Which Statistical Test for an Environmental Science Dissertation? The 2026 Comparison Guide

Environmental data breaks the assumptions of the tests most students were taught. Species counts are not normally distributed and contain a great many zeros. Quadrats within a transect are not independent. Water quality readings taken monthly at the same site are autocorrelated and seasonal. Contaminant concentrations arrive with values recorded as “below detection limit”. Running a t test on any of that will produce a number, and the number will be wrong.

This guide compares the analytical approaches that environmental science dissertations actually need, organised by the kind of data your fieldwork produced, and states plainly which one to pick.

The comparison at a glance

If you want the general logic behind matching outcome type to model family, our decision guide to choosing a statistical test covers the underlying tree; what follows is the environmental-specific layer on top.

Count data: why GLMs beat transformed ANOVA

Ecological counts — individuals per quadrat, invertebrates per kick sample, seedlings per plot — are the single most common data type in environmental dissertations, and the single most commonly mis-analysed. The old approach was to log-transform or square-root-transform the counts and run ANOVA. The modern approach is a generalised linear model with an appropriate error family.

Start with a Poisson GLM. Then check for overdispersion by comparing the residual deviance to the residual degrees of freedom; a ratio much above 1 means the Poisson assumption that variance equals the mean has failed, which is the normal state of affairs in field data because organisms aggregate. Switch to a negative binomial GLM, which adds a dispersion parameter to accommodate that clustering. Report the model family, the dispersion check, the coefficients on the response scale (as incidence rate ratios) as well as the link scale, and the confidence intervals.

If your counts contain far more zeros than even a negative binomial predicts, ask whether the zeros have two origins: sites where the species genuinely cannot occur, and sites where it occurs but was not detected. That distinction is what zero-inflated and hurdle models formalise, and articulating it well is a genuine mark-earner in the discussion.

Community composition: PERMANOVA, ANOSIM and ordination

If your response is a site-by-species matrix rather than a single variable, univariate tests are the wrong tool. The standard workflow is:

1. Choose a dissimilarity measure. Bray-Curtis is the default for abundance data; Jaccard for presence-absence.
2. Visualise with NMDS. Report the stress value — below 0.1 is good, below 0.2 usable, above 0.2 should not be interpreted as a map.
3. Test with PERMANOVA (the adonis routine in the vegan package for R), which partitions variation in the dissimilarity matrix and tests group differences by permutation.
4. Check the dispersion assumption with a test of multivariate homogeneity of group dispersions. A significant PERMANOVA can reflect differing spread rather than differing centroids; if dispersion also differs, say so.
5. Identify which taxa drive the difference with SIMPER, treating it as descriptive rather than inferential.

PERMANOVA is generally preferred to the older ANOSIM because it accommodates complex and unbalanced designs, partitions variance, and gives an interpretable effect size in R². If your grouping structure resembles cluster membership rather than a designed factor, the methods in our guide to cluster analysis for a dissertation cover how grouping solutions are chosen and validated.

Environmental time series and trend detection

For a monitoring series — river nitrate, air particulate concentrations, groundwater levels — the question is usually “is there a trend?” Ordinary least squares regression on the date is a poor answer because the residuals are autocorrelated, seasonal and typically skewed.

The environmental standard is the Mann-Kendall test, a rank-based test of monotonic trend, paired with Sen’s slope estimator for the magnitude. Where the data are strongly seasonal, use the Seasonal Kendall variant, which blocks by month before combining. Where consecutive observations are autocorrelated, use a modified Mann-Kendall with variance correction. Report the direction, the Sen’s slope with units per year and its confidence interval, and the p value — that combination gives a reader everything they need.

If instead you are looking for a step change at a known date, such as a discharge consent tightening, an intervention analysis or a change-point test is more appropriate than a monotonic trend test.

Before-after-control-impact designs

Environmental impact questions are rarely a simple comparison of two groups. If you sampled a site before and after a road opened, and also a control site over the same period, your hypothesis is not “did the impact site change” — it is “did the impact site change differently from the control site.” That is a time × site interaction term in a mixed model or two-way ANOVA, and the interaction is your result. Reporting the main effects and ignoring the interaction is the standard failure mode of BACI dissertations.

Where you have multiple control sites or several sampling occasions per period, the design becomes beyond-BACI and belongs in a mixed model with random effects for site and time.

Nested field designs and pseudoreplication

Field sampling is almost always hierarchical: quadrats within transects within sites; leaves within trees within stands; samples within reaches within rivers. If you analyse 200 quadrats as 200 independent observations when they came from eight sites, you have replicated the measurement rather than the treatment. Your standard errors will be far too small and your p values misleadingly convincing.

The correct answer is a mixed-effects model with a random intercept for the clustering unit, or in simple designs the summary-measures approach — average up to the level at which treatment was applied, and analyse those eight values. Our introduction to multilevel models for nested data covers the intraclass correlation, when the extra complexity is warranted, and how to report the random effects.

Correlated predictors and model selection

Environmental predictors are notoriously collinear: temperature, elevation, aspect and soil moisture all covary. Check variance inflation factors and drop or combine variables above a VIF of about 5. Principal component analysis is a legitimate way to reduce a set of correlated physicochemical measurements to a few interpretable axes before regression.

For selecting among candidate models, information-theoretic comparison using AIC or AICc is standard in ecology and generally preferable to stepwise selection, which capitalises on chance. Report the model set, the AIC values and the delta-AIC weights rather than presenting a single “best” model as though it were the only one considered. Where your interest is prediction of a binary outcome such as species presence, our comparison of linear versus logistic regression covers the choice of link and what the coefficients mean.

Censored data: the non-detect problem

Environmental chemistry produces left-censored data — the analyte was present below the limit of detection. Substituting zero, the detection limit, or half the detection limit is common in student work and biases both the mean and the variance, sometimes severely. Use methods designed for censored data: Kaplan-Meier estimation adapted for left censoring, regression on order statistics, or maximum-likelihood estimation. State the proportion of non-detects and the detection limit in your methods; a reader cannot judge your results without both.

Small-sample field studies

Dissertation fieldwork is constrained by time, access and weather, so samples are often small and unbalanced. Three practical consequences:

• Non-parametric alternatives such as Kruskal-Wallis and Mann-Whitney become genuinely useful; our overview of non-parametric tests explains what each one actually tests.
• Permutation tests are well suited to ecological data because they make no distributional assumption and are readily available in R.
• Non-significant results are extremely likely and are not a failure. Report the effect size and confidence interval, discuss the detectable effect your design supported, and interpret honestly — our guide on what to do when results are not significant shows how to write that as a finding.

Resist deleting inconvenient observations. Extreme values in environmental data are frequently real — a storm event, a pollution incident, a rare species record — and removing them can delete your most interesting result. Our discussion of whether to remove outliers covers the difference between an error and an extreme true value.

What to write in your analysis section

Name the software and version, including R packages such as vegan, lme4 or glmmTMB with their versions. State each model’s error family and link function. Report the assumption checks you performed and what they showed, including overdispersion and spatial autocorrelation where relevant. Give effect estimates with confidence intervals on an interpretable scale, and for multivariate tests report the permutation count. Keeping a written record of variable definitions, units, detection limits and cleaning decisions as you go makes this section far easier to write — our template for building a data dictionary and cleaning log is designed for exactly this.

Frequently asked questions