The most consequential readings may be the ones not recorded

Imagine a study tracking skin conductance or heart-rate features during simulated trading decisions. The researcher notices that several minutes of data are absent around the most demanding tasks. Participants might have moved their wrists, adjusted electrodes, paused the device or removed it because the session became uncomfortable. If researchers simply delete incomplete records and average the remainder, the apparent physiology of the group could describe the easiest-to-record moments rather than the moments of greatest interest. This is not proof that the missing segments contain higher stress. It is a reason to study why the observations were unavailable.

The missing-data problem differs from ordinary random noise. Noise makes an observed value uncertain; missingness determines which participants and times remain in the analysis at all. When absence depends on activity, fatigue, device contact or an outcome relevant to the hypothesis, a complete-case analysis may distort both the level and relationship of variables. The critical research question is therefore not merely how many observations are missing. It is whether the process generating missingness is related to the quantities the study is trying to estimate.

What the methodological evidence established

Roderick Little developed a statistical test of the hypothesis that multivariate values are missing completely at random [1]. The work provides a formal way to examine one restrictive missingness assumption under specified conditions. Passing such a test does not prove that every missing-data mechanism is harmless, and failing it does not automatically identify why records are missing. The distinction matters because a wearable study might have missingness related to observed activity logs, unobserved distress or both. Each possibility places different demands on inference and sensitivity analysis. Collins, Schafer and Kam used simulation to compare how including auxiliary variables can affect missing-data methods [2]. They argued for an inclusive strategy when suitable additional information can improve the handling of missingness. In an NSI-style protocol, such variables might include device wear status, motion flags, time of day, contact-quality measures and task segment. Adding variables is not a substitute for accurate measurement or a guarantee of unbiased results, but it can make assumptions more plausible and reveal what differs between complete and incomplete recordings.

Enders and Bandalos performed Monte Carlo simulations comparing full-information maximum likelihood estimation with common deletion and imputation strategies under the conditions they examined [3]. Their results illustrate why listwise deletion is not a universal default for incomplete behavioral datasets. Yet methods that perform well when missingness is ignorable may still fail if participants stop recording specifically because of an unmeasured state related to the outcome. A research article should identify the assumed mechanism, describe the estimator and explain what would happen if that assumption were false.

The difference between missing completely at random and missing not at random

Three common categories provide a vocabulary for analysis. Missing completely at random refers to absence unrelated to observed or unobserved values within the relevant model. Missing at random allows missingness to depend on variables that have been observed and included appropriately. Missing not at random refers to missingness related to values that remain unobserved even after available conditioning. These labels describe assumptions, not qualities that can be inferred with certainty from a simple percentage of lost records. A quality dashboard saying that 92 percent of signals were present would not tell us which mechanism applies.

A well-documented recording system should maintain an event log: device removed, contact interrupted, segment corrupted, participant withdrew, software failed or collection was intentionally paused. The reasons should not be inferred after looking at the desired association. Report completeness by person, time, task and signal quality rather than only as a total. Examine whether missingness is more common during high-demand conditions, and publish sensitivity analyses showing how results change under plausible alternative assumptions. The article Temporal Leakage warns of a different hazard: letting unavailable future information contaminate historical model training.

What a trustworthy NSI study would report

Before data collection, specify primary outcomes and a missing-data plan. State which variables will be used to explain missingness, how imputation or likelihood methods will be evaluated, when a session is too incomplete for a declared analysis and which conclusions depend on assumptions that cannot be tested directly. Researchers should retain an untouched record of the original gaps. Imputed values are estimates with uncertainty, not secretly recovered physiology. If a model is trained on filled-in data, the evaluation pipeline must prevent information leakage between training and held-out sessions.

The next test is whether conclusions survive plausible assumptions about data that were never observed. If a purported physiological effect disappears when difficult sessions are given greater weight or when attrition is modeled differently, the original conclusion should be weakened rather than promoted. Cross-Participant Validation is relevant because exclusions can also change which people reach the test set. The resulting lesson is simple but demanding: absence is part of the evidence-generation process. NSI is a proposed interdisciplinary research framework, not an independently validated medical or financial predictor. Accurate handling of missing observations is a necessary methodological condition, not proof that a nervous-system indicator works.

Scientific bibliography & sources

  1. [1] Little RJA. A test of missing completely at random for multivariate data with missing values. Journal of the American Statistical Association (1988). https://doi.org/10.1080/01621459.1988.10478722
  2. [2] Collins LM, Schafer JL, Kam CM. A comparison of inclusive and restrictive strategies in modern missing data procedures. Psychological Methods (2001). https://pubmed.ncbi.nlm.nih.gov/11778676/
  3. [3] Enders CK, Bandalos DL. The relative performance of full information maximum likelihood estimation for missing data in structural equation models. Structural Equation Modeling (2001). https://doi.org/10.1207/S15328007SEM0803_5

Research note

NSI is a proposed research framework, not an independently validated medical or financial prediction system. This article is research and education, not individualized medical or investment advice. Market observations are timestamped historical context, not live prices.