A causal claim concerns a counterfactual contrast, not merely an observed association, temporal sequence, or predictive relationship.
Part IV · Chapter 10
Cause, Comparison and the Counterfactual
A causal claim concerns a counterfactual contrast, not merely an observed association, temporal sequence, or predictive relationship.
Learning objectives
What this chapter asks you to be able to do.
- Distinguish description, association, prediction, causal effect, causal mechanism, and normative judgment.
- Explain the counterfactual and potential-outcomes logic of causal inference without treating notation as a substitute for design.
- Define a causal estimand in terms of intervention, outcome, population, time horizon, and comparison condition.
- Explain why randomized assignment can support causal inference and why attrition, noncompliance, treatment versions, spillovers, and implementation failure can still weaken an experiment.
- Analyze the ethical and institutional authorization requirements of real-world public experimentation.
- Identify confounding, selection into treatment, and the risks of conditioning on post-treatment variables.
- Explain the identifying logic and principal assumptions behind natural experiments, instrumental variables, regression discontinuity, difference-in-differences, matching, and regression adjustment.
- Use comparative case analysis and process tracing to investigate mechanisms and rival explanations.
- Distinguish average effects from heterogeneous and distributional effects, and internal validity from transportability.
- Design robustness, sensitivity, placebo, falsification, and negative-control checks that test assumptions rather than merely decorate a preferred result.
- Evaluate AI assistance in causal research without mistaking generated diagrams, code, or methodological confidence for identification.
- Extend the cumulative Asterbridge Decision Dossier with a Causal Claim and Identification Record.
Chapter summary
The argument in inspectable form.
The same unit does not ordinarily reveal both potential outcomes at the same time; causal designs construct informative comparisons under assumptions.
A causal estimand must specify intervention, treatment version, outcome measure, population, comparison condition, and time horizon before an estimator is selected.
Random assignment can protect an assignment contrast, but randomization does not eliminate attrition, noncompliance, measurement error, treatment variation, spillovers, or external-validity problems.
Research usefulness does not create ethical or legal authority to experiment on people.
Observational causal inference must address selection into treatment and confounding. More controls do not automatically reduce bias, especially when variables are post-treatment or colliders.
Natural experiments are credible only when the actual assignment process supports the causal comparison; natural does not mean random by definition.
Instrument relevance is not instrument validity. Exclusion and independence assumptions remain load-bearing, and IV estimands may be local.
Regression discontinuity can identify a local effect near a threshold under continuity and non-manipulation conditions; it is not automatically a global effect.
Difference-in-differences relies on a defensible counterfactual trend. Pre-trend diagnostics inform but do not prove parallel trends, and staggered timing requires care.
Matching and regression adjustment can improve observed comparability but cannot guarantee balance on unmeasured confounders.
Comparative and process-tracing approaches contribute causal evidence by using cases, sequences, and diagnostic observations to test rival explanations.
Effect and mechanism are different claims. A credible average effect does not prove the mechanism, and mechanism evidence does not automatically establish a population-average effect.
Interference is common in public policy because one person's or district's treatment can affect another's outcome.
Average effects can hide treatment-effect heterogeneity, distributional losses, and subgroup-specific mechanisms.
Internal validity is not transportability. Causal effects must carry population, jurisdiction, time, scale, and treatment-version limits.
Robustness and falsification should test assumptions transparently rather than serve as a specification search for agreement.
Large administrative datasets can improve causal analysis but scale does not solve confounding or policy-induced selection.
AI can assist causal research, but generated code, diagrams, and methodological language require human verification of estimands, causal structure, variable timing, and specification search.
Current NousPolis N0 architecture supports typed causal claims, methods, assumptions, experiments, evidence, and provenance, but Chapter 10 identifies a live design challenge: counterfactual assumptions must remain load-bearing dependencies of the causal estimate rather than optional prose.
Causal finding, normative judgment, legal authority, and democratic authorization remain separate.
Key terms
Concepts to carry forward.
- causal effect
- A contrast between outcomes under different interventions or exposure states, defined for a specified unit, population, and time horizon.
- counterfactual
- An outcome that would have occurred under an alternative intervention or exposure condition that was not actually realized for the same unit at the same time.
- potential outcome
- The outcome a unit would exhibit under a particular treatment or exposure condition in the potential-outcomes framework.
- causal estimand
- The precise causal quantity a study seeks to learn, such as an average effect for a defined population or treatment group.
- treatment / exposure
- The intervention, policy, condition, event, or state whose causal effect is being studied; a label is inadequate if materially different versions are bundled together.
- confounder
- A pre-treatment factor associated with both treatment or exposure and the outcome in a way that can distort a causal comparison.
- identification
- The set of design and substantive conditions under which an observed data distribution can support a particular causal estimand.
- random assignment
- An assignment mechanism that uses chance to allocate treatment, permitting probabilistic comparison under correct implementation.
- intention-to-treat
- A causal contrast based on assignment to treatment rather than treatment actually received.
- natural experiment
- A study that exploits externally generated treatment variation or an assignment process not designed by the researcher; its causal credibility depends on the specific assignment logic, not on the word natural.
- instrumental variable
- A variable used to generate variation in treatment under strong assumptions connecting the instrument to treatment and restricting other pathways to the outcome.
- regression discontinuity
- A design that uses a rule-based treatment threshold and compares units close to the cutoff under a continuity-type assumption.
- difference-in-differences
- A design that compares outcome changes over time between treated and comparison groups, relying on a defensible counterfactual trend assumption.
- overlap / common support
- The condition that relevant types of units have meaningful comparison possibilities across treatment states within the target population.
- interference
- A condition in which one unit's treatment or exposure can affect another unit's outcome.
- treatment-effect heterogeneity
- Variation in causal effects across people, groups, contexts, treatment versions, or time.
- internal validity
- Credibility of the causal claim for the studied population, setting, and design.
- external validity / transportability
- The extent to which a causal claim can be credibly extended to other populations, jurisdictions, times, scales, or treatment versions.
- mechanism
- A causal process or sequence through which an intervention contributes to an outcome.
- placebo / falsification test
- A test using outcomes, times, groups, or relationships that should not exhibit the claimed causal pattern if key assumptions are correct.
Review & discussion
Questions for seminar, revision, or assessment.
- Why does a before-and-after comparison fail to identify a causal effect without further assumptions?
- Write a potential-outcomes description of the effect of the Asterbridge shuttle without using mathematical notation.
- Why must a causal estimand specify population and time horizon rather than simply asking whether a policy worked?
- What does random assignment protect? List four things it does not automatically protect.
- Distinguish assignment to treatment, receipt of treatment, and exposure to spillovers in the shuttle trial.
- Why can a policy experiment be scientifically informative yet ethically or legally impermissible?
- Construct an example in which controlling for a post-treatment variable creates bias or changes the estimand.
- What makes a natural experiment causal? Why is the label itself insufficient?
- Explain why a strong first stage is necessary but insufficient for instrumental-variable validity.
- Why is an RD estimate generally local to units near the cutoff? What forms of manipulation threaten the design?
- Explain the parallel-trends assumption in DiD using Asterbridge districts. Why do pre-trends provide evidence rather than proof?
- How can staggered treatment timing and heterogeneous effects complicate a conventional two-way fixed-effects DiD regression?
- Why does observed covariate balance after matching not establish absence of unobserved confounding?
- Give a causal question for which process tracing would add information that an average treatment effect would not provide.
- Create one effect claim and one mechanism claim about the same Asterbridge intervention. What evidence would each require?
- Give a public-policy example of interference. How would it change the definition of treatment exposure?
- When should treatment-effect heterogeneity be treated as an exploratory finding rather than a confirmatory claim?
- Construct a case in which a highly internally valid estimate should not be transported to another jurisdiction.
- What is the difference between robustness to specifications and robustness of the identification assumptions?
- Design one placebo or falsification test for the Asterbridge shuttle study and explain what a failure would imply.
- How can a million administrative observations produce a precise but causally unidentified result?
- An AI proposes a DAG and writes correct DiD code. What must a human analyst still verify?
- Which current NousPolis objects can store the pieces of a causal argument, and what causal relationships should become mandatory dependencies before N5?
- Why should an experimentally identified result not automatically outrank qualitative mechanism evidence for every political question?
- At what point does a causal estimate become an authorized public decision? Explain why the answer must remain compatible with 'AI analyses. Humans authorise.'
Further reading
Continue into the literature.
For the counterfactual logic of causal inference, begin with Holland (1986) and Hernan and Robins (2020); Imbens and Rubin (2015) provides a systematic potential-outcomes treatment. Dunning (2012) is especially valuable for political scientists because it emphasizes design, assignment processes, and the role of qualitative evidence in natural experiments. Lee and Lemieux (2010) gives an accessible guide to regression discontinuity, while Angrist, Imbens, and Rubin (1996) is foundational for the causal interpretation of instrumental variables with imperfect compliance. For modern difference-in-differences, read Callaway and Sant'Anna (2021) alongside Goodman-Bacon (2021). Stuart (2010) provides a broad review of matching methods. Collier (2011) is a concise introduction to process tracing and diagnostic evidence. Hudgens and Halloran (2008) shows why interference requires causal estimands beyond the no-spillover assumption. For the ethics of social experimentation, pair the Belmont Report (National Commission 1979) with Humphreys (2015) and Desposato (2018). On AI and causal reasoning, Liu et al. (2024), Jin et al. (2024), and Saklad et al. (2026) provide useful benchmarks that make clear why causal fluency should not be treated as established causal competence.
Companion, not replacement
The textbook carries the complete argument.
This page reproduces the chapter's study and navigation layer from the current living manuscript. The Asterbridge narrative, historical and methodological argument, figures, Political Science Lens boxes, NousPolis Canon crosswalks, labs, red-team exercises, and N5 Practice boundaries remain in the canonical textbook rather than being republished wholesale here.