Applying Scientific Cycle to Social Science
---
name: applying-scientific-cycle-to-social-science
description: Applies the scientific cycle, induction, and deduction to economics and social science research. Use to develop research questions, hypotheses, study designs, evidence assessments, and iterative research plans, or to distinguish causal, descriptive, predictive, and interpretive claims.
---
# Applying the Scientific Cycle to Social Science
Turns a question about people, markets, institutions, or policy into a testable, revisable research argument. Treats conclusions as conditional on measurement, context, design, and assumptions. Adapts to the user's study stage — does not force completed research back through every step.
User asks: "How would I study whether remote work reduces employee productivity?"
- Frame it: Causal question. Population = knowledge workers; unit = employee-task; outcome = output per hour; comparison = remote vs. in-office.
- Hypothesize: Remote work changes productivity via reduced commute time (+) and reduced informal coordination (−). Rival: selection — who chooses remote work differs systematically.
- Predict: If coordination loss dominates, teams with high interdependence should show larger productivity declines than solo-task roles.
- Design: Natural experiment around a mandated return-to-office policy; difference-in-differences comparing high- vs. low-interdependence teams, pre/post. State parallel-trends assumption explicitly.
- Deliver: Compact study brief (see Default Deliverable below), flagging that any numbers are illustrative, not found data.
No preamble — go straight into framing the question type before proposing designs.
Progress:
- Classify the question: descriptive, predictive, causal, or interpretive
- Separate empirical claims from normative judgments
- State the observation/pattern and its evidentiary status (established vs. assumed)
- Form hypothesis + explicit rival explanations
- Operationalize abstract constructs (trust, welfare, inequality, etc.)
- Deduce specific, falsifiable predictions (direction, horizon, subgroup)
- Choose design matched to data availability; state identifying assumptions
- Specify ethics/consent/privacy considerations for any new data collection
- Analyze: report effect sizes + uncertainty, not just significance
- Revise hypothesis/scope and propose the next discriminating test
1. Frame the inquiry
Identify population, setting, period, unit of analysis. Classify the question type. Flag where normative judgment is being smuggled into an empirical claim.
2. Observe and question
State the documented pattern or anomaly being explained. Check whether sampling, measurement, or shifting definitions could produce it artifactually. If no real data was given, say so explicitly rather than inventing a motivating pattern.
3. Form a hypothesis
Propose a mechanism. List at least one serious rival explanation. Operationalize any abstract construct — define how it's measured and discuss construct validity across groups.
4. Deduce predictions
State auxiliary assumptions explicitly. Specify: outcome, comparison, direction, horizon, subgroup. Prefer predictions that discriminate between the hypothesis and its rivals over predictions both would produce.
5. Design and gather evidence
Match method to question: survey, interview, ethnography, archival, administrative, experiment, natural experiment, or formal model. For causal claims, define treatment, outcome, target effect, counterfactual, and what makes the comparison credible.
6. Analyze and assess
Report effect size and uncertainty, not a bare significance threshold. Address missingness, multiple testing, dependence. For qualitative work, explain case selection, coding, negative cases, reflexivity. Never fabricate data, estimates, or citations.
7. Revise and repeat
State whether evidence supports, challenges, or leaves the hypothesis unresolved — within scope. Distinguish "imprecise test" from "evidence of no effect." Propose the next discriminating study.
Example 1: Input: "Does minimum wage increase unemployment?" Output: Reframes as a causal question requiring a comparison group and counterfactual. Notes classic identification problem (simultaneity: wages and employment both respond to local economic conditions). Proposes border-discontinuity design (Card-Krueger style) comparing adjacent counties across a state line with differing minimum wage, stating the parallel-trends/no-spillover assumptions needed. States that estimated effects are local to low-wage, tradable-sector contexts and may not generalize nationally or to large wage hikes.
Example 2: Input: "Users report our app feels 'untrustworthy' — investigate why." Output: Classifies as interpretive/descriptive, not causal. Recommends operationalizing "trust" (e.g., via validated survey scale plus behavioral proxy like abandonment rate). Suggests mixed-method design: structured interviews for mechanism discovery (induction), followed by a pre-registered survey testing specific hypothesized drivers (deduction). Flags that interview-derived hypotheses tested on the same interview data are exploratory, not confirmatory.
Example 3: Input: "Here's a dataset — found a correlation between social media use and reported loneliness." Output: Flags reverse causality and selection as rival explanations before treating the correlation as informative. States that adding controls doesn't establish identification, and some "controls" (e.g., current mood) may be mediators/colliders that bias the estimate. Proposes panel design with within-person fixed effects or an instrument (e.g., plausibly exogenous rollout variation) as a stronger test, and specifies what evidence would distinguish "social media causes loneliness" from "lonely people use social media more."
- State identifying assumptions out loud (parallel trends, exclusion restriction, continuity at cutoff) — don't just name the method.
- Treat a null or imprecise result differently from a precisely estimated zero effect.
- Mark any hypothesis formed after seeing the data as exploratory, even if it sounds like a clean prediction.
- Scope causal claims: local vs. average, short-run vs. long-run, this population vs. generalizable.
- Address ethics (consent, privacy, harm) whenever proposing new data collection on people.
- Scale the deliverable to what's asked — a one-paragraph answer doesn't need the full study-brief template.
- Don't present simulated or hypothetical results as if they were observed data.
- Don't claim a control variable "fixes" a causal estimate without checking it isn't a mediator or collider.
- Don't treat statistical significance alone as evidence of a meaningful effect size.
- Don't confuse reproducing an analysis (same data, same code) with replication (new data).
- Don't let normative conclusions ("policy X is good") follow automatically from an empirical estimate without stating the value judgment involved.
- Don't force a completed, published study back through the full design cycle — assess it on its own terms instead.