Choosing a method
When you upload a dataset, Celeus suggests methods that fit its structure. This page explains the questions behind those suggestions, so you can check that a suggestion matches your study. Each method links to its page in the method reference.
Five questions decide most of the choice. Answer them about your study design, not only about the columns in the file.
1. What kind of outcome do you have?
Section titled “1. What kind of outcome do you have?”The outcome is the variable you want to explain or compare.
- A numeric measurement, such as a test score, a weight or a yield: the t-tests, ANOVA family and linear regression.
- Two categories, such as passed or failed: logistic regression, or a test of association if you are comparing groups.
- Three or more unordered categories: multinomial logistic regression.
- Ordered categories, such as a five-point rating: ordinal logistic regression.
- A count of events, such as the number of defects per batch: Poisson regression, or negative binomial regression when the counts vary more than a Poisson model allows.
- Time until an event, where some units have not had the event by the end of follow-up: see question 4.
- No single outcome, because you want to describe the dataset first: the descriptive table.
2. How many groups are you comparing?
Section titled “2. How many groups are you comparing?”- Two groups: Welch’s t-test by default, Student’s t-test when the variances are equal, or the Mann-Whitney U test when the outcome is far from normally distributed or is itself ordinal.
- Three or more groups: one-way ANOVA, Welch’s ANOVA when the variances differ, or the Kruskal-Wallis test on ranks.
- Two grouping factors at once, such as fertiliser and soil type: two-way ANOVA.
- Groups plus one or more variables to adjust for, such as a baseline score, age or sex: ANCOVA, which takes any number of numeric or categorical covariates. If you want a coefficient for every predictor rather than the adjusted group comparison, use linear regression.
For two categorical variables, the comparison is a test of association: the chi-squared test, or Fisher’s exact test when the expected counts are small. For two numeric variables, it is a correlation: Pearson for a linear relationship, Spearman for any consistently increasing or decreasing one.
3. Are the observations independent or paired?
Section titled “3. Are the observations independent or paired?”Two measurements on the same unit, such as before and after, or two raters scoring the same items, are paired. Treating them as independent groups ignores that pairing and usually wastes information.
- Paired numeric measurements: the paired t-test or the Wilcoxon signed-rank test.
- Paired yes/no outcomes: McNemar’s test.
- Two methods measuring the same quantity, where the question is whether they agree: Bland-Altman analysis. A correlation answers a different question, since two methods can be highly correlated and still disagree.
4. Is the outcome censored?
Section titled “4. Is the outcome censored?”An outcome is censored when, for some units, you know only that the value lies beyond a limit.
- Time to an event, where some units leave the study or reach its end without the event: Kaplan-Meier with the log-rank test to compare groups, Cox regression to adjust for several predictors, or Weibull accelerated failure time regression when you want time ratios (how much a predictor stretches or shrinks the time to the event) instead of hazard ratios.
- Values below or above a detection limit, such as a concentration reported only as “below 0.5”: the Peto-Peto test to compare groups, or censored regression.
Dropping censored units, or replacing them with the limit, biases the results. These methods use the censored values for what they are.
5. Are the observations clustered?
Section titled “5. Are the observations clustered?”Observations are clustered when several come from the same unit: repeated measurements on one subject, pupils within one school, plots within one field. Observations in a cluster tend to resemble each other, and a method that ignores this overstates how much information the data hold.
- A numeric outcome measured repeatedly under each condition: repeated-measures ANOVA.
- More general clustered designs: a linear mixed model, or its versions for a yes/no outcome and for a count outcome.
Classification and accuracy questions
Section titled “Classification and accuracy questions”Some questions are about how well a score or a test classifies, not about a difference between groups: the ROC curve and its area for a numeric score against a known reference, a comparison of two ROC curves, and sensitivity and specificity for a yes/no test.
When the assumptions do not hold
Section titled “When the assumptions do not hold”Every method’s page lists what you provide and its options. After a run, the results page shows the assumption checks that matter for the chosen method, and a failed check points to the alternative to try, such as a rank test instead of a t-test. A suggestion is a starting point; the design of your study, which Celeus cannot see in the file, decides the method.