Technical detail · Missing-Data Repair Filter Contracts
Under the Hood: MCAR, MAR, MNAR
MCAR, missing completely at random, means missingness is unrelated to the value or other variables, as with a random packet collision. Missing points resemble observed ones, so simple imputation faces less selection bias; it still does not recover each missing measurement.
MAR, missing at random, means observed variables explain missingness—for example, a gateway overload flag, queue length, or low battery reading. A model using those variables and correlated sensors can produce useful estimates with little bias when its assumptions hold.
MNAR, missing not at random, means missingness depends on the unobserved value itself. A pressure sensor that saturates and stops reporting at high pressure selectively hides extremes. Mean fill or interpolation from normal readings can erase them.
Suppose an incubator temperature sensor reports every minute and usually sits near 37 degrees C. In an MCAR test, randomly hide 20 samples and interpolate from neighbours: expected selection bias is small because the hidden points resemble visible ones. For MAR, hide samples when gateway queue length is high; a model including queue length and neighbouring sensors can still recover useful estimates. For MNAR, hide samples only above 39 degrees C. Interpolation from normal neighbours systematically replaces overheating evidence with ordinary values.
The distinction changes the repair decision. Random loss does not systematically omit the high tail, but a constant mean fill still cannot reconstruct short-lived peaks. With value-dependent dropout, the missing interval may be the evidence that an alarm needs. Record the observed evidence behind the assumed mechanism before approving automation.