2. Formulating testable hypotheses from crash test objectives

A crash test objective is rarely stated in a form that statistics can consume directly. "Verify that the new bumper absorbs energy as well as the baseline" or "confirm the restraint system reduces head injury risk" are engineering intentions, not testable statements. Before any data are collected, the analyst must translate each objective into a pair of competing claims about a population parameter — a null hypothesis and an alternative hypothesis — and decide whether the evidence should be judged in one direction or two. These choices are not bookkeeping. They determine the sample size, the acceptance criterion, and, ultimately, whether a real safety improvement is recognized or a regression slips through.

From objective to parameter

The first move is to identify the measurable quantity the objective actually concerns. Crash engineering objectives usually reduce to one of a small set of parameter types: a mean (average peak deceleration, mean HIC across a test series), a proportion (fraction of tests exceeding a threshold, pass rate of a sled configuration), a variance or standard deviation (repeatability of a pulse), or a difference between two conditions (baseline vs. modified design). The hypothesis must name that parameter explicitly and attach a numerical value to it.

Consider a concrete example. A program wants to demonstrate that a revised front structure lowers the mean peak chest acceleration in a 56 km/h frontal rigid-barrier test relative to the current design, whose historical mean is 42.0 g. The objective "lower chest acceleration" becomes a claim about the population mean μ of the revised design. The null hypothesis encodes the conservative default — no improvement, or worse:

H0: μ ≥ 42.0 g      (revised design is no better than baseline)
H1: μ < 42.0 g       (revised design reduces mean peak chest acceleration)

Note the direction. The burden of proof sits on the improvement claim, which is exactly what a safety engineer should want: the new design is not credited with a benefit until the data overcome the assumption that it offers none. A common error is to write H0 as "μ = 42.0 g" and H1 as "μ ≠ 42.0 g" out of habit. That formulation answers a different question — "is the design different?" — and would treat a higher chest acceleration as equally interesting as a lower one. For a one-directional engineering objective, the one-sided form is the honest one.

One-sided versus two-sided tests

The choice between one-sided and two-sided alternatives follows from the objective, not from statistical convenience. Use a one-sided test when only one direction of departure matters and a result in the opposite direction would lead to the same engineering action as "no effect." Use a two-sided test when departures in either direction are consequential and must be detected — for example, when qualifying a component whose performance must stay within a tolerance band, or when a change could plausibly help or harm and both outcomes require a response.

A useful rule of thumb: if a significantly worse result would trigger a redesign just as a significantly better result would trigger a rollout, and you need to know which, a two-sided test is appropriate. If a worse result simply means "keep the current design" and only improvement justifies a change, a one-sided test is defensible and more powerful in that direction. The trade-off is real: a one-sided test at α = 0.05 places all 5% of false-alarm risk in the direction you care about, whereas a two-sided test at the same α splits it 2.5% per tail, making it harder to detect an improvement of a given size.

Decision errors and their trade-offs

Every hypothesis test is a decision rule that can fail in two ways. A Type I error (false positive, probability α) occurs when H0 is true but is rejected — here, concluding the revised design improves chest acceleration when it does not. A Type II error (false negative, probability β) occurs when H0 is false but is not rejected — failing to recognize a genuine improvement. The power of the test is 1 − β.

In crash engineering the two errors carry asymmetric costs. A Type I error can release an ineffective or unsafe design change into production, with recall, liability, and injury consequences. A Type II error wastes a real improvement and may cause a program to abandon a valid countermeasure. Neither is free, and the analyst must set α and the target power deliberately, then size the test series accordingly. The relationship among them is governed by the effect size Δ, the variability σ, the sample size n, and the chosen α:

For a one-sided test of a mean, required n ≈ ( (z_{1-α} + z_{1-β}) · σ / Δ )²

Example (illustrative): σ = 3.0 g, Δ = 1.5 g, α = 0.05, power = 0.80
  z_{0.95} = 1.645, z_{0.80} = 0.842
  n ≈ ( (1.645 + 0.842) · 3.0 / 1.5 )² ≈ (4.974)² ≈ 24.7 → 25 tests

These numbers are a worked illustration, not measured results. The message they carry is the practical one: detecting a 1.5 g improvement against 3.0 g of test-to-test scatter demands roughly 25 valid tests. If the program can only afford eight, then either the detectable effect must be larger, the variability must be reduced through tighter build control, or the analyst must accept lower power and state that limitation openly. Setting α = 0.05 with power 0.80 is a common default, but a safety-critical qualification may justify α = 0.01 to guard against false positives, accepting the larger n that follows.

Writing the hypothesis pair correctly

Three habits keep hypothesis statements defensible. First, state the parameter and the threshold numerically — "μ < 42.0 g," not "the new design is better." Second, put the claim you want to establish in the alternative, so that the null represents the status quo or the conservative assumption. Third, fix the direction and the α level in the test plan before data collection, and record the target effect size Δ and assumed σ that drove the sample-size calculation.

For proportion-based objectives the same logic applies with a different parameter. If a program must show that the pass rate p of a new sled setup exceeds 0.90, the pair becomes H0: p ≤ 0.90 versus H1: p > 0.90, and the sample size follows from the binomial power calculation rather than the normal approximation above. For variance objectives — demonstrating that a pulse is more repeatable — the parameter is σ² and the test is typically one-sided on the variance ratio, since only reduced scatter is the goal.

What the decision rule delivers

Once the pair is fixed, the analysis becomes mechanical: compute the test statistic from the sample, compare its p-value to α, and state the decision in engineering terms. If the one-sided test on the chest-acceleration example returns p = 0.03 with α = 0.05, the decision is to reject H0 and conclude the revised structure reduces mean peak chest acceleration — subject to the caveats that the test series met the build and instrumentation controls and that the effect size is practically meaningful, not merely statistically detectable. If p = 0.18, the correct statement is that the data do not provide sufficient evidence of improvement at the chosen power; it is not proof that the designs are equivalent. That distinction — failing to reject versus proving no difference — is the single most common misinterpretation in crash data review, and it is settled at the hypothesis-writing stage, long before any accelerometer trace is examined.