Understanding Correlation: Measuring Relationships in Data

Numbers tell stories, but rarely in isolation. A researcher studying maternal health in Bihar or fertility patterns in Kerala quickly realises that one variable hardly ever moves alone. Income shifts alongside education. Age at marriage changes with literacy. Family size responds to a dozen factors at once. Correlation is the statistical tool that helps us measure exactly how strongly two variables move together, and in which direction. It is one of the most widely used techniques in population and family health research, precisely because human behaviour is so interconnected.

What is correlation? 🔗

Correlation is a statistical measure that quantifies the strength and direction of the relationship between two variables. When two characteristics tend to change together in a predictable pattern, we say they are correlated. The classic textbook example is height and weight: taller people generally weigh more, though not perfectly so. In population studies, researchers routinely look at pairs like age and blood pressure, household income and child nutrition, or contraceptive use and fertility rates.

The National Library of Medicine defines correlation as a statistical measure of the extent to which two variables relate to one another. Importantly, correlation describes association, not causation. Two variables can move together for many reasons, including coincidence, a shared third cause, or a genuine influence of one on the other. A researcher who confuses these can draw spectacularly wrong conclusions.

Why correlation matters in population and family health 🔗

Public health rarely lets researchers run clean laboratory experiments. We cannot randomly assign women to be literate or illiterate, nor can we force families to have a certain number of children. Instead, researchers observe what already exists in the population and try to detect patterns. Correlation is the first analytical step in almost any such study. It tells us whether two indicators move together strongly enough to be worth investigating further with more advanced techniques like regression or causal modelling.

For example, when planners want to understand why infant mortality remains high in certain districts, they often begin by correlating it with factors like female literacy, antenatal care coverage, and household sanitation. The correlations point them toward the variables that deserve a closer look.

Positive vs. negative correlation 🔗

Every correlation has two basic properties: direction and strength. Direction tells us whether the variables move together or in opposite directions. Strength tells us how tightly they track each other.

Positive correlation 🔗

A positive correlation exists when two variables move in the same direction. As one increases, the other tends to increase too. As one decreases, the other tends to decrease.

Classic examples in population health include:

• Height and weight: Taller individuals usually weigh more than shorter ones.
• Maternal age and pregnancy complications: Risk of complications generally rises as maternal age increases beyond a certain point.
• Household income and life expectancy: Wealthier households tend to live longer, on average.
• Years of schooling and age at first marriage: More years of education usually mean later marriage.

None of these relationships is perfect. A short person can weigh more than a tall person; a young woman can face complications while an older one does not. The correlation simply tells us about the general pattern across many observations.

Negative correlation 🔗

A negative correlation, also called an inverse correlation, exists when the two variables move in opposite directions. As one rises, the other falls. The relationship between female literacy and family size is one of the most studied negative correlations in population studies.

A large analysis using Indian state-level data found that crude birth rates are inversely related to female literacy rates, with the relationship for female literacy being significantly stronger than for male literacy. States like Kerala and Himachal Pradesh, with very high female literacy, show some of the lowest total fertility rates in the country, while states with lower female literacy tend to report higher fertility. The Sample Registration System data consistently shows this pattern across multiple years.

Other examples of negative correlation in family health include:

• Age and resting metabolic rate: As age increases, metabolic rate tends to decrease.
• Contraceptive prevalence and total fertility rate: Higher contraceptive use is associated with smaller family size.
• Maternal education and infant mortality: More educated mothers tend to have lower infant mortality.
• Physical activity and risk of cardiovascular disease: More activity, lower risk.

Zero or no correlation 🔗

Sometimes two variables show no consistent pattern at all. For example, the shoe size of mothers and the birth weight of their babies would likely show no meaningful correlation. When researchers find zero correlation, it suggests that one variable carries no useful information about the other in a linear sense.

The correlation coefficient (r) 🔗

To move beyond saying “these two variables are related,” researchers need a single number that captures both direction and strength. That number is the correlation coefficient, almost always denoted by the letter r for a sample and the Greek letter ρ (rho) for a population.

The most widely used version is the Pearson product-moment correlation coefficient, developed by Karl Pearson in the 1890s from earlier work by Francis Galton. It measures the linear association between two continuous variables.

The range of r: -1 to +1 🔗

The correlation coefficient is always a number between -1 and +1. This fixed range is what makes it so useful. Whatever the units of the original variables, the coefficient itself is unitless and directly comparable across studies.

• r = +1: A perfect positive correlation. Every increase in one variable is matched by a proportional increase in the other. All points lie exactly on a straight upward-sloping line.
• r = -1: A perfect negative correlation. Every increase in one variable is matched by a proportional decrease in the other. All points lie exactly on a straight downward-sloping line.
• r = 0: No linear correlation. The variables show no consistent linear pattern.
• Values between: Anything between these extremes represents an imperfect relationship, with strength depending on how close r is to 1 or -1.

Interpreting the strength of r 🔗

Real-world data rarely produces perfect correlations. A common rule of thumb in health research, as outlined by the NIH guide to health statistics, treats coefficients greater than 0.7 or less than -0.7 as strong, with weaker values indicating moderate or weak relationships. A typical interpretive scheme looks like this:

• 0.00 to ±0.30: Weak or negligible correlation.
• ±0.30 to ±0.50: Moderate correlation.
• ±0.50 to ±0.70: Substantial correlation.
• ±0.70 to ±1.00: Strong to very strong correlation.

These thresholds are guidelines, not laws. What counts as a strong correlation depends on the field. In physics, researchers may expect r values above 0.95. In behavioural and population studies, where measurements are noisier and human variation is huge, an r of 0.4 may be considered meaningful and useful.

The formula and how it works 🔗

For a sample of paired observations (x, y), the Pearson correlation coefficient is calculated as:

r = Σ[(xᵢ – x̄)(yᵢ – ȳ)] / √[Σ(xᵢ – x̄)² × Σ(yᵢ – ȳ)²]

The numerator captures how the deviations of x and y from their respective means move together. When both variables are simultaneously above or below their means, the product is positive, pushing r toward +1. When one is above its mean while the other is below, the product is negative, pulling r toward -1. The denominator normalises the result, ensuring r stays between -1 and +1 regardless of the units of measurement.

In practice, almost no one calculates r by hand. Statistical software like SPSS, R, Stata, or even Microsoft Excel (using the CORREL or PEARSON function) does the work in seconds. The researcher’s real job is to choose the right variables, check that the data meets the assumptions of the test, and interpret the result correctly.

Assumptions behind Pearson’s r 🔗

Pearson’s correlation has a few important assumptions that researchers in family health studies must respect:

• Both variables are continuous, measured on an interval or ratio scale.
• The relationship is linear. Pearson’s r only detects straight-line associations. A perfectly U-shaped or curved relationship can produce r close to zero even when the variables are strongly related.
• The data is approximately normally distributed, with no extreme outliers.
• Observations are independent of one another.

When these assumptions are violated, particularly for ordinal data or non-linear relationships, researchers turn to alternatives like Spearman’s rank correlation, which uses ranks instead of raw values and can detect any monotonic relationship.

Common pitfalls in interpreting correlation 🔗

Correlation is powerful but also frequently misused. A guide published in the Malawi Medical Journal notes that the misuse of correlation is so widespread among researchers that some statisticians have wished the method had never been invented. Three errors come up again and again.

Correlation is not causation 🔗

This is the most repeated warning in statistics, and yet it is constantly forgotten. The fact that ice cream sales and drowning deaths rise together does not mean ice cream causes drowning. Both rise in summer. In population health, regions with more hospitals often have higher reported disease rates, not because hospitals cause disease, but because better facilities detect more cases.

Restricted range weakens correlation 🔗

If a study only looks at a narrow slice of a variable, the correlation will appear weaker than it really is. Measuring height and weight only among adults aged 25 to 30 will produce a weaker r than measuring across the full population, because the natural variation has been artificially restricted.

Outliers can distort r 🔗

A single extreme observation can dramatically inflate or deflate the correlation coefficient. Before trusting any r value, a scatter plot is essential. The plot reveals whether the relationship is genuinely linear, whether outliers exist, and whether the assumptions are reasonable.

Correlation in family health research: a practical view 🔗

Correlation is most useful when it generates good questions. When researchers found a strong negative correlation between female literacy and infant mortality across Indian states, the next step was not to declare that teaching women to read directly saves babies. It was to ask: through what pathways does education matter? Better health-seeking behaviour? Higher household income? Greater autonomy over reproductive choices? Later age at marriage? Each pathway becomes a hypothesis for further study.

A systematic review of female education and fertility in South Asia confirms a negative relationship between literacy levels and total fertility rates, but the authors are careful to point out that the relationship operates through many intermediate factors. Correlation flagged the pattern; deeper research explained why.

This is correlation at its best: a starting point, a flashlight pointed at a relationship worth exploring, never the final word.

What do you think? If you discovered a strong positive correlation between two variables in a population health study, what additional analyses would you want to run before drawing any conclusions about cause and effect? And can you think of two variables in your own community whose relationship would surprise people if it were carefully measured?

References

1. https://www.nlm.nih.gov/oet/ed/stats/02-300.html
2. https://pmc.ncbi.nlm.nih.gov/articles/PMC4649870/
3. https://factly.in/data-srs-2018-shows-the-strong-correlation-between-education-fertility-rate/
4. https://en.wikipedia.org/wiki/Pearson_correlation_coefficient
5. https://pubmed.ncbi.nlm.nih.gov/23638278/
6. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6054002/