Life Data Analysis · Chapter 6

Common Pitfalls

The mistakes seen in practice and the guards against them.

Life data analysis is the easiest analysis in this knowledgebase to run and one of the easiest to run wrongly, because the software always returns a β and an η and neither of them looks suspicious.

PitfallWhat it looks likeThe guard
An MTBF read as a lifeA quoted mean used to set a limit, an interval or a warrantyThe mean fixes one number and leaves the behaviour free; ask for β and a B-life
Suspensions droppedOnly the failed units in the datasetEvery unit with a known age is data; survivors carry evidence of survival
Mixed failure modesOne fit across everything that came off the itemOne distribution per mode; a dogleg in the plot is two modes arguing
Age measured in the wrong unitsCalendar time on an item whose damage accumulates with cyclesUse the clock the damage runs on: hours, cycles, landings, starts
Age since whenTime since manufacture used for a repaired itemTime since the last renewal of that mode, or the fit is meaningless
Method unstatedA β quoted with no estimator and no sample sizeSay which method, how many failures, and how many suspensions
Small-sample bias ignoredAn MLE β taken at face value on ten failuresApply the correction, and report the interval as well as the point
The three-parameter fit reached forA location parameter added because it improves r²Only where a physical reason exists for failure-free life, never to flatter the fit
r² read as validationA fit declared good on r² = 0.99 aloneA straight line through the wrong model is still straight; check the residual shape
Extrapolating far past the dataA B1 life quoted from thirteen failures, none of them earlyThe tails are where the data is thinnest and the confidence widest
Confidence intervals omittedA single β, presented as a factOn small samples the interval on β routinely spans a factor of two
Constant rate assumed foreverAn item with β = 4 carried in a model as a fixed λA constant rate on a wear-out mode is true at one fleet age and expires
Fleet age ignoredA field rate quoted with no statement of the fleet's age profileThe same item gives a different rate on a young fleet and an old one
Left truncation missedUnits that entered the record part-worn treated as newTheir early life was never observed, and ignoring that biases the fit
Interval censoring flattenedAn inspection finding recorded as a failure at the inspection timeIt failed somewhere in the interval, and the method for that exists
Weibull assumedThe distribution chosen before the data was looked atWeibull is flexible, not universal; a lognormal repair-time or a bimodal wear pattern is not Weibull
A repairable system fitted as livesA Weibull through the gaps between one machine's successive repairsThose gaps are not independent lives; a repairable system's arrivals belong to the NHPP models in reliability growth
The fit never revisitedβ and η from year two, still in the model in year eightRefit as the fleet ages; the tail is exactly what new data adds

Six deserve more.

An MTBF does not describe a life. It is the mean of the distribution, and the mean is one number out of a two-parameter family: hold it at 100 hours and the age by which a tenth of the population has failed can be anywhere from about half an hour to eighty hours, depending only on β. The worked example does this with two real datasets, both ten failures in a thousand hours. Neither report was dishonest and neither number was wrong; the mean simply does not carry the information a life limit needs, so asking for it more precisely does not help.

Dropping suspensions is the error the arithmetic punishes hardest. It feels harmless, because the survivors have not done anything yet. What they have done is survive, and an estimator shown only the units that broke sees a sample of the unlucky. The bias has a fixed direction: the estimated life can only come out short, so every interval derived from it comes out short too, and the programme spends money being early.

Mixed modes produce a fit that describes nothing. A bearing that fails from grease breakdown at one age and from cage fatigue at another produces a plot with a bend in it, and a single Weibull through the bend fits neither population. The tell is visual and it is why the plot matters more than the r². Split the data by mode, and if the mode cannot be determined from the record, that is a finding about the FRACAS, not about the bearing.

Successive repairs of one machine are not a sample of lives. This is the misuse that looks most like the method and is furthest from it. Life data analysis needs independent and identically distributed lives, which is what a population of items each running once to failure gives you. Take one repaired machine instead and record the gaps between its failures, and the gaps are neither independent nor identically distributed: each one starts from whatever state the last repair left the machine in. Fitting a Weibull to them returns a β and an η that mean nothing, and the β is routinely misread as wear-out when what it is describing is a system whose failure intensity is drifting. The right model there is a process rather than a distribution, and it lives in reliability growth. The test is one question: are these the lives of different items, or the intervals of one?

A good r² is not a validated model. Rank regression will happily draw a straight line through points that curve gently, and the correlation stays high while the model is wrong in exactly the region a maintenance decision cares about. Look at where the points sit relative to the line, not at the summary statistic, and if the fit has to be defended formally, use a goodness-of-fit test built for estimated parameters or a likelihood comparison against the lognormal. The foundations chapter sets out all three.

The result expires. Every β greater than one says the hazard is a function of age, so the item's failure rate is a function of the fleet's age profile. Quoting it as a constant is legitimate only alongside the age profile it was true at, and only until the fleet moves on. On the worked example the constant everything downstream carried was 55 failures per million hours; averaging the fitted hazard across the ages the fleet actually has gives 401, and nothing about the hardware changed.


Want to see this on a live system model? Request a walkthrough.