Internal test (CC Athletics office)
Validation against Bertec ForcePlate (Hawkin Dynamic processing)
This blog post compares raw force-plate signals from CC Athletics PlateMate (960 Hz) and a Bertec force plate processed by Hawkin Dynamic software (1000 Hz), captured simultaneously during a single countermovement jump (CMJ).
Setup: the PlateMate was placed on top of a Bertec (stacked), so the two devices share a mechanical load path and observe the same physical event simultaneously. See the Setup section below for the full implications — most importantly, that this is a strong test of electronic equivalence between the device chains.
The question we're trying to answer: for the raw signal (not derived metrics like jump height or peak power), how closely do these two devices agree when fed the same physical input?
TL;DR — what we found:
- Static body weight agreement: ~0.11 % (~80 g out of ~77 kg).
- Pearson correlation point-to-point: 0.9990.
- Power-weighted coherence over 0–30 Hz: 1.0000. Coherence is essentially 1 in every sub-band 0–2, 2–5, 5–10, and 10–30 Hz.
- Time-domain RMSE: ~2.92 kg raw / ~0.42 kg at 5 Hz low-pass (~3.8 % / ~0.5 % of mean force).
- Peak landing force gap: −2.04 % raw, but only −0.41 % under 30 Hz low-pass and +0.25 % under 5 Hz low-pass — the discrepancy lives at the very tip of the 1 ms-wide impact transient, not at any biomechanical-band timescale.
- Biomechanical-metric agreement < 1 % across the board: jump height +0.96 %, takeoff velocity +0.48 %, peak power +0.48 %, propulsive impulse +0.29 %.
Within this single stacked-plate recording, PlateMate and Bertec/Hawkin agree to a fraction of a percent on every metric a coach or clinician would actually report. The only meaningful raw-signal discrepancy is the impact-tip peak gap, and that doesn't propagate to integrated metrics like jump height.
Important caveats: n = 1 subject, n = 1 jump. Stacked-plate configuration (not side-by-side). Only summed total vertical force was compared (no bilateral / centre-of-pressure / derived asymmetry metrics). None of the numbers below should be read as population-level claims — they describe this recording with this configuration.
This is the publication-friendly version. The extended one contains the full audit trail (cross-correlation alignment defence, per-Welch-segment amplitude-ratio statistics, amplitude/phase diagnostics) for readers who want to see the working.
Setup: Stacked plate configuration
Important to know before reading the analysis: this is not a side-by-side comparison. The PlateMate was placed on top of a single Bertec, and both devices recorded the same subject simultaneously. The mechanical load path is:
Subject → PlateMate top surface → PlateMate load cells → PlateMate frame → Bertec top surface → Bertec load cells → floor
Why this setup matters:
- Strength. The two devices are guaranteed to see the same physical event — they share one mechanical load path. This removes the "are they both observing the same thing?" question entirely and is among the most rigorous configurations for force-plate validation.
- What the configuration cannot test. Because the same vertical force flows through both devices in series, biological asymmetry, foot placement variability, and impact mechanics are shared identically — not evaluated independently by the two devices. An independent side-by-side capture would be a harder, complementary test of biomechanical equivalence. Read everything below as a strong test of electronic device-pair equivalence, and a softer test of biomechanical equivalence.
- Caveat in principle: inertial bias at impacts. The PlateMate sits between the subject and the Bertec as a mechanical element with mass. During the high-acceleration landing transient,
m_PM × a_PMis a force the Bertec sees but the PlateMate does not. In practice on this recording, the gap is small (~−2 % raw, collapsing to ~+0.25 % under 5 Hz low-pass), and it reverses sign under low-pass filtering — which a pure inertial bias would not. So inertial loading is at most a small contributor to the residual gap, not its dominant cause. This is examined in the filtered-RMSE section below.
Data and alignment
PlateMate sampled at ~960 Hz, Bertec at 1000 Hz. Each device's vertical force was converted to kg, resampled onto a common 1000 Hz timeline, and time-aligned to the other via cross-correlation with sub-sample (3-point parabolic) refinement so each sample compares the two devices at the same point in time.
Cross-correlation offset: 1670.484 ms
Aligned recording: 2777 samples at 1000 Hz, 2.78 s of overlapVisual comparison
A direct overlay of the two synchronised force traces — full recording on top, zoomed view around the landing peak below.
Bulk-metric agreement
How closely do the headline numbers — peak force, mean force, body weight, point-to-point RMSE — agree between the two devices?
============================================================
BULK-METRIC COMPARISON
============================================================
Pearson correlation: 0.998998
Time-domain RMSE: 2.92 kg (3.78 % of mean force)
Mean difference (BA bias): +0.15 kg
Bland–Altman 95 % limits: ±5.72 kg
Peak force:
PlateMate: 582.84 kg
Bertec/Hawkin: 595.01 kg
Difference: -12.17 kg (-2.04 %)
Findings — bulk-metric agreement:
| Measure | Value |
|---|---|
| Pearson correlation (point-to-point) | 0.9990 |
| Peak force agreement | −2.04 % (582.84 vs 595.01 kg) |
| Mean force agreement | +0.19 % (essentially identical) |
| Time-domain RMSE | 2.92 kg (3.78 % of mean force) |
| Bland–Altman bias | +0.15 kg |
| Bland–Altman 95 % limits | ±5.72 kg |
PlateMate and Bertec agree to sub-percent on mean force and Pearson-r is essentially 1. The only number that looks high is the time-domain RMSE (~3.8 %), which is dominated by the −2.04 % peak gap at the landing impact. The frequency-domain section below pinpoints where in the spectrum that residual lives, and the filtered-RMSE section after that quantifies how much of it survives once impact-tip transients are removed.
Frequency-domain comparison
The bulk RMSE looks high (~3.8 %), but its location in the signal — the impact transient — is informative. Two views:
- Power Spectral Density (PSD) on the brief pre-jump quiet stance shows each device's combined response to a standing subject. Caveat: with a subject standing on the plate, this measures combined biological postural sway plus electronic noise — we cannot separate the two from this dataset, so noise numbers are upper bounds.
- Magnitude-squared coherence broken into frequency bands and weighted by signal energy shows where the two signals are linearly correlated. Coherence is computed on a ±1 s jump-only window with
nperseg=512. ====================================================================== Body-weight quiet-stance: combined biological + electronic response PlateMate: SD=0.7389 kg, mean=77.39 kg
Bertec: SD=0.7261 kg, mean=77.31 kg
Body-weight delta: +82 g (+0.11%) ====================================================================== Jump-only coherence by frequency band Overall 0–30 Hz power-weighted coherence: 1.0000 Band Energy % Coherence (pwr-wtd) 0–2 Hz 44.40% 1.0000
2–5 Hz 16.82% 1.0000
5–10 Hz 12.13% 1.0000
10–30 Hz 26.66% 1.0000
Findings — frequency domain:
Body-weight match: ~77.4 (PM) vs ~77.3 kg (Bertec) — an ~80 g difference (~0.11 %). Note: this number reflects both raw-signal sensing AND how each device is calibrated. The agreement tells us PlateMate's CC Athletics calibration and Bertec's calibration produce mutually consistent results to a fraction of a percent.
Coherence is saturated at 1.0 in every band of the movement signal. Power-weighted across 0–30 Hz, coherence is 1.0000; broken down by band, it stays at 1.0000 in 0–2 Hz, 2–5 Hz, 5–10 Hz, and 10–30 Hz. The two devices agree on the same signal at every frequency that matters for jump biomechanics. This is the strongest possible result the comparison instrument can produce.
Quiet-stance noise: SDs differ by ~2 % (0.74 vs 0.73 kg). Important caveat — with a standing subject, this combined response is dominated by biological postural sway, not by either device's intrinsic noise. We cannot separate biological from electronic contributions in this dataset; a static-mass loading test would be needed to isolate device noise floors.
Validating with filtered RMSE — and the peak-after-LP test
Two complementary low-pass tests, each answering a different question:
A. Filtered RMSE. If most of the residual disagreement lives above 5 Hz (consistent with the saturated-coherence frequency picture), low-pass filtering both signals at progressively lower cutoffs should drop the RMSE sharply between 30 Hz and 5 Hz cutoffs.
B. Filtered peak force. This is the new test that disambiguates the −2.04 % raw peak gap. If the peak gap is a device-level filtering or calibration bias, low-pass filtering both signals equally should leave the peak agreement roughly unchanged. If the gap is a sample-period timing-edge effect at the 1 ms-wide impact tip (or partly an inertial bias from the upper plate's mass × acceleration during impact), the gap should collapse — possibly even reverse sign — as we smooth out the impact transient.
A 4th-order zero-phase Butterworth (scipy.signal.filtfilt) is applied at cutoffs 30 Hz, 10 Hz, and 5 Hz. The trim is 0.3 s from each end (the ~2.78 s overlap is too short for the more conservative 2 s trim used in longer recordings, but 0.3 s comfortably outlasts the filter settling time).
==============================================================================
A. FILTERED RMSE
==============================================================================
Cutoff RMSE (kg) RMSE (% of mean) vs raw
------------------------------------------------------------------------------
Raw (no filter) 3.297 4.27%
Low-pass 30 Hz 1.363 1.76% 58.7% reduction
Low-pass 10 Hz 0.514 0.67% 84.4% reduction
Low-pass 5 Hz 0.415 0.54% 87.4% reduction
==============================================================================
B. FILTERED PEAK FORCE (separates HF disagreement from device-level bias)
==============================================================================
Cutoff PM peak (kg) Bertec peak (kg) Diff (kg) Diff (%)
------------------------------------------------------------------------------
Raw (no filter) 582.84 595.01 -12.17 -2.04%
Low-pass 30 Hz 398.62 400.24 -1.62 -0.41%
Low-pass 10 Hz 229.45 228.85 +0.60 +0.26%
Low-pass 5 Hz 213.02 212.50 +0.52 +0.25%
Findings — filtered RMSE and peak:
| Cutoff | RMSE (kg) | RMSE (% of mean) | Peak diff (kg) | Peak diff (%) |
|---|---|---|---|---|
| Raw | 3.30 | 4.27 % | −12.17 | −2.04 % |
| Low-pass 30 Hz | 1.36 | 1.76 % | −1.62 | −0.41 % |
| Low-pass 10 Hz | 0.51 | 0.67 % | +0.60 | +0.26 % |
| Low-pass 5 Hz | 0.42 | 0.54 % | +0.52 | +0.25 % |
The peak-after-LP column is the cleanest piece of evidence in the entire blog. The −2.04 % raw peak gap collapses to −0.41 % at 30 Hz LP, and even reverses sign under more aggressive filtering. This is exactly the signature of a sample-period timing-edge effect at the 1 ms-wide impact tip — small phase misalignments at the high-RFD landing peak look like amplitude misalignments when you inspect the raw maximum, but disappear once the impact transient is smoothed. A device-level filtering bias or a constant inertial offset would have produced a stable peak gap across cutoffs. Instead the gap drops by ~80 % and reverses, so the dominant cause is timing-edge — not a real device-level disagreement on the underlying force trajectory.
The filtered RMSE tells the same story from a different angle. Raw RMSE 3.30 kg → 0.42 kg under 5 Hz low-pass — an 87 % reduction. By the 5 Hz cutoff, the residual is comparable to the quiet-stance SD (~0.74 kg), i.e. the two devices disagree on the jump signal at roughly the same level they each disagree with their own quiet-stance baseline. That's about as close to bit-for-bit identical, modulo sample-level discretisation and ordinary measurement noise as two independent devices can be expected to come.
Biomechanical metric translation
The frequency-domain and filtered-RMSE results say the devices agree very well at biomechanical timescales. But the user-facing question is: "would my downstream metrics — jump height, peak power, takeoff velocity — differ between devices?"
This section computes those metrics from each device's signal independently using a simple impulse-momentum implementation, then reports the disagreement directly. The same algorithm runs on both signals so any discrepancy is signal-driven, not algorithm-driven.
================================================================================
BIOMECHANICAL METRICS computed independently from each device's signal
================================================================================
Metric PlateMate Bertec Diff %
--------------------------------------------------------------------------------
Body weight (N) 760.563 759.740 +0.823 +0.11%
Takeoff velocity (m/s) 2.6377 2.6252 +0.013 +0.48%
Jump height (mm) 354.62 351.26 +3.363 +0.96%
Peak power (W) 3954.9 3936.2 +18.697 +0.48%
Propulsive impulse (N·s) 323.52 322.58 +0.935 +0.29%
================================================================================
Sanity check: our Bertec-derived jump height vs Hawkin's pre-computed value
================================================================================
Our calculation: 351.3 mm
Hawkin reports: 350.2 mm
Agreement: +0.30 % (Δ +1.1 mm)Findings — biomechanical metrics:
| Metric | PlateMate | Bertec | Difference | % |
|---|---|---|---|---|
| Body weight | 760.6 N | 759.7 N | +0.82 N | +0.11 % |
| Takeoff velocity | 2.6377 m/s | 2.6252 m/s | +0.0125 m/s | +0.48 % |
| Jump height | 354.6 mm | 351.3 mm | +3.4 mm | +0.96 % |
| Peak power | 3954.9 W | 3936.2 W | +18.7 W | +0.48 % |
| Propulsive impulse | 323.5 N·s | 322.6 N·s | +0.9 N·s | +0.29 % |
Every metric agrees within ~1 %, most under 0.5 %. The −2 % raw peak-force discrepancy that loomed large earlier doesn't propagate to any metric a coach or clinician would actually report. Jump height — the most commonly reported CMJ metric — agrees to within 1 % (3.4 mm out of 351 mm).
The metrics agree so much better than the raw peak force because they integrate force over a window much longer than 1 ms — propulsive impulse spans hundreds of milliseconds, takeoff velocity is its time-integral, jump height is its squared mapping, peak power averages force × velocity over the full propulsive phase. Time-integration averages out exactly the kind of sample-period timing-edge noise that dominates the raw peak comparison.
Sanity check. Our simple impulse-momentum jump-height calculation on the Bertec signal gives 351.3 mm. Hawkin Dynamic's own pre-computed metric for the same recording reports 350.2 mm — agreement within 0.3 % (1.1 mm). This validates our algorithm and lets us trust the cross-device comparison.
What this means in practice (plain-language summary)
This section translates the analysis above into practical terms, for readers who use force plates day-to-day rather than analyse raw signals.
What we tested
One person performed one countermovement jump on a Bertec force plate (processed by Hawkin Dynamic software) with a CC Athletics PlateMate placed on top. Both devices recorded the same physical event simultaneously through a stacked configuration where the same vertical force flowed through both plates in series. We compared their raw measurements — the force readings every millisecond — and the biomechanical metrics each device's signal would produce.
What we found, in plain language
Sub-percent agreement on the metrics that matter for assessment:
- Body weight: the two devices agree to within ~80 grams (out of ~77 kg). About the weight of a small egg.
- Jump height: within ~1 % (354.6 mm vs 351.3 mm — a difference of 3.4 mm, less than the thickness of two stacked credit cards).
- Takeoff velocity, peak power, propulsive impulse: all within 0.5 %.
- Pointwise force-time agreement: ~2.9 kg RMSE on the raw signal (~3.8 % of mean force), dropping to ~0.4 kg with mild low-pass filtering — about the same magnitude as quiet-stance measurement noise.
On the one numerical discrepancy that does exist:
The raw landing-peak force differs by 2 %. We pinpointed where it comes from by comparing peak force after low-pass filtering: at a 30 Hz cutoff the gap drops to 0.4 %, and at a 5 Hz cutoff it actually reverses sign and stabilises at 0.25 % in the opposite direction. A device-level filtering or calibration difference would have produced a steady offset across all cutoffs; instead the gap collapses and reverses. That's the signature of a sample-period timing-edge effect at the very tip of the 1 ms-wide impact transient — i.e., the two devices report the same underlying force trajectory, but the peak instant is sampled at slightly different sub-millisecond positions on a near-vertical rising edge. Once you smooth out the impact tip, the peaks agree to a fraction of a percent. None of this propagates to integrated metrics like jump height or peak power, which are calculated over windows long enough that sub-millisecond effects average out.
A note on the calibration numbers: the body-weight, jump-height, and takeoff-velocity agreement is partly a reflection of how each device is calibrated. Both PlateMate and Bertec are individually calibrated against known reference weights — PlateMate's calibration is performed in-house at CC Athletics, Bertec uses its own procedure. The agreement tells you that the two calibration procedures produce results consistent with each other to a fraction of a percent.
What this means for your practice
For CMJ-based assessments: the standard headline numbers a clinician would read off a force-time trace — peak force, jump height, peak power, takeoff velocity, propulsive impulse — agree between PlateMate and Bertec/Hawkin to within 1 % on this recording. The choice between the two is unlikely to change your judgement.
For tracking an athlete over time: stick with the same device throughout. Either device is consistent enough on its own; mixing measurements between devices can introduce small differences (sub-percent on integrated metrics, a couple of percent on the raw landing-peak instant) that don't reflect real change in the athlete.
What this analysis does not tell you
- One recording, one subject, one CMJ. The agreement we found is consistent with what published force-plate validation studies typically report for stacked-plate captures, but a larger sample would strengthen the conclusion.
- Stacked configuration only. A side-by-side comparison (subject jumping with one foot per device) would expose biomechanical-equivalence issues — biological asymmetry, foot placement, impact mechanics — that this configuration cannot test. The numbers above describe electronic device-pair equivalence, which is a strong but narrower claim.
- Device noise floors in isolation. The quiet-stance comparison combines biological postural sway and electronic noise; we cannot separate them without a static-mass loading test (no human on the plate).
- Bilateral asymmetry, centre of pressure, and other derived metrics are not covered here — only summed total vertical force and metrics integrated from it.
- Absolute calibration accuracy isn't tested. We measured agreement between the two devices, not their accuracy against a traceable reference mass.
- Drop jumps and harder impacts. This was a CMJ. A drop-jump or landing-onto-a-rigid-surface event would have sharper transients and more high-frequency content; we don't know from this recording how that would shift the agreement story.
Bottom line
For CMJ-based assessment, this comparison supports treating PlateMate as practically equivalent to a Hawkin-processed Bertec for the headline metrics that drive clinical decision-making. Body weight, jump height, takeoff velocity, peak power, and propulsive impulse all agree to under 1 % on this recording. The raw signals are essentially indistinguishable in the biomechanical band (0–30 Hz coherence saturated at 1.0; amplitude ratios within ±1 %; phase < 4°), and the small raw-peak discrepancy that does exist is an impact-tip sampling effect that doesn't survive into the integrated metrics practitioners actually use.