The Sensor Network That Disagrees With Itself
A school science project monitors classroom air quality with three identical low-cost particulate sensors mounted at different heights. Over a week, sensor A on the teacher's desk reads a mean of 38 µg/m³, sensor B near the ceiling reads 21 µg/m³, and sensor C by the door reads 55 µg/m³. The students conclude that the door is a pollution source. Their teacher notices the sensors were calibrated once at purchase, the room is heated by a radiator under the window, and the door opens onto a corridor where students congregate at break. The data logger records one reading per minute but the sensors were swapped between positions halfway through the week.
1. Evaluate the students' conclusion and identify the confounders that must be ruled out.
Read the worked answer
The conclusion that the door is a pollution source is premature and probably confounded. Three issues undermine it:
- Position confounders. Height alone changes readings: warm air rises from the radiator, so the ceiling sensor may sit in cleaner, stratified air while the door sensor samples the breathing zone and incoming corridor air. The difference may reflect location, not source strength.
- Instrument confounders. The sensors were calibrated only once and are not shown to agree with each other. Without a co-location check, a 17 µg/m³ gap between units could be sensor bias rather than real air.
- Design flaw. Swapping sensors midway is good practice for separating sensor from position, but only if the swap is analysed. If the door position reads high regardless of which unit is there, position matters; if one unit always reads high, the sensor is at fault.
The students need co-location data and a proper before/after comparison before attributing the reading to the corridor.
2. Propose an analysis and a follow-up experiment that would settle the question.
Read the worked answer
Analysis: Treat the week as two periods (before and after the swap). For each sensor, compare its mean in each position. A two-way breakdown - sensor identity versus position - separates instrument bias from location effect. If sensor C reads high in every position, it is biased; if the door position reads high for every sensor, location is implicated.
Follow-up experiment:
- Co-locate all three sensors in one spot for 24 hours and compute pairwise offsets; apply correction factors.
- Then run a controlled comparison: keep sensors fixed and measure during break versus non-break periods, with the door closed and open on matched days.
- Log door-open duration and corridor occupancy so the exposure variable is measured, not assumed.
- Repeat over several days to average out weather and heating cycles.
Settling the question requires demonstrating that readings change with the suspected cause while other factors are held constant, which is the core logic of a controlled comparison.