What an Outlier Is and What to Do With It
Why this matters
Somebody in the shop is going to look at a set of readings, see one that does not fit, and quietly leave it out. It feels like tidying. It is the single most consequential thing anyone does to a data set, because the point that does not fit is disproportionately likely to be the one carrying the fault, and once it is gone nobody can tell it was ever there.
The word "outlier" sounds like a property of a number. It is not. It is a claim about a population: this point does not belong with those points. That claim can only be made once you have said which population you meant, and most field arguments about outliers are actually unstated disagreements about the population. The case below is one route, twelve units, and three points that each needed a decision, and the decision is different for every one of them.
The call
A commercial customer with twelve nominally identical rooftop units, one site, annual preventive maintenance, one tech doing all twelve in two days. Part of the routine is a supply fan current reading on each unit, recorded against last year's.
This year's numbers come in with a tight cluster and two points sitting outside it. The tech's instinct is to note the two odd ones as suspect units and move on. That instinct got both of them wrong and nearly missed a third point that was not odd at all.
Before the readings: the two hazards the route creates
Both hazards here come from the instruction, not from the equipment misbehaving.
Twelve rooftop units means twelve trips near a roof edge. In general industry, 29 CFR 1910.28 requires protection at an unprotected side or edge 4 feet or more above a lower level; on a construction project the trigger under 29 CFR 1926.501 is 6 feet, and a service shop can fall under either Part depending on what the job is, so establish which one the day is before you go up rather than after. Set the work position, the access and the fall protection before the meter comes out, because the meter is the thing that pulls attention off the edge.
Taking current at an energized disconnect is work on or near live parts. 29 CFR 1910.333(a)(1) requires deenergizing first unless the employer can demonstrate that deenergizing introduces additional or increased hazards or is infeasible due to equipment design or operational limitations, which is the demonstration a running-load reading normally rests on; the boundary and PPE come from the employer's electrical safety program, most often built on NFPA 70E in the edition the employer or site has adopted, binding through that program rather than on its own. Inspect leads, cables, probes and connectors for external defects and damage before use, as 29 CFR 1910.334(c)(2) requires, and confirm the instrument's measurement category marking covers the point, that marking coming from IEC 61010-1 through the listing on the instrument.
What the word is actually claiming
An outlier is a point that does not belong to the population you assumed. So before any point can be called one, two things have to be said out loud:
The population. Not "the units." Same model, same configuration, same operating state, same load, same measurement point, same instrument. Every one of those is a wall around the set, and a point outside a wall is not an outlier, it is a member of a different set.
The mechanism you expect. What would have to be true, physically, for a member of that population to land there? If you can name a mechanism, the point is evidence about that mechanism. If you cannot, you have no basis for excluding it and no basis for keeping it either, which means the honest next move is another reading, not a decision.
A point set aside without both of those is not an outlier removal, it is data falsification with a tidy motive, and it is the reason so many "we have always seen this" patterns in shop records are artifacts of years of quiet deletion.
The first anomaly: 4.1 against a cluster of 9.6 to 11.2
Ten of the twelve came in between 9.6 and 11.2 A. Unit 4 read 4.1 A.
The mechanism question first: what would make a supply fan motor of this size draw well under half of what its identical siblings draw, while running? A drive problem that lets the motor turn without moving air would do it, and so would a badly slipping belt, since less power transmitted is less current drawn. Both are real, and both are worth a look. So the tech went to look.
What was there was neither. The clamp had been placed on the conductor feeding the condenser fan rather than the supply fan. Retaken at the right conductor, the unit read 10.4 A, inside the cluster.
That is one of only three legitimate dispositions, which are worth stating before the cases work through them: reclassify the point because it belongs to a different population, name a measurement fault and retake it, or keep it and let it be the finding. This one is the second: a named measurement fault, so the reading is retaken rather than interpreted. The point that matters is the ordering. The mechanism question is what sent the tech back up rather than into the spreadsheet, and "wrong conductor" is a mechanism you find by going and looking, never by staring at numbers.
The second anomaly: 13.8 above the cluster
Unit 11 read 13.8 A, well above the top of the cluster. Mechanism candidates are the ordinary ones for a fan motor drawing more than its siblings, and the tech was ready to write it up.
The service history closed it first. Unit 11's motor had been replaced two years earlier with one carrying a different nameplate. It is the same rooftop unit model and it is not the same population, because the population was defined by configuration and this one's configuration changed. Its own reading the prior year was 13.5 A, so 13.8 A is 0.3 A above its own history, which is where the rest of the cluster also sits year over year.
That is the first disposition: reclassify. The point was never in the population, so it was never an outlier. It gets its own series and its own comparison, and it stays in the record with a note saying why it is being compared against itself rather than against its siblings.
The point nobody flagged
Unit 6 read 11.2 A, at the top of the cluster and inside it. Nothing about it looked odd across the twelve.
Against its own prior year of 9.8 A, it had moved 1.4 A, a rise of about 14 percent. Every other unit in the population had moved under 0.3 A, which against a base near 10 A is under 3 percent. Unit 6 was the only real finding on the route, and the outlier hunt had been looking in the wrong direction the entire time: across units, when the information was within units.
This is the third disposition, and it is the one that gets skipped: keep the point, because it is the finding. Nothing was wrong with the measurement and nothing was wrong with the population. The number was simply telling the truth about a machine that had changed.
It also names the general trap. A cross-sectional comparison, twelve units on one day, and a longitudinal comparison, one unit across years, answer different questions and have different sensitivities. The cross-section hides a change that is still inside the spread of the group. A companion article works through what each mode of reading forgives; the short version here is that a set of readings taken on one day cannot detect a slow change, and no amount of outlier analysis will make it do so.
The three dispositions
There are exactly three defensible things to do with a point that does not fit, and none of them is deletion.
- Reclassify it. The point belongs to a different population. Say which, and give it its own comparison.
- Attribute it to a named measurement fault and retake it. Wrong point, wrong function, bad coupling, wrong instrument. The name is required; "probably a bad reading" is not a name, and a retake without one just gives you a second unexplained number.
- Keep it as the finding. No population problem, no measurement problem. The machine did that. Report it.
If none of the three applies, the point stays in the record marked as unexplained, and unexplained is a legitimate final state. A record with an honest unexplained point is stronger than a clean record that got that way through judgment nobody wrote down.
What the shop changed
Two changes came out of this route, and both are record changes rather than instrument changes.
The reading form got a field for the measurement point by name, because the 4.1 A retake showed that a reading with no stated point cannot be audited by anyone who was not standing there. And the annual comparison got flipped: each unit against its own prior year first, the cross-unit spread second. That reordering is what would have surfaced unit 6 on the first pass, and it costs nothing, because both numbers were already in the file.
How to verify a disposition was legitimate
Read the anomalous point's disposition to someone who was not on the job, and see whether they can restate the reason without using the words "looked wrong." A disposition that survives that is either a named population, a named measurement fault, or a finding. A disposition that does not survive it was a judgment, and judgments belong in the notes as judgments, marked, never applied silently to the data underneath.
One more check, on the record rather than the point: search your own historical data for gaps. A month with no reading on a route that runs monthly is either a missed visit or a deleted point, and the two look identical afterward. If you cannot tell them apart, the shop has been losing information it does not know it lost.
References
- 29 CFR 1910.28 and 29 CFR 1926.501, OSHA fall protection on walking-working surfaces in general industry and in construction, which set different trigger heights
- 29 CFR 1910.333(a)(1) and 29 CFR 1910.334(c)(2), OSHA general industry, energized work and inspection of test instruments and leads before use
- Manufacturer documentation for the equipment's configuration history and rated values
- See related: Reading a Trend Against Reading a Value; What to Write Down So a Reading Survives