A missing value can interrupt an otherwise convenient analytical workflow. The table is ready, the intended comparison is understood and a large group of records lacks one field. The pressure is then to choose a usable representation and continue. I want that preparation decision examined carefully when the proposed value would imply a clinical finding. A blank in the available research material does not, by itself, establish normality or any other state of the participant. Giving it such a value changes the evidence before the formal analysis begins.
The fictional Seravune Therapeutics registry review of 2021 covered 18000 records for a baseline field. Eleven thousand two hundred had a recorded value; 6800 did not. The proposal was to code the missing portion as normal so the internal comparison could retain the whole set. I understood the practical concern about excluding a substantial portion of the material. I objected to the clinical interpretation being supplied by a data-preparation choice. The relevant experts needed to understand why the values were missing and what that meant for the intended use.
We rejected the normal-value coding and supported S$2.2 million source and fitness review. The account retained 11200 recorded values and 6800 missing entries. It did not claim that all missingness had one cause or that every value would later be recovered. The decision made the work less convenient, but it preserved the distinction between an observation and an assumption. Scientists, statisticians and clinical owners remained responsible for the analysis and interpretation, with consent, privacy and permitted use handled through the appropriate processes. The data office's contribution was to keep the record honest about what it contained.
The missing entry is a question about the record
A value can be absent for several reasons. It may not have been collected, may not have reached the source or may not be accessible in the material being reviewed. Understanding that origin can matter to the proposed analysis. The data team should not fill the gap with the explanation that makes the table easiest to use. It should investigate what can be established and state what remains uncertain. Where the cause cannot be resolved, the uncertainty still belongs in the account the responsible expert receives.
This is different from declaring that no analysis can ever use incomplete material. That is not a decision I would make from the missing count alone. The relevant scientific and statistical assessment must consider the question, the source and the consequences of the available evidence. My concern is that a preparation step should not quietly decide that question by assigning a clinical meaning without support. A clearly represented missing state allows the expert to examine it. An invented normal value can make the issue disappear from view while leaving its effect in the analysis.
The 6800 count therefore describes the reviewed field in the specified material. It does not state that those participants had normal findings, abnormal findings or no relevant observation in every possible source. I want the public and management account to retain that scope. A concise story can still distinguish absence in a dataset from a conclusion about the person represented. The difference is especially important when a label may later be copied into a portfolio summary without the preparation discussion that originally surrounded it.
Preparation can change the claim before analysis begins
Data preparation is sometimes described as a neutral stage before the substantive work. Many preparation choices are technical, but some alter what the material appears to say. A code can represent missingness, or it can assert a finding. A transformation can preserve the original observation, or it can replace it with an assumption that needs explicit review. I want the data leader to notice when that boundary is being crossed. The point is not to prevent transformation; it is to make consequential interpretation accountable.
In the registry case, the proposed normal code would have made the column complete in a superficial sense. It would not have added an observed baseline value for the 6800 records. That is why I resisted treating completeness as the programme's immediate success measure. A fully populated field can be less informative if readers cannot distinguish actual observations from supplied assumptions. The selected response retained an inconvenient state because the inconvenience reflected something the institution still needed to understand. Removing the blank would not remove the underlying question.
Colleagues who wanted to proceed had a reasonable concern about delay and the usefulness of the comparison. The S$2.2 million review consumed time and resources, and the outcome was not predetermined. I do not present the decision as costless caution. I considered the cost justified because the proposed coding would change the claim without the necessary basis. A strong case account should keep both sides visible: the work the institution hoped to do promptly and the reason the current preparation did not support doing it in that form.
Make unresolved source work visible to the decision-maker
The review's status should remain part of the programme account. A decision to investigate missingness is not evidence that it has been resolved. I did not report recovery of every absent value or a completed scientific conclusion. The source work might clarify some records while leaving others uncertain. The responsible decision-maker needs to know the actual position, not an optimistic ending inferred from the fact that a programme has been funded. A data leader should make the next decision assessable even when the record remains incomplete.
For a board, I would frame the issue around the relationship between the source and the claim. What does the available material contain? Which interpretations were introduced during preparation? Who assessed them for the intended use? What remains unresolved? These are questions about accountability and the evidential basis of a recommendation. They do not require directors to choose a statistical treatment themselves. The board should be able to understand why management's conclusion is supported and where specialist judgement remains necessary.
For a leadership website, this subject can demonstrate the substance of a research-data role without making clinical claims. A real leader can describe a permitted example of a preparation decision, the disagreement and the selected response. Participant information and restricted study details should remain within approved processes. The account should avoid naming a treatment benefit or regulatory outcome that the work did not establish. The professional contribution is visible in how the leader preserved the question for the appropriate experts rather than allowing a convenient code to answer it silently.
I want a research dataset to be usable because its meaning is clear, not because every field has been made to look complete. In the 2021 example, preserving 6800 missing entries made the work harder in the near term. It also preserved the difference between what was recorded and what had yet to be understood. That distinction is a practical form of data leadership. It leaves the institution able to examine uncertainty rather than building its next conclusion on a finding the source never supplied.