Which Statistical Test Should I Use? Let Stratum Ask

Every which-test flowchart asks you things your data file already knows. Guide Me asks only what it doesn't, then builds the analysis and the chart.

Follow along: process_measurements.csv — 2,000 wafer measurements with Conductivity, Thickness and Resistivity, plus Tool, Shift, Recipe and a Pass / Rework / Scrap Disposition. Download the CSV.
Requires Stratum 1.5, where Guide Me arrives on Mac and Windows.

The classic answer to "which statistical test should I use?" is a flowchart, and most of its boxes — continuous or categorical, how many groups — are questions about columns the software already holds. What it can't know is what you're trying to find out, which column is the outcome, and whether two columns of numbers are the same units measured twice. Guide Me asks exactly those, and nothing else.

One question, then your columns

Click Guide Me in the toolbar beside Organize (it is also pinned at the top of the + menu). The first panel asks What do you want to find out? and offers five answers: Describe what's in this data, Compare groups, See how two things relate, Predict or explain one column, and Watch a process over time.

Stratum's Guide Me sheet with five illustrated tiles: describe, compare, relate, predict, watch a process — Five answers to one question. Everything else is read off the file.

Pick Compare groups and the next panel opens with its pickers already filled in — Stratum's best guess for this file, chosen by finding a measurement that actually separates by some grouping. Set Compare to Conductivity and Split by to Tool. Before you commit, the panel names the analysis — One-Way ANOVA — and what it read to get there: Reading Conductivity (Continuous) by Tool (Nominal).

That sentence is the routing, out in the open. Nobody asked whether Tool is a category or how many levels it has — the file says so. If Stratum had read a column wrong, say a 1/2/3 machine code taken for a measurement, you'd see it here before anything runs.

Guide Me panel with Compare set to Conductivity and Split by set to Tool, suggesting One-Way ANOVA shown as a violin — The analysis, the chart, and the levels it read — all on the panel before you run anything.

A follow-up only where the file is silent

There are exactly four follow-up questions, and each appears only when its answer would change the result:

Are these two sets of values matched? — when you compare two measurement columns (Or compare with (optional)). Before/after on one unit and two independent samples look identical on disk. A split by a group is unpaired by construction, so it's never asked there.
What do you want out of it? — predicting a number: a model you can read (Linear Regression) or the most accurate one (Random Forest).
One column at a time, or how they group? — describing four or more measurements.
Watching it move, or checking it against a specification? — a control chart or Process Capability.

The panel then names what the other answer would have done. It reports what it picked; it never calls a test the correct one.

The chart is chosen for the data, not the analysis

For a group comparison, rows per group pick the picture: ten or fewer gives a Dot/Strip of the actual values, a hundred or more a Violin (a density needs rows), a Box-Whisker in between. About 500 rows per tool earns the violin here. A date as the second column gives a Trend line; past 2,000 rows a scatter becomes a Heatmap/HexBin; a paired comparison is a Slopegraph. Where the data can't choose, Shown as lists the alternatives.

What Guide Me routes to

You want to	Your columns	It asks	It runs	Chart
Describe	fewer than 4 measurements	—	Summary	Histogram
Describe	only categories	—	Frequencies	Bar
Describe	4+ measurements	each column, or how they group?	Summary / Principal Components / Clustering	Histogram / PCA Biplot / Parallel Plot
Compare	measurement split by a 2-level group	—	Two-Sample (unpaired)	by group size
Compare	measurement split by 3+ levels	—	One-Way ANOVA	by group size
Compare	… and a second factor	—	Two-Way ANOVA	by group size
Compare	two measurement columns	matched?	Two-Sample, paired or unpaired	Slopegraph / Box-Whisker
Compare or relate	two categories	—	Contingency Table; Correspondence Analysis if both have 3+ levels	Mosaic / Correspondence Map
Relate	two measurements	—	Correlation	Scatter; Trend for a date; Heatmap/HexBin past 2,000 rows
Relate	a category, no second column	—	Differences (scans the groupings)	Mosaic
Predict	a categorical outcome	—	Decision Tree	the tree
Predict	a numeric outcome	readable or most accurate?	Linear Regression / Random Forest	Scatter / —
Watch a process	a measurement	over time or against spec?	I-MR (X̄-R with subgroups) / Process Capability	— / Histogram

Relate a categorical outcome to a measurement and Stratum compares the measurement across the groups instead, and says so under the result rather than quietly swapping your columns. A few analyses stay off the panel because no answerable question separates them — Boosted Trees beside Random Forest, CUSUM and EWMA beside I-MR, the P and C charts that need a sample-size column. They're in the menus.

Add the report, the chart, or both

Add decides what gets built: the analysis, the Chart, or Both (the default). Run it and the cards arrive configured, report selected — the answer first, the picture as its check. The ANOVA here is not subtle (F near 99 on 3 and 1,996 df), and the violin shows which tool pulls away.

Stratum window with the One-Way ANOVA report for Conductivity by Tool selected and a violin chart card beside it — Run it builds both cards, configured, with the report in front.

Now the categorical side: See how two things relate, First column = Disposition, Second column = Tool. Both have three or more levels, so Guide Me goes past a contingency table to a Correspondence Analysis, shown as a Correspondence Map. With two levels on either side the map collapses to a line, and the table stays the answer.

Guide Me relating Disposition and Tool, suggesting Correspondence Analysis shown as a correspondence map — Two roomy categories: a map of where the association lives, not just a grid of counts.

A second opinion, too. Even when you know the test, the reading line checks that Stratum has your levels right. If it calls a group code Continuous, fix it once in Column Definition… and every picker follows.

Download Stratum

Frequently asked questions

Which statistical test should I use to compare two groups?

If the two groups are different units, an unpaired two-sample test; if each row measures the same unit twice (before and after, left and right), a paired one. That is the one fact your file cannot tell software, so Guide Me asks it directly whenever you compare two measurement columns.

When should I use ANOVA instead of a t-test?

When the grouping has three or more levels. Guide Me routes a measurement split by a two-level group to Two-Sample, by three or more levels to One-Way ANOVA, and adds Two-Way ANOVA when you name a second factor.

Does Guide Me check the test's assumptions?

No. Guide Me chooses the analysis and the chart; the report is where you check it. The One-Way ANOVA report, for example, offers Levene's test and Welch's ANOVA as sections.

What if my outcome is a category?

Under Predict, a categorical outcome gets a decision tree. If you relate a categorical outcome to a measurement, Stratum compares the measurement across the groups instead and says so on the panel, rather than silently swapping the roles.

Is Guide Me available on Windows?

Yes. The routing lives in Stratum's shared engine, so the Mac and Windows apps ask the same questions and reach the same analysis.