring, accountability and reconsideration triggers.
What belongs in a responsible analytics record?
Record the purpose, target population, data period, variables, exclusions, design assumptions, model version, evaluation evidence and known limitations. Identify who authorised collection, who can access the data, who approves action and who can stop use. Include retention and deletion rules, subgroup error checks, an explanation or correction route where appropriate, monitoring metrics and a review trigger.
State the approved use and prohibited extensions so a later team does not apply the result outside its evidence boundary. Retirement should cover operational data, downstream systems and the audit information needed for accountability.
Assessment move
For every recommendation, state the intervention, comparison, target population and identification assumptions. Rewrite observational causal claims as precise associations, then identify the stronger design needed to estimate an intervention effect. Draw plausible confounding paths and ask whether each proposed control is measured before treatment and appropriate for the estimand.
Build a governance checklist covering purpose, necessity, consent or notice, minimisation, access, retention and deletion. Add an impact table naming affected groups, benefits, error costs, appeal or correction routes and the person accountable for review. Distinguish average accuracy from subgroup performance and consider whether a technically correct model is being used outside its validated context.
Write monitoring metrics and a trigger that would pause or revise deployment. Preserve a decision record with data period, exclusions, model version, evidence boundary and approved use. Before submission, examine every verb: describes for patterns, predicts for forecasts, and estimates a causal effect only when the design and assumptions justify that level of claim.
Close with a realistic review and retirement plan rather than a vague promise to monitor. Practise a feature-rollout case by listing self-selection mechanisms before proposing any adjustment. Write the supported observational sentence first, then a research plan that could identify an intervention effect.
For privacy, compare precise raw data with an aggregated alternative and explain whether the decision actually needs the extra detail. For fairness, examine who experiences false positives, false negatives and delayed correction. For accountability, name a person or role with authority to stop use. Create a retirement checklist covering deletion, archived audit evidence, communications and downstream systems.
Review the analysis again when the population, purpose, data source or policy environment changes, even if headline accuracy remains stable.
Working through Causality, Privacy and Data Ethics in COMM5000? Sia is AskSia’s AI Statistics tutor — ask any COMM5000 Causality, Privacy and Data Ethics question and get a clear, step-by-step explanation grounded in how COMM5000 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
