A soil pH of 5.2 correlates with lower maize yield across the Barind Tract. Rainfall correlates with river sediment load in the Brahmaputra basin. Fertilizer application rate correlates with nitrate concentration in shallow groundwater. These are statements a researcher can support with data, publish in a journal, and defend in a thesis defence. They are also statements that do not tell you why the relationship exists, whether changing one variable will change the other, or what would happen if you intervened.
Correlation and causation are different things. This is widely repeated in statistics textbooks and equally widely ignored in environmental research practice. The problem appears most often not in what researchers claim directly, but in the language they reach for when interpreting their results.
In this series: This post pairs with Linear Regression in R for Environmental Data — regression quantifies relationships, but correctly interpreting what those relationships mean requires the framework covered here.
What correlation measures
A correlation coefficient measures the strength and direction of a linear relationship between two variables. Pearson's r of 0.85 between soil organic matter and water retention tells you that fields with high organic matter tend to hold more water, and that the relationship is strong. It tells you nothing about mechanism, nothing about the direction of influence, and nothing about what would change if you added organic matter to a low-retention soil.
The p-value attached to a correlation tells you the probability of observing a relationship this strong if no relationship actually existed. Statistical significance is not the same as practical importance, and it is not evidence of causation. A p-value of 0.001 tells you the correlation is very unlikely to be a chance result. It does not tell you what produced it.
Four alternative explanations for any correlation
When two variables are correlated, four explanations are possible.
The correlation reflects a real directional influence. Soil pH dropping below 5.0 activates aluminium toxicity, which damages root architecture, which reduces water and nutrient uptake, which reduces yield. This is a plausible causal chain with a known mechanism.
The direction is reversed. Less common in environmental soil-water research but not impossible — declining yields could drive land management changes that alter soil pH over time.
The most common source of spurious environmental correlations. Organic matter content correlates with both higher soil pH and higher crop yield in acidic Bangladeshi soils. If you correlate pH with yield without measuring organic matter, you may attribute to pH an effect that organic matter is producing. The correlation between pH and yield is real. The causal inference is wrong.
Spurious correlations between unrelated variables appear whenever you have enough data. Researchers who mine large environmental datasets without prior hypotheses will find relationships that are statistically significant and biologically meaningless.
How it appears in environmental research writing
The clearest sign of correlation-causation confusion in a results section is the verb. "Soil pH significantly affected grain yield" is a causal claim — it implies that changing pH would change yield, and requires experimental evidence: an intervention, a control, a randomised design. "Soil pH was significantly correlated with grain yield" is an associative claim. It requires only that the data showed co-variation.
Most observational environmental studies generate the second kind of evidence and use language that implies the first. This matters because policy recommendations built on correlational evidence, presented as if causal, can lead to interventions that don't work.
The confounding problem in soil and water research
Confounders are variables that correlate with both the predictor and the outcome you are studying. In soil science, organic matter is a confounder in almost every agronomic correlation because it affects soil pH, water retention, nutrient cycling, microbial activity, and yield through multiple pathways at once.
The practical test: ask whether the correlation disappears, weakens, or reverses when you control for candidate third variables. In multiple regression, adding organic matter as a predictor alongside pH often substantially reduces the pH coefficient. If it does, organic matter was confounding the pH-yield relationship. If the pH coefficient holds after controlling for organic matter, the causal inference is stronger — though still not confirmed without experimental manipulation.
What causation actually requires
Three conditions, met together, build a causal argument.
The cause must precede the effect. In longitudinal soil studies, this means showing that pH dropped before yield declined, not just that both are low at the same time.
There must be a biological, chemical, or physical process that explains how X produces Y. Aluminium toxicity is a well-characterised mechanism linking low pH to root damage. A correlation between soil colour index and crop yield has no known mechanism — it might reflect organic matter, drainage class, or mineralogy, any of which could be the actual driver.
Ideally through randomisation, or through statistical control in observational studies — measuring and adjusting for the most likely alternative explanations.
Experimental manipulation is the most reliable path to establishing causation because it controls confounders by design. Observational data can support causal inference when the mechanism is well-understood and confounders are measured, but it cannot prove it.
When correlation is enough
Not every research question requires causal inference.
Estimating crop yield from remote sensing data, flagging high-nitrate risk areas from soil survey variables. A model that predicts accurately doesn't need to explain the mechanism.
Identifying which variables are worth investigating in a controlled experiment. High correlations in observational data are the right starting point for designing the next study.
Recommending lime application, changing fertilizer regulations, designing water treatment systems. The intervention assumes causation. Make sure the evidence supports that assumption.
Language to use and avoid
The shift from causal to associative language is not hedging — it is accurate reporting. If your study design can establish causation, use causal language. If it cannot, the associative framing is the correct one.
| ✗ Instead of | ✓ Write |
|---|---|
| "X significantly affected Y" | "X was significantly associated with Y" |
| "Application of X increased Y" | "Higher X values corresponded with higher Y" |
| "X caused a decline in Y" | "X was negatively correlated with Y" |
| "Results show that X improves Y" | "Results suggest a positive association between X and Y" |