Which DOE Design Should I Use? A Decision Guide

Part of our complete guide to DOE for biologists.

Ask five different DOE textbooks "which design should I use?" and you'll get five flowcharts full of terms like Resolution IV, central composite, and Plackett-Burman before you've even described your experiment. None of that is the actual first question. The actual first question is much simpler: what do you already know about your system?

Start with what you know, not the design's name

Every DOE project falls into one of three situations, and the situation — not a textbook chapter — is what should pick your design:

You have a list of candidate variables and don't know which ones matter. Screen. You're not trying to find the exact optimum yet, just to cheaply rule out the variables that don't do anything so you can stop spending runs on them.

You already know your key variables and want to understand how they interact, or whether there's a peak inside your tested range. Characterize. This adds enough runs to estimate interactions cleanly and check for curvature.

You know your key variables and want the exact best settings. Optimize. This fits a full curved model and solves for the settings that maximize or minimize your outcome — or balance two outcomes against each other.

If that's all you needed, the complete guide covers it. The rest of this article is the questions people actually get stuck on.

"But I think I already know which variables matter"

Screen anyway, unless you're certain. The cost of screening a variable that turns out not to matter is small — one column in a design you were already running. The cost of skipping straight to characterization and being wrong is much larger: you've spent a bigger run budget confirming interactions between variables, only to find out later that one of the "obviously important" ones you didn't include was actually driving the result.

The exception is when you've already screened this system before, on a previous batch of a similar process, and the important variables genuinely haven't changed. Then it's reasonable to start at characterization.

Does the number of factors change the answer?

At one to three factors, not really — a screening design at that size is just a full factorial (every combination, no shortcuts), so there's no real downside to always screening first even if you're fairly confident.

At larger factor counts, the honest tradeoff shows up: a standard screening design tells you which variables matter but not whether any of them curve — you'd need a separate characterization step to find that out. A Definitive Screening Design closes that gap for a lot of factor counts, detecting curvature at a run count close to a standard screen instead of a full second experiment. See Definitive Screening Designs, Explained for how that works — and note you don't have to ask for one specifically; BioDOE uses one automatically when it's a real improvement for your factor count.

What if a factor is expensive to reset, or I need to split the work across two days?

Neither of those changes which of the three designs you need — they change how the design is built once you've picked one. See Hard-to-Change Factors and Split-Plot Designs and How to Block an Experiment After You've Already Designed It.

The honest shortcut

You don't actually need to memorize any of this to use BioDOE. The three situations above are the three goal cards you pick from — "Find what matters," "Map the shape," "Find the peak" — in plain language, no design terminology required. BioDOE picks the underlying construction method for you, including newer options like Definitive Screening Designs when they're a strict improvement, and steps you through Hard-to-Change and blocking as separate questions if they apply. Start free and pick your goal — the rest of this decision tree happens under the hood.