A numeric conversion can finish without an error while replacing some entries with missing values. That behaviour may be intentional, but the replacements still need to be counted and understood.
In pandas to_numeric, the errors setting can request coercion of invalid numeric input to missing values. It is a handling rule, not a diagnosis of why the source entry could not be converted.
Imagine the text values “12”, “13 kg”, “unknown” and an empty field. They may need different treatment: one is directly numeric, one includes a unit, one is an explicit response and one may be absent. Converting first and inspecting later risks making those distinctions disappear.
Keep a before-and-after audit
Retain the raw column and write converted values into a separate column. Count missing values before conversion, then identify entries that became missing only afterwards. Group those entries by their original text or documented reason, taking care not to expose private data in logs.
Do not repair the example by stripping every non-digit character. That could remove a minus sign, a decimal mark or information about a unit. Define the valid format before choosing a cleaning rule.
For a small table, manually inspect every newly missing value. For a large one, use counts and representative cases, then investigate unexpected groups. Record the accepted rule and its effect on the number of usable observations.
The aim is not to ban coercion. It is to make a convenient software option produce a visible, reviewable change in the dataset.
The accompanying visual is an editorial illustration.