Integrity and missing records in observation data
In one sentence
A missing observation is an unknown activity; it is not zero waiting. If the plan, the actual records and the correction trail are all kept, the denominators of the proportions can be audited.
By the end of this lesson
- Calculate the missing rate of 24 missing records in 480 planned observations as 5% and the valid count as 456.
- Calculate the valid-observation proportion of 31.58% for 144 waiting records and separate it from the planned denominator.
- Find the bounds of 30% and 35% for the whole plan without filling in the missing activities.
- Choose the appropriate correction or re-observation decision for duplicate identifiers, late records and not-visible records.
Why it matters
An observation table having no blank cells does not mean the data are complete. If a moment that was not seen has been filled in as processing, the table looks orderly but information has been lost. Integrity in observation data requires being able to trace what each planned observation was. Was it done, was it done on time, was the resource visible, on what evidence was the class chosen? The answers to these questions come before the proportion calculation. The observation identifier, the planned and actual time, the resource code, the class and the record status are stored together. The relationship between the raw information and the information corrected later is preserved. That a result closes numerically does not show that this relationship is right.
A missing record and a zero event are different things. A valid observation with no waiting seen is zero for the waiting indicator; when the resource was not visible, this indicator is unknown. Turning the two into the same value can pull the proportion down. When missing records arise especially during busy periods, equipment interruptions or safe-access problems, the observed moments may not represent the whole plan. A small-looking missing rate does not remove this relationship. Even 5% missing can change a result close to the decision threshold. For this reason, besides the amount of missing data, the day, shift and conditions in which it occurred are examined too. A reason record is not an activity estimate; it only helps explain how the unknown arose.
It is possible to calculate a proportion from the valid observations, but that proportion belongs first to the observed portion. If an interpretation is to be made about the whole plan, the way the missing records were selected and lost is assessed. Without any additional assumption, two extreme cases can be calculated: none of the missing records is waiting, and all of them are. These are not a confidence interval; they are logical bounds produced by our ignorance of the missing values. They do not include random sampling error. Even so, they show openly whether the decision is sensitive to the missing records. If the threshold falls within these bounds, instead of producing a firm decision by filling the missing records into a convenient class, the information requirement that remains open is stated.
An integrity check is not limited to missing records. Writing the same observation twice, replacing the planned time with the actual time, taking an out-of-scope record inside, and changing the class dictionary after seeing the result also spoil the calculation. Corrections can be made; what matters is that it can be traced what was changed and why. Writing down what the observer recalls afterwards as if it were an exact measurement is not a reliable correction. If there is no source evidence, the unknown state is preserved. Such a record is more useful than a table filled in to look tidy, because it honestly shows which result is supported and which question needs new observation.
Formula
Number of valid observations (observations) = Number of planned observations − Number of missing observations
Missing rate (%) = Number of missing observations ÷ Number of planned observations × 100
Waiting proportion in the observed portion = Number of waiting ÷ Number of valid observations
Lower bound for the whole plan = Number of waiting ÷ Number of planned observations
Upper bound for the whole plan = (Number of waiting + Number of missing) ÷ Number of planned observations
Unlock the rest of this lesson with an Academy account
Create a free Academy account with your work email address; it can be used once it is approved.
- Step-by-step instructions
- Worked example with real numbers
- Where the method misleads
- Common mistakes