Measures two or more variables as they naturally occur and tests whether — and how strongly — they are related. It cannot, by itself, establish causation.
Research designQuantitative / non-experimental
What it is
Correlational research examines relationships among variables that the researcher measures but does not manipulate: Does job satisfaction relate to turnover intention? Is advertising spend associated with brand recall? Variables are measured as they naturally occur — usually in a survey or from records — and statistical techniques quantify the direction and strength of association, from simple correlations to full structural equation models.
The design's great strengths are realism and reach. It studies variables that cannot ethically or practically be manipulated (leadership style, income, personality), handles many variables at once, and scales to large samples. Most published survey research in management — including the typical PhD questionnaire study testing a model of predictors, mediators and moderators — is correlational, whatever its statistical sophistication.
Its defining limitation is causal ambiguity. If X and Y correlate, X may influence Y, Y may influence X, or a third variable Z may drive both. Statistical controls, theory and temporal separation can strengthen a causal argument, but no analysis of non-experimental cross-sectional data — not even SEM — turns correlation into demonstrated causation. Disciplined correlational researchers write “is associated with” and “predicts”, never “causes”.
✓ When to use
The question is whether variables are related, and how strongly, in natural settings
The presumed cause cannot be manipulated — ethically (stress, layoffs) or practically (personality, firm size)
You are testing a theoretical model with multiple predictors, mediators or moderators on survey data
You need to identify predictors for screening or forecasting (e.g. which applicant characteristics predict performance)
As a bridge study: relationships found here justify the expense of a later experiment
✗ When NOT to use
You must establish causation and manipulation is feasible — run an experiment instead
The constructs lack validated measures — measurement error attenuates every correlation you compute
Variables have little natural variance in your setting (everyone scores high on safety climate) — restriction of range flattens correlations
The expected sample is too small for the model — complex path models on 60 respondents produce noise
You would collect predictor and outcome from the same respondents, same method, same moment, on a topic where common-method bias is a known threat — and cannot add any remedy
Typical research questions
RQ — Is perceived organizational support associated with employees' turnover intention, controlling for tenure and pay satisfaction?
RQ — Does social-media engagement predict purchase intention among Gen-Z consumers, and is the link mediated by brand trust?
RQ — Is transformational leadership related to team innovation, and does psychological safety moderate the relationship?
RQ — Which service-quality dimensions best predict overall customer satisfaction in retail banking?
Key characteristics
Purpose
Detect, quantify and model associations among naturally occurring variables; test theoretical models; identify predictors
Typical data
Quantitative — multi-item survey scales, archival metrics, test scores; often several constructs per respondent
Researcher control
No manipulation; control is exercised statistically (covariates) and through design (measurement quality, temporal separation)
Temporal aspect
Usually cross-sectional; lagged or longitudinal panels strengthen directional claims
Typical sample
Moderate to large (commonly 150–500+); complex models such as SEM need more
Variables & measurement
Even without manipulation, theory assigns roles: predictors (treated as independent variables), outcomes (dependent variables), mediators that transmit an effect, moderators that change its strength, and control variables that rule out rival explanations. Define each role before data collection — a variable's role determines where it sits in the analysis and what its coefficient means.
Measurement quality is the design's engine room. Use validated multi-item scales where they exist, check reliability and construct validity in your own data, and remember that unreliable measures attenuate observed correlations: a true r of .50 measured with two scales of reliability .70 shows up as roughly .35.
Choose established scales and cite their sources; adapt wording minimally and report any changes
Select control variables from theory and prior evidence, not by throwing in everything available
Decide levels of measurement early — an ordinal single-item outcome closes off many analyses
Where possible, separate sources or time points for predictor and outcome to blunt common-method bias
Sampling approaches that fit
Probability sampling (simple random, stratified) when a frame exists — it protects generalizability of the estimated relationships
Purposive/criterion sampling of a defined population (e.g. B2B sales staff with 1+ year tenure) — common and defensible if inclusion criteria are explicit
Convenience and snowball samples dominate student research; acknowledge the limitation and describe the sample thoroughly
Ensure variance on the key variables — deliberately include diverse organizations or roles rather than one homogeneous site
Size the sample by planned analysis: power analysis for regression (detecting the smallest effect of interest), and higher benchmarks for SEM (often 200+, more with many parameters)
Data collection methods that fit
Structured questionnaires with validated scales — online or paper; the default vehicle
Archival and records data — performance ratings, sales figures, absenteeism — excellent for outcomes measured independently of the survey
Multi-source designs — predictor from employees, outcome from supervisors or systems — the strongest remedy for common-method bias
Time-lagged collection — measure predictors at T1 and outcomes at T2 to add temporal separation
Match the technique to the question and the measurement level. Start with bivariate association, add statistical control, then model structure. All links open the Statistical Methods Atlas.
Each link below opens the matching method page (opens the Statistical Methods Atlas) with assumptions, worked examples and reporting guidance.
Whatever the technique, interpret coefficients as associations. “POS predicted lower turnover intention (β = −.32)” is a statement about the model, not proof that raising POS will lower turnover.
Quality criteria
Construct validity — reliability (alpha/omega), convergent and discriminant validity evidence for every latent construct
Internal validity is inherently limited — strengthen the causal argument with theory, controls, temporal separation and multi-source data, and calibrate claims accordingly
Statistical conclusion validity — adequate power, assumption checks, no p-hacking across dozens of untheorized paths
External validity — describe the sample and setting honestly; one company's employees are not “employees in general”
Common-method bias assessment — report procedural remedies and any statistical diagnostics used
A worked mini-example
A PhD scholar tests whether perceived organizational support (POS) relates to turnover intention via affective commitment, with supervisor incivility as a moderator. She surveys 384 employees across 14 IT firms (criterion sampling: 1+ year tenure), using established scales (POS 8 items, commitment 6, intention 3), with predictor and outcome blocks separated and demographics collected last.
CFA supports the four-factor measurement model; omega ranges .84–.91. SEM shows POS negatively associated with turnover intention, partially mediated by commitment (indirect effect −.14, 95% CI [−.20, −.09]); the moderation is nonsignificant. She reports the model as consistent with — not proof of — the theorized causal chain, notes the cross-sectional single-source design as the key limitation, and proposes a two-wave replication.
Common pitfalls
Causal language from cross-sectional data — “engagement drives performance” when nothing was manipulated or even temporally separated
Common-method variance ignored — all constructs self-reported in one sitting with no remedy or diagnostic
Controlling for variables that are themselves outcomes or mediators, which can bias the focal association either direction
Fishing across correlation matrices and reporting the survivors as hypothesized findings
Interpreting a nonsignificant correlation in a small sample as “no relationship” — check power and confidence intervals
Restriction of range from a single homogeneous site flattening every correlation
Running SEM on inadequate samples and over-interpreting fit indices
What to report
Theoretical model with hypothesized roles (predictor, mediator, moderator, controls) stated before results
Sample: population, inclusion criteria, recruitment, n, response rate, and descriptives
Measures: source, item count, sample item, reliability in this sample; measurement-model results (CFA/validity) where applicable
Correlation matrix with means and SDs for all study variables
Model results with coefficients, confidence intervals and effect sizes (R², f²) — not p-values alone
Common-method-bias remedies and diagnostics; assumption checks
Limitations that name the causal ambiguity explicitly, and the design that could resolve it