Causal and prospective analytics
Know which action changes the outcome before you act. Subconscious brings experimental design, synthetic populations, effect estimation, and scenario analysis into one research workflow. The result can inform a price, product, message, segment, or competitive decision before the business commits to it.
The useful unit is a testable decision: change a defined input, measure a defined outcome, and state what evidence would overturn the recommendation. The eight-decision study matrix maps business questions to methods, inputs, and executable contracts.
Three questions, different evidence
| Question | Evidence it needs | What it supports |
|---|---|---|
| What happened, and to whom? | Observed transactions, survey responses, or historical records. | Description and associations; a causal interpretation needs an identification strategy. |
| What changes when we intervene? | A controlled comparison, such as randomized attribute levels or a randomized A/B test. | An estimated causal contrast within the experiment's design and population. |
| What should we do next? | Intervention estimates combined with a decision model, feasible scenarios, external calibration where needed, and an objective. | A prospective recommendation with explicit assumptions and a way to test it. |
A/B testing is itself experimental and can be prospective. A controlled multi-attribute design extends the question from “Which of these two complete offers wins?” to “Which features change choice, by how much, for whom, and which feasible combination should we test next?” Interactions require a design and model that identify them; they do not follow automatically from a main-effect fit. The identification logic is developed in Hainmueller, Hopkins and Yamamoto (2014).
The Subconscious workflow
- Define the intervention. Name the action, population, alternatives, outcome, and decision rule before collecting responses.
- Run a controlled design. Vary the relevant attributes across synthetic respondents. Preserve the realized tasks and settings so the comparison can be inspected and repeated.
- Estimate effects and heterogeneity. Read the relevant effect estimates, estimator assumptions, uncertainty, and differences across respondents or segments. Keep a respondent utility distribution distinct from an interval on a population effect.
- Compare future actions. Use the fitted choice model to evaluate feasible configurations and competitive scenarios. Add costs and calibrated market inputs when the decision requires financial or volume outputs.
- Validate and update. Compare with human evidence that matches the task and population. Record discrepancies and revise the next study accordingly.
The advantage is the connected workflow: a population definition, a controlled intervention, an inspectable estimate, and a reusable decision model. Synthetic execution makes repeated studies across designs and populations practical; scientific reliability comes from the design and validation of each result. More simulated responses alone do not establish greater human fidelity.
See Methodology, reproducible runs, and the supported analysis contract.
What “causal” means here
Randomization can identify the effect of changing a feature on responses in the synthetic experiment, under the design's assumptions. Applying that result to people requires evidence that the simulated population reproduces the human response to the same intervention. A convincing persona or an accurate average survey response does not establish that transfer.
A prospective use does not require every validation exercise to occur before the historical event. Replication of completed human studies can test a method; a prospective recommendation then specifies the intervention and prediction before the new outcome is observed. Keep those two records distinct.
Evidence attached to the decision
The Mintel case study reports 0.79–0.88 Jensen–Shannon similarity across synthetic and human attribute-choice distributions in a moisturizer conjoint, using demographics without human choice data as model input. It also records a mismatch at higher prices. That is evidence for that study and metric, not a universal accuracy percentage or a competitor comparison.
The causal-fidelity working paper distinguishes agreement in effect direction and rank from calibrated effect magnitudes and realized intervention lift. Its reported rank fidelity does not establish the stronger quantities. The paper is not peer reviewed; inspect the human-baseline records for study-level evidence.
For any new decision, report the intervention, outcome scale, interval type, population, held-out comparison, and known misses. Claims of superior accuracy require a matched evaluation on the same tasks, populations, information access, and scoring rules.
Compare complete research workflows
Prospective analysis and choice research are not exclusive to one provider. Qualtrics documents conjoint and MaxDiff workflows and a simulator. Aaru describes scenario planning, and Simile describes testing decisions before acting. Those public descriptions establish overlap; they are not a matched validation of competing systems.
Evaluate the complete contract: can the workflow specify the intervention, measure its effect, represent heterogeneity, compare future actions, explain uncertainty, and provide a relevant human validation record? Subconscious makes that contract explicit across its business decision matrix, research guides, REST reference, and MCP tools.