The Illusion of the Independent Variable
In classical one-factor-at-a-time testing, changing a single parameter produces what seems like a clean, predictable line. During a recent composite curing trial, the engineering group increased the dwell temperature from 140°C to 180°C and recorded a steady rise in resin cross-linking. Every single-variable chart indicated that higher thermal input yielded superior polymer density. Confidence was high, the process window was documented, and the team prepared to lock in the recipe across production autoclaves without testing alternative pressure boundaries.
The flaw became obvious the moment vacuum bag pressure dropped from 6.0 bar to 4.2 bar during an auxiliary pump switch. Instead of producing the expected slight dip in density, the high-temperature specimens suffered catastrophic delamination and void creation. Temperature was not acting as an autonomous driver of quality. Its influence was deeply coupled with compaction force, meaning the slope of the temperature response line swung from strongly positive to severely negative depending on the pressure level.
-
01
Non-Parallel Trajectories When interaction plots display crossing or non-parallel response lines, main effects cannot be interpreted independently without misleading production operators.
-
02
Confounded Resolution Hazards Fractional factorial designs below Resolution V fold two-factor interactions into main effects, disguising mutual dependencies as primary drivers.
-
03
Synergistic & Antagonistic Shifts Evaluating interactions uncovers positive synergy where variables amplify performance, or antagonistic degradation where optimal settings clash.
Mapping the Curvature: From Linear Lies to 3D Surfaces
Calculating main effect averages across all experimental runs effectively washes out the interaction by blending positive and negative extremes into a bland, flat midpoint. An operator reading the summary report sees a modest main effect slope and assumes the parameter has minor impact. In reality, the parameter possesses enormous impact, but its direction inverts midway through the operating space.
“A main effect is merely an average across conditions. If those conditions change the sign of the effect, the average tells you nothing true about the process.”
Resolving these hidden mechanisms requires stepping beyond basic screening matrices into full factorial blocks or central composite designs. By evaluating the interaction term (A × B) in the regression model, engineers isolate the coefficient of joint variance. When that p-value drops below 0.01, standard single-factor operating guidelines must be abandoned in favor of contour plots that define bounded operational sweet spots.
Add Engineering Note
Share your observations or validate experimental factor assumptions.