Weight and Body-Size Variables

Why Body-Size Records Need More Than a Number

A weight value can affect interpretation of an exposure model, but a number alone is not a usable covariate. “70” could mean kilograms or pounds; it could be a measurement from screening or from the day of a concentration sample; and it could be measured, converted, carried forward, or imputed. Those differences matter. A value with the wrong unit can be seriously wrong in magnitude, while a correctly measured value from the wrong time can misrepresent a participant’s body size at the event being analyzed.

The practical task is to prepare each body-size record so that its meaning and origin are recoverable. Preserve the source value and unit, record when the measurement was taken, and distinguish that observation from any value created during data preparation. This is particularly important for weight, which can vary over time, and for other measures such as height, waist circumference, or body-surface measurements.

Represent the Measurement and Its Time

At minimum, a prepared record should identify the subject, body-size variable, numeric value, unit, measurement time, and source. If the source supplies a status or method, preserve that too. Keep the source record or a traceable link to it, particularly when the analysis-ready dataset uses a converted or otherwise changed value. A useful structure separates the original value and unit from the standardized value and unit rather than overwriting the source.

The relevant time is the time the measurement was taken, not automatically the time it was entered into a database, the nominal visit label, or the time of the nearest dose. Retain the most precise valid measurement time available. If only a date is known, do not manufacture an exact time; record the date-level precision or an explicit time-precision indicator. Apply the project’s established time-zone and time-origin conventions consistently.

When a weight is used as a time-varying covariate for a dose or sample event, the analysis rule must say which measurement is assigned to that event. For example, it may use a measurement taken at the event, or the latest eligible measurement before it. The rule should specify how old a measurement may be and what happens when no eligible measurement exists. A future measurement should not silently be treated as if it were already known at an earlier event.

Check Changes Without Treating Flags as Verdicts

Repeated body-size measurements provide an opportunity to identify transcription, unit, subject-linkage, and timing problems. A useful first screen compares consecutive records for the same subject and variable. The absolute change is the new value minus the preceding value. A relative change expresses that difference in relation to the preceding value. If the interval is known, the change per unit time can also be calculated.

These summaries are interpretable only when the records refer to the same subject and measurement, use comparable units, and have correctly ordered times. A percent change also requires a nonzero reference value. The rate requires a meaningful elapsed time; a date-only timestamp does not justify pretending that the measurements were exactly 24 hours apart. Always check unit consistency before comparing values: a switch from pounds to kilograms can resemble a dramatic biological change.

A change flag is a prompt to investigate, not proof that a record is invalid. The appropriate alert rule depends on the measurement, population, time interval, and data-collection context. There is no single percentage threshold that establishes an impossible weight change for every participant. Acute fluid shifts or clinical changes may be genuine; a mistaken unit, decimal place, date, or subject identifier may create a similar pattern. The review should consult source records and context, document its outcome, and avoid silently deleting or replacing the observation.

Keep Observed, Derived, and Imputed Values Distinct

A weight measured at a particular time is evidence from the source. A standardized-unit value is a transformed representation of that evidence. A derived body-size variable is calculated from inputs under a defined formula, and an imputed value substitutes for an unobserved value under an explicit rule or statistical procedure. These distinctions should survive the data-preparation steps, not just appear in a note written after analysis.

For a derived variable, retain its inputs, formula or algorithm version, and the time basis of those inputs. For an imputed value, record which value was missing, the imputation method, and any applicable imputation identifier or uncertainty representation. Never encode missing weight as zero: zero would be a numeric value with a different meaning, not a marker for absence. If the analysis uses a substitute for missing weight, retain the observed data separately and label the substitute so it cannot be mistaken for a measurement.

Carrying a previous weight forward is one possible assignment rule, but it does not create a new measurement. It also assumes that the earlier value is suitable for the later event under the analysis plan. Record the source measurement time and the age of that measurement at the event, and identify the carried-forward value as substituted. Other choices, such as leaving the covariate missing or using a defined imputation method, answer different analytic needs and require their own documentation.

Practical Alternatives and Failure Modes

There is no universally best way to represent body size for every analysis. A measurement at each event is direct but may not exist. Using the latest prior measurement is practical when the analysis plan allows it, but its relevance can diminish as it ages. A visit-level or baseline value is simpler, but it does not represent subsequent changes. Interpolation between two measurements uses information from both; if the later measurement occurs after an event, that approach uses future information and may be inappropriate for an analysis intended to represent what was known at the event. A formal imputation approach can address missing values under stated assumptions, but its outputs remain imputed, not observed.

The choice should be defined before assigning values and applied consistently. Useful diagnostics include checking for mixed units, duplicate or out-of-order measurement times, unexplained abrupt changes, missing source status, and event-time assignments that use a later measurement. These checks help locate possible problems; they do not determine by themselves whether a value is clinically plausible or analytically appropriate.

• Wrong reasoning: “The value is extreme, so it must be a data error.” Correction: Treat an extreme change as a review signal; verify the source and context before deciding how to handle it.
• Wrong reasoning: “The standardized column says kilograms, so the source was measured in kilograms.” Correction: Preserve the original unit and conversion provenance; a converted value is not evidence of the source unit.
• Wrong reasoning: “The last available weight is the participant’s measured weight at every later event.” Correction: It remains a measurement from its original time; any later assignment must follow an explicit rule and be labelled accordingly.
• Wrong reasoning: “A blank weight can be replaced by zero so the dataset is complete.” Correction: Zero is a value, not a missingness marker. Preserve missingness or use a documented method that distinguishes substituted values from observations.