To translate selection bias, confounding, adjustment and causal interpretation correctly, a translator must preserve the mechanism that could distort an observed association. These terms are not generic synonyms for “error.” Selection bias concerns how people or observations enter, remain in or contribute information to a study. Confounding concerns a third factor that distorts an exposure–outcome association. Information bias concerns how variables are measured or classified. Statistical adjustment is an attempt to account for specified differences under assumptions; it does not automatically make an observational result causal.
This world-facing guide explains how to translate research bias, confounders, adjusted estimates, stratification, standardisation, inverse-probability weighting, collider bias, residual confounding and causal language in epidemiology, social science, education, economics and data analysis. It is designed for research papers, policy summaries, evidence reviews and educational nonfiction. The examples are fictional and are not medical, legal or policy advice. Their purpose is to show how a seemingly harmless translation choice can alter what a study can legitimately claim.
The safest translation method is to build a causal reading of the sentence before rewriting it. Identify the exposure or predictor, outcome, study population, comparison, timing, candidate confounders, selection process and any adjustment method. Then ask whether the source reports an association, an adjusted association, an estimate intended to have a causal interpretation, or a limitation that weakens such an interpretation. Translation should preserve that evidence ladder rather than upgrading every statistically adjusted result into cause and effect.
A fifty-second orientation
The CDC Field Epidemiology Manual describes confounding as distortion of an exposure–disease association by a third factor and advises investigators to consider chance, selection bias, information bias, confounding and investigator error before concluding that an association is causal. CDC guidance on observational vaccine-effectiveness studies also illustrates how selection processes and confounding can bias effect estimates. These distinctions are useful far beyond health research because they map different pathways by which evidence can be distorted.
A compact translation rule is: name the distortion mechanism, preserve the direction of the relationship, preserve the adjustment status, and preserve the strength of the causal claim. “Associated with” is not automatically “caused by.” “Adjusted for age” is not automatically “free of confounding.” “Selected participants” is not automatically “selection bias.” The target language must retain the conditions under which the source uses each concept.
1. Bias is systematic distortion, not simply any mistake
In research methods, bias usually refers to systematic distortion in design, conduct, measurement, analysis or reporting that can move an estimate away from the target quantity. A typographical error is an error, but it is not necessarily a statistical bias mechanism. Random sampling variation is uncertainty, but it is not the same as systematic bias.
A target-language word meaning prejudice may be correct in some contexts but misleading in technical research prose if it suggests intentional human unfairness. Research bias can occur without conscious intent. The translator should choose the established methodological term or explain that the issue is systematic distortion.
Do not remove the adjective when the source names a specific bias. Selection bias, information bias, publication bias and recall bias describe different pathways. Turning them all into “possible bias” loses the mechanism that allows readers to judge the study.
2. Selection is not automatically selection bias
Every study selects participants or observations somehow. Selection bias arises when the selection process creates systematic differences relevant to the association or estimate being studied. A sample can be highly selective without necessarily producing bias for every research question, and a large sample can still be biased if the selection mechanism is problematic.
A translation saying “selection occurred, therefore the study is biased” may overstate the source. Preserve whether the authors say selection could have introduced bias, likely introduced bias, or was controlled by design. The strength of the methodological judgement matters.
When the source says “selection bias due to loss to follow-up,” keep the pathway. Participants first entered the study, then differential follow-up affected who contributed outcome data. This is different from a recruitment process that excludes certain people at baseline.
3. Recruitment bias begins before the analysis
Suppose a fictional study compares study habits and examination performance but recruits volunteers through an advanced study-skills club. The sample may overrepresent learners already motivated to improve. That recruitment pattern can affect estimates about the broader student population.
A target-language summary should not automatically say the study “proves study clubs improve marks.” First, recruitment may limit representativeness; second, club membership may be associated with motivation and other factors. The source may discuss both selection and confounding.
Keep words such as volunteer sample, convenience sample, population-based sample and probability sample distinct. These are sampling descriptions. Whether they create bias depends on the estimand and selection mechanism, but the translation should preserve the design facts that readers need to evaluate that question.
4. Loss to follow-up can create selection after enrolment
Imagine a cohort in which people with both high exposure and poor outcomes are more likely to stop responding. If analysis includes only participants with complete follow-up, the remaining sample may no longer represent the association that would have been observed had everyone remained.
Translation must distinguish “lost to follow-up” from “excluded at baseline,” “withdrew treatment,” and “missing one questionnaire item.” These events occur at different stages and can have different implications.
Do not call every loss to follow-up biased. The potential for bias depends on why missingness occurs and how it relates to exposure, outcome and analysis. Preserve cautious words such as may, could and potentially when the source is cautious.
5. Healthy-user and healthy-worker effects are selection patterns with context
Some observational studies encounter systematic differences between people who choose or remain in particular behaviours, jobs or programmes and those who do not. Terms such as healthy-user effect or healthy-worker effect describe specific patterns in certain research settings.
A literal translation can sound like a moral judgement about “healthy people.” The methodological point is that participation or employment status may select people with characteristics related to the outcome. The term should be explained in relation to the study mechanism.
Do not apply these labels to a new dataset merely because participants appear healthier. If the source uses the named bias, translate it. If the translator notices a possible mechanism independently, that belongs in an editorial query, not silently in the target text.
6. Confounding requires a third-variable relationship
Confounding occurs when the observed association between an exposure and an outcome is distorted by another factor related to both under the causal structure. The precise definition depends on the analytic framework, but the central idea is that part of the observed association is attributable to an alternative pathway rather than the exposure effect of interest.
A confounder is not merely any variable associated with the outcome. Nor is it any variable that improves prediction. The source should identify why the variable threatens the causal comparison. Translation should preserve that role rather than converting “adjusted covariate” automatically into “confounder.”
For general readers, “a third factor that can distort the exposure–outcome relationship” is often useful. But in technical writing, keep the established term confounder because the source may later distinguish confounders from mediators, moderators and colliders.
7. Association with exposure and outcome is not enough by itself
A common teaching rule says a confounder is associated with both exposure and outcome and is not on the causal pathway. This can be helpful but is not a complete causal definition in all settings. Modern causal analysis focuses on whether conditioning on a variable helps block a noncausal path without blocking the causal effect or opening a new path.
Translation should preserve the source framework. If an introductory textbook uses the traditional criteria, do not rewrite it as a full graphical-model discussion. If a research article uses directed acyclic graphs, do not simplify every variable into “a factor linked to both.”
This principle—translate the methodological level actually present—prevents both oversimplification and accidental invention. The target should be as advanced as necessary and no more authoritative than the source.
8. A worked confounding example: study time and marks
Imagine a fictional observational study finding that students who attend optional study sessions earn higher examination marks. Prior academic preparation may influence both attendance and later marks. If better-prepared students are more likely to attend, part of the association between attendance and marks could reflect prior preparation.
A source might say “the crude association was attenuated after adjustment for prior achievement.” Translate attenuated as reduced in magnitude, not eliminated, unless the adjusted estimate actually approaches the null and the authors say so.
Do not change “association after adjustment” into “effect after removing prior achievement.” Statistical adjustment does not physically remove students’ prior achievement. It changes the comparison produced by the model under assumptions.
9. Crude and adjusted estimates answer related but different comparisons
A crude estimate is calculated without the specified adjustment. An adjusted estimate incorporates a model, stratification, standardisation or weighting procedure intended to account for particular differences. The two values can differ substantially.
A target-language phrase such as “corrected estimate” can overstate what adjustment achieves. “Adjusted estimate” is usually safer because it says what was done without declaring that all bias has been corrected.
If the source lists variables included in adjustment, keep them. “Adjusted for age, baseline score and school” is more informative than “adjusted for confounders,” particularly when some included variables were selected for precision or design rather than because they were proven confounders.
10. Adjustment does not guarantee causal identification
An adjusted regression coefficient can still be biased if important confounders are unmeasured, variables are measured with error, the model is misspecified, selection bias is present or inappropriate variables are conditioned on. Translation should not turn “adjusted association” into “causal effect” unless the source makes and supports that causal claim.
The difference often rests on one verb. “X was associated with Y after adjustment” is not the same as “X caused Y after adjustment.” A translator seeking energetic prose should resist replacing associated with caused, led to or resulted in unless the evidence statement warrants it.
For public communication, “the association remained after accounting statistically for measured differences in…” can be clearer than jargon. Keep the word measured when it matters; the study cannot adjust for an unmeasured factor merely by acknowledging it exists.
11. Stratification can control confounding by comparing within strata
Stratification separates observations into levels of another variable and examines associations within those strata. If an apparent crude association changes substantially within strata, that pattern can reveal confounding or effect modification depending on the structure.
A translator should preserve whether stratification is a design, analysis or reporting device. “Results stratified by age group” does not necessarily mean the original sampling was stratified by age. Timing and purpose matter.
If the source calculates a pooled adjusted estimate across strata, keep the distinction between stratum-specific estimates and the combined estimate. Do not average numbers casually in the target text. The statistical method determines the weighting.
12. Standardisation changes the comparison population
Direct or indirect standardisation can compare rates after accounting for differences in a population structure such as age. The resulting standardised rate may be a constructed comparison quantity rather than the directly observed rate in either population.
A target phrase like “actual rate after correction” can mislead. Standardised rates are often intended for comparison, not as literal observed proportions. Preserve the reference or standard population when the source provides it.
The words crude, age-specific and age-standardised should remain distinct. A table can contain all three. Swapping labels while preserving numbers creates a serious interpretation error because readers will attribute the rate to the wrong population structure.
13. Inverse-probability weighting creates a weighted pseudo-population
Inverse-probability weighting assigns observations weights based on estimated probabilities of exposure, treatment or selection under a model. In causal analyses, the aim may be to create a weighted comparison with particular balance properties. The resulting dataset is not a new set of physically duplicated people.
A translation saying “participants were multiplied according to probability” would be misleading. Better language explains that observations contribute different statistical weights. The term pseudo-population should be treated as a conceptual weighted population, not a literal extra population.
Keep extreme-weight trimming, stabilisation or truncation terminology if the source reports those procedures. They alter the analysis and sometimes its target estimand. Do not simplify every weighting choice to “normalisation.”
14. Propensity scores are probabilities under a model, not proof of comparability
A propensity score is typically the estimated probability of receiving an exposure or treatment given measured covariates under the specified model. It can be used for matching, weighting, stratification or adjustment.
Translation should preserve whether a study matched on the propensity score, weighted by it or included it as a covariate. These are different procedures. “Propensity-adjusted” can be too vague if the method is central to interpretation.
Do not say propensity-score methods “make observational groups randomized.” They can improve measured covariate balance under assumptions, but unmeasured confounding may remain. A good translation keeps the methodological ambition and limitation together.
15. Matching can improve comparability but changes the analysed sample
Matching pairs or groups observations with similar values of selected variables. Some observations may remain unmatched and therefore be excluded from a matched analysis. This changes the population represented by the estimate.
A source reporting “1:1 propensity-score matching” should not become “participants were equally randomised.” Matching is an observational adjustment strategy unless the study separately uses random assignment.
Preserve the matching ratio, caliper or replacement rule if the source treats it as important. Those details affect which observations contribute and how close the matched groups are on the specified score.
16. Residual confounding means confounding may remain after adjustment
Residual confounding can remain because a confounder is measured imperfectly, modelled too crudely, omitted entirely or represented with insufficient detail. A study can therefore adjust for a variable and still acknowledge residual confounding.
A target phrase meaning “small remaining error” may understate the concept. Residual confounding is specifically the confounding that remains after the attempted adjustment. Its magnitude may be uncertain.
When the source says residual confounding cannot be excluded, retain that uncertainty. Do not rewrite it as “residual confounding was present” unless the authors actually establish that. Cannot exclude is an evidential limitation, not a confirmed diagnosis of bias.
17. Unmeasured confounding identifies a missing information problem
Unmeasured confounding refers to potential confounding by variables not measured adequately in the data. Statistical adjustment cannot directly condition on a factor that is entirely absent, though study design, proxy variables, instrumental-variable methods, sensitivity analysis or other strategies may address the problem under additional assumptions.
Translation should not suggest that regression automatically accounts for “all other factors.” Statements such as “controlling for other variables” are often too broad unless the variables are named. A more accurate target says “adjusting for the measured covariates listed in the model.”
If the source performs a sensitivity analysis for unmeasured confounding, preserve whether it estimates how strong such a confounder would need to be, explores scenarios or uses another method. Do not turn a sensitivity analysis into direct measurement of the missing variable.
18. Mediators are on a causal pathway, not ordinary confounders
A mediator lies on a causal pathway through which an exposure may affect an outcome. If the research question concerns the total effect, adjusting for a mediator can remove part of that effect from the estimate. A confounder, by contrast, creates a noncausal pathway that the analysis may need to block.
Translation should preserve whether a variable is described as mediator, intermediate, confounder or covariate. Replacing all with “control variable” can hide the causal role and make the analysis impossible to interpret correctly.
Suppose a fictional study asks whether a training programme improves scores partly by increasing practice time. Practice time could be a mediator. If the model adjusts for practice time, it may estimate a direct effect rather than the total effect. The target text should not call practice time a confounder unless the source does.
19. Effect modification is not confounding
Effect modification or interaction occurs when an association or causal effect differs across levels of another variable. This is a feature of the relationship being studied, not necessarily a bias to eliminate.
If an intervention appears more effective in one age group than another, the source may investigate interaction. A translation saying age “confounded the effect” can be wrong if age instead modifies the effect.
Preserve whether the authors describe subgroup differences as exploratory, statistically supported, prespecified or uncertain. Interaction claims can be sensitive to sample size and multiple testing. Do not strengthen a tentative subgroup pattern into a definitive difference.
20. Collider bias can be created by conditioning
A collider is a variable influenced by two other variables in a causal structure. Conditioning on a collider—or sometimes on its descendants—can open a noncausal association between causes that were otherwise independent along that path. This is why “adjust for everything available” is not a safe universal rule.
The term collider is highly technical and may not have an intuitive everyday equivalent. A good translation can keep the established technical term and explain that it is a common effect of two variables. What matters is preserving the direction of arrows in the causal structure.
Do not translate collider bias as simply “confounding.” Both can create misleading associations, but the mechanism differs. Confounding is typically addressed by controlling appropriate common causes; collider bias can be created by controlling the wrong common effect.
21. Conditioning includes more than regression adjustment
Researchers can condition on a variable by stratifying, matching, restricting, selecting, including it in a regression model or otherwise analysing within levels of that variable. Translation should not assume that conditioning always means adding a covariate to regression.
This broader meaning is important for collider bias. A study can induce selection bias by restricting analysis to a subgroup even without fitting a multivariable model. The selection itself conditions on eligibility.
When the source uses conditioning in a causal-inference sense, an everyday translation like “taking into account” may be too vague. Keep the technical term or specify the actual method used.
22. Information bias concerns measurement and classification
Information bias arises when information about exposure, outcome or other variables is measured, recorded or classified systematically incorrectly. Examples can include recall bias, interviewer bias or differential misclassification.
This is different from selection bias. A perfectly representative sample can still have poor measurements; a perfectly measured sample can still be selected in a biased way. Translation should keep mechanism names attached to the right stage of the study.
If a source says outcome assessors knew exposure status and might classify outcomes differently, preserve that directional concern. Do not reduce it to “data may be inaccurate.” The mechanism matters because it affects how bias may enter the estimate.
23. Recall bias is about differential remembering or reporting
Recall bias can arise when groups remember or report past exposures differently, often because of current outcome status or awareness. It is a specific information-bias mechanism, not simply human memory being imperfect.
A target phrase meaning “memory error” may be too broad. Random forgetting can add noise without producing the same systematic distortion as differential recall between comparison groups.
Preserve the comparison. If cases may search their memories more intensely than controls, the issue is not just that memories are inaccurate but that accuracy or reporting differs by study group.
24. Misclassification can be differential or nondifferential
Misclassification means a categorical variable is assigned incorrectly. Differential misclassification varies according to another relevant variable, while nondifferential misclassification follows a different error pattern under the study design. Their effects on estimates can differ.
Do not translate nondifferential as harmless or unbiased. Nondifferential measurement error can still distort estimates, and the direction is not universally toward the null in every complex setting.
A strong target sentence preserves the classification rule and which groups may be measured differently. This lets readers understand the mechanism without relying on an oversimplified slogan about the direction of bias.
25. Reverse causation is a temporal interpretation problem
Reverse causation arises when the outcome, or an early stage of it, influences the exposure rather than the exposure causing the outcome as initially hypothesised. Cross-sectional and observational studies can be especially vulnerable when timing is unclear.
Suppose a fictional study finds that students receiving more tutoring have lower baseline marks. It would be wrong to conclude tutoring causes low marks if low performance prompted families to seek tutoring. The direction of causation may run partly from outcome-related need to exposure.
Translation should preserve temporal words such as preceded, followed, predicted, was associated with subsequent, and may reflect reverse causation. Replacing these with generic cause-and-effect verbs can erase the core limitation.
26. Temporality is necessary for many causal claims but not sufficient
For an exposure to cause a later outcome in the ordinary causal sense, the exposure must precede the outcome. But temporal order alone does not eliminate confounding, selection bias or measurement error.
A longitudinal study can establish timing better than a cross-sectional snapshot while still remaining observational. Do not translate longitudinal as causal. It describes repeated or time-ordered observation, not automatic freedom from bias.
Preserve how strongly the source speaks: “prospectively associated with” is more cautious than “led to.” An editor may prefer vivid verbs, but causal vocabulary is a substantive claim, not a style choice.
27. Directed acyclic graphs preserve assumptions visually
Directed acyclic graphs, often abbreviated DAGs, represent assumed causal relationships with nodes and arrows. They can help identify adjustment sets, confounders, mediators and colliders under the specified causal model.
Translation of a DAG must preserve arrow direction. Reversing one arrow changes the assumed causal structure. Labels can be translated, but the graph topology is data-like content and should not be rearranged for aesthetic symmetry without methodological authority.
A caption saying “assumed causal diagram” should not become “proven causal pathway.” DAGs encode assumptions and subject-matter knowledge used for analysis. They are tools for reasoning, not empirical proof that every arrow is correct.
28. Negative controls and falsification checks probe alternative explanations
Some observational studies use negative-control exposures or outcomes expected not to have the causal relationship under study. Associations with a negative control can suggest residual bias or uncontrolled structure. These methods rely on their own assumptions.
A target phrase like “control group” may be misleading because negative controls are not necessarily groups of participants. They can be exposures, outcomes or analyses. Preserve the type of control.
Likewise, a falsification check does not “falsify the study” merely because it is performed. It tests whether patterns appear where the causal explanation predicts they should not. Translate the intended diagnostic role rather than the ordinary emotional force of the word false.
29. Sensitivity analysis asks how conclusions respond to assumptions
A sensitivity analysis changes assumptions, definitions, models or analytic choices to see how conclusions respond. In bias analysis, it may examine how strong unmeasured confounding would need to be to explain an observed association.
Translation should preserve the distinction between “the result remained similar under tested alternatives” and “the result is proven robust.” Sensitivity analysis explores a defined set of alternatives. It cannot establish stability against every unknown source of error.
If the source reports a tipping point, threshold or E-value, keep the exact quantity and interpretation. Do not turn a sensitivity measure into a probability that confounding exists unless the method actually defines such a probability.
30. Association, prediction and causation are three different claims
A variable can be associated with an outcome without causing it. It can predict an outcome accurately without being a causal driver. It can be causally important while adding little predictive power beyond other variables. These are different scientific roles.
Translation often goes wrong when strong predictive language is mistaken for causal language. “X predicts Y” in a regression or machine-learning context may mean X contains information useful for predicting Y. It does not automatically mean changing X would change Y.
Keep verbs disciplined: associated with, predicts, precedes, is linked to, estimates the effect of, increases risk, causes and mediates all make different claims. A precise article may repeat the same cautious verb many times because that repetition protects the evidence level.
31. A worked example: crude association weakens after adjustment
Imagine a fictional observational study in which students attending an optional workshop score ten points higher on average than non-attendees. After adjustment for baseline score, attendance motivation and school year, the estimated difference is three points, with substantial uncertainty.
A flawed target summary says: “The workshop improved marks by ten points, but after correcting bias the true effect was three points.” This makes two unsupported upgrades. The crude difference was not necessarily causal, and the adjusted estimate is not automatically the true effect.
A better translation says: “Workshop attendance was associated with a ten-point higher average score before adjustment. After statistical adjustment for the measured baseline differences, the estimated association was three points.” If the authors make a causal claim under explicit assumptions, that causal language can then be translated with those assumptions attached.
The example shows why adjusted and true are not synonyms. Statistical models construct comparisons under specified data and assumptions. Translation should not turn a modelling step into a guarantee.
32. A worked example: selection after enrolment changes the analysed population
A fictional cohort enrols 2,000 participants. Follow-up is complete for 1,500. Participants with the highest exposure and worst outcomes are disproportionately missing at follow-up. A complete-case analysis therefore excludes many observations from the combination most relevant to the exposure–outcome relationship.
A target sentence saying “500 participants were randomly missing” would be unsupported. Missing data may follow a systematic pattern. Conversely, saying “the result is biased” may be stronger than the source if investigators only say bias is possible.
A faithful translation can say: “Differential loss to follow-up may have introduced selection bias because missingness was more common among participants with both high exposure and poor outcomes.” The mechanism, direction and uncertainty are all visible.
33. A worked example: adjusting for a collider creates a new association
Consider a simplified fictional causal structure in which both academic talent and intensive tutoring affect admission to a selective programme. Suppose talent and tutoring are otherwise independent in the source population. If analysis is restricted only to admitted students, the restriction conditions on admission, a common effect of both variables.
Within the admitted group, lower talent may be statistically associated with more tutoring because either route can help produce admission. A translator should not describe this induced relationship as evidence that tutoring reduces talent. The selection process created the association under the simplified model.
This example is a useful warning against translating “adjusted for more variables” as “more accurate.” Conditioning can sometimes reduce bias and sometimes create it, depending on causal structure. The variable’s role matters more than the number of covariates.
34. Practice clinic with explained answers
Practice one. A study says “selection bias may have occurred because response rates differed by exposure group.” Preserve may. Do not translate this as confirmed bias unless the source establishes it.
Practice two. A variable is adjusted for in regression. That does not automatically make it a confounder. Preserve the source label—covariate, predictor, confounder or adjustment variable.
Practice three. The crude risk ratio is 2.0 and the adjusted ratio is 1.3. Do not call 1.3 the “true ratio” unless the source does and supports that interpretation. It is the estimate after specified adjustment.
Practice four. A source says “residual confounding cannot be ruled out.” This means remaining confounding is possible; it does not mean investigators measured and confirmed a residual confounder.
Practice five. A study is longitudinal. Do not translate longitudinal as causal. The design gives temporal information but does not eliminate all alternative explanations.
Practice six. A mediator lies on the pathway from exposure to outcome. Do not rename it a confounder merely because it is statistically associated with both.
Practice seven. An interaction term suggests effects differ across groups. Do not call the interacting variable a bias. Effect modification can be scientifically meaningful heterogeneity.
Practice eight. A DAG shows X → M → Y. Reversing an arrow in a translated figure changes the assumed causal model. Translate labels, not topology, unless the source is being corrected with authority.
Practice nine. A propensity-score matched study is observational. Do not call it randomised after matching. Matching attempts to improve comparability on measured variables but does not create random assignment.
Practice ten. A study adjusts for age, sex and baseline score but not family income because it was not measured. Do not write “all confounding factors were controlled.” The adjustment set is limited to measured variables included in the model.
35. Frequently asked research-bias translation questions
Is every association biased? No. Bias refers to systematic distortion. An association can be unbiased under assumptions, biased upward, biased downward or affected mainly by random uncertainty. Translate the source’s assessment.
Does statistical adjustment remove confounding? It can reduce confounding from appropriately measured and modelled variables under assumptions, but residual and unmeasured confounding may remain. “Adjusted” is safer than “corrected.”
Are selection bias and confounding the same? No. They can interact, but they arise through different mechanisms. Selection concerns who contributes to the analysed data; confounding concerns alternative pathways that distort the exposure–outcome comparison.
Can an observational study make causal claims? Some observational analyses are designed for causal inference under explicit assumptions and methods. Translation should preserve those assumptions and the strength of the authors’ claim rather than applying a blanket yes or no.
What is the strongest final check? Draw a small verbal map: who entered the study, what was measured, what was compared, which variables were adjusted for, what could distort the comparison, and how strongly the source interprets the result. The target should give the same answers.
36. Connect this guide to the existing eduKate translation architecture
This specialist article sits under the established Master Art of Translation owner rather than replacing it. Nearby `Translate` articles cover randomised trials, regression coefficients, p-values, confidence intervals, risk ratios, incidence and survival analysis. This guide focuses on the validity pathways that determine what those estimates can mean.
The Vocabulary Learning Hub supports precise distinctions among confounder, mediator, collider, covariate, predictor and modifier. How English Works supports the grammar of hedging, modality and causation: may, could, is associated with, predicts, leads to and causes are not stylistic substitutes when evidence strength differs.
The final rule is to preserve the pathway by which the source thinks distortion might occur. A research translation is not accurate merely because every estimate and p-value survives. The causal roles, selection process, measurement process, adjustment set and evidential caution must survive too. Good translation keeps the limits of knowledge as carefully as it keeps the findings.