aStat manual Open aStat

Every analysis

Each entry says when to use the analysis, what data it needs, what aStat reports, and the SAS procedure and R functions it matches.

Summary statistics

Use when
You want a first look at any sheet, or a table of means for a report.
Data
One or more measurement columns, optionally split by a grouping column.
You get
n, mean, SD, SE, median, quartiles, minimum, maximum, and coefficient of variation; box plots or mean bars.
Matches
SAS PROC MEANS (quartiles use the same definition, R's type 2).

One-sample t-test

Use when
You compare one group's mean with a target value, such as a breed standard.
Data
One measurement column and the target value.
You get
The t-test with a confidence interval, Cohen's d, a normality check, and the Wilcoxon signed-rank test as a backup.
Matches
SAS PROC TTEST with H0=; R t.test().

Two-sample t-test

Use when
Two groups of different animals, such as treated and control.
Data
A measurement column and a group column with exactly two values.
You get
The pooled and Welch t-tests (Welch does not assume equal spread), the folded F test of equal variances, Cohen's d, and the Wilcoxon rank-sum test.
Matches
SAS PROC TTEST; R t.test() and wilcox.test().

Paired t-test

Use when
The same animals were measured twice, such as before and after weaning.
Data
Two measurement columns, one row per animal.
You get
The paired t-test on the differences, Cohen's dz, and the Wilcoxon signed-rank test.
Matches
SAS PROC TTEST with PAIRED; R t.test(paired = TRUE).

One-way ANOVA

Use when
Three or more groups of different animals in a completely randomized design.
Data
A measurement column and a group column.
You get
The ANOVA table, eta squared, Tukey comparisons of every pair with confidence intervals, Levene's test of equal spread, and Welch's ANOVA and the Kruskal–Wallis test as backups.
Matches
SAS PROC GLM with LSMEANS / ADJUST=TUKEY; R lm() and emmeans.

Randomized complete block design (RCBD)

Use when
Animals or plots were grouped into blocks (barns, litters, fields, years) and every treatment appears in every block.
Data
A measurement, a treatment, and a block column.
You get
Type III tests for block and treatment, the relative efficiency of blocking (how many more animals a completely randomized design would have needed), treatment means adjusted for blocks with Tukey comparisons, and the Friedman test when each block holds one animal per treatment.
Matches
SAS PROC GLM, MODEL y = block trt.

Two-factor ANOVA (factorial)

Use when
Two treatment factors were crossed, such as fertilizer by variety.
Data
A measurement and two factor columns. Testing the interaction needs every combination, with replication.
You get
Type III tests for both factors and their interaction, means of every combination, an interaction plot, Tukey comparisons, and simple effects (one factor within each level of the other) when the interaction matters.
Matches
SAS PROC GLM with SS3, MODEL y = A|B.

ANCOVA

Use when
Groups differ in a measurement taken before treatment, such as starting weight, and you want a fairer comparison.
Data
A measurement, a group column, and a covariate measured before treatment could affect it.
You get
A test that the covariate's slope is the same in every group, Type III tests, the common slope, group means adjusted to the average covariate, and Tukey comparisons.
Matches
SAS PROC GLM, MODEL y = trt x.

Latin square

Use when
Treatments were blocked in two directions, such as row and column in a poultry house, with each treatment once in every row and column.
Data
A measurement, a treatment, a row, and a column column.
You get
Type III tests for rows, columns, and treatments, the relative efficiency compared with a completely randomized design (for a standard t × t square), adjusted means, and Tukey comparisons. A 3 × 3 square leaves only 2 error degrees of freedom; aStat warns when error df are few.
Matches
SAS PROC GLM, MODEL y = row col trt.

Balanced incomplete block design (BIBD)

Use when
Blocks are too small to hold every treatment, such as litters of three piglets testing four creep feeds.
Data
A measurement, a treatment, and a block column. Each treatment appears at most once per block.
You get
Treatment tests within blocks (the intra-block analysis), the design constants (block size k, replications r, and how often each pair shares a block, λ), the efficiency factor, and means adjusted for blocks with Tukey comparisons. Raw means are biased in these designs, so aStat reports adjusted ones.
Matches
SAS PROC GLM, MODEL y = block trt, with LSMEANS.

Linear regression

Use when
You want to predict or explain one measurement from one or more others, such as body weight from heart girth.
Data
An outcome and 1 to 10 predictor columns, all numeric.
You get
Coefficients with confidence intervals, R² and adjusted R², the overall F test, a check for influential points (Cook's distance), and variance inflation factors when predictors overlap.
Matches
SAS PROC REG; R lm().

Correlation

Use when
You want to know which measurements rise and fall together, without choosing an outcome.
Data
2 to 20 numeric columns. Rows missing a value are left out pair by pair.
You get
Pearson, Spearman, and Kendall correlations with p-values, described as weak to very strong.
Matches
SAS PROC CORR (PEARSON, SPEARMAN, KENDALL).

Split-plot

Use when
One factor was applied to whole units and a second within them, such as ration fed to each steer and aging time applied to cuts from that steer.
Data
A measurement, the whole-plot factor, the sub-plot factor, and the whole-plot unit (or block) column.
You get
Type III F tests with Satterthwaite degrees of freedom, so the whole-plot factor is tested against the variation between whole plots and the sub-plot factor against the smaller variation within them; variance components; means; and comparisons within each whole-plot level.
Matches
SAS PROC MIXED with RANDOM unit(A) and DDFM=SATTERTH; R lme4 and lmerTest.

Repeated measures

Use when
Each animal stays on one treatment and is measured several times, such as body weight every two weeks.
Data
One row per animal per time: a measurement, a treatment, a time, and an animal ID column (up to 20 times).
You get
Tests for treatment, time, and their interaction with Kenward–Roger degrees of freedom; four covariance structures (unstructured, AR(1), compound symmetry, Toeplitz) compared by AIC, or the one you choose; means over time; and treatment differences at each time.
Matches
SAS PROC MIXED with REPEATED / TYPE= and DDFM=KR (aStat uses the linear form of Kenward–Roger that SAS uses); R mmrm.

Crossover

Use when
Each animal received the treatments in turn over periods, with a washout between them.
Data
One row per animal per period: a measurement, a treatment, a period, and an animal ID column. aStat finds each animal's sequence from the data.
You get
Tests for sequence, period, and treatment, with animals as a random effect; an optional carryover test (3 or more periods); variance components; and treatment means with Tukey comparisons. In a two-period design, a significant sequence effect can signal carryover.
Matches
SAS PROC MIXED, MODEL y = sequence period trt; RANDOM animal(sequence).

Repeatability (variance components)

Use when
Each animal (or herd, or pen) has several records and you want to know how much of the variation is lasting differences between them, such as litter size over several farrowings.
Data
A measurement and a group column with at least 3 groups.
You get
REML variance components, repeatability (the intraclass correlation), a likelihood ratio test of the group variance, the overall mean, and each group's predicted (BLUP) mean beside its raw mean.
Matches
SAS PROC MIXED with RANDOM group.

Random coefficients

Use when
You want the average growth line, and how much animals or pens differ in their own slopes.
Data
A measurement, a numeric predictor (such as day), and a group column, with each group measured at 2 or more predictor values.
You get
The average intercept and slope, the variances of intercepts and slopes and their correlation, a test of whether slopes differ, and each group's own line.
Matches
SAS PROC MIXED, RANDOM intercept x / SUBJECT=group TYPE=UN.

Sire evaluation (BLUP)

Use when
You have progeny records by sire, and possibly dam, and want heritability and a ranking of sires.
Data
A measurement, a sire column, and optionally a dam column (dams within sires). One row per offspring.
You get
Variance components, heritability (four times the sire share of variance), and each sire's EPD (half his breeding value), breeding value, predicted progeny mean with its standard error, and accuracy, best sires first. Predictions come from Henderson's mixed model equations.
Matches
SAS PROC MIXED, RANDOM sire dam(sire).
Heritability from fewer than about 20 sires is very imprecise; aStat warns when there are fewer.

One proportion

Use when
You want a rate, such as the dystocia rate in a herd, with a confidence interval, or to compare it with a target rate.
Data
A column of outcomes, and which value counts as the event.
You get
The proportion, the exact (Clopper–Pearson) and Wald intervals, an exact test against the target if you give one, and a frequency table.
Matches
SAS PROC FREQ with BINOMIAL.

Two-way table

Use when
You want to know whether two categorical columns are related, such as dystocia by sire breed, without fitting a model.
Data
Two columns with 2 to 20 values each.
You get
Counts with row percentages, Pearson's chi-square, and Fisher's exact test (which the answer uses when expected counts are small). For a 2 × 2 table, the risk in each row, risk difference, relative risk, and odds ratio with its confidence interval.
Matches
SAS PROC FREQ with CHISQ, EXACT, and MEASURES.

Logistic or probit model

Use when
The outcome is yes or no (pregnant or open, cleared or not) and you have treatments, blocks, or a dose.
Data
Either one row per animal with a yes/no column, or rows of events out of trials (lambs cleared out of lambs dosed). Add a treatment, a covariate, or both, and optionally a block, fixed or random.
You get
Likelihood ratio tests for each term, the estimated chance of the event in each group, odds ratios between groups (logit) or differences in probability (probit), the effect of each unit of the covariate, a test of equal slopes, and, for a dose alone, the dose at which half respond (ED50). Checks cover all-or-none groups, few events, and overdispersion.
Matches
SAS PROC GENMOD (fixed blocks) or PROC GLIMMIX (random blocks) with DIST=BINOMIAL; R glm() or lme4::glmer().

Count model

Use when
The outcome is a count: fecal egg counts, lesions, piglets born alive.
Data
A column of whole numbers of 0 or more, plus a treatment, a covariate, or both, and optionally a block.
You get
Poisson and negative binomial fits compared by AIC (or the one you choose), likelihood ratio tests, estimated mean counts, ratios between groups, and an overdispersion check. Negative binomial tests refit the spread parameter for each smaller model.
Matches
SAS PROC GENMOD with DIST=POISSON or DIST=NEGBIN; R glm() and MASS::glm.nb().

Ordinal scores

Use when
The outcome is an ordered score, such as locomotion 1 (sound) to 5 (severely lame), where the steps between scores need not be equal.
Data
Numeric scores with 3 to 10 values (code text categories as numbers in order first), plus a treatment, a covariate, or both, and optionally a fixed block.
You get
Likelihood ratio tests, the expected mean score in each group, the chance of each score in each group, and odds ratios of a higher score between groups.
Matches
SAS PROC GENMOD with DIST=MULTINOMIAL LINK=CUMLOGIT; R MASS::polr().

MANOVA

Use when
You measured several traits on each animal, such as milk yield, fat, and protein, and want to test them together.
Data
2 to 10 measurement columns and one or two factors.
You get
Wilks' lambda, Pillai's trace, the Hotelling–Lawley trace, and Roy's greatest root with Type III tests; each trait's own F test; means; and how strongly the traits are related within groups. The answer leads with Pillai's trace, which holds up best when assumptions are not quite met.
Matches
SAS PROC GLM with MANOVA H=; R summary.manova() F approximations.

Power and sample size

Use when
You are planning a study: how many animals do you need, or what chance does a planned study have of finding the effect?
Data
None. Choose the design (one sample, paired, two groups, three or more groups, or two rates), then enter the effect you expect and its standard deviation (or the expected rates).
You get
The number of animals needed (rounded up) or the power, the standardized effect size, and a power curve over study sizes.
Matches
SAS PROC POWER (two-sided t power counts both tails, as SAS does); R power.t.test(), power.anova.test(), and power.prop.test().