3. Experimental design principles

An occupant injury experiment is only as credible as the design that produced it. A sled test, a drop tower impact, or a cadaveric rib loading series can generate terabytes of clean data and still fail to answer the question it was built around — because the hypothesis was vague, the variables were confounded, or the sample was too small to detect the effect of interest. This section covers the four design pillars that determine whether an injury study produces defensible conclusions: hypothesis formulation, variable structure, control strategy, and statistical power.

From Research Question to Testable Hypothesis

A usable hypothesis in injury biomechanics must specify three things: the population or specimen type, the loading or exposure condition, and the measurable injury or response outcome with a stated direction. "Seat belt pretensioning reduces injury" is not testable. A testable version reads: "In Post-Mortem Human Surrogate (PMHS) thoraces subjected to 40 km/h frontal sled pulses, a 4 kN pretensioner force applied 15 ms before belt loading reduces peak sternal deflection relative to an identical pulse without pretensioning." Every clause maps to a design decision — specimen selection, pulse profile, actuator timing, and the response channel that will be analyzed.

Two hypothesis types dominate the field. A superiority hypothesis (H₁: μ_treatment < μ_control) suits countermeasure evaluation, where the goal is to show a restraint or airbag configuration outperforms a baseline. An equivalence or non-inferiority hypothesis suits validation work, where a finite element (FE) model or an Anthropomorphic Test Device (ATD) must be shown to reproduce cadaveric response within an acceptable band. These two framings demand different sample sizes and different statistical tests, so the choice must be made before any specimen is ordered.

Variables: Independent, Dependent, and the Ones That Bite

The independent variable is what the experimenter manipulates — pretensioner force level, pulse severity (Δv), impact angle, or restraint type. The dependent variable is the measured outcome: peak chest deflection, AIS severity, rib fracture count, head injury criterion (HIC), or peak lumbar axial force. Between them sit controlled variables (held constant: pulse shape, seat geometry, specimen posture) and nuisance variables (uncontrolled but measurable: bone mineral density, specimen age, soft tissue condition).

Injury studies carry a variable structure that most engineering experiments do not: the specimen itself is a major source of variance. PMHS age, sex, and bone quality can shift fracture tolerance by a factor of two or more. This is why blocking and randomization matter. If you test all young specimens with the treatment and all elderly specimens as controls, age is perfectly confounded with treatment and no statistical method can rescue the result. Randomize specimen assignment to groups, or block by age decade and analyze within blocks.

Controls in Injury Experiments

Controls establish the counterfactual — what would have happened without the intervention. Four control strategies are relevant here:

• Negative control: the same specimen type and pulse without the countermeasure. This is the baseline against which treatment effect is measured.
• Positive control: a known-injurious or known-safe condition that confirms the test setup can actually produce the expected outcome. If a positive control fails to injure, the rig or instrumentation is suspect.
• Matched control: a specimen pair matched on age, sex, and bone density, with one receiving treatment and one not. This reduces between-specimen variance but is expensive and often impractical with PMHS.
• Repeated-measures control: the same specimen tested at multiple non-injurious load levels, using each specimen as its own control. This is powerful for sub-failure response studies but cannot be used once injury occurs.

Controls must also address order effects and operator effects. If the same technician always sets up the treatment condition, subtle setup differences can masquerade as treatment effects. Rotate operators or blind the analyst to group assignment where feasible.

Statistical Power: Sizing the Experiment Before You Cut Metal

Power is the probability of detecting a true effect of a given size. It depends on four quantities: effect size (Δ), standard deviation (σ), significance level (α, typically 0.05), and sample size (n). For a two-sample t-test, the approximate relationship is:

n per group ≈ 2 · (z_{1-α/2} + z_{1-β})² · σ² / Δ²

For α = 0.05 (two-sided), power = 0.80:
  z_{1-α/2} = 1.96,  z_{1-β} = 0.84
  n per group ≈ 2 · (1.96 + 0.84)² · σ² / Δ²
             ≈ 15.68 · (σ/Δ)²

Example: peak chest deflection, σ = 8 mm, minimum
detectable effect Δ = 5 mm
  n per group ≈ 15.68 · (8/5)² ≈ 40 specimens per group

That number — 40 PMHS per group — is why full-factorial cadaveric studies are rare. The practical response is to increase Δ (test a larger, more aggressive intervention), reduce σ (better specimen matching, tighter posture control, repeated measures), or accept lower power and report it honestly. A study with n = 6 per group and σ/Δ = 1.6 has power near 0.35; a non-significant result there is uninformative, not evidence of no effect.

Design Flow and the Power–Feasibility Trade-off

The figure below traces how a research question narrows into a testable hypothesis, then into variable and control decisions, and finally into a power calculation that feeds back to constrain the design. The feedback loop is the critical feature: power analysis is not a final checkbox but a design driver that may force you to revise the hypothesis, tighten controls, or change the outcome measure.

Putting It Together: A Design Checklist for Injury Studies

The following table summarizes the decisions that must be locked before data collection begins. Each row corresponds to a design element that, if left ambiguous, will surface as an unexplainable result later.

Two failure modes recur in occupant injury research. The first is underpowered confirmation: a study with n = 4 per group reports "no significant difference" and the authors conclude equivalence, when the study simply could not detect the effect. The second is confounded treatment: the treatment group differs systematically from the control in a way that was not measured, so the observed effect cannot be attributed to the intervention. Both are design failures, not analysis failures, and neither can be corrected after the fact.

The practical discipline is to write the analysis plan — including the statistical test, the α level, the primary endpoint, and the MDE — before the first specimen is tested. If the power calculation says the experiment is not feasible at the required effect size, the correct response is to redesign the study, not to run it underpowered and hope for a significant result.