Correlation analysis guide
Correlation analysis: find and interpret relationships in survey data
A correlation analysis shows whether two variables rise and fall together and how closely they are linked – say, satisfaction with customer service and willingness to recommend. This guide explains how Pearson and Spearman work, how to judge the size of r, why correlation does not prove cause, and what 121surveys calculates for you.
Building and testing is free. To publish, you choose a paid plan.
The basics: what the correlation coefficient tells you
What does the correlation coefficient r measure?
The correlation coefficient r sums up the relationship between two variables as a number between −1 and +1. The sign gives the direction: positive means high values of one go with high values of the other; negative means one rises as the other falls. The absolute value gives the strength: at 0 there is no linear relationship, at ±1 every point sits on a straight line. Keep in mind that r only captures linear (Pearson) or monotonic (Spearman) patterns. An inverted U-shape can therefore produce an r close to 0 – a scatter plot shows you whether that is the case.
Pearson correlation: formula and assumptions
Pearson's r relates the joint variation of two variables to their individual variation: r = Σ(x − x̄)(y − ȳ) / √(Σ(x − x̄)² · Σ(y − ȳ)²). The numerator is the sum of the cross-products of deviations; the denominator makes the result independent of the unit, so measuring in pounds or pence does not change r. Pearson suits interval or near-interval data: numeric answers, sliders and, in practice, usually rating scales with five or more points. The relationship should be roughly linear, and even a few outliers can shift r a lot.
Spearman rank correlation: for ranks and outliers
Spearman's rho does not use the values themselves but their ranks: each answer gets a rank, tied values share the average rank, and Pearson's r is then calculated on those ranks. This lets Spearman pick up any monotonic relationship, even one that is not a straight line, and makes it far less sensitive to outliers. For short rating scales with few points, skewed distributions or genuine rankings, Spearman is often the more cautious choice.
How strong is strong? Putting r in context
A widely used rule of thumb from Cohen calls r around 0.1 small, around 0.3 medium and 0.5 or more large. It is a convention, not a law: in attitude surveys, values around 0.3 are often meaningful already, while two measurements of the same thing should agree far more closely. The square r² helps: it gives the share of variation the two variables have in common. At r = 0.3 that is only 0.09, or nine per cent – the relationship is real but leaves most of the differences between respondents unexplained.
How many responses
To reliably detect a moderate correlation (r = 0.3), you need about 85 responses. Weaker correlations need considerably more.
Correlation is not causation
Why a relationship does not prove a cause
A correlation only says that two variables vary together. It cannot show that one causes the other. People who are happy with support recommend you more often – but perhaps satisfied customers simply contact support more and so have more chances for good experiences. To argue for a cause you need more: an order in time, a plausible mechanism and ideally an experiment in which you deliberately change one variable and watch the other.
Third variables, spurious correlation and direction
Often a third variable influences both. If satisfaction with a piece of software goes with team size, the real reason may be the industry in which large teams are more common. Split the data by such groups and a relationship can weaken, vanish or even reverse. The direction is open, too: does trust lead to satisfaction, or satisfaction to trust? A single survey cannot answer that on its own.
Key driver analysis in surveys
In customer and employee surveys, correlation is a popular quick driver analysis: which individual rating – price, delivery time, advice – is most closely linked to overall satisfaction or NPS? That is a good first look. The drivers, however, are usually linked to each other as well; people who like the advice often rate delivery more highly too. What each driver contributes on its own only shows up in a multiple regression, which you can run in SPSS, for example.
Correlation analysis in 121surveys
A correlation matrix across all numeric questions
The analysis calculates the Pearson correlation for every pair of numeric questions in your survey and shows it as a matrix: rating scales, star ratings, NPS, eNPS, CSAT, CES, pulse check, sliders and number fields. The tab appears as soon as your survey has at least two such questions.
Calculated pairwise, with the count per cell
For each pair of questions, everyone who answered both questions counts. The matrix shows how many people that is in every cell, so you don't confuse a value based on a handful of answers with one based on many. Too few shared answers produce no value.
Filters and date range of the analysis
The correlations follow the filters and date range you have set in the analysis. That way you can check, for instance, whether the link between service and recommendation looks different for new customers than for long-standing ones. These in-depth analyses are available from the analyst role upwards.
Export for Spearman and regression, MCP for AI assistants
Spearman correlations, significance tests and regressions are done in your statistics package after exporting the raw data as an SPSS file, Excel, CSV, ODS or JSON. An AI assistant can query the matrix through the MCP tool get_correlations.
Frequently asked questions
What is a correlation analysis?
A correlation analysis tests whether and how closely two variables are related. The result is a coefficient between −1 and +1: the sign shows the direction, the size shows the strength. In surveys you use it, for example, to see how closely ratings of individual services relate to overall satisfaction.
When should I use Pearson and when Spearman?
Pearson suits numbers and longer scales with a roughly linear relationship. Spearman works with ranks and is a good fit for short ordinal scales, skewed distributions and data with outliers. If both values are close, the result is robust; if they differ a lot, take a look at the scatter plot.
Does 121surveys also calculate Spearman correlation?
No. The correlation matrix in 121surveys shows Pearson correlations. For Spearman, significance tests or partial correlations, export the raw data – for example as an SPSS, Excel or CSV file – and continue in your statistics software.
How many responses do I need for a correlation analysis?
It depends on how strong a relationship you want to detect: weak relationships need far more responses than strong ones. Also make sure that enough people answered both questions of each pair – the matrix in 121surveys shows the count in every cell. To estimate how many responses you need for a given precision, use the sample size calculator.
Does a high correlation mean that is where I should act?
It is a pointer, not proof. A strong link between an individual rating and overall satisfaction makes that area a candidate. Whether improving it actually raises satisfaction only becomes clear from a before-and-after comparison or a targeted test – and a regression separates the contribution of overlapping drivers.
Find relationships, then test causes
The correlation matrix in 121surveys shows at a glance which ratings rise and fall together. Use it as a map for your next questions – and test causes with group comparisons, a regression or a targeted experiment.
