--- name: cscw-experiments description: Use when designing or auditing the empirical work behind a CSCW paper — interview and ethnographic rigor, trace and log analysis, surveys, deployments, and mixed methods — matching each method's own validity standard and the ethics of studying real communities. --- # CSCW Empirical Work "Experiments" is the wrong word for most CSCW evidence, and that is the point. The venue is deliberately methods-pluralist: interview studies, ethnography, large-scale trace analysis, surveys, field deployments, controlled experiments, and mixed designs all publish here — **each judged by its own tradition's standard of rigor, not by a quantitative default.** The commonest reviewing disaster is a paper that borrows a method without its discipline. ## Rigor, by method family | Method | What rigor means here | What reviewers flag | | --- | --- | --- | | Interviews | Purposeful sampling with a rationale; saturation or a defended stopping rule; a described analysis process (coding approach, memoing, disagreement handling) | "We interviewed 12 people and themes emerged" with no analytic trail | | Ethnography / field observation | Duration and depth of engagement; researcher's relationship to the setting; thick description that earns the interpretation | Drive-by observation labeled ethnography | | Trace / log analysis | Construct validity (does the log field measure the practice claimed?); an identification strategy for any causal wording; robustness to platform quirks (bots, deleted content, API sampling) | Correlational results narrated causally; metrics inherited from the platform unexamined | | Surveys | Instrument provenance or validation; sampling frame vs. claimed population; nonresponse handling | Convenience sample generalized to "users" | | Deployments / experiments | Genuine group-level conditions; power analysis where inference is statistical; contamination between conditions addressed | N = groups treated as N = individuals | | Mixed methods | An explicit integration logic — which strand leads, which bounds, where they may disagree | Two mini-studies stapled together, each too thin to stand | Two pluralism rules cut across all rows: - **Do not apply one tradition's checklist to another's method.** Demanding inter-rater reliability statistics from an interpretivist analysis, or accepting vibes in place of identification from a causal claim, are the same category of error. State *which* tradition your analysis works in, then meet that tradition fully. - **Qualitative sample sizes are justified by purpose, not by envy.** Twenty-four well-chosen moderators can ground a concept; two million log rows cannot rescue a construct that measures the wrong thing. ## Group-level design decisions - **Unit of analysis and unit of observation must be named separately.** You may observe individuals (interviewees, accounts) while claiming about collectives (teams, communities); the analysis section must say how the aggregation is licensed. - **Sample communities, not just people.** For multi-community studies, describe how communities were chosen and what variation the set covers — community selection is the qualitative analogue of a sampling frame. - **Time matters.** Cooperative practices are rhythms (shift rotations, release cycles, norm renegotiations). A snapshot design should say what it cannot see. ## Ethics as design, not paperwork CSCW evidence usually comes from real communities with stakes in the findings. Reviewers read the ethics description as part of the method: - State IRB/ethics-review status and the consent posture for each data source — including whether "public" trace data was treated as fair game and why that is defensible for *this* community. - Plan quote handling at design time: verbatim quotes from small or hostile-scrutiny communities can be reverse-searched; commit to paraphrase or alteration policies and disclose them. - Consider the community's exposure, not only the individual's: naming a small community can harm it even with every user anonymized (see `cscw-artifact-evaluation` for release-time handling). ## Evidence-plan skeleton ```text [Claim] [Tradition] interpretivist / positivist / computational / mixed (lead strand: ___) [Observation] who or what is observed, at what unit and timescale [Aggregation] how individual observations license collective claims [Validity] the ONE threat most likely to sink this design + mitigation [Ethics] consent posture per data source; quote policy; community exposure [Stop rule] what tells you data collection is done ``` Draft this skeleton before collecting anything; paste the filled version into the methods section as its outline. Under Revise and Resubmit, new data collection is often infeasible — a design that anticipates the obvious objection is the cheapest insurance the journal model offers. Method norms are stable venue culture; submission-mechanics facts elsewhere in this pack carry the 2026-07-08 access date and should be re-verified independently.