StratumStratum vs Excel for Statistical Analysis
Excel is a spreadsheet, not a statistics tool — and using it as one is where analyses quietly go wrong. Here's where it fails at real statistics, the failure modes named plainly, and what changes when you open the same file in a tool built for the job — native to Mac and Windows.
Excel is a spreadsheet — a brilliant, general-purpose grid of cells, and probably the most-used piece of software on the planet. What it is not is a statistics tool. You can compute a t-test in it, and for a quick average that's perfectly fine. But the moment an analysis gets real, the spreadsheet stops being the right instrument — and the failures are quiet ones. A wrong range, a blank read as a zero, a missing value that never announces itself: the number still looks fine, and it's wrong. Excel isn't bad software; it's the wrong software for this, and it's worth being precise about exactly where it stops.
Where Excel runs out of road
- It has no real concept of missing data. "Not measured" and "zero" look identical to a formula, so a blank cell silently distorts a mean, a count, or a regression unless you hunt every one down by hand. A statistics tool treats missing values as first-class and excludes them the way the method requires.
- Results ride on hand-built formulas that fail silently. Your analysis lives in thousands of hand-entered cell references; drag one a row too far, or leave a range that didn't grow when you added data, and the result looks fine and is wrong — with nothing to flag it, and painful for anyone else to verify.
- There are no diagnostics. The Analysis ToolPak gives you an F and a p-value and stops. No VIF, no leverage, no Cook's distance, no residual or normality checks, no Tukey post-hoc — the exact outputs that tell you whether a model is trustworthy or a fluke. Without them you're reporting a number you can't defend.
- The charts aren't statistical. Excel does bar and line charts well, but box plots, violins, Q-Q plots, ECDFs, mosaics, Pareto charts, and SPC control charts are awkward or simply absent — and those are exactly the charts that explain a distribution.
- There's no reproducibility or audit trail. A finished workbook is a snapshot of values and formulas, not a record of what you did. Re-run it a month later, or hand it to a reviewer, and there's no re-editable analysis to inspect — just cells to re-trace.
- It mangles data on the way in. Excel is documented to auto-convert values it thinks it recognizes — gene names and part numbers turned into dates, long IDs rewritten in scientific notation, leading zeros stripped from codes. The corruption happens on open, before you've done anything, and it's why some scientific fields renamed their identifiers to escape it.
- It strains at scale. Past a few hundred thousand rows Excel slows noticeably; its hard ceiling is 1,048,576 rows. Plenty of real datasets are larger than that.
- The Mac version lags. Excel for Mac has historically trailed the Windows build on the very analysis features you'd reach for, the ToolPak included.
A concrete example: a one-way ANOVA with post-hoc
Say you're comparing a measurement across four groups. In Excel you'd run the ToolPak's single-factor ANOVA, get an F and a p-value — and then stop, because the ToolPak has no post-hoc test. To find which groups differ you'd hand-build Tukey HSD across cells, with no built-in critical values, and hope you referenced the right ranges. In a dedicated tool you pick the response and the factor, read the F and p, then right-click for Tukey HSD and Levene's test. Same analysis; one is a chore you can get subtly wrong, the other is three clicks with the diagnostics attached.
Feature-by-feature
| Excel for Mac | Stratum | |
|---|---|---|
| Inferential tests | Limited (ToolPak) | t-tests, ANOVA + Tukey, MANOVA, correlation, chi-square, regression with full diagnostics |
| Post-hoc / diagnostics | None built in | Tukey HSD, Levene's, VIF, leverage, Cook's distance |
| Missing data | Blank = zero to most formulas | First-class; excluded per method |
| Data integrity | Auto-converts IDs / values on import | Types columns; never rewrites your data |
| Statistical charts | Bar / line; box plots awkward | 30+ incl. box, violin, density, Q-Q, Pareto, mosaic, heatmap, control charts |
| Machine learning | None built in | PCA, k-means & hierarchical clustering, decision trees, random forests, boosted trees |
| Scale | ~1M-row ceiling, sluggish before | Smooth to ~5 million rows |
| Reproducibility | Hand-built cell formulas, no audit trail | Re-editable analyses saved inside the document |
| Privacy | Often cloud-tied | 100% native and offline |
| Platform | Cross-platform; Mac build trails | Native to Mac & Windows |
Two more traps that bite quietly
Even the everyday summary statistics have sharp edges in a spreadsheet:
- Quartiles depend on which function you picked. Excel ships two percentile/quartile methods (
.INCand.EXC) that give different answers; grab the wrong one and your box plot's box moves without warning. - Blanks and text in a range are handled inconsistently. Some functions skip them, some treat a blank as zero — and a single stray non-numeric cell can quietly change a mean or a count.
Stratum sidesteps all of this: it types every column at import, treats missing values as first-class, and computes each statistic with one documented method — so the result is the same every time and there's no cell arithmetic left to audit.
Where Excel is genuinely the right tool
To be fair, that's a real list. Excel is ubiquitous, familiar, and unbeaten for data entry and storage, for quick arithmetic and what-if models, for sharing a table with someone who'll open it anywhere, and for financial and business spreadsheets full of bespoke formulas. If your "analysis" is a sum, an average, a lookup, or a pivot table, Excel is the fastest path — use it. None of that was ever the argument. The point is narrower and factual: Excel was never built for statistics, and the moment your question turns inferential it's working outside its design.
There's nothing to migrate
You don't convert anything, and you don't relearn from scratch. Stratum opens your existing .xlsx and .csv files directly — drag one in and it imports as a typed dataset, numeric and categorical columns detected automatically, missing values kept as missing. From there you get diagnostics, 30+ modern statistical charts, machine learning, and datasets up to roughly 5 million rows, native on Mac and Windows alike. Your spreadsheets stay exactly where they are; Stratum is simply where you take them the moment a real analysis question comes up — Excel for the ledger, Stratum for the statistics.
Download Stratum See all features →
Frequently asked questions
Can Excel for Mac do ANOVA and regression?
Barely. Via the Analysis ToolPak it will run the test and hand you an F and a p-value, but with no diagnostics at all — no VIF, leverage, Cook's distance, residual or normality checks, or post-hoc tests — which are exactly the outputs that tell you whether the model holds up. For real inferential work a purpose-built stats app like Stratum is both more capable and far less error-prone.
How many rows can Excel handle?
A worksheet maxes out at 1,048,576 rows, and it slows well before that. Stratum stays responsive on datasets up to roughly 5 million rows, native on Mac and Windows alike.
Do I have to convert my Excel files to use Stratum?
No — there's nothing to migrate. Stratum opens .xlsx and .csv files directly. Drag the file in and it's imported as a typed dataset, columns detected and missing values kept as missing, ready to chart and test.
Is Excel ever the better choice?
Yes. For data entry, quick arithmetic, what-if models, pivot tables, and sharing a plain table, Excel is excellent. Reach for a stats app when the question becomes inferential or the charts need to be statistical.
What's the best statistics software for Mac if not Excel?
It depends on your work, but a native app for Mac & Windows that covers descriptive stats, inferential tests with diagnostics, statistical charts, and machine learning in one place — like Stratum — avoids both Excel's limits and the overhead of R or Python.
Is Stratum's math independently validated?
Yes. Stratum publishes a per-analysis validation record that names the outside tool each statistic is checked against — R, SciPy, statsmodels and scikit-learn — and states plainly which analyses are only partly covered. Linear regression passes all nine NIST Statistical Reference Datasets least-squares cases, including Filip, a tenth-degree polynomial that is a known stress case for least-squares solvers.