Screening Design vs. Response Surface: Which Do You Need?
Part of our complete guide to DOE for biologists.
Open most DOE software and you're asked to pick a design type before you've defined a single variable. If you don't already know the difference between a screening design and a response surface design, that's not a helpful question yet — it's a name for a decision you haven't made the inputs for. The good news is the choice maps directly onto a question you can actually answer: what do you already know about your system?
If you don't know which variables matter: screen
A screening design exists to answer one question cheaply: of everything you could be testing, which few things actually move the outcome? It tests each variable at just two levels (a low and a high setting), which keeps the run count low even with five, six, or more candidate variables. What it deliberately doesn't do is tell you the exact optimum — with only two levels per variable, it can tell you the direction that helps, not where a peak sits inside that range.
That's the right trade for this stage. You're not ready to optimize six variables at once — you're ready to find out that three of them don't matter, so you stop spending runs on them.
If you know your key variables and want interactions and curvature: characterize
Once you're down to a handful of variables that clearly matter, a characterization design adds enough runs — usually a few extra center points — to do two things a screening design can't: estimate how variables interact with each other cleanly, and check whether the relationship is a straight line or bends somewhere inside your tested range. That bend, called curvature, is the tell that a true optimum exists between your low and high settings rather than just outside them.
If you know your variables and want the exact peak: optimize
A response surface (optimization) design fits a full curved model across your variables and solves for the settings that maximize or minimize your response — or balance two responses against each other, like maximizing yield while keeping a side effect below a threshold. This needs more runs than screening because it's modeling curvature deliberately, not just checking for its presence.
The natural path through all three
Most real projects don't pick one of these and stop — they move through them in order. Screen a long list down to the vital few. If curvature shows up, characterize or move straight to optimization on just those. The reason this works without wasted effort is that a curvature finding isn't a dead end: BioDOE can extend a screening design into a full response surface design instead of making you throw it out and start over, because the runs you already have stay valid — you're adding to the picture, not replacing it.
→ Related: Reading Your ANOVA Table and Pareto Chart covers exactly how to spot that curvature signal when your results come back.
Not sure which stage your experiment is at? Start free — pick your goal in plain language and BioDOE chooses the design type for you.