Causal Assumption Check Prompt

Public field comparison

Causal Assumption Check Prompt

Causal Assumption Check Prompt

Test whether an analysis supports causal claims by checking confounders, selection bias, timing, and alternatives.

Assess whether an analysis can support a causal claim by examining identification assumptions, design-specific threats, diagnostics, and alternative explanations.

—

Use this prompt to audit causal claims, distinguish causal evidence from association, identify unsupported assumptions, and define the analyses needed for a defensible conclusion.

Causal Assumption Check
Requirements Clarification
Output Quality Review
Review Checklist Building
Implementation Planning

Auditing causal claims in observational studies
Reviewing quasi-experimental identification strategies
Checking claim-to-estimand alignment
Building a causal assumption and bias register
Planning design-specific robustness and sensitivity analyses

Goal or task
Current context
Constraints
Files, data, or examples
Definition of done

Causal claim and decision
Study design and estimand
Data and analysis evidence
Domain context and causal assumptions
Constraints and review stakes

Replace every bracketed placeholder before running. Give the model enough context to inspect assumptions, ask only blocking questions, and produce a concrete deliverable. For code prompts, include relevant files, errors, logs, and test commands.

In ChatGPT, replace every bracketed variable with the study-specific information. Provide the causal claim, estimand and design, methods or protocol, causal diagram if available, data dictionary, code, model outputs, diagnostics, sensitivity analyses, and identifiable source locations. Remove or aggregate sensitive personal data, then run the prompt. If required evidence is unavailable, instruct ChatGPT to produce the limited intake assessment rather than a causal verdict.

Use this when you need a production-ready analytics review result in Data Analysis, not a generic brainstorm. The expected output should include findings, implementation steps, risks, and verification checks.

Review a difference-in-differences analysis claiming that a policy reduced hospital admissions. Supply the estimand, treatment and comparison regions, adoption dates, model specification, event-study output, sample rules, concurrent policy history, and robustness results. The review will test parallel-trends support, anticipation, composition changes, spillovers, timing issues, and whether the reported claim matches the identified effect.

Advanced

Advanced

ChatGPT

ChatGPT

analytics review

analytics review

chatgpt
data-analysis
causal analysis
statistics

chatgpt
data-analysis
causal-inference
study design
statistical review

Causal Assumption Check Prompt | AMO.ng

Causal Assumption Check Prompt | AMO.ng

Test whether an analysis supports causal claims by checking confounders, selection bias, timing, and alternatives.

Audit causal claims using estimand alignment, design-specific assumptions, bias checks, diagnostics, sensitivity evidence, and qualified verdicts.

Prompt-body line comparison

Removed Added Unchanged context

−Act as a senior Data Analysis specialist using ChatGPT. Your task is: [Goal or task].
+Review whether the supplied analysis justifies its causal claim.
 
−Context:
−- Current situation: [Current context]
−- Constraints: [Constraints]
−- Available materials: [Files, data, examples, URLs, logs, notes]
−- Success criteria: [Definition of done]
+Inputs
+- Causal claim and decision: [Causal claim and decision]
+- Study design and estimand: [Study design and estimand]
+- Data and analysis evidence: [Data and analysis evidence]
+- Domain context and causal assumptions: [Domain context and causal assumptions]
+- Constraints and review stakes: [Constraints and review stakes]
 
−Workflow:
−1. Restate the objective in operational terms and identify any missing information that would block a reliable answer.
−2. Make reasonable assumptions only when they are low risk, and label them clearly.
−3. Produce the main deliverable for "Causal Assumption Check Prompt" with enough detail that a skilled operator can execute it immediately.
−4. Include edge cases, failure modes, dependencies, and tradeoffs that a junior prompt would usually miss.
−5. Add a verification checklist with concrete tests, review questions, metrics, or acceptance criteria.
−6. End with the smallest safe next action.
+Minimum input contract
+A full review requires: the causal claim; treatment or exposure; outcome; target population; relevant time window; intended estimand; study design or identification strategy; analysis methods; and the results or diagnostics on which the claim relies. Useful additional materials include a protocol, causal diagram, data dictionary, eligibility rules, code, model specifications, balance tables, attrition and missingness reports, robustness analyses, and cited domain evidence.
 
−Output format:
−- Executive summary
−- Detailed plan or implementation
−- Risks and mitigations
−- Verification checklist
−- Next action
+If the claim, estimand, identification strategy, or supporting analysis is missing or internally contradictory, ask only the questions needed to unblock review. If answers are unavailable, produce a limited intake assessment containing the known scope, conflicts, critical unknowns, and a preliminary threat map; do not issue a causal verdict. Make bounded progress on non-blocked sections and preserve unresolved details as unknown rather than filling them in.
 
−Do not give generic advice. Optimize for a production-quality analytics review outcome.
+Tool and evidence boundaries
+Use ChatGPT to inspect only content actually available in the conversation, including uploaded documents, visible tables, code, and reported outputs. Do not imply access to private systems, external links, datasets, software, or analysis runs that were not made available. You may explain calculations or draft diagnostic code, but label it proposed unless execution output is supplied. Never fabricate citations, coefficients, sample sizes, p-values, confidence intervals, diagnostics, or study procedures.
+
+Classify each material statement as one of: supplied fact, observed result, reported execution evidence, assumption, hypothesis, unknown, conflict, or unsupported claim. Cite supplied evidence using available document, section, page, table, figure, code, or output identifiers. State when no precise locator exists. Separate empirical diagnostics from identification assumptions that cannot be proven from the observed data alone.
+
+Causal review procedure
+1. Normalize the claim. Express the treatment or exposure, comparator, outcome, target population, time horizon, unit of analysis, and estimand. Note whether the requested effect is total, direct, controlled direct, intention-to-treat, treatment-on-the-treated, local, or another stated effect. Flag mismatches between the public claim and the estimated quantity.
+2. Reconstruct the causal model. Use any supplied causal diagram; otherwise provide a provisional directed edge list and explain its evidentiary basis. Identify candidate confounders, mediators, colliders, instruments, selection variables, proxies, and post-treatment variables. Mark uncertain edges and plausible alternative diagrams.
+3. Evaluate core assumptions. Address temporal ordering, consistency and treatment-version ambiguity, exchangeability, positivity or overlap, interference and spillovers, measurement validity, missingness, attrition, selection into the sample, model specification, and transportability. Check for reverse causality, conditioning on colliders, inappropriate mediator adjustment, unmeasured confounding, differential measurement error, and outcome-driven model selection.
+4. Apply the relevant design branch:
+   - Randomized study: randomization integrity, allocation concealment where relevant, noncompliance, attrition, contamination, interference, baseline imbalance, analysis population, and outcome switching.
+   - Matching, weighting, stratification, or regression adjustment: covariate timing, confounder selection rationale, propensity specification, overlap, extreme weights, post-adjustment balance, functional form, and residual confounding.
+   - Difference-in-differences or event study: parallel-trends justification, pre-trend evidence and power, no anticipation, stable composition, treatment timing, staggered-adoption estimator issues, concurrent shocks, spillovers, and outcome-specific trends.
+   - Instrumental variables: relevance, independence, exclusion restriction, monotonicity when invoked, weak-instrument evidence, compliance population, and interpretation of the local effect.
+   - Regression discontinuity: assignment rule, cutoff manipulation, continuity, bandwidth and polynomial sensitivity, density and covariate checks, sorting, and local estimand limits.
+   - Interrupted time series or panel/time-series design: baseline trend, seasonality, autocorrelation, structural breaks, concurrent interventions, lag structure, and sufficient pre- and post-intervention observations.
+   - Natural experiment or other observational design: the source of as-if random variation, manipulation risks, institutional details, exposure assignment, and credible counterfactual.
+   Mark non-applicable branches explicitly rather than forcing them into the review.
+5. Examine analysis integrity. Check unit-of-analysis and standard-error alignment, clustering or repeated measures, multiple outcomes and researcher degrees of freedom, missing-data handling, influential observations, transformations, treatment timing, heterogeneous effects, and whether uncertainty estimates match the design. Distinguish statistical significance, effect magnitude, practical relevance, and causal identification.
+6. Build an alternative-explanations register. Include plausible confounding, selection, measurement, secular trends, regression to the mean, concurrent events, differential attrition, reverse causation, spillovers, and specification dependence. For each, describe the causal path, expected bias direction if knowable, available evidence, and a diagnostic or design remedy.
+7. Assess robustness and sensitivity. Reconcile the primary result with reported negative controls, placebo tests, falsification outcomes, pre-trends, balance checks, alternate specifications, bandwidths, samples, lag structures, missing-data analyses, sensitivity to unmeasured confounding, and effect bounds. Do not treat robustness across variations of the same flawed identification strategy as independent confirmation.
+8. Adjudicate the claim using one status: supported within the stated scope, conditionally supportable, associational only, indeterminate, or contradicted by supplied evidence. A supported status requires a clearly matched estimand, a credible identification argument, no unresolved critical design failure, and supplied evidence for relevant diagnostics. Explain that observational diagnostics can increase or decrease credibility but cannot prove all causal assumptions.
+9. Recommend proportionate remediation. Prioritize changes that affect identification before cosmetic reporting changes. Separate required corrections, recommended sensitivity analyses, optional improvements, and issues that cannot be repaired with the available design. Provide appropriately qualified replacement wording for the causal claim.
+
+Safety and authority boundaries
+Treat this as methodological decision support, not approval of a study, publication, policy, clinical action, financial action, or legal conclusion. Do not alter data, execute code, contact participants, submit findings, publish claims, or approve consequential decisions. Those actions require an authorized human. Request de-identified or aggregated evidence where possible; do not reproduce credentials, direct identifiers, or unnecessary sensitive records. Stop and request a safer input if review would require exposing protected data. High-stakes conclusions require review by an appropriately qualified domain expert and statistician. Preserve source materials and analysis provenance; propose changes rather than overwriting evidence.
+
+Required output
+A. Claim and estimand map
+- Original claim
+- Normalized causal question
+- Treatment, comparator, outcome, population, time window, unit, and estimand
+- Claim-to-estimand mismatches
+
+B. Evidence inventory
+Create a table with: evidence item, source locator, evidence class, relevance, reliability limitation, and conflict or missing-status note.
+
+C. Causal structure
+Provide the supplied or provisional causal diagram as an edge list, followed by confounders, mediators, colliders, selection variables, instruments, uncertain edges, and alternative structures.
+
+D. Assumption register
+Create a table with: assumption, why required, design branch, empirically assessable or fundamentally untestable, expected evidence, actual supplied observation, status, consequence if violated, and remedy. Use statuses supported, partially supported, unsupported, violated, unknown, or not applicable.
+
+E. Bias and alternative-explanations register
+Create a table with: threat, causal mechanism, affected estimate, likely bias direction or unknown, severity, supplied evidence, diagnostic, and mitigation.
+
+F. Design and analysis diagnostics
+Report each applicable check with its expected observation, actual supplied observation, evidence locator, interpretation, and status. Clearly mark diagnostics that were merely recommended, not run.
+
+G. Robustness and sensitivity reconciliation
+List each analysis as reported, proposed, unavailable, or blocked. For reported analyses, state the primary and sensitivity results, whether they agree, and what threat the comparison addresses. Explain unresolved discrepancies.
+
+H. Causal claim adjudication
+Give the status, confidence level with rationale, strongest supporting evidence, decisive weaknesses, scope limits, and qualified replacement wording. Do not use causal proof language.
+
+I. Remediation and handoff
+Separate required corrections, recommended analyses, optional improvements, and non-repairable design limitations. For every proposed analysis, specify its purpose, required inputs, expected diagnostic pattern, decision rule, and responsible human reviewer.
+
+J. Verification and acceptance matrix
+Create a table with: acceptance check, expected evidence, actual evidence, pass, fail, blocked or not applicable status, and unresolved action. At minimum verify claim-estimand alignment, temporal ordering, adjustment-set validity, overlap where relevant, selection and attrition handling, design-specific assumptions, uncertainty estimation, robustness reconciliation, evidence traceability, and wording proportionality. The review is acceptable only if every material claim is traceable, every critical assumption has a status, reported and proposed work are distinct, contradictions are reconciled or left explicit, and the adjudication follows from the registers. Otherwise label the review incomplete or blocked.
+
+Completion integrity
+Use completed, tested, measured, verified, approved, or executed only when the supplied materials contain corresponding evidence. Otherwise use proposed, reported but not independently verified, unavailable, or blocked. End with the smallest safe next action that would reduce the most consequential unresolved causal uncertainty.