Pearson Correlation

Pearson's correlation coefficient measures whether two continuous variables depart from their means in step with one another. It is the covariance divided by the product of the two standard deviations, which makes it unit-free and therefore readable on the same scale whether the variables are scores, minutes or amounts of money. As the coefficient approaches +1 the points settle onto an upward line, as it approaches -1 onto a downward one, and around 0 no linear pattern is visible. The word to hold on to is linear: a strong but curved relationship, one that reverses past a certain point for instance, can produce an r close to zero. YouReply Analyze reports the coefficient, the two-tailed p value, the number of observations used and a strength label derived from the absolute magnitude.

• Do brand trust scale scores and repurchase intention scores rise together?
• Is there a linear relationship between weekly screen time and total sleep quality score?
• How far do employees' job satisfaction scores move together with their organisational commitment scores?
• Is survey completion time related to a consistency score computed from the answers?

• When both variables are measured at interval or ratio level, such as scale totals, durations, amounts or counts.
• When the points in a scatter plot gather roughly around a straight line rather than a curve.
• When the question concerns the strength and direction of joint variation rather than predicting one variable from the other.
• When scale development work calls for a description of the relationships among items or subdimensions.
• After outliers have been cleaned, or at least examined with their reasons understood, since a single distant point can move the coefficient noticeably.

• Both variables have to be continuous: numeric columns whose measurement level is set to 'scale' on the Variable tab.
• On an ordinal scale, and especially on a single Likert item, equal spacing between steps cannot be assumed, and a rank-based coefficient is then easier to defend.
• A variable coded as 0 and 1 yields a computable coefficient, but its interpretation is that of a point-biserial correlation, which is worth knowing before you use it.
• The analysis works pairwise: a row missing a value in either column drops out entirely, so the number of observations used can be smaller than the row count in your file.

A linear relationship

The coefficient is defined only for linear association. If the points follow a curve, flatten out past a threshold, or trace a U shape, an r near zero means the relationship is not linear rather than that it is absent.

Bivariate normality

The distribution behind the p value assumes the two variables are jointly normal. The coefficient itself is computed without that assumption; what is affected is the accuracy of the significance test, and tolerance for moderate departures grows with sample size.

Independent pairs

Each row should be an independent pair of observations. Where the same person recurs, or where clusters such as classrooms or households are present, the standard error of the coefficient looks smaller than it is.

Outliers examined

Because Pearson's r rests on products of deviations from the mean, one distant point can exert considerable influence. A single unusual observation can lift a coefficient that belongs near zero, or conceal a relationship that is really there.

An unrestricted measurement range

If the sample covers only part of a variable's range, high scorers alone for example, the coefficient comes out smaller than it should. This is a limitation of the sample's coverage, not of the computation.

• A linear relationship: The panel runs no test of linearity. The scatter plot drawn when scatter data are present in the result is the most direct way to make this judgement, so look at the chart before reading the coefficient.
• Bivariate normality: No normality test is run within this method. If you want to see the distributions, run the normality tests as a separate method.
• Independent pairs: The panel does not check this assumption automatically; the researcher evaluates it.
• Outliers examined: No automatic outlier detection or exclusion takes place; cleaning decisions stay with you on the Data and Variable tabs. The scatter plot is the most practical way to spot distant points.
• An unrestricted measurement range: The panel does not check this assumption automatically; the researcher evaluates it.

1. 1Drop your data file onto the panel's upload area; a CSV or XLSX file opens into the grid on the Data tab.
2. 2On the Variable tab check the measurement level of the two numeric columns. Fixing numbers stored as text and declaring missing-value codes here keeps incomplete pairs from dropping out unnoticed.
3. 3On the same tab build a derived column if you need one, such as a scale total; computing the total in advance rather than working with item averages keeps the result readable.
4. 4On the Analysis tab choose the method from the correlation group or use the search box; the 'Only ones that fit my data' switch narrows the picker.
5. 5In the parameter form name the two columns to be related and start the analysis; the order does not matter, since the coefficient treats both columns symmetrically.
6. 6Read the result starting from the chart: inspect the shape of the points where scatter data are available, then open the coefficient, the p value and the number of observations in the collapsible sections.
7. 7Export the table to Excel and the chart to PNG; the computation credits card names the library call and the version behind the numbers.

Correlation coefficient

r, ranging from -1 to +1. Its sign gives the direction of the relationship and its absolute magnitude the strength of the linear association.

Significance value

The two-tailed p value computed under the hypothesis that the population coefficient is zero. As the sample grows, even small coefficients can reach significance.

Number of observations used

The count of pairs with a valid value in both columns. It belongs in your report, because depending on how missing values fall it can sit well below the row count of the file.

Strength and direction label

A verbal label derived from the absolute value of the coefficient, given together with the direction, and written as no correlation when the coefficient is exactly zero.

Relationship charts

Correlation methods get a correlation bar chart showing the coefficient, joined by a scatter plot of the points when scatter data are present in the result.

The coefficient itself

r is already a standardized effect size, so no further measure is needed. The panel's labels read off the absolute value: below 0.2 negligible, below 0.3 weak, below 0.5 moderate, below 0.7 strong and above that very strong. These are not the only defensible cut-points in the literature: a coefficient of 0.30 can be notable in behavioural research while looking weak in laboratory data with high measurement reliability.

Coefficient of determination

Squaring r gives the share of variability the two variables hold in common. The panel does not return this as a field of its own, so compute it from r; seeing that a coefficient of 0.50 corresponds to only a quarter of the variability shared keeps interpretation sober.

No confidence interval is returned for the coefficient. No interval based on the Fisher z transformation is computed, and r, p and the number of observations all arrive as point estimates. Where a journal asks for an interval you will have to produce it outside the panel, and that is more than a formality: in small samples the interval around r is surprisingly wide.

The numbers below are a representative example, not data from a real study or a real user.

Research question

In an illustrative customer study, is there a linear relationship between the total score on a brand trust scale and a repurchase intention score?

Variables

First variable: brand trust scale total, range 10-50 (continuous) · Second variable: repurchase intention score, range 0-100 (continuous)

Example result

312 participants with valid values on both variables entered the analysis. The coefficient was r = 0.48, p < 0.001. Because the absolute value stays below 0.5, the label describes a moderate positive relationship. Squared, it puts the shared variability at 23 per cent. In the scatter plot the points form an upward-spreading cloud with no visible curvature or detached cluster.

Interpretation

Participants scoring high on trust tend to report higher repurchase intention. With shared variability at 23 per cent, most of the variation in intention is left unexplained by this scale: price, habit and availability all sit outside the picture. The coefficient says nothing about how much a one-point rise in trust moves the intention score, which requires a regression that reports a slope. Nor can direction be read from these data, since trust feeding intention and repeat buyers growing more trusting would produce the same coefficient. The figures are illustrative.

Research question: Is there a linear relationship between satisfaction and burnout?

First variable

satisfaction

Second variable

burnout

Computation credits: scipy 1.18.0 · statsmodels 0.14.6 · scikit-learn 1.9.0 · numpy 2.5.1 · pandas 3.0.5 · semopy 2.3.11 · 89fc29a · Data seed: 20260912

Brand trust was moderately and positively related to repurchase intention, r(310) = .48, p < .001, n = 312.

An example sentence close to APA style; the numbers are representative.

• A coefficient near zero is not evidence of no relationship; in curved or threshold-shaped patterns the linear component is small and r follows it down.
• A single extreme observation can shift the coefficient in either direction, which makes an r reported without a look at distant points fragile.
• The coefficient establishes no causation. Two variables moving together can equally reflect a third variable driving both, and where that worry is concrete, a partial correlation is more informative.
• r is not a slope: being unit-free, it cannot be read as a one-point rise producing a given increase, a reading reserved for regression coefficients.
• When the sample spans only a narrow range of a variable, the coefficient is systematically attenuated and a restricted range hides the true relationship.
• Pooling distinct subgroups can mislead: a relationship absent within every group can appear in combined data purely because the group means differ.

• Spearman Rank CorrelationWhy: When the variables are ordinal, the distributions are clearly skewed, or extreme values are pulling the coefficient around; it measures monotonic association on ranks.
• Linear Regression (OLS)Why: When you need change per unit, that is a slope and an intercept, or want to predict one variable from the other.
• Partial CorrelationWhy: When you want the relationship between two variables while holding a third one constant.
• Exploratory Factor Analysis (EFA)Why: When the question concerns the dimensional structure underlying many items rather than their pairwise relationships.

• Spearman Rank Correlation
• Linear Regression (OLS)
• Exploratory Factor Analysis (EFA)

The coefficient came out at 0.48. How much does a one-point rise change the other variable?

A correlation cannot answer that. r is standardized and unit-free: dividing by both standard deviations strips out the scale information. To learn the change per unit you need to run a linear regression, and the same r can coexist with very different slopes depending on how the variables are spread out.

The relationship is significant. Can I draw a causal conclusion?

No. A correlation states that two variables vary together, not that one produces the other. The reverse direction yields the same coefficient, and a third variable affecting both can create the same pattern. A causal claim needs an experimental design, or at least a model that accounts for temporal order and confounders; where the concern is one specific confounder, a partial correlation is a reasonable first step.

Does the panel check linearity and normality?

It does not. Neither a linearity test nor a normality test is run in this method; the coefficient, the p value and the number of observations are computed directly. The scatter plot is the practical route to judging linearity, and the normality tests are available as a separate method. When the assumptions fail, the coefficient is still produced, but its interpretation weakens.

Can one outlier really change the result?

It can. A single observation sitting far out on both scales can visibly raise or lower the coefficient even in a dataset of several hundred, and a mistyped age or income is the most common source. If you spot a distant point in the scatter plot, first check whether the value is genuine. If the observation is valid and its influence persists, reporting the Spearman coefficient alongside, which responds less to extremes because it works on ranks, is a transparent way out.

• Field, A. (2018). Discovering Statistics Using IBM SPSS Statistics
• Anscombe, F. J. (1973). Graphs in Statistical Analysis
• Rodgers, J. L., & Nicewander, W. A. (1988). Thirteen Ways to Look at the Correlation Coefficient
• scipy.stats.pearsonr documentation