Why a Blank Dose Record Is Not a Zero Dose
A dose history is an input to a pharmacokinetic model, not simply a list of numbers. A scheduled dose that was documented as withheld contributes no drug input. A dose record that is absent, however, does not establish that the patient received no dose. The distinction matters because a model uses the administration history to predict concentrations: an unsupported zero can make a plausible exposure history look like underdosing, while an assumed dose can create exposure that may not have occurred.
The earlier material on representing missed doses addressed how to encode a dose known to have been missed. The present task is the evidence-based decision that comes first: determine whether the data support a known administration, a known nonadministration, or neither. In particular, an absent database row and a documented zero are not interchangeable.
Classify the Evidence, Not the Empty Cell
A zero amount is meaningful only when its source and coding definition establish what it means. For example, a dose row with amount 0 may be a deliberate code for “not administered,” or it may be a placeholder inserted during data processing. Conversely, no row at a scheduled time might reflect a missed dose, a delayed data transfer, a failed record linkage, or a dose that was never scheduled. The data structure alone does not resolve these possibilities.
Useful evidence includes the medication administration record, infusion or device logs, contemporaneous patient diaries, dose accountability, pharmacy dispensing records, and documentation of a hold or discontinuation. These sources answer different questions. A dispensing record establishes that medication was supplied, not that it was swallowed. An electronic container opening supports access to medication, not necessarily ingestion. A prescription establishes intent, not administration. The strength of an inference should match what the source actually observes.
For each planned dose opportunity, keep two questions separate: Is there a usable record? and Is actual administration known? A source may contain a record that explicitly says “not taken,” so the record exists while the dose is zero. A database may have no dose row even though a second source verifies administration. Or the prescription may be known while actual adherence remains uncertain.
| Observed pattern | What it supports | What it does not establish by itself |
|---|---|---|
| Administration record says given, with time and amount | Administration, subject to the reliability of that source | That every recorded dose was fully delivered if the record does not verify delivery |
| Record says withheld or not taken | Zero input for that opportunity, if the record is credible and refers to the actual dose | That no other unrecorded dose was taken |
| No dose row at a scheduled time | A gap in that data source | Nonadministration |
| Medication dispensed or container opened | Availability or access | Ingestion or complete administration |
| Conflicting records or uncertain recall | Uncertainty requiring adjudication or explicit representation | A uniquely known dose history |
How Dose Status Enters an Exposure Model
A simple model makes the consequence of classification explicit. Suppose, for illustration, that each dose is an instantaneous input to a one-compartment system with first-order elimination, and that bioavailability \(F\), volume \(V\), and elimination rate constant \(k\) are constant. Let \(d_j\) be the prescribed amount at time \(t_j\), and let \(Z_j=1\) if that dose was administered and \(Z_j=0\) if it was not. For times after those dose opportunities, the dose-related concentration is
This expression is an illustration of input-history dependence, not a general absorption model. It assumes instantaneous input and linear elimination. In it, a confirmed zero sets \(Z_j=0\); a confirmed dose sets \(Z_j=1\). When the administration status is unknown, \(Z_j\) is unknown too. Setting it to zero is an additional assumption, not a neutral way to store a blank.
The distinction also clarifies why concentration data do not always settle an uncertain dose history. A concentration reflects the combined effects of prior input and disposition. Different dose histories can produce predictions that overlap once parameter values, measurement error, and other doses are considered. A concentration can inform plausibility, but it does not by itself turn an unobserved administration into a verified event.
Worked Example: A Documented Hold Versus an Absent Row
Consider a hypothetical oral dose of 100 mg scheduled at 08:00. The administration record for Monday states “withheld by clinician.” Tuesday’s administration table has no row at 08:00, and there is no accompanying hold, patient report, or administration log. Classify the two opportunities.
For Monday, the explicit hold supports a confirmed zero, assuming the record is valid and applies to that scheduled dose. Thus \(Z_{\text{Mon}}=0\), and the dose contributes \(\frac{F(100\text{ mg})(0)}{V}\exp\{-k(t-08{:}00)\}=0\) to the model’s dose-related concentration at a later time.
For Tuesday, the absent row establishes a missing record, not a confirmed zero. The dose status is unresolved, so the analyst must not set \(Z_{\text{Tue}}=0\) solely because the row is absent. Further source review could establish administration or nonadministration; without it, actual adherence remains unknown.
The identical scheduled amount and time do not make the two data situations equivalent. Monday has affirmative evidence of nonadministration; Tuesday has no affirmative evidence about what happened.
A Short Derivation: Why Coding an Unknown Dose as Zero Changes the Prediction
For one dose at time \(t_d\), the expression above reduces to \(C_d(t)=\frac{F d Z}{V}\exp\{-k(t-t_d)\}\). If administration is known, \(Z\) is fixed at 0 or 1. If it is unknown, retaining the two possible histories shows exactly what is at stake. The difference in predicted concentration between the administered and not-administered histories is
For positive \(F,d,V\), and \(t\geq t_d\), this difference is positive. The dose’s contribution decays over time because \(k>0\), but it does not become zero immediately after administration. Therefore, replacing an unknown \(Z\) with zero removes this entire contribution from the predicted history. This justification depends on the stated simple model; in a more detailed model, the input’s time course changes, but the need to distinguish unknown input from known absence remains.
Worked Example: Quantifying the Consequence of an Uncertain Dose
Suppose, hypothetically, that a 100 mg oral dose may or may not have been taken at hour 8. Use \(F=0.80\), \(V=40\) L, and \(k=0.10\ \text{h}^{-1}\). Calculate the dose’s predicted concentration contribution at hour 12 under each status.
If taken, \(Z=1\) and the elapsed time is \(12-8=4\) h:
If confirmed not taken, \(Z=0\), so the contribution is \(\frac{(0.80)(100\text{ mg})(0)}{40\text{ L}}e^{-0.40}=0\text{ mg/L}\). For an unknown status, the data do not justify choosing either contribution as the known one. These values are conditional predictions under the stated assumptions, not evidence that the patient did or did not take the dose.
Worked Example: Resolving a Record Gap With Multiple Sources
In a hypothetical study, a participant is prescribed 50 mg at each of three evening dose opportunities. The central study table has a usable row for the first opportunity (50 mg given), no row for the second, and a zero-amount row for the third. Source review finds that the second opportunity’s infusion-device log records delivery of 50 mg; the third opportunity’s signed nursing record says “dose withheld.” The data dictionary defines a zero amount as “not administered.”
The first opportunity is confirmed administration from the study table. The second has a central-table gap, but the independent device log supports confirmed administration; the gap is a missing record in one source, not a missing dose in the reconciled history. The third is confirmed nonadministration, supported by both the defined zero code and the nursing record.
The total amount supported as administered across these opportunities is \(50\text{ mg}+50\text{ mg}+0\text{ mg}=100\text{ mg}\). This is the known administered total for these three opportunities, not the prescribed total of 150 mg. If the device log only showed that the device was accessed rather than that delivery occurred, the second classification would need to remain uncertain instead.
Practical Adjudication and Diagnostics
A defensible classification can be made systematically. First identify the scheduled opportunity and the relevant time window from the dosing plan. Then inspect the source field definitions and determine whether the record is a dose, a status code, or an administrative placeholder. Next reconcile independent sources, checking subject identifiers, dates, time zones, and duplicate entries. Finally record both the classification and its evidence. The classification should not be more certain than the underlying documentation.
Specify the intended dose time and amount using the dosing plan; do not infer an opportunity just from a gap between recorded events.
Check data dictionaries and source procedures to establish whether blank, zero, withheld, and given have distinct meanings.
Compare records that measure administration, patient behavior, or medication availability, and note conflicts rather than silently selecting one.
Retain a dose amount only when administration is supported; encode confirmed nonadministration as zero input and preserve unresolved status as unresolved.
Useful diagnostics include counts of scheduled opportunities by status, dose totals by source, and timelines that overlay recorded doses with planned opportunities and concentration samples. Look for impossible or suspicious patterns: repeated exact doses despite a source that records only prescriptions, zero amounts that occur only after a data conversion, administration times after a concentration sample when the event ordering says otherwise, or long gaps that appear only for one data source. Compare the original source extract with the analysis-ready event history. A clean-looking model input is not evidence that the underlying status was known.
When two sources conflict, report the conflict and apply a prespecified source hierarchy only if the hierarchy is justified by how each source was collected. If classification cannot be resolved, keep the uncertainty visible in the data record and state what assumption the primary analysis uses. Do not disguise an assumption as an observed fact.
Common Mistakes, Alternatives, and Limits
- Wrong reasoning: “There is no dose row, so the patient missed the dose.” Correction: no row shows only that the source lacks a usable record. Seek corroborating evidence or retain uncertain status.
- Wrong reasoning: “The prescribed amount is 100 mg, so 100 mg entered the model.” Correction: prescription is not proof of administration. Use the prescribed amount as an input only when administration is supported or when the analysis explicitly assumes adherence.
- Wrong reasoning: “A pharmacy refill proves that all doses were taken.” Correction: dispensing supports availability, not ingestion on a particular date.
- Wrong reasoning: “Coding unknown status as zero is conservative.” Correction: it systematically removes possible input and can distort exposure predictions and parameter estimates; its direction and magnitude of impact depend on the data and model.
The main practical alternatives are to use only source-confirmed doses, to make an explicit adherence assumption for unresolved opportunities, or to represent more than one plausible dose history when evaluating conclusions. These are different analytical choices, not interchangeable data-cleaning steps. Their suitability depends on the study question, source quality, and how much information the concentration data provide. No approach can recover a uniquely true dose history when the records do not contain enough evidence.
Finally, status classification does not solve every dosing-history problem. A record may confirm that a dose occurred but not establish its exact time or delivered amount. Those are distinct uncertainties and should not be forced into the categories “zero” and “missing.” Likewise, the simple concentration calculations here omit absorption delays, time-varying parameters, and other complexities; they illustrate why dose status matters, not a universal exposure calculator.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A scheduled dose has no row in the central dosing file, but a validated administration log records that the dose was given. What is the status of the dose and of the central-file record?
- A medication was dispensed and the patient’s diary says “dose probably taken,” but there is no administration confirmation. Does this establish a confirmed dose? Explain.
- In the illustrative model, a 40 mg dose at time \(t_d\) has unknown status. Write the concentration contribution at time \(t>t_d\) for confirmed administration and confirmed nonadministration.
- Why is a recorded zero amount not automatically proof of nonadministration?
Solutions. (1) The dose is supported as administered by the validated log; the central-file record is missing. These are distinct statements about the event and one data source. (2) No. Dispensing establishes availability and “probably taken” signals uncertainty, but neither confirms that this particular dose was taken. Classify adherence as uncertain unless other reliable evidence resolves it. (3) With administration, \(C_d(t)=\frac{F(40\text{ mg})}{V}\exp\{-k(t-t_d)\}\); with confirmed nonadministration, \(C_d(t)=0\). The values of \(F,V,\) and \(k\) are needed for a numerical result. (4) The field may be a placeholder or have a different coding meaning. Its definition and source documentation must establish that zero denotes an actual nonadministration before it is treated as a confirmed zero.