# Causal Reasoning and Research Design Distinguish a predictive association from a defensible causal claim. Use this workbook alongside the course. Write your answers before opening the solutions. Practical work is self-reviewed; scored knowledge checks are in the Academy. ## 1. Causal framing ### Treatment and outcome A causal question asks about the effect of a specified intervention or exposure, not merely whether two variables move together. Define treatment, outcome, unit and horizon before choosing a method. ### Confounders A confounder influences both the treatment and outcome and can create a misleading association. A causal diagram makes the assumed relationships explicit. The diagram is a statement of assumptions, not proof that they are correct. ### Mediators A mediator lies on a proposed causal pathway. Controlling for it can change the estimand from a total effect to another quantity. Adding every available variable to a regression is not automatically better causal analysis. ### Selection mechanisms Selection mechanisms determine which observations are included. Conditioning on a variable influenced by several causes can introduce bias. Explain why the sample and controls correspond to the causal question rather than relying on correlation-based selection alone. ### Worked example A policy announcement coincides with both currency and gold movement. Treating the currency move as the sole cause of gold's move ignores the possibility that the announcement directly influenced both. ### Independent exercise Draw a simple diagram with the announcement, currency and gold. State two different causal stories consistent with the observed co-movement. My inputs and assumptions: My calculation or decision: Evidence that would change my conclusion: ## 2. Designs ### Natural experiments A natural experiment uses a source of variation argued to be sufficiently unrelated to other outcome determinants for a defined question. The argument must be explained; calling an event natural does not make assignment random. ### Before-after limitations Before-after comparisons can confound the event with trends, seasonality or simultaneous changes. A change after an announcement is not necessarily caused entirely by the announcement. ### Difference-in-differences assumptions Difference-in-differences compares changes across treated and comparison groups under assumptions such as an appropriate counterfactual trend. Group selection, timing and heterogeneous effects matter. A formula alone does not validate the design. ### Instrumental-variable assumptions An instrumental-variable design requires a justified relationship with treatment and restrictions on other paths to the outcome. A variable that predicts treatment strongly is not automatically a valid instrument. ### Worked example Gold rises after an announcement, but another major release occurs in the same window. A before-after difference cannot uniquely assign the rise to the first announcement. ### Independent exercise Describe what additional design evidence would be needed before making a causal attribution, and identify one assumption that may remain untestable. My inputs and assumptions: My calculation or decision: Evidence that would change my conclusion: ## 3. Diagnostics ### Pretrends Pretrend checks examine whether groups behaved differently before a proposed treatment. They can reveal problems but cannot prove the unobserved post-event counterfactual. Low power in a short preperiod is also a limitation. ### Placebo tests Placebo tests apply the design to times or outcomes where the claimed effect should not appear under the theory. A failure can challenge the design; a pass is supportive evidence, not a universal guarantee. ### Sensitivity to omitted variables Sensitivity analysis asks how conclusions change under plausible omitted influences or alternative specifications. Choose variations for substantive reasons and report them rather than searching only for stable-looking results. ### Measurement error Measurement error can distort treatment, outcome and timing. Misrecorded event times may smear responses across windows. Audit data provenance before interpreting sophisticated estimates. ### Worked example A study finds a supposed effect before the treatment occurred. This may indicate anticipation, timing error or an invalid design. It cannot be ignored solely because the post-event estimate is attractive. ### Independent exercise Write an investigation plan for that pre-event effect with at least two competing explanations. My inputs and assumptions: My calculation or decision: Evidence that would change my conclusion: ## 4. Communication ### Identify untestable assumptions Identify assumptions that cannot be directly tested and explain why they may be plausible or doubtful. A reader should know where the conclusion depends on judgment rather than observed evidence. ### Separate prediction and explanation Separate prediction from explanation. A variable useful for forecasting need not be a manipulable cause, and a causal effect may not yield a profitable forecast once timing and costs are considered. ### Report bounded conclusions Bound conclusions to the studied population, period and intervention. An estimate in one institutional setting does not automatically transfer to every gold market or future policy regime. ### Avoid unsupported policy claims Avoid converting a limited market study into a broad policy claim without the necessary welfare, mechanism and external-validity analysis. State the narrower question actually addressed by the evidence. ### Worked example A well-supported effect in one event setting is presented as proof that every similar announcement will move gold in the same direction. The generalization exceeds the identified sample and intervention. ### Independent exercise Rewrite that conclusion with a bounded claim and a concrete follow-up study for transferability. My inputs and assumptions: My calculation or decision: Evidence that would change my conclusion: ## Course project Draw a causal diagram and explain the identification limits of a proposed gold-market study. ### Self-review rubric - Concepts and reasoning: 25% - Calculations, data and evidence: 30% - Process and risk controls: 25% - Limitations and communication: 20% Record one correction and one next practice task. This rubric is not automatically graded. ## Worked solutions ### Exercise 1 One story has the announcement affect both directly; another includes an additional currency-to-gold path. The same observed association does not distinguish them without further assumptions or design evidence. ### Exercise 2 Seek a credible comparison or source of variation, account for simultaneous information and specify the counterfactual argument. Assumptions about how the outcome would have behaved without the event cannot generally be observed directly for the same unit and time. ### Exercise 3 Check public information timing and anticipation, verify timestamps and assess whether the comparison groups or trends are appropriate. Report what the investigation supports and retain uncertainty where the alternatives cannot be separated. ### Exercise 4 State the estimated effect for the defined setting under the listed assumptions, then propose a separately specified sample of other events or regimes. Treat transfer as a new question rather than a guaranteed property. ## Further reading - https://www.itl.nist.gov/div898/handbook/ - https://miguelhernan.org/whatifbook