Most people do not need to read research the way a journal reviewer reads research.
A parent may need to know whether a proposed therapy has evidence for a child like theirs. An autistic adult may need to know whether a study about burnout included adults with similar support needs. A clinician may need the size and certainty of a treatment effect. A school team may need to know whether a classroom intervention was actually tested in schools. A legislator may need to know whether a policy claim rests on statewide data, a small convenience sample, or expert opinion.
Those are different reading jobs.
The fastest way to overload yourself is to open a 22-page article and treat every paragraph as equally important.
Read in passes, not from top to bottom
A four-pass reading system
Each pass earns the next one. A weak or irrelevant study does not deserve 45 minutes just because it is published.
60–90 seconds
5 minutes
10–20 minutes
Deep dive
This is not cutting corners. It is triage.
A paper about a mouse model should not consume the same attention when your question is “What accommodation helps an autistic adult complete a medical visit?” A survey of 80 university students should not be read as though it settled a statewide disability policy. A well-conducted qualitative study should not be discarded because it does not have a p-value when the question is about lived experience or implementation.
Before you judge the study, identify the claim
Different claims need different evidence
“How common is this?”
Look for representative sampling, surveillance, population data, clear definitions, response rates, and whether the denominator matches the population being described.
“Does this intervention change the outcome?”
Look for an appropriate comparison, randomization when feasible, baseline balance, adherence, missing data, effect size, harms, and whether the outcome was prespecified.
“What happens to people using this system?”
Qualitative, mixed-methods, participatory, and implementation research may be exactly the right evidence when the question concerns meaning, barriers, acceptability, communication, workflow, or context.
There is no universal evidence hierarchy that makes one design “best” for every question.
A randomized trial can be powerful for estimating intervention effects. It is not designed to tell you everything about why an autistic adult stopped attending the clinic. A qualitative interview study can reveal barriers a trial never measured. It cannot by itself estimate how common that barrier is across Arkansas.
The method should fit the question. Then the conclusion should stay inside the method.
Know what kind of evidence is in your hands
Systematic review / meta-analysis
A systematic review uses explicit methods to find and assess multiple studies. A meta-analysis statistically combines results when appropriate. Neither is automatically reliable: the search, inclusion criteria, risk-of-bias assessment, heterogeneity, publication bias, and quality of the included studies still matter.
Randomized trial
Participants are allocated to comparison groups by a random process. Good randomization can reduce confounding, but poor allocation, missing outcomes, unblinded outcome assessment, selective reporting, small samples, or short follow-up can still weaken the result.
Observational study
Cohort, case-control, cross-sectional, registry, and other observational designs can show patterns and associations in real populations. Confounding, selection, measurement, and reverse-causation questions become especially important when someone turns an association into a causal headline.
Qualitative study
Interviews, focus groups, observation, and other qualitative methods can examine experience, meaning, implementation, barriers, communication, and context. Judge sampling, analytic transparency, researcher reflexivity, richness of data, and whether interpretations are grounded in participants’ accounts.
Mixed-methods study
Combines quantitative and qualitative approaches. The useful question is not merely whether both kinds of data appear, but whether the study integrates them in a way that explains the problem better than either strand alone.
Case report / case series
Can identify unusual events, generate hypotheses, or document a new implementation experience. It cannot establish prevalence or prove a general causal effect by itself.
Do not confuse a reporting checklist with a quality score
CONSORT, PRISMA, STROBE, and other reporting guidelines help authors report important information readers need. CONSORT 2025 is the current major reporting guideline for randomized trials; PRISMA 2020 remains the central reporting guideline for systematic reviews; EQUATOR maintains a searchable library of reporting guidelines across study designs.
Good reporting makes appraisal possible. It does not guarantee the study itself was well designed.
A beautifully reported biased study is still biased. A poorly reported study may be impossible to appraise even if the underlying work was competent.
Do not let the paper’s order control your reading order
Abstract
Introduction
Methods
Results
Discussion
You do not need to become a statistician to catch the biggest errors
Five number checks
Statistical significance is not the finish line.
Methodology literature has repeatedly emphasized that p-values do not provide effect magnitude; effect sizes and confidence intervals give readers information about the size and precision of an estimate. Cochrane’s current guidance likewise emphasizes uncertainty, heterogeneity, risk of bias, indirectness, imprecision, and publication bias when judging a body of evidence.
A p-value cannot answer “Is this worth doing?”
Suppose a school intervention produces a statistically detectable change of 0.8 points on a 100-point scale.
That could be statistically significant with enough data and still be practically trivial. Or a modest average effect could be meaningful if the outcome is important, harms are low, implementation is easy, and some students benefit substantially.
The research-reading question is therefore broader:
- How large was the effect?
- What scale or outcome changed?
- Is that change meaningful in real life?
- How certain is the estimate?
- Were harms measured?
- How long did the effect last?
- Who actually benefited or dropped out?
Bias is not the same as “the researchers were dishonest”
Risk of bias means the study could systematically lean away from the truth
CHECK THE PROCESSSelection
Who could enter the study? Who could not? Did recruitment favor people with time, transportation, speech, technology, stable housing, English fluency, lower support needs, or proximity to a research center?
Measurement
Was the outcome measured reliably? Did the tool measure what the authors claim it measured? Were observers aware of which intervention participants received?
Missing data
Did people leave the study because the intervention was burdensome, inaccessible, ineffective, or harmful? How did the analysis handle that loss?
Selective reporting
Were many outcomes measured but only favorable ones highlighted? For trials, compare the paper with a protocol or registry when available.
Confounding
In non-randomized research, another factor may differ between groups and partly explain the observed association. Good analysis can address some confounding, but not every unmeasured factor.
Interpretation / spin
Does the conclusion sound stronger than the data? Watch for causal language from observational data, broad claims from narrow samples, or subgroup findings promoted without appropriate caution.
Cochrane’s risk-of-bias frameworks treat bias as a property of a particular result and examine design, conduct, analysis, and reporting. That is more useful than labeling an entire paper “good” or “bad.”
Registration and protocols help you see whether the target moved
For many clinical trials, a public registry such as ClinicalTrials.gov records planned study information before or during the trial. ICMJE recommends prospective registration of clinical trials as a condition of publication consideration in member journals. CONSORT 2025 also emphasizes access to trial protocols and transparency around prespecified methods and outcomes.
When the stakes are high, compare:
- What outcomes were registered?
- Which one was called primary?
- Did the final paper report the same outcomes?
- Did the analysis change?
- Were unfavorable or null outcomes missing?
A change is not automatically misconduct. Sometimes protocols change for legitimate reasons. The important issue is whether the change is transparent and whether it alters how confidently you should interpret the result.
Funding and conflicts are context—not a shortcut to belief or disbelief
ICMJE’s current recommendations require transparent disclosure of financial and non-financial relationships that could be perceived as conflicts. Funding source, employment, patents, consulting, advocacy roles, intellectual commitments, and other relationships may matter.
A conflict does not automatically invalidate a study. Lack of a declared conflict does not make the study automatically correct.
Use the disclosure to ask better questions: Who designed the study? Who controlled the data? Who chose the outcomes? Who performed the analysis? Did the sponsor have publication rights? Are independent studies consistent with the result?
A systematic review is not “all the evidence” just because it has a forest plot
Read synthesis with the same skepticism you use for a single study
Search
Did the review search broadly enough to find relevant studies? Were databases, dates, languages, unpublished studies, and registers handled transparently?
Eligibility
Were inclusion/exclusion criteria defined before results were known? Are the included studies truly answering the same question?
Bias + certainty
Did reviewers assess risk of bias and certainty, or merely count studies? GRADE considers risk of bias, inconsistency, indirectness, imprecision, and publication bias.
Heterogeneity
If studies differ substantially in participants, interventions, outcomes, or effects, one pooled number can conceal meaningful variation.
Missing evidence
Publication bias can distort a literature when positive or interesting studies are more likely to appear than null findings.
Review date
A rigorous review can become outdated. Check the search end date and whether important newer trials have appeared.
PRISMA 2020 helps systematic-review authors report how studies were identified, screened, included, and excluded. Its flow diagram is useful because it shows what entered the review and what was removed. But PRISMA compliance is not the same as high certainty of evidence.
Autism research needs an extra applicability audit
“Autistic participants” can describe a much narrower group than the headline suggests
Intellectual disability
A 2019 cross-sectional review and meta-analysis of 301 autism studies published in 2016 estimated that about 94% of autistic participants in those studies did not have intellectual disability, despite intellectual disability being much more common in the broader autistic population.
Ask: Were autistic people with intellectual disability actually represented—or does the article generalize beyond its sample?Communication
Studies may require spoken responses, reading, computer use, lengthy questionnaires, or tolerance of unfamiliar testing environments. Those requirements can systematically exclude people who communicate differently or need more support.
Ask: What communication ability did participation itself require?Age
Evidence from preschool children cannot automatically answer an adult employment question. Research on college students cannot stand in for older adults, nonspeaking adults, or people with substantial daily support needs.
Ask: How close is the age and life context to the person or population I care about?Recruitment
Clinic, university, online, registry, school, and community samples reach different people. A study requiring repeated travel to a major research center can under-represent rural families and people with fewer resources.
Ask: Who had a realistic chance of getting into this study?Whose outcome?
Some autism research measures outcomes selected by clinicians, teachers, or caregivers. Other work includes autistic priorities directly. Those perspectives can overlap, but they are not interchangeable.
Ask: Who decided what “better” means?Participation in research design
Participatory autism research involves autistic people and allies in decisions about research rather than limiting involvement to being subjects. Community-priority and participatory studies can reveal questions, outcomes, burdens, and interpretations traditional research teams may miss.
Ask: Were autistic people involved in shaping the question, measures, interpretation, or dissemination?Under-representation does not make every autism study useless. It changes the boundary of the claim.
A result can be valid for a narrowly defined sample and still be misused when the discussion turns that sample into “autistic people” as a whole.
Ask whether the outcome matters to the person, not just whether it moved
A study can successfully reduce an observable behavior while failing to improve—and possibly worsening—the participant’s distress, autonomy, communication, pain, fatigue, or quality of life.
A program can improve appointment completion while making appointments more exhausting.
A school intervention can reduce classroom disruption while increasing shutdowns later at home.
A workplace intervention can improve supervisor ratings while increasing masking and burnout.
Those are not claims that any particular intervention causes those outcomes. They are reminders that outcome selection controls what the study is capable of noticing.
Before accepting the word “improvement,” identify the exact outcome, who measured it, when it was measured, and what tradeoffs were tracked.
Community relevance is evidence about implementation
Autistic co-led research priority work has repeatedly emphasized practical issues such as inclusion, skills, public understanding, health, education, relationships, safety, employment, and services. Participatory research literature likewise argues that autistic people can contribute to decisions about research questions, design, analysis, and dissemination.
That does not make lived experience a substitute for controlled outcome research. It makes lived experience part of choosing the right question, the right outcome, and the right interpretation.
Arkansas adds another applicability layer
A study conducted in a metropolitan academic clinic may assume:
- a specialist is nearby;
- families can return weekly;
- reliable broadband exists;
- transportation is available;
- clinicians have protected training time;
- schools have specialized staff;
- insurance covers the intervention;
- the person can tolerate the clinic environment.
Those assumptions are implementation variables.
For Arkansas, a useful reading ends with questions such as: Could this work in a rural county? Which parts could be remote? What training is required? What would it cost? How much travel does it create? What happens when the recommended specialist has a six-month wait? Does the intervention survive ordinary staffing turnover?
Not every source is the study
The paper
Best place to inspect methods, sample, results, limitations, tables, and what was actually measured.
News / press release
Useful for finding a study or understanding context. Not enough by itself for a high-stakes decision.
Preprint
A manuscript shared before formal peer review. It can be valuable and timely, but its status should remain visible because the paper may change.
Social post / commentary
Can raise important criticism or context. Trace factual claims back to the study, data, correction, or other primary source.
Peer review is a filter, not a truth machine
Publication means a paper passed a journal’s editorial and peer-review process. It does not mean every method is flawless, every analysis is correct, every conclusion is justified, or every result will replicate.
Research can later receive corrections, expressions of concern, or retractions.
PubMed explicitly indexes retraction notices and retracted publications, and PMC marks retracted articles and links them to notices when the structured publication data supports that connection.
When a claim matters, check the current article record—not just a PDF someone saved years ago.
“No evidence” and “evidence of no effect” are different statements
A field can have no good studies. A study can be too small to distinguish benefit from no benefit. A review can find inconsistent results. Or high-certainty evidence can show that an intervention produces little or no meaningful effect.
Those are four different situations.
Look at the confidence interval and certainty assessment. If plausible values include meaningful benefit and meaningful harm, “not statistically significant” does not magically turn the result into proof of no effect.
When studies disagree, do not average the headlines
Ask why they disagree.
- Different population?
- Different definition of autism, burnout, anxiety, communication, success, or recovery?
- Different intervention dose?
- Different comparator?
- Different follow-up length?
- Different outcome scale?
- Different risk of bias?
- Different setting?
- Chance variation?
- Selective publication?
Cochrane’s guidance on meta-analysis emphasizes heterogeneity precisely because pooled effects can vary across studies. When variation is real, the average effect may not describe every setting.
AI can help you navigate a paper. It should not become the paper.
Use automated summaries as an index, not as evidence
Useful jobs
- define unfamiliar statistical or clinical terms;
- turn a methods paragraph into a checklist of things to verify;
- identify where sample, outcomes, exclusions, and limitations appear;
- compare terminology across papers;
- help create a plain-language question list.
Do not outsource these
- whether the model quoted the paper accurately;
- whether the result was primary or secondary;
- whether an association was turned into causation;
- whether numbers, confidence intervals, or participant counts were altered;
- whether a paper has been corrected or retracted;
- whether a source the model names actually exists.
For a high-stakes decision, the verification chain should end at the original paper, registry, guideline, official dataset, or other primary source—not at an AI-generated sentence about it.
Learn the difference between evidence, guidance, and policy
A systematic review summarizes research. A clinical guideline may combine evidence with values, feasibility, harms, costs, patient preferences, and expert judgment. A school policy tells a district what its rules are. A statute or regulation establishes legal requirements. A professional position statement expresses an organization’s recommendation.
All can matter. They are not the same kind of authority.
When someone says “the research says we have to,” ask whether the source is actually research, a guideline, a law, or a local policy.
A reporting guideline can tell you what should be visible
If you are unsure what information a well-reported study should contain, the EQUATOR Network is useful because it links reporting guidelines by study design. CONSORT 2025 covers randomized trials; PRISMA 2020 covers systematic reviews; STROBE addresses observational studies, with many extensions available for specialized designs.
You do not need to complete a 30-item checklist for every paper. Use the guideline when something important is missing and you need to know whether that absence matters.
Use a stop rule so research does not become an endless tunnel
Know when you have enough
Green · enough for orientation
You know what was studied, what was found, the major limit, and whether it is relevant. Good enough for general learning or a low-stakes question.
Yellow · compare before acting
The result is relevant but small, uncertain, conflicting, indirect, preprint-only, narrow in population, or important enough that one paper should not carry the decision.
Red · do not rely on this alone
Wrong population, wrong design for the claim, retraction/correction concern, severe bias, unexplained missing data, absent methods, unsupported causal claim, or high-stakes decision without corroboration.
Core terms without the fog
Research Reading Sheet
One paper. One page of notes.
Print this when a study matters enough to inspect but not enough to become your entire day.
My actual question
What decision could this research change?
Study identity
- Original research
- Systematic review / meta-analysis
- Randomized trial
- Observational
- Qualitative
- Mixed methods
- Guideline / policy / commentary
Year:
Who was studied?
Age / population:
Sample size:
Setting / location:
Important exclusions:
What was compared?
Primary outcome:
Follow-up length:
What did they find?
Effect / result:
Uncertainty / CI:
Harms / tradeoffs:
What could distort it?
- Selection
- Measurement
- Missing data
- Confounding
- Selective reporting
- Funding / conflict
- Short follow-up
- Other
Autism / disability applicability
- Communication needs represented
- Intellectual disability represented
- Support needs described
- Age fits
- Rural / access context fits
- Outcome matters to participants
My confidence right now
- Enough for orientation
- Need another study
- Need a systematic review
- Need protocol / registry
- Need professional interpretation
- Not applicable to my question
What would change my mind?
One-sentence evidence note
In , researchers found . The biggest uncertainty is . This may / may not apply to my situation because .
Use the right ANCHOR research tool for the right job
ANCHOR Research is the deeper source route: studies, source checking, evidence comparisons, research questions, and supporting material.
Plain-Language Evidence Notes is the companion reading format for turning a source or body of evidence into a clear public explanation that preserves uncertainty.
ARCHIE and the Resource Map answer a different question: where can a person actually go, and what current service fits the need?
TASAT is a testing/support-access system. Research can inform its methods and interpretation, but a published paper should not be used to turn a support tool into a diagnosis or claim validation it has not earned.
Academy is where evidence can become training, scenarios, field practice, and role-specific instruction.
Reading research is the bridge between a claim and those next actions.
Research-reading standards + primary references
Sources
- Cochrane Handbook for Systematic Reviews of InterventionsCurrent methodological handbook for systematic reviews, including risk of bias, meta-analysis, synthesis, and interpretation.
- Cochrane Handbook — Summary of Findings and GRADEDescribes certainty domains including risk of bias, inconsistency, indirectness, imprecision, and publication bias.
- Cochrane Handbook — Risk of Bias in Randomized TrialsCurrent framework for assessing bias in particular randomized-trial results.
- PRISMA 2020Current central reporting guideline for systematic reviews, including checklist and study-flow reporting.
- CONSORT 2025 / SPIRIT 2025 Published StatementsCurrent reporting standards for randomized trials and trial protocols.
- EQUATOR NetworkSearchable library of reporting guidelines across research designs.
- ClinicalTrials.gov — Clinical Trial Reporting RequirementsCurrent federal information on clinical-trial registration and summary-results reporting.
- ICMJE — Clinical Trial RegistrationCurrent ICMJE policy recommending public registration of clinical trials at or before first participant consent.
- ICMJE — Disclosure of Financial and Non-Financial RelationshipsCurrent recommendations for transparency around relationships and activities that may be perceived as conflicts.
- National Library of Medicine — Errata, Retractions, and Other Linked Citations in PubMedExplains how corrections, retractions, and linked publication notices are represented in PubMed.
- National Library of Medicine — Confidence IntervalsPlain-language statistical training resource on interpretation and precision.
- Sullivan & Feinn — Using Effect Size—or Why the P Value Is Not EnoughExplains why statistical significance alone does not provide the magnitude of an effect.
- Russell et al. — Selection Bias on Intellectual Ability in Autism ResearchCross-sectional review and meta-analysis documenting major under-representation of autistic people with intellectual disability in sampled autism research.
- Pickard et al. — Participatory Autism ResearchExamines participatory approaches in which autistic people and allies contribute to research decisions.
- Autistic Co-Led Community Priorities for Future Autism ResearchAutistic co-led research identifying community priorities and demonstrating the importance of who helps shape the research agenda.
Continue with ANCHOR