Studies of the long-term effects of repetitive head impacts usually learn who was exposed by asking people decades later. In the most precise recent work, older adults complete a tablet-based reaction-time battery, and their impact history comes from a questionnaire about past sports and occupations. If some unexposed people say they were exposed, and some exposed people say they were not, the "exposed" group is diluted. The measured cognitive cost of head impacts is then wrong, usually too small. This note proposes to estimate that misreporting from the cognitive data itself. People truly exposed to head impacts show a distinctive pattern over repeated days of testing: slower executive responses, more day-to-day wobble, and less improvement with practice. That multi-part signature can serve as a set of noisy witnesses against which the self-report is checked. Both the true exposure and the rates of false "yes" and false "no" answers become estimable without any gold-standard records.
The core statistical move is a latent-class measurement model, meaning a model in which true exposure is a hidden yes/no variable and every observed measure is a noisy readout of it. It allows misreporting in both directions. That directly relaxes the "no false positives" assumption of the closest existing method. The secondary ingredient comes from surgical navigation. When the view is narrow and frames are noisy, estimating a hidden state with a temporal model beats trusting single snapshots. We borrow that idea to read each person's multi-day test trajectory rather than one session mean, and we test explicitly whether it earns its place.
Core Claim
Retrospective head-impact exposure can be corrected for misreporting in both directions. The repeated-testing dynamics of a digital cognitive battery supply several conditionally independent indicators of true exposure. Tasks known to be unaffected by head impacts supply a negative control that removes general slowing, so it is not mistaken for exposure. The corrected exposure effect on executive response time should be larger than the naive effect. It should be null on the control tasks. The estimated over-reporting rate should be measurably above zero. If it is not, the existing one-directional method is enough.
What the Prior Work Gives Us
Two studies on the Boston University Alzheimer's Disease Research Center cohort motivate the problem. Both used the Defense Automated Neurobehavioral Assessment (DANA, a six-task tablet battery). Digital Neuropsychological Measures in Older Adults Exposed to Repetitive Head Impacts found slower Code Substitution (2091.26 vs 1915.36 ms, P = 0.017) and Procedural Response Time (739.55 vs 688.69 ms, P = 0.026) in 97 cognitively intact adults, 22 of them labeled exposed. It found no difference on Simple Response Time or Match-to-Sample. Exploring the Effect of Repetitive Head Impacts on the Stability of Digital Cognitive Measures in Cognitively Intact Older Adults: A Pilot Study studied 58 adults, 10 exposed. Exposed participants had greater day-to-day variability on Go/No-Go, Code Substitution and Simple Response Time, and they gained less from practice on Procedural Response Time. Both studies classify exposure from recalled history under the 2021 traumatic encephalopathy syndrome criteria and name recall bias as a limitation, but neither models it. Our proposal treats their labels as one fallible indicator rather than ground truth. We also reuse their finding that some tasks are unaffected and turn those tasks into a negative control.
The method we extend is Learning Models from Data with Measurement Error: Tackling Underreporting. It proves that a binary exposure reported only by under-reporting can be identified from a single error-prone source when exposure depends nonlinearly on covariates. On the Boston Birth Cohort it recovered smoking under-reporting of 15.7% to 20.2% and a heroin–obesity risk difference of 0.227. That work constrains our design: its identification rests on "no false positives," which is exactly what we drop. We replace curvature-based identification with identification from multiple outcome indicators. This needs repeated measurements but tolerates errors in both directions.
Sports-Related Concussion: Advances in Diagnosis, Biomarkers, and Long-Term Outcomes documents that concussion epidemiology relies on recalled histories and is shaped by selection. Examples include the gap between elite-athlete risk and the null community PROTECT-TBI result, and brain-bank selection. It calls for objective measures but offers no correction model. We supply one for the retrospective setting where sensors cannot reach back in time.
From surgical navigation, A Combined EM and Visual Tracking Probabilistic Model for Robust Mosaicking: Application to Fetoscopy showed something specific. Frame-to-frame registration spikes to 50-pixel errors and drifts linearly. A square-root unscented Kalman filter, a nonlinear filter that tracks a hidden state through time, stays under 10 pixels. The analogy we take is narrow. A single DANA session is a noisy "frame," and practice effects are a systematic drift that differs by exposure. A temporal model of each person's trajectory should separate level, practice slope and variability better than per-session averages. We do not claim the surgical machinery transfers wholesale. The ablation below decides whether it helps at all.
The Model
Let $A_i \in \{0,1\}$ be person $i$'s true exposure and $\tilde A_i$ the self-report. The over-report rate is $\alpha = P(\tilde A=1 \mid A=0)$ and the under-report rate is $\beta = P(\tilde A=0 \mid A=1)$. Response time on an executive task on testing day $t$ is modeled as
$$ y_{it} = \mu_{A_i} + \lambda g_i - \gamma_{A_i}\log t + \epsilon_{it}, \qquad \epsilon_{it}\sim \mathcal N(0,\sigma^2_{A_i}), $$
where $\mu_A$ is the exposure-specific level, $\gamma_A$ the practice gain, and $\sigma^2_A$ the day-to-day variability. The term $g_i$ is a general-speed factor estimated only from the negative-control tasks, Simple Response Time and Match-to-Sample, and $\lambda$ is its loading. The logarithmic practice curve is a standard diminishing-returns form, and we will compare it against a free-form alternative.
Conditional on $A_i$ and $g_i$, the level, the practice slope, the variability and the self-report are treated as separate witnesses. With three or more such witnesses, the classical latent-class argument identifies the class-specific parameters together with $\alpha$ and $\beta$, up to label swapping. We fix the labels by requiring the exposed class to be slower. Conditional independence is the weakest assumption, because tasks share general speed. Conditioning on $g_i$ is our defense, and its adequacy is tested directly through the negative-control prediction. Fitting is by full Bayesian inference, which keeps uncertainty honest given the very small number of exposed participants.
Research Program
The program has four stages. The first is a simulation study calibrated to the reported effect sizes and sample sizes. It will map when $\alpha$ and $\beta$ are recoverable as a function of cohort size, days of testing and true misreport rates. With 10 exposed people in the pilot, we expect wide posteriors, and the study will say how many participants and testing days are needed. The second is a reduction check: with $\alpha$ fixed at zero, our model must reproduce the one-directional method's behavior on its own synthetic designs. The third is a semi-synthetic test on the real DANA data. We deliberately flip known fractions of self-report labels in each direction and check whether the model recovers the flip rates. The fourth is the substantive analysis: corrected exposure effects and misreport rates on the full BU ADRC DANA data.
Baselines
The naive reference point is the analysis in the original DANA papers. It compares mean response times between self-reported groups and takes the labels as truth. The strongest reported result is Code Substitution at 2091.26 vs 1915.36 ms (P = 0.017) and Procedural Response Time at 739.55 vs 688.69 ms (P = 0.026), from Digital Neuropsychological Measures in Older Adults Exposed to Repetitive Head Impacts. We rerun it on identical participants and sessions.
The stronger conventional competitor is a linear mixed-effects growth model. It has a random intercept and slope per person and self-reported exposure as a covariate. This is what a careful analyst would fit to multi-day data, and it already uses trajectories, so it isolates whether misclassification correction matters beyond trajectory modeling.
The strongest existing correction method is the one-directional latent exposure model of Learning Models from Data with Measurement Error: Tackling Underreporting. It is identified through the same demographic covariates (age, education, sex). No reported result was found on Lacuna for this method applied to head-impact exposure. Its only reported figures are the Boston Birth Cohort smoking under-reporting of 15.7% to 20.2% and the risk difference of 0.227.
Two ablations isolate our ingredients. The first removes the temporal trajectory and uses per-person session means only. The second removes the negative-control speed factor $g_i$. All methods get the same participants, covariates and priors where applicable, and a fixed number of posterior samples or optimizer restarts. Beating the competitors means two things in the semi-synthetic test: lower error in recovered flip rates and recovered exposure effect, with calibrated intervals. On real data it means satisfying the negative-control prediction where competitors do not.
Why not just…?
Why not just keep the self-reported labels and run a sensitivity analysis over assumed misreport rates? On the Boston Birth Cohort, Learning Models from Data with Measurement Error: Tackling Underreporting found that grid sensitivity analysis produced intervals too wide to decide whether the heroin effect was nonzero. The latent model gave a tight 0.227. A grid also fixes the misreport rates rather than estimating them. Our semi-synthetic flips test this directly.
Why not just apply the existing under-reporting model as is? It may be enough. If retrospective recall in older adults errs almost only toward forgetting, $\alpha$ will be near zero. The one-directional model is then simpler and identified with less data, and we would recommend it. Our model nests it, so a posterior for $\alpha$ concentrated near zero on real data would refute the bidirectionality claim. We treat that as a live possibility, not a strawman. Over-reporting by people who already notice cognitive slippage and search their past for a cause is plausible but unproven.
Why not just obtain objective records, such as rosters or service files, for a validation subsample? Where such records exist, standard validation-sample correction is simpler and more trustworthy, and it is a real threat to this proposal's usefulness. For community-dwelling older adults, records of youth contact sport or occupational impacts mostly do not exist. That is the regime the method targets. Wherever records exist for any participants, we will use them only as held-out ground truth for scoring estimated exposure, which also tests our model against the validation-sample approach.
Why not just test more days and average? Averaging shrinks session noise but does nothing to misclassified labels. The mixed-model baseline already captures that benefit, so any gain beyond it is attributable to the correction.
Decisive Checks
The first check is the negative control. After correction, the exposure effect on Simple Response Time and Match-to-Sample should be null, and the effect on Code Substitution and Procedural Response Time should be at least as large as the naive estimate. Misclassification that does not depend on cognitive state attenuates effects, so correction should enlarge them. A corrected effect that appears on the control tasks would mean the model is relabeling generally slow people as exposed, and the approach fails.
The second check is semi-synthetic recovery. Flipped label rates of 10–20% in each direction must be recovered within posterior intervals of reasonable width at the 97-person cohort size. The simulation will state in advance the minimum detectable $\alpha$.
The third check is the reduction. With $\alpha$ fixed at zero, our model must match the one-directional method on its synthetic settings.
The fourth check is the trajectory ablation. If session means perform as well as the trajectory model, we drop the surgical-filtering analogy and report that the misreporting correction alone carries the result.
References
- Digital Neuropsychological Measures in Older Adults Exposed to Repetitive Head Impacts
- Exploring the Effect of Repetitive Head Impacts on the Stability of Digital Cognitive Measures in Cognitively Intact Older Adults: A Pilot Study
- Learning Models from Data with Measurement Error: Tackling Underreporting
- Sports-Related Concussion: Advances in Diagnosis, Biomarkers, and Long-Term Outcomes
- A Combined EM and Visual Tracking Probabilistic Model for Robust Mosaicking: Application to Fetoscopy