Absent Values

## Three states that look the same on screen Nothing useful here is not one state in a data column but several, and an analyst is expected to name at least three of them on sight. **A measured zero** is a completed observation whose answer was none. The meter was read and it had not moved; the shop was open and sold nothing; the count was taken and came to zero. A zero is data. It was produced by the same process that produced every other number in that column, and it is exactly as trustworthy as they are. **A present value of length zero** belongs to text columns. A value was supplied, it arrived, it is stored, and it contains no characters. Somebody submitted a form field without typing into it, or a producing system wrote the field with nothing between its delimiters. It is a real value in the same sense that a value of five characters is a real value; it simply has no characters in it. **A cell with nothing recorded** is the only one of the three that is absence. No observation stands behind it. Whatever the tool uses internally to mark such a cell, the meaning is the same: the table is telling you that it does not have this value. On screen the three are nearly indistinguishable. A zero renders as a zero, a length-zero text value renders as nothing, and an absent cell renders as nothing or as whatever placeholder the display layer picked. Looking at a table is therefore the least reliable way to tell them apart, and it is how most people first try. ## Why only the third one is absence Because two of the three are values, every operation that takes values takes them. Absence is not a value, which is why tools hand you a dedicated **absence test** - the predicate that asks, cell by cell, whether the value is absent - instead of expecting you to compare a cell against something. Why a plain comparison is unreliable depends on the design in front of you. Where absence is carried as a reserved bit pattern inside the number format, that pattern compares unequal to everything, including a copy of itself, so an equality test answers false and tells you nothing. Where the tool defines its own typed absence marker, the comparison can yield a third outcome that is neither true nor false, which then has to be coerced somewhere before it can select rows. The dedicated predicate answers the question directly under either design, and it answers no for a zero and no for a length-zero text value. | State | What produced it | Flagged by the absence test | Removed by dropping incomplete rows | |---|---|---|---| | A measured zero | an observation that came out as none | no | no | | A present value of length zero | a supplied text value with no characters | no | no | | A cell with nothing recorded | no observation at all | yes | yes | ## What each state does downstream - A zero is a term in every sum and a member of every average, and it pulls a mean toward zero exactly as any other small reading would. - A length-zero text value is a legitimate distinct value. It forms its own category in a value-frequency listing, its character length is zero rather than absent, and it survives any step that removes rows containing holes. - A cell with nothing recorded is the only one the absence machinery sees. What an aggregate then does with it is a genuine fork: some designs step over the holes, so a number comes back computed from fewer values; others propagate, so one hole makes the whole answer absent until a flag says otherwise. Both directions exist in tools people actually use, and the skipping one is the more dangerous because it is silent. - The number of cells the absence test flags is a property of the third state alone. A column full of zeros and length-zero text values reports no holes whatsoever. ## How the confusion reaches a published number Take a monthly figure by shop. One shop was open all month and sold none of the product; another closed for refurbishment and filed no return at all. If both arrive as zero, the monthly total is right and the average per shop is wrong, because a shop that did not report has been counted in the denominator as a shop that sold nothing. If both arrive as absent, the outage is visible but the genuine zero has been thrown away, and the product looks better than it is. Only keeping them apart lets you report both the demand figure and the coverage figure, and coverage is usually the more interesting finding. The same confusion in text is quieter. A length-zero value in a name column means the form was submitted with the box untouched, which is a fact about the form. Treating it as absence hides how often that happens; treating absence as a length-zero value invents submissions that never occurred. ## What an interviewer is listening for 1. That you name three states rather than two, and say which single one is absence. 2. That you reach for the absence test rather than a comparison, and can say why a comparison is unreliable. 3. That you turn the distinction into a consequence - a denominator, a coverage figure, an outage - instead of leaving it as a definition.