Which statistical test should I use?

Most test-selection agony disappears once you ask four questions in order. This guide walks them with the answers researchers actually need.

1. What kind of outcome did you measure?

• A continuous measurement (yield, concentration, mass, absorbance) → the t-test / ANOVA / regression family below.
• Binary outcomes (yes / no) — germinated or not, infected or not → binary logistic regression.
• True counts (insects per plant, colonies per plate) need Poisson-family models, and contingency tables chi-square or Fisher’s exact test — not offered in MinutePlot yet.

Questions 2 and 3 cover the continuous case, which is most lab data; question 4 the relationships.

A word on words. A factor is a variable you vary (fertiliser type); its levels are the options it takes (control, A, B); the replicates measured at one level form a group. With two factors, each combination of levels is a cell — one group. This is how MinutePlot’s templates are laid out: factor A down the rows, factor B across grouped columns, replicates within each group.

2. How many factors did you vary?

Count the variables you deliberately varied, not the groups they make: a 2 × 3 design has six groups but two factors, and it is the factors that choose the test.

• One factor (fertiliser type) → one-way ANOVA.
• Two crossed factors (fertiliser and watering regime) → two-way ANOVA, which also tests whether the factors interact.

  	Daily				Weekly
  	r1	r2	r3	r4	r1	r2	r3	r4
  Control	42.1	40.8	43.0	41.5	38.2	37.5	39.1	38.8
  High N	52.3	53.1	51.6	52.8	44.9	45.8	44.1	45.3

  r = replicate: r1 is replicate 1, r2 replicate 2, and so on.

  Example data from the Two-way ANOVA and interactions guide.

• Three crossed factors → three-way ANOVA: three main effects, three two-way interactions and the three-way interaction, in one analysis.
• Repeated measurements of the same subjects (each subject measured under every level of one factor, e.g. over time) → one-way repeated-measures ANOVA.

  	Timepoints
  	Day 0	Day 7	Day 14
  Plant 1	5.0	6.1	7.2
  Plant 2	4.8	5.9	7.0
  Plant 3	5.3	6.4	7.6
  Plant 4	5.1	6.0	7.1
  Plant 5	4.9	6.2	7.3
  Plant 6	5.2	6.3	7.4

  Example data from the Repeated-measures ANOVA guide.

• Two within-subject factors (every subject measured under every combination) → two-way repeated-measures ANOVA.
• One between-subjects factor and one within (groups of subjects, each subject measured over time) → two-way mixed ANOVA.

3. With one factor: how many groups?

If question 2 landed on one factor, the number of its levels decides the test.

• Two groups → a t-test:
  • independent (unpaired) when the two groups contain different subjects;
  • paired when each subject was measured twice (before and after, left and right).

  	Values
  	r1	r2	r3	r4	r5	r6	r7	r8
  Control	5.1	5.4	4.9	5.6	5.0	5.3	4.8	5.2
  Treated	6.0	6.4	5.8	6.6	6.1	5.9	6.3	6.2

  r = replicate: r1 is replicate 1, r2 replicate 2, and so on.

  Example data from the Independent t-test guide.

• Three or more groups → one-way ANOVA — never a chain of t-tests, because many pairwise tests inflate the chance of a false positive; ANOVA asks the question once, properly, and letters then show which groups differ.

  	Values
  	r1	r2	r3	r4	r5	r6
  Control	41.2	39.8	43.1	40.5	42.0	38.9
  Low N	45.6	47.2	44.9	46.8	45.1	46.3
  High N	50.3	52.1	49.7	51.4	50.8	48.9
  N + P	51.0	49.5	52.6	50.2	51.7	53.0

  r = replicate: r1 is replicate 1, r2 replicate 2, and so on.

  Example data from the One-way ANOVA with post-hoc letters guide.

4. Is a relationship, not a difference, the question?

• A continuous predictor (temperature, concentration) and a continuous outcome → linear or polynomial regression.
• A dose series with a sigmoid response → dose-response regression (4PL: the EC50 and the plateaus).
• A time series and the rate of change over a chosen window → time-series slope over a range.
• A binary outcome (yes / no) and one or more predictors → binary logistic regression.
• The strength of association between measured variables → the correlation matrix.
• Many variables at once — to see the structure, or to predict one outcome from many measurements → PCA and PLS; to see samples and variables clustered together → the clustered heatmap.

New to the app itself? The Using MinutePlot pages walk the workbench click by click — data entry, figure styling, panels, exports.

The assumptions, briefly

The parametric tests above assume residuals are roughly normal and group variances similar. MinutePlot checks both automatically with every analysis — Shapiro–Wilk for normality and Levene’s test, centred on the median, for equal variances (the two-group comparison is Welch’s t-test, which does not assume equal variances) — and tells you in plain words whether the assumptions held.

When a check fails. Mild violations are usually tolerable: with groups of similar size, ANOVA and the t-test are robust to moderate non-normality and unequal variances. A transformation (log for right-skewed measurements, square root for counts, arcsine for proportions) often restores both. MinutePlot offers Dunn’s non-parametric pairwise comparisons as a post-hoc option but no non-parametric alternative to the omnibus tests themselves; for severely non-normal data with small groups, run a rank-based test elsewhere and say so in the methods.