brycewang-stanford/Auto-Empirical-Research-Skills / skills/17-DAAF-Contribution-Community-daaf/dot-claude/skills/science-communication
science-communication
VerifiedTranslating technical findings for non-technical audiences. Narrative frameworks (Pyramid Principle, SCQA), plain-language translation, executive summaries, policy briefs, causal language. Use when presenting to stakeholders or reviewing deliverables
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 12.4 KB |
| references/accessibility-equity.md | markdown | 17.5 KB |
| references/audience-analysis.md | markdown | 16.0 KB |
| references/communication-review.md | markdown | 16.5 KB |
| references/deliverable-templates.md | markdown | 19.3 KB |
| references/narrative-frameworks.md | markdown | 18.5 KB |
| references/plain-language.md | markdown | 22.2 KB |
references/plain-language.md
View on GitHub# Plain-Language Translation
How to communicate statistical and analytical findings in language your audience can understand. Plain language is not dumbing down — it is precise communication calibrated to the reader's vocabulary. As the NIH states: "People with average reading skills also appreciate messages conveyed simply and clearly."
This reference covers four areas: translating statistical jargon, calibrating reading level, expressing uncertainty through hedging, and choosing appropriate causal language.
## Statistical Jargon Translation Table
When writing for non-technical audiences, replace statistical terminology with accessible equivalents. The goal is to preserve meaning while removing the barrier of unfamiliar vocabulary.
| Technical Term | Plain-Language Equivalent | Notes |
|----------------|--------------------------|-------|
| Statistically significant | The difference is unlikely to be due to chance alone | Never say "significant" without context — readers assume "important" |
| p-value | How likely we'd see this result if there were really no difference | Avoid reporting exact p-values to non-technical audiences |
| p < 0.05 | We can be fairly confident this pattern is real, not just random noise | Frame as confidence, not probability |
| Confidence interval | A range of plausible values; we're [X]% confident the true value falls within this range | "Between X and Y" is often sufficient |
| 95% confidence interval | If we repeated this study 100 times, about 95 of those ranges would contain the true value | Only use this precise definition for technical audiences |
| Standard deviation | How spread out the values are; a measure of variation | "Typical spread" works for most audiences |
| Standard error | How precisely we've estimated the average | Usually omit entirely for non-technical audiences |
| Variance | How much the values differ from each other | Usually use "variation" or "spread" instead |
| Mean / average | Average (arithmetic mean); the typical value | "Average" is universally understood |
| Median | The middle value — half are above, half below | More intuitive than mean for skewed data; explain the concept |
| Percentile | Ranking relative to others; "at the 75th percentile" means scoring higher than 75% of the group | Use concrete framing: "higher than 3 out of 4" |
| Correlation | Two things tend to move together | Emphasize: "move together" does not mean one causes the other |
| Regression | A statistical method for measuring the relationship between factors while holding other factors constant | "Our analysis accounts for differences in X, Y, and Z" |
| Coefficient | The estimated relationship between two factors | "For each additional X, we see Y more/fewer Z" |
| Effect size | How big the difference actually is (not just whether it exists) | "The gap is equivalent to roughly..." with concrete comparison |
| Odds ratio | How much more (or less) likely something is in one group compared to another | "People in group A were X times as likely to..." |
| Relative risk | How much higher (or lower) the risk is in one group vs. another | Same as odds ratio for non-technical audiences |
| Hazard ratio | How much faster (or slower) one group experiences an event compared to another | "Group A experienced the event X times as fast" |
| Regression to the mean | Extreme values tend to move back toward average over time | "Unusually high (or low) results tend to bounce back" |
| Confounding variable | Another factor that could explain the apparent relationship | "There might be a hidden factor driving both patterns" |
| Omitted variable bias | Our estimate may be off because we couldn't account for an important factor | "We couldn't measure everything that matters" |
| Selection bias | The people in our study may not represent the full population | "Our sample may not reflect everyone" |
| Endogeneity | The thing we're studying and the outcome may influence each other | "The cause and effect may run in both directions" |
| Heterogeneity | The pattern differs across subgroups | "The effect isn't the same for everyone" |
| Interaction effect | The relationship between A and B changes depending on C | "The impact of A depends on C" |
| Fixed effects | Accounting for unchanging differences between groups | "After controlling for stable characteristics of each group" |
| Instrumental variable | A technique for isolating the causal effect when direct measurement is contaminated | "We used a natural experiment to separate cause from correlation" |
| Difference-in-differences | Comparing trends before and after a change, between affected and unaffected groups | "We compared how things changed for the affected group vs. a similar unaffected group" |
| Regression discontinuity | Comparing outcomes just above and below a threshold | "We compared students just above and just below the cutoff" |
| Propensity score matching | Comparing similar individuals who received different treatments | "We compared people who were similar in every measurable way except the one factor we're studying" |
| Robustness check | Testing whether our findings hold up under different assumptions | "We checked our results multiple ways and they held up" |
| Sensitivity analysis | Testing how much our conclusions depend on our assumptions | "We tested what would happen if our assumptions were wrong" |
| Power (statistical) | The ability of a study to detect a real effect if one exists | "Whether our study was large enough to detect a meaningful difference" |
| Sample size | The number of observations in the study | "We studied N people/schools/districts" |
| Weighting | Adjusting the analysis so that the sample better represents the population | "We adjusted so that our results reflect the full population" |
| Bootstrapping | A technique for estimating uncertainty by resampling the data | Usually omit entirely; say "we estimated the uncertainty" |
| Cross-validation | Testing a model on data it wasn't trained on | "We verified the model works on new data" |
| Overfitting | A model that learns noise instead of signal | "The model memorized the specific data rather than learning the general pattern" |
| Null hypothesis | The assumption that there is no real effect or difference | "The possibility that the pattern is just random noise" |
| Type I error | Concluding there's an effect when there isn't one (false positive) | "A false alarm — seeing a pattern that isn't really there" |
| Type II error | Missing a real effect (false negative) | "Missing a pattern that is actually there" |
| Multicollinearity | Predictors in the model are too closely related to each other | "Some of our factors are so closely related that we can't separate their individual effects" |
## Reading Level Guidance
The NIH recommends writing for the actual audience, not for a universal grade level. Match complexity to the reader:
| Audience | Target Reading Level | Guidance |
|----------|---------------------|----------|
| General public (health literacy context) | 6th-8th grade | NIH Clear & Simple guidance for limited-literacy audiences |
| General public (educated) | 9th-11th grade | Appropriate for most public-facing research communication |
| Policy audiences | 10th-12th grade | Can handle more complex sentences but not jargon |
| Academic audiences | Graduate level | Full technical vocabulary appropriate |
### Practical Reading Level Tips
Based on NIH and federal plain language guidelines:
1. **Sentence length:** Average 15-20 words per sentence. One idea per sentence.
2. **Paragraph length:** 2-3 sentences maximum. One topic per paragraph.
3. **Word choice:** Pick the familiar word over the unusual one. "Use" not "utilize." "Before" not "prior to." "About" not "pertaining to."
4. **Active voice:** "We found that X causes Y" not "It was found that Y is caused by X." Active voice clarifies who did what.
5. **Concrete over abstract:** "1 in 5 students" not "20% of the student population." Concrete examples stick; abstract percentages slide off.
6. **Define once, use consistently:** If you must introduce a technical term, define it clearly the first time, then use it consistently. Don't alternate between synonyms.
### Common Word Replacements
From federal plain language guidelines (PlainLanguage.gov), selected for research communication:
| Avoid | Use Instead |
|-------|-------------|
| additional | more |
| assist, assistance | help |
| commence | begin, start |
| currently | now (or omit) |
| demonstrate | show |
| determine | find, decide |
| disseminate | share, send |
| due to the fact that | because |
| endeavor | try |
| facilitate | help, ease |
| finalize | finish, complete |
| implement | carry out, start |
| in accordance with | under, per |
| in order to | to |
| in the event of | if |
| methodology | method |
| modify | change |
| necessitate | need, require |
| numerous | many |
| parameters | limits |
| pertaining to | about |
| prior to | before |
| provide | give |
| regarding | about |
| subsequent | later, next |
| sufficient | enough |
| terminate | end, stop |
| utilize | use |
| validate | confirm |
| whereas | because, since |
## Hedging Language Scale
Hedging communicates your confidence level to the reader. The challenge: too little hedging overstates your evidence; too much hedging buries the finding. Match your hedge to your inferential strength.
### Five-Level Hedging Scale
| Level | Confidence | Example Phrases | When to Use |
|-------|-----------|-----------------|-------------|
| **Very high** | Near-certain | "The evidence clearly shows..." / "We can confidently state..." / "The data demonstrate..." | Well-established findings with robust evidence from multiple independent sources or RCTs |
| **High** | Strong evidence | "The evidence strongly suggests..." / "Our analysis provides strong support for..." / "The data consistently show..." | Strong observational evidence with appropriate controls and robustness checks |
| **Moderate** | Substantial evidence | "The evidence suggests..." / "Our findings indicate..." / "The data point toward..." | Solid evidence with some limitations or caveats |
| **Low** | Preliminary evidence | "The evidence is suggestive of..." / "Preliminary findings point to..." / "There are indications that..." | Limited data, early-stage research, or important caveats that weaken conclusions |
| **Very low** | Speculative | "It is possible that..." / "One interpretation is..." / "The data are consistent with, but do not confirm..." | Exploratory findings, small samples, many competing explanations |
### Hedging Anti-Patterns
| Anti-Pattern | Problem | Better |
|--------------|---------|--------|
| "The data prove that..." | Nothing in observational research is "proven" | "The evidence strongly suggests..." |
| "It is perhaps possible that there might be some indication of a potential..." | Over-hedging buries the finding | "The evidence suggests..." |
| "We found a significant relationship..." | "Significant" means different things to technical and lay readers | "We found a [strong/meaningful/consistent] relationship..." |
| "More research is needed." (as the conclusion) | This is always true and therefore meaningless | State what you found, then note specific open questions |
## IPCC Calibrated Uncertainty Language
The Intergovernmental Panel on Climate Change developed a formal framework for expressing uncertainty that has become a model for science communication across disciplines. It uses two independent scales.
### Likelihood Scale (Probability of Outcome)
These terms express quantified probability. In formal usage, they appear in *italics* to signal calibrated meaning.
| Term | Probability Range |
|------|-------------------|
| Virtually certain | 99-100% |
| Extremely likely | 95-100% |
| Very likely | 90-100% |
| Likely | 66-100% |
| More likely than not | >50-100% |
| About as likely as not | 33-66% |
| Unlikely | 0-33% |
| Very unlikely | 0-10% |
| Extremely unlikely | 0-5% |
| Exceptionally unlikely | 0-1% |
**Key design features:**
- The scale is asymmetric — terms overlap (e.g., "very likely" is a subset of "likely")
- "About as likely as not" (33-66%) is the pivot point between affirmative and negative likelihood
- The everyday meaning of "likely" (~70-75% to most people) roughly aligns with the calibrated meaning (66-100%), but "unlikely" is often misread as near-zero rather than the actual 0-33%
### Confidence Scale (Epistemic Quality)
Confidence reflects the author team's judgment about the validity of a finding, based on **evidence quality** (limited/medium/robust) and **degree of agreement** (low/medium/high):
| Confidence Level | Interpretation |
|------------------|---------------|
| Very high | Robust evidence + high agreement |
| High | Strong evidence + substantial agreement |
| Medium | Medium evidence + medium agreement, or mixed |
| Low | Limited evidence or low agreement |
| Very low | Minimal evidence or fundamental disagreement |
### Using Both Scales Together
Confidence certifies the epistemic quality of the assessment; likelihood communicates the probability. They can be stacked: "It is *very likely* (high confidence) that X is occurring."
**Decision rule:** Use likelihood terms only when confidence is high or very high. Expressing a precise probability for a finding with low confidence creates false precision.
### Adapting for Non-Climate Research
While the IPCC framework was designed for climate science, its two-dimensional approach (probability + confidence) is valuable for any research communication that needs to express uncertainty clearly:
- **For policy audiences:** Use the likelihood terms directly (they're intuitive) and translate confidence as "strength of evidence"
- **For general audiences:** Map to everyday language: "very likely" → "the evidence strongly points to..."; "about as likely as not" → "the evidence is mixed"
- **For academic audiences:** Use the framework as a supplement to standard statistical reporting
## Causal vs. Correlational Language
Choosing the right linking language is one of the most consequential decisions in research communication. The wrong word can either overstate your evidence (implying causation from correlation) or understate it (hedging a well-identified causal effect into meaninglessness).
### The Core Distinction
| Study Design | What You Can Claim | Appropriate Language |
|--------------|-------------------|---------------------|
| Randomized controlled trial (RCT) | Causal effect | "X caused Y" / "X increased Y" / "X reduced Y" |
| Strong quasi-experimental (DiD, RDD, IV) | Causal effect under stated assumptions | "X led to Y" / "X resulted in Y" / "The evidence indicates X caused Y" |
| Observational with controls | Association, potentially causal | "X is associated with Y" / "X predicts Y" / "X is linked to Y" |
| Cross-sectional correlation | Association only | "X and Y tend to occur together" / "X correlates with Y" |
### Linking Language Spectrum
Adapted from Haber et al. (2022), who systematically evaluated how readers interpret different linking phrases. Words are arranged from strongest to weakest causal implication:
| Causal Strength | Linking Words | Appropriate For |
|----------------|---------------|-----------------|
| **Strong causal** | causes, produces, leads to, results in, prevents | RCTs, strong quasi-experiments with clearly stated assumptions |
| **Moderate causal** | increases, reduces, improves, affects, influences, impacts | Quasi-experimental designs; use with explicit design disclosure |
| **Weak causal / ambiguous** | contributes to, is a factor in, is a determinant of, explains, predicts | Observational studies with strong controls; flag the design explicitly |
| **Associational** | is associated with, is related to, is linked to, correlates with | Observational studies; the safe default |
| **Descriptive** | differs between, is higher among, varies across | Descriptive comparisons without causal intent |
### The "Schrödinger's Causal Inference" Problem
Haber et al. (2022) identified a pervasive pattern in published research: papers use hedged associational language in their results ("X was associated with Y") but then imply causal sufficiency in their conclusions and recommendations ("clinicians should consider X to reduce Y"). The causal claim simultaneously exists and doesn't exist depending on which paragraph you read.
**Prevention:** Ensure that the strength of language in your recommendations does not exceed the strength of language in your findings. If you found an "association," your recommendation should say "further investigation is warranted" — not "policymakers should adopt X."
### Practical Causal Language Rules
1. **Default to associational language** unless your study design explicitly supports causal claims. "Is associated with" is almost always safe.
2. **State the study design when using causal language.** "Our difference-in-differences analysis shows that the policy *reduced* dropout rates by 4 percentage points" is honest because it names the design. "The policy *reduced* dropout rates" without design context invites misinterpretation.
3. **Separate findings from implications.** Use causal language in findings only if the design warrants it. In the implications section, you can discuss causal *possibilities* with appropriate hedging: "If this association is causal, it would suggest..."
4. **Watch for direction-of-causation assumptions.** "Poverty affects test scores" assumes a direction. "Poverty is associated with lower test scores" is neutral about the direction. Only claim directionality if your design supports it.
5. **Avoid "weasel causality."** Words like "impact," "influence," and "affect" are verbs that imply the exposure *did something* to the outcome — they are causal in ordinary English even if the writer thinks they're being cautious. If your design is observational, these words overclaim.
6. **Never let a word list substitute for clear thinking.** Per Haber et al., policing individual words fails because context, sentence structure, and rhetorical positioning all contribute to how readers interpret causal claims. The solution is structural transparency about your design and estimand, not vocabulary restrictions.
## Number Formatting Standards
How you present numbers affects comprehension as much as how you present words. Match formatting to the audience.
### General Rules
| Guideline | Example | Rationale |
|-----------|---------|-----------|
| **Round aggressively for non-technical audiences** | "About 14 percentage points" not "13.7 percentage points" | False precision distracts; rounded numbers are easier to remember |
| **Use natural frequencies over percentages for general audiences** | "1 in 5 students" not "20% of students" | Natural frequencies are processed faster and remembered longer |
| **Use percentages for technical audiences** | "20.3% (95% CI: 18.1-22.5)" | Technical readers expect precision |
| **Abbreviate large numbers** | "$14M" or "$14 million" not "$14,000,000" | Readers process abbreviated forms faster |
| **Provide baselines for differences** | "Scores rose by 14 points, from 52% to 66%" | A difference without a baseline is uninterpretable |
| **Use absolute AND relative comparisons together** | "3x more likely (15% vs. 5%)" | Relative comparisons alone can mislead: "3x more likely" sounds dramatic but may mean 0.3% vs 0.1% |
| **Avoid decimal places for general audiences** | "About 3 in 10" not "29.7%" | Decimal places imply false precision for lay readers |
| **Use consistent formatting within a deliverable** | Don't switch between "20%", "one-fifth", and "1 in 5" for the same quantity | Inconsistency creates confusion about whether these are different numbers |
### Framing Effects to Watch
The same number can feel very different depending on how it is framed:
| Framing | Example | Perception |
|---------|---------|------------|
| Relative risk (high anchor) | "3x more likely to drop out" | Alarming — emphasizes the multiplier |
| Absolute risk | "15% vs. 5% dropout rate" | More measured — shows actual rates |
| Natural frequency | "15 out of every 100 students" | Most intuitive for general audiences |
| Inverted frame | "85% of students completed" vs. "15% dropped out" | Same data, different emotional impact — choose deliberately |
**Rule of thumb:** Present both absolute and relative comparisons when the finding is important. Lead with whichever framing is most appropriate for your audience and purpose, but always include the other for context.
## Domain-Specific Jargon
The jargon translation table above covers statistical and methodological terms. However, every research domain has its own specialized vocabulary (e.g., "Title I," "FRPL eligibility," "AYP" in education; "incidence," "comorbidity," "morbidity" in public health).
**Guidance for domain-specific translation:** When working with a specific data domain, consult the relevant domain skill (e.g., `education-data-context`, `education-data-source-ccd`) for the authoritative definitions and caveats of domain-specific terms. Then apply the same plain-language translation principles from this file: replace with familiar equivalents, define on first use, and provide concrete examples.
The jargon translation table in this file is deliberately domain-agnostic — it handles the statistical and analytical language that appears in *every* research project. Domain-specific translation is handled by domain skills to avoid duplication and ensure accuracy.
## References and Further Reading
Haber, N.A. et al. (2022). "Causal and Associational Language in Observational Health Research." *American Journal of Epidemiology*, 191(12), 2084-2097. https://doi.org/10.1093/aje/kwac137
NIH. "Plain Language at NIH." https://www.nih.gov/institutes-nih/nih-office-director/office-communications-public-liaison/clear-communication/plain-language-nih
NIH. "Clear & Simple: Developing Effective Print Materials for Low-Literate Readers." https://www.nih.gov/institutes-nih/nih-office-director/office-communications-public-liaison/clear-communication/clear-simple
Federal Plain Language Guidelines. https://www.plainlanguage.gov/guidelines/
IPCC. "Guidance Note for Lead Authors of the IPCC Fifth Assessment Report on Consistent Treatment of Uncertainties." https://www.ipcc.ch/site/assets/uploads/2017/08/AR5_Uncertainty_Guidance_Note.pdf
Mastrandrea, M.D. et al. (2011). "The IPCC AR5 Guidance Note on Consistent Treatment of Uncertainties: A Common Approach Across the Working Groups." *Climatic Change*, 108, 675-691.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| receiving-code-reviewobra/superpowersUse when receiving code review feedback, before implementing suggestions, especially if feedback seems unclear or technically questionable - requires technical rigor and verification, not performative agreement or blind implementation | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| requesting-code-reviewobra/superpowersUse when completing tasks, implementing major features, or before merging to verify work meets requirements | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| using-git-worktreesobra/superpowersUse when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallback | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| verification-before-completionobra/superpowersUse when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |