Sensors yield only proxy measures of context
Aliases: operationalization · proxy variables
What it is
No sensor measures "context" directly. Sensors measure physical quantities — infrared radiation, acceleration, reed-switch closure, Wi-Fi signal strength; context — at home, asleep, has company — is an inferred construct sitting above those quantities. The mapping between the two is a proxy measure: a PIR sensor "detecting motion" is a proxy for "someone is home"; a wristband reporting "low heart rate + stillness" is a proxy for "asleep".
Take "proxy" literally: it stands in for the construct without being the construct. Every proxy has failure scenarios — PIR is blind to a motionless person, heart rate misreads an anxious person lying on their side. Most context-system error is not broken hardware but the proxy relation not holding in that scenario.
Why it happens
Two layers make proxies inevitable and fallible:
- A behavioural assumption separates construct from signal. "Motion → someone present" assumes "people move when in a room". That holds for someone walking, not for someone reading, napping or meditating. A proxy's validity equals the coverage of its hidden behavioural assumption, and human behavioural variety guarantees counterexamples. Sensor fusion (door contact + PIR + electricity) mitigates but does not eliminate: each sensor carries a different assumption, and fusion assumes "the counterexamples do not coincide" — yet another breakable assumption.
- The constructs themselves are socially defined. Sleep is physiologically a continuum (light, deep, lying still with eyes closed); binarising it into asleep/awake is an engineering decision, not a physical fact. The other half of proxy failure comes precisely from forced discretisation: the boundary band (the twenty minutes of falling asleep) accumulates systematic misjudgements — error concentrates where definitions blur, which is a definitional problem, not a precision problem.
The corollary is the correct design posture: treat every proxy as shipping with a failure-condition list, and hold "when this proxy goes blind" to the same level of specification as "what it measures".
Studying it
- Validity studies: with self-report or human annotation as ground truth, compute sensitivity and specificity per proxy. Activity recognition has mature paradigms — public datasets such as Opportunity and WISDM score sensor combinations against human-annotated activity labels.
- Failure-scenario mining: deliberately collect edge cases (motionless occupants, pet triggers, overlapping occupants) and quantify how error distributes across scenarios — average accuracy conceals failure concentration; per-scenario error reporting is what informs design.
- Field contrasts: proxy validity measured in staged lab settings differs sharply from real homes (pets, occlusion, domestic habits); re-calibration in the field is a standing finding.
One methodological caution: validity is not a property of the sensor but of sensor × scenario × population. The same sensor's validity can differ entirely across households; report validity together with its scenario conditions or not at all.
Where it stops holding
- Proxy failure is systematic, not random — it biases toward particular people and scenarios. Passive infrared drifts for people with low skin temperature (older adults) or high (fever); motion recognition fails outright for wheelchair users. The error distribution is uneven, which makes this an equity question.
- Fusion has diminishing returns while exposure grows linearly. The third and fourth sensors usually add less confidence than exposure — the trade-off between proxy quality and privacy cost must be evaluated explicitly per fusion design.
- High-confidence proxies exist. A reed switch on "the door opened" or GPS on "outdoors" are near-direct measurements — indirectness is a spectrum. Trust should be tiered by how many steps separate each proxy's physical quantity from its construct.
Applying it
- Maintain a proxy spec sheet per sensing channel: physical quantity measured, construct it stands for, hidden behavioural assumption, known failure scenarios, and failure mode (miss or false alarm). This sheet is the risk capital for designing automations.
- The heavier the consequence, the shorter the proxy chain required (physical quantity one step from construct) or the more independent proxies must agree; a single long-chain proxy ("Wi-Fi dropped → the user left home") must never trigger a high-consequence action.
- Turn known failure scenarios into explicit design constraints: a PIR room sensor must compensate for the stillness blind spot (periodic ultrasound, a camera, or extended timeouts) — treating "no motion" as fact is a design error, not a sensor one.
- How to check: after deployment, sample weeks of self-reported events (home / asleep / company) against corresponding sensor readings, computing misses and false alarms per scenario; for concentrated failure scenarios, decide proxy swap, redundancy, or a more conservative trigger — and if none is answerable, that judgement should not be automated.