What is a Correlational Research Design in Research Methodology?
A correlational design measures variables as it finds them and reports how they move together. The hard part is not computing r — it is knowing what r can and cannot carry, and being honest about whether the question behind the study was causal all along.
- What separates a correlational design from an experiment, and from a cross-sectional one
- Which coefficient to use, what it assumes, and why the usual strength bands are arbitrary
- The three failure modes that are specific to correlational work
- “Writing ‘associated with’ instead of ‘causes’ makes the study non-causal.” It does not. It only hides what the study was for. Hernán argues that swapping causal language for association language “just sweeps it under the rug”, that the methods for estimating causal effects differ from those for estimating associations, and that stating the causal objective explicitly “reduces ambiguity in the scientific question”. Note also that every estimate from a randomised trial is an association too — what makes it causal is the design and the assumptions, not the verb.
What makes a design correlational
In an experimental study an investigator allocates the intervention. Cochrane’s EPOC group defines both randomised and non-randomised trials as “an experimental study in which people are allocated to different interventions”, differing only in whether the allocation method is random.3
A correlational design does none of that. The investigator allocates nothing. Both variables are measured as they occur, and the analysis reports how they vary together. That single fact — no allocation — is what places it outside the experimental family and determines everything that follows about what it can support.
What you get in exchange is reach. Questions that cannot be randomised for ethical, practical or political reasons can still be studied, and existing data can be used. What you give up is the mechanism that makes groups comparable.
The methods literature places this inside a wider category. Nonexperimental research is “research that lacks the manipulation of an independent variable, random assignment of participants to conditions or orders of conditions, or both”, and correlational research is the type “in which the researcher measures two variables and assesses the statistical relationship (i.e., the correlation) between them with little or no effort to control extraneous variables”.11
| Feature | What it means | Why it matters |
|---|---|---|
| No manipulation | “The defining feature of correlational research is that neither variable is manipulated”11 | Nothing is done to participants, so nothing can be attributed to what was done |
| No random allocation | Participants are not assigned to conditions11 | Groups are not made comparable, so differences between them remain available as explanations |
| Little control of extraneous variables | Variables are measured “with little or no effort to control extraneous variables”11 | Confounding is addressed in the analysis, if at all, rather than by design |
| Quantified direction and strength | “Correlation coefficients describe the strength and direction of an association between variables”1 | The output is a signed number, not a group difference |
Nonexperimental work is therefore not a weaker substitute chosen out of convenience. It is “appropriate — even necessary — when these conditions are not met”.11

Correlational, cross-sectional, observational
These three words are routinely used as though they were interchangeable. They describe different things, and a study can be all three at once.
| Term | What it describes | The axis it sits on | Opposite or alternative |
|---|---|---|---|
| Observational | No allocation by the investigator | Who decides what happens to participants | Experimental — someone allocates3 |
| Cross-sectional | Everything is measured at one point in time | When measurements are taken | Longitudinal, cohort, case-control4 |
| Correlational | The analysis reports association between measured variables | What the analysis does with the data | Comparative, predictive or causal analysis of the same data |
A correlational analysis can sit on top of cross-sectional data, cohort data, or even trial data. Calling a study “correlational” says what was computed; it does not by itself say when the data were collected or who decided the exposure. Saying all three explicitly is more informative than picking one as a label.
Simple, multiple and partial correlation
These three names describe how many variables are in play. They are a separate question from which coefficient you use, which is covered in the next section.
Simple correlation
Two variables, one coefficient. The correlation coefficient is “a measure of the linear relationship between two variables”, in which “a perfect linear relationship yields a correlation coefficient of +1 (or −1 for a negative relationship)” and “no linear relationship yields a correlation coefficient of 0”.9 This is the form the rest of this page mostly discusses, and the form the worked example uses.
Multiple correlation
One outcome against a set of predictors taken together. The multiple correlation coefficient “can be defined as the Pearson Correlation between the outcome variable, y,” and “the predicted values,” “from regressing y onto a set of x variables”.10
Because it correlates the outcome with its own predicted values, “the Multiple R is always greater than or equal to zero (i.e., always positive)”.10 A simple r of −0.70 and one of +0.70 are different findings. A multiple R of 0.70 carries no direction at all, so the direction of each predictor has to be read from the regression coefficients rather than from R.
Partial correlation
Two variables, with the influence of others taken out. “Partial correlation is the correlation between two variables after removing the effect of one or more additional variables.”9 Where a third variable plausibly drives both, this is the form that asks whether anything is left once it is accounted for.
| Form | What it relates | What you get | What to watch |
|---|---|---|---|
| Simple | Two variables9 | A signed coefficient between −1 and +11 | Sign gives direction, magnitude gives strength |
| Multiple | One outcome and a set of predictors10 | R, never negative10 | No direction can be read from R itself |
| Partial | Two variables, holding one or more others constant9 | A signed coefficient answering a narrower question | It answers a different question from the simple coefficient, so the two are not interchangeable |
Choosing the right coefficient
Pearson’s coefficient measures “a linear relationship between 2 continuous, random variables”.1 Two words in that sentence do most of the work: linear, and continuous.
| Coefficient | Measures | Assumes | Use it when | Breaks when |
|---|---|---|---|---|
| Pearson product-moment (r) | A linear relationship between two continuous random variables1 | “Both variables are continuous, jointly normally distributed, random variables”1 | Data are jointly normal and the relationship looks straight1 | The relationship is curved, or extreme outliers are present — these “may have undue influence”1 |
| Spearman rank correlation | A monotonic association1 | Only that the data can be ranked | “For nonnormally distributed continuous data, for ordinal data, or for data with relevant outliers”1 | The relationship is not monotonic — rises then falls |
A correlation can be near 1 while two measurements disagree systematically — if one instrument reads consistently 5 units high, the two still move together perfectly. As Schober and colleagues put it, “correlations also do not describe the strength of agreement between 2 variables”.1 Method-comparison questions need an agreement analysis, not a correlation.
Interpreting r
Almost every page on this subject publishes a table mapping ranges of r onto words like weak, moderate and strong. It is worth knowing what the methodological literature says about those tables.
Schober and colleagues present such a scheme as “an example” of a conventional approach, and then state that these cutoffs “are arbitrary and inconsistent and should be used judiciously”.1 They are a convention, not a finding — which is why different sources publish different bands, and why no band table should be presented as though it settled anything.
Three things are more informative than a label:
- The scatterplot. “Visual inspection of scatter plots is always advisable.”1 Figure 1 below shows why this is not optional advice.
- The confidence interval. An r of 0.45 from 20 participants and an r of 0.45 from 2,000 are different claims; only the interval shows it.
- The observed range of both variables. Correlation depends on the variance available to it, so the range you sampled is part of the result.
Correlation also “fails to adequately describe nonlinear or nonmonotonic relationships”.1 A coefficient near zero is not evidence of no relationship; it is evidence of no linear one.
How r is computed, and the values it can take
Pearson’s coefficient is “a dimensionless measure of the covariance, which is scaled such that it ranges from −1 to +1”.1 The National Institute of Standards and Technology sets out the computation in sums of squares and cross-products.9
The two extremes and the midpoint have exact meanings. “A perfect correlation of −1 or +1 means that all the data points lie exactly on the straight line”, while 0 “indicates that there is no linear or monotonic association”.1 A value of +1 is a perfect positive relationship, −1 a perfect negative one.
The conventional strength bands
The band table below is the convention most often reproduced. It is set out here applied to the absolute value of r, so that −0.55 and +0.55 are read the same way: the sign gives the direction of the relationship, and the magnitude gives its strength.
| Absolute value of r | Strength of relationship |
|---|---|
| 0.00 ≤ |r| < 0.20 | Very weak |
| 0.20 ≤ |r| < 0.40 | Weak |
| 0.40 ≤ |r| < 0.60 | Moderate |
| 0.60 ≤ |r| < 0.80 | Strong |
| 0.80 ≤ |r| ≤ 1.00 | Very strong |
The bands are written as continuous intervals on purpose. Versions that run 0.00–0.19, 0.20–0.39 and so on leave every value between them unassigned, so a coefficient of 0.195 belongs to no band at all.
Schober and colleagues publish a different set of cutoffs for the same purpose, treating “a coefficient of <0.1” as indicating “a negligible” relationship and “>0.9 a very strong” one, under the caption “Example of a Conventional Approach to Interpreting a Correlation Coefficient”.1 On that scheme a coefficient of 0.15 is negligible; on the table above it is very weak but not negligible. The same paper states plainly that such cutoffs “are arbitrary and inconsistent and should be used judiciously”.1 Report the coefficient, its confidence interval and the plot. A band label may accompany them, but it cannot stand in for them.
A worked example
The following example puts the coefficient, the sign and the band table together, and shows where the reading has to stop.
- Question
- Whether smartphone use and sleep quality are associated among college students
- Variables
- Smartphone use in hours per day, and a sleep quality score. Both are measured as they occur. Neither is allocated by the investigator, so neither is an independent variable in the experimental sense
- Data collection
- A questionnaire completed by 200 students
- Analysis
- Pearson product-moment correlation, after inspecting the scatterplot1
Where the reading stops. The design cannot establish which of the two came first, nor rule out a third factor — a later bedtime, shift work, an undiagnosed sleep disorder — that raises one and lowers the other. A coefficient of −0.55 is equally compatible with smartphone use affecting sleep, with poor sleep prompting more smartphone use, and with neither. The coefficient also describes a linear association only, which is why the scatterplot is inspected before the number is reported.1
What the data look like
Correlation is the one statistic that most rewards plotting. The first figure is the classic demonstration of why, and it uses real published data.
Four datasets with the same mean, the same regression line and the same correlation coefficient.
All four have r = 0.82, the same means (x̄ = 9.0, ȳ = 7.5) and the same least-squares line, y = 3 + 0.5x.2 Only one of them is a dataset you would willingly summarise with a correlation coefficient. This is why “visual inspection of scatter plots is always advisable”.1
Real data, plotted from the values published by Anscombe in 1973.2
The same relationship measured across groups and across individuals can point in opposite directions.
Robinson’s 1930 census example, as reported by Freedman. Across the 48 US states, the correlation between percent foreign-born and percent literate was +0.53. Computed across individuals it was −0.11.5 The aggregate figure suggests the foreign-born were more literate; they were not. They lived in states where the native-born happened to be more literate. “The ecological correlation gives the wrong inference.”5
Schematic scatter; the two correlation values are the published figures.5
Measuring a relationship over a narrow slice of one variable shrinks the correlation you observe.
What to look for. A weak correlation in a narrow sample is not evidence of a weak relationship. Selecting on the predictor — studying only high achievers, only severe cases, only one clinic — compresses the variance that correlation is computed from. Report the observed range of both variables so a reader can judge this for themselves.
Schematic. Illustrative shapes, not data from any study.
The causation question, properly
“Correlation does not imply causation” is true, universally repeated, and not much help. The useful version of the argument is more uncomfortable.
Schober and colleagues state the limit plainly: “an observed correlation (ie, association) does not assure that the relationship between 2 variables is causal”.1 That is the floor. The question is what to do about it.
The common response is linguistic — write “associated with”, never “causes”, and consider the problem handled. Hernán argues this fails. Journals ask authors to swap causal terms for association terms, so causal aims are described indirectly, which “impedes a frank discussion about methodology”. Because the methods for estimating causal effects differ from those for estimating associations, the ambiguity produces confusion rather than caution.6
Every estimate from a randomised trial is also an association. What licenses the causal reading is the design and the assumptions behind it, not the choice of verb. Dropping causal language “just sweeps it under the rug” rather than resolving anything.6 Hernán’s recommendation is the opposite of euphemism: state the causal objective explicitly, because doing so “reduces ambiguity in the scientific question” — which in turn means fewer analysis errors and less overinterpretation.6
For a correlational study this cashes out as: say what you were trying to find out, say what would have to be true for the association to be causal, and say which of those conditions your design cannot establish. If the aim is genuinely causal and randomisation is impossible, a quasi-experimental design may get closer than a correlation will — the methods exist precisely for that situation.7
Threats specific to correlational designs
Three failure modes are characteristic of this design rather than general research hazards.
The ecological fallacy
“The ecological fallacy consists in thinking that relationships observed for groups necessarily hold for individuals.”5 A correlation computed on groups is an ecological correlation, “because the unit of analysis is not an individual person but a group of people”.5 Figure 2 shows the canonical case, where the correlation does not merely weaken at the individual level — it changes sign.
Restriction of range
Correlation is computed from variance. Sample a narrow slice of one variable and you remove the variance the coefficient needs, and the correlation shrinks toward zero whether or not the underlying relationship did. Figure 3 shows the mechanism. This is why a near-zero correlation in a highly selected sample is weak evidence of anything.
Confounding and selection
Because nobody allocated anything, the groups being compared differ in every respect that led them to differ in the first place. In the language of the evaluation literature, “the bias potentially of concern here is ‘selection’ bias”,7 and statistical adjustment can only touch the characteristics that were measured. A correlational study is appraised on exactly these grounds: ROBINS-I, the tool for non-randomised studies, opens with confounding and selection of participants among its seven bias domains.8
Reporting a correlational study
The relevant guideline is STROBE — the Strengthening the Reporting of Observational Studies in Epidemiology statement, “a checklist of 22 items” covering “three main study designs: cohort, case-control, and cross-sectional studies”.4
It exists because “the reporting of such research is often inadequate, which hampers the assessment of its strengths and weaknesses”, and because in published observational research “important information is often missing or unclear”.4
It is a reporting checklist, not a design manual. Its authors say so directly: “these recommendations are not prescriptions for designing or conducting studies”, and the statement “should not be interpreted as an attempt to prescribe the reporting of observational research in a rigid format”.4 Clearing STROBE means your study is described completely, not that it was well designed.
Common mistakes
Most of these are decided before any coefficient is computed.
| Mistake | Stage | What it costs | What to do instead |
|---|---|---|---|
| Calling one variable the “independent variable” | Design | Contradicts the design — nothing was manipulated, so nothing is independent in the experimental sense | Name them both as measured variables, or state the causal direction you are proposing and why6 |
| Sampling a narrow slice of the predictor | Sampling | Restriction of range drags the correlation toward zero regardless of the real relationship | Sample across the range, and report the observed range of both variables |
| Using Pearson on skewed or ordinal data | Analysis | The assumption of continuous, jointly normal variables fails1 | Use Spearman for non-normal, ordinal, or outlier-affected data1 |
| Reporting r without looking at the scatterplot | Analysis | Anscombe’s quartet is the whole argument: four radically different datasets, one coefficient2 | Plot first. “Visual inspection of scatter plots is always advisable”1 |
| Reading r near zero as “no relationship” | Interpretation | Correlation “fails to adequately describe nonlinear or nonmonotonic relationships”1 | Say no linear relationship, and show the plot |
| Quoting a strength band as though it were a finding | Interpretation | The cutoffs “are arbitrary and inconsistent and should be used judiciously”1 | Report r with a confidence interval and the plot; use labels loosely if at all1 |
| Using correlation to show two methods agree | Interpretation | Correlations “do not describe the strength of agreement between 2 variables”1 | Use an agreement analysis for method comparison |
| Inferring individual behaviour from group-level data | Interpretation | The ecological fallacy — Robinson’s correlation flipped from +0.53 to −0.115 | Match the unit of analysis to the unit you want to make claims about5 |
| Hiding a causal aim behind “associated with” | Write-up | “Just sweeps it under the rug” and impedes frank discussion of methodology6 | State the causal objective, then state what the design cannot establish6 |
| Presenting a hypothetical example as a study | Write-up | An illustrative figure read as evidence | Label worked examples as illustrative, and cite real studies as real |
Where correlational designs are used
The design appears wherever allocation is impossible, unethical or beside the point. The studies below are real, and each is cited so the coefficient can be checked against its source.
| Field | What was correlated | Reported coefficient | How the authors describe the design |
|---|---|---|---|
| Psychology | Loneliness and depression among 654 young adults aged 18 to 29, recruited in March 202012 | “loneliness in young adults was positively correlated with depression (r = 0.531, p < 0.001)”12 | “by adopting a cross-sectional research approach”12 |
| Medicine | HbA1c and fasting plasma glucose in 3,523 adults aged 30 and above13 | “There was a correlation between HbA1c and FPG (r = 0.84; P < 0.0001)”13 | “designed as a population based investigation”, described elsewhere in the paper as cross-sectional13 |
| Education | Teacher self-efficacy and student academic achievement, 60 teachers and 100 students in Quetta14 | “r ═ .253, p ≤ .05, n ═100”14 | “an exploratory quantitative research design was chosen”14 |
| Business and organisations | Transformational leadership and work engagement among nurses in five Iranian public hospitals15 | “transformational leadership has positive relationship with work engagement (r = 0.70)”15 | “A descriptive, correlational, cross-sectional design was used”15 |
The last column is worth reading across. Only one of the four calls itself correlational. The others call themselves cross-sectional, population-based or exploratory quantitative, and one paper uses two different labels for itself. The design is defined by what was done, not by the word the authors chose.
Advantages and limitations
The advantages are practical and real: correlational designs study questions that cannot ethically or practically be randomised, work with data that already exist, handle many variables at once, and are often the only feasible first step when a field is young.
The limitations are structural rather than fixable:
- No allocation means no guaranteed comparability. Selection bias is the standing threat, and adjustment reaches only the measured characteristics.7
- r describes linear association only, and says nothing about agreement or about nonlinear relationships.1
- The unit of analysis constrains the claim. Group-level correlations do not transfer to individuals.5
- The observed range shapes the coefficient, so the sampling frame is part of the result.
- Causal questions need causal designs. Where randomisation is impossible, quasi-experimental methods are built for the gap.7
Take-home message
A correlational design measures variables rather than assigning them, so r carries association and nothing more. Plot the data before you trust the coefficient, match the unit of analysis to the claim you want to make, report the interval and the range rather than a strength label — and if the question was causal, say so instead of hiding it behind “associated with”.
How to apply this to your work
- State whether anyone allocated anything. If nobody did, the study is observational and no phrasing changes that.3
- Say whether your aim is causal. If it is, say so and name the assumptions; ambiguity causes analysis errors, not caution.6
- Match the unit of analysis to the claim. Group data support group claims.5
- Sample across the range of both variables, and report the range you achieved.
- Plot before you compute. Then pick Pearson or Spearman from what the plot shows.1
- Report r with a confidence interval, not a strength label on its own.1
- Write up against STROBE’s 22 items, remembering it checks description, not design.4
Common questions
Is a correlational study the same as a cross-sectional study?
No. “Observational” describes who allocates the exposure: no one does. “Cross-sectional” describes when measurements are taken — all at once. “Correlational” describes what the analysis does. A study can be all three, and saying all three is more informative than choosing one as a label.34
Should I use Pearson or Spearman?
Pearson measures “a linear relationship between 2 continuous, random variables” and assumes both are “continuous, jointly normally distributed, random variables”. Spearman measures a monotonic association and is appropriate “for nonnormally distributed continuous data, for ordinal data, or for data with relevant outliers”.1 Decide from the scatterplot, not from habit.
What counts as a “strong” correlation?
There is no principled answer, and that is the honest reply. The conventional band tables are presented in the methodological literature as “an example” of a convention, with the explicit caveat that such cutoffs “are arbitrary and inconsistent and should be used judiciously”.1 Report the coefficient with its confidence interval and the plot, and let the reader judge.
Does a correlation near zero mean there is no relationship?
It means there is no linear one. Correlation “fails to adequately describe nonlinear or nonmonotonic relationships”.1 Set II of Anscombe’s quartet is a perfect curve with a correlation of 0.82, which makes the converse point just as sharply.2
Can a correlational study ever support a causal claim?
Not on its own, and the limit is stated plainly: “an observed correlation (ie, association) does not assure that the relationship between 2 variables is causal”.1 But the right response is not to avoid causal words. Hernán argues that doing so “just sweeps it under the rug”, and that stating the causal objective explicitly “reduces ambiguity in the scientific question”.6 Where the aim is causal and randomisation is impossible, quasi-experimental designs are built for that situation.7
Why did my correlation shrink in a subgroup?
Often restriction of range. Correlation is computed from variance, so narrowing the sampled range of either variable compresses the coefficient toward zero whether or not the underlying relationship changed. Report the observed range of both variables so a reader can tell the difference.
Declared interest: Pubrica provides study design, statistical analysis and manuscript support commercially. Nothing on this page should be read as evidence that any design or analysis choice, or paid assistance with one, improves the likelihood of publication — no claim of that kind is made here.
References and sources
- Schober P, Boer C, Schwarte LA. Correlation coefficients: appropriate use and interpretation. Anesth Analg. 2018;126(5):1763–1768. doi:10.1213/ANE.0000000000002864 read 20 October 2026.View on Anesthesia & Analgesia
- Anscombe FJ. Graphs in statistical analysis. The American Statistician. 1973;27(1):17–21. data values read 20 October 2026 from a published reproduction of the quartet.View the reproduced dataset
- Cochrane Effective Practice and Organisation of Care (EPOC). What study designs can be considered for inclusion in an EPOC review and what should they be called? EPOC Resources for review authors; 2017. read 20 October 2026.View on Cochrane EPOC
- von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP, et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. PLoS Med. 2007;4(10):e296. doi:10.1371/journal.pmed.0040296 read 20 October 2026.View on PLOS Medicine
- Freedman DA. Ecological inference and the ecological fallacy. Technical Report No. 549, Department of Statistics, University of California, Berkeley; 1999. Prepared for the International Encyclopedia of the Social & Behavioral Sciences. Reporting Robinson’s 1950 analysis of the 1930 US census. read 20 October 2026.View the technical report
- Hernán MA. The C-word: scientific euphemisms do not improve causal inference from observational data. Am J Public Health. 2018;108(5):616–619. doi:10.2105/AJPH.2018.304337 read 20 October 2026.View on the American Journal of Public Health
- White H, Sabarwal S. Quasi-Experimental Design and Methods. Methodological Briefs: Impact Evaluation No. 8. Florence: UNICEF Office of Research; 2014. read 20 October 2026.View on BetterEvaluation
- Sterne JAC, Hernán MA, Reeves BC, Savović J, Berkman ND, Viswanathan M, et al. ROBINS-I: a tool for assessing risk of bias in non-randomised studies of interventions. BMJ. 2016;355:i4919. doi:10.1136/bmj.i4919 read 20 October 2026, via the National Collaborating Centre for Methods and Tools registry entry.View on NCCMT
- National Institute of Standards and Technology, Information Technology Laboratory. Correlation; and Partial Correlation. Dataplot Reference Manual, Auxiliary Chapter. read 20 October 2026.View the NIST reference pageView the NIST partial correlation page
- Beasley TM. Multiple R IS the square root of R2: multiple correlation coefficient using matrix formulation. General Linear Model Journal. 2019;45(1):23–32. read 20 October 2026.View the article PDF
- Price PC, Jhangiani R, Chiang I-CA. Research Methods in Psychology. 2nd Canadian edition. Victoria, BC: BCcampus; 2015. Chapter 7, Nonexperimental research. read 20 October 2026.View the chapter
- Lv F, Yu M, Li J, Tan J, Ye Z, Xiao M, et al. Young adults’ loneliness and depression during the COVID-19 pandemic: a moderated mediation model. Frontiers in Psychology. 2022;13:842738. doi:10.3389/fpsyg.2022.842738 read 20 October 2026.View on Frontiers in Psychology
- Ho-Pham LT, Nguyen UDT, Tran TX, Nguyen TV. Discordance in the diagnosis of diabetes: comparison between HbA1c and fasting plasma glucose. PLOS ONE. 2017;12(8):e0182192. doi:10.1371/journal.pone.0182192 read 20 October 2026.View on PLOS ONE
- Shahzad K, Naureen S. Impact of teacher self-efficacy on secondary school students’ academic achievement. Journal of Education and Educational Development. 2017;4(1):48–72. read 20 October 2026.View the full text on ERIC
- Hayati D, Charkhabi M, Naami A. The relationship between transformational leadership and work engagement in governmental hospitals nurses: a survey study. SpringerPlus. 2014;3:25. doi:10.1186/2193-1801-3-25 read 20 October 2026.View on SpringerPlus
How to cite this article
Pick your referencing style and copy it. Your university or target journal may vary the punctuation — check its guide if it has one.
This article is written to be learned from. Cite it if you are referring to what it says; if you are using the evidence behind it, cite the sources listed above instead.
Settle these before you report a coefficient.
[ ] Nobody allocated anything — confirm the design is observational [ ] The aim is stated: descriptive, predictive, or causal [ ] Unit of analysis matches the unit the claim is about [ ] Both variables sampled across their range, range reported [ ] Scatterplot inspected before choosing a coefficient [ ] Pearson only if continuous, jointly normal, linear [ ] Spearman if non-normal, ordinal, or outliers present [ ] Outliers identified and their influence checked [ ] r reported with a confidence interval, not a label alone [ ] Nonlinearity considered before reading r near zero as "no relationship" [ ] Correlation not used as evidence of agreement between methods [ ] Confounders named, and what adjustment cannot reach stated [ ] Causal language matched to the design, not softened to hide the aim [ ] Write-up checked against STROBE's 22 items
We help you match the analysis to what the data can support, choose the coefficient from the scatterplot rather than from habit, and write the result so the causal language matches the design.