Notation For Longitudinal Treatments And Covariates

Why Visit Indices Matter

In a longitudinal study, a participant can contribute several measurements over time. A notation that simply labels a participant as treated or untreated loses the order of treatment decisions and measurements. To describe a two-visit study clearly, we need to show which variable was recorded at which point and which earlier values are included when we refer to a history.

Consider a study with a baseline measurement, a first treatment decision, a later covariate measurement, a second treatment decision, and an outcome recorded after the second decision. The order matters for organizing the data: the later covariate is not another name for the baseline covariate, and the second treatment is not the entire treatment history.

For example, \(L_{1i}\) might be a disease-severity measurement for participant \(i\) taken between the first and second treatment decisions. The notation does not say that this measurement has a particular value, distribution, or causal role. It identifies when the measurement belongs in the record. A covariate can contain several measurements, so \(L_{0i}\) and \(L_{1i}\) may be vectors. For simplicity, \(A_{0i}\) and \(A_{1i}\) denote scalar treatments here; if a treatment consists of several components, each \(A_{ti}\) can also be a vector.

Current Values and Histories

A current value is indexed by its own visit: \(A_{1i}\) is the second treatment alone. A history collects values from earlier visits through the named visit. Use an overbar to denote such a collection. Thus \(\bar A_{1i}\) contains both treatments, while \(\bar L_{1i}\) contains both covariate measurements.

The subscript on a history marks its last included visit. Consequently, \(\bar A_{0i}\) does not include \(A_{1i}\), and \(\bar L_{0i}\) does not include \(L_{1i}\). In particular, the bar does not mean “all variables in the study”; it means the sequence of the indexed variable through the indicated visit. If a history is discussed without a participant index, such as \(\bar A_1\), the index \(i\) is suppressed for readability, not replaced by a different convention.

These definitions are bookkeeping conventions, not assumptions about a statistical model. They do not require that every participant receive the same treatment, that covariates be scalar, or that the values be independent. They do require a study schedule that lets the analyst identify the relevant occasions. If a measurement is taken after a treatment decision rather than before it, the record must reflect that actual timing; using the symbols above without clarifying the schedule could misrepresent what is observed at each point.

Two Useful Bookkeeping Results

History notation is useful because it compresses a sequence without changing its contents. The first result makes explicit how the sequence grows from the first visit to the second.

Derivation. By definition, the treatment history through visit 0 contains only \(A_{0i}\), and the history through visit 1 contains the treatment values at visits 0 and 1 in chronological order. Therefore \(\bar A_{1i}=(A_{0i},A_{1i})\), which is precisely \((\bar A_{0i},A_{1i})\). The same definition applied to the covariates gives \(\bar L_{1i}=(L_{0i},L_{1i})=(\bar L_{0i},L_{1i})\). If \(L_{0i}\) has \(p_0\) components and \(L_{1i}\) has \(p_1\), concatenating the vectors gives \(p_0+p_1\) components. Thus the identities and the entry counts follow directly from the definitions. \(\square\)

The second result is a practical uniqueness property: once the two treatment values and two covariate values are known, the histories are determined, and the current second-visit values can be recovered from the histories.

Proof. By definition, \(\bar A_{1i}\) is the ordered tuple \((A_{0i},A_{1i})\), so its first entry is \(A_{0i}\) and its second is \(A_{1i}\). Likewise, \(\bar L_{1i}=(L_{0i},L_{1i})\), so its first and second entries are \(L_{0i}\) and \(L_{1i}\). The order and entry labels are fixed by the visit indices; there is no second assignment of these four values consistent with the same ordered histories. \(\square\)

This result is simple, but it guards against a real data-management error: treating a history as an unordered set. If the order were discarded, the pair \((A_{0i},A_{1i})\) would no longer indicate which treatment came first. A history is an ordered sequence, not merely a list of values.

Putting the Full Record in Chronological Order

A complete record for one participant can be written as \((L_{0i},A_{0i},L_{1i},A_{1i},Y_i)\). This order is chosen to match the study schedule: baseline covariates, first treatment, later covariates, second treatment, and final outcome. It does not imply that every possible variable must be observed, nor does it specify a model for the relationship among the values.

When covariates are vectors, it helps to check dimensions explicitly. For instance, if \(L_{0i}=(L_{0i}^{(1)},L_{0i}^{(2)})\) measures age and baseline severity, and \(L_{1i}=(L_{1i}^{(1)},L_{1i}^{(2)},L_{1i}^{(3)})\) measures severity, symptoms, and laboratory status later, then \(\bar L_{1i}\) contains five scalar components in the order baseline age, baseline severity, later severity, later symptoms, and later laboratory status. Parenthesized superscripts here identify components, not additional visit indices.

Common Indexing Errors and Checks

Most notation errors in a two-visit record are errors of timing or scope. They can be caught by asking two questions: what occasion does this subscript identify? and does this history include only values through the stated visit? A quick audit should verify that every participant-level variable is indexed consistently and that the ordering matches the data-collection schedule.

• Using \(A_{1i}\) to mean the full treatment history. \(A_{1i}\) is the second treatment value; \(\bar A_{1i}=(A_{0i},A_{1i})\) is the history.
• Putting the second treatment into \(\bar A_{0i}\). The history through visit 0 contains only \(A_{0i}\).
• Using \(L_{1i}\) for baseline and later covariates together. \(L_{1i}\) is the covariate value at the later measurement; the history is \(\bar L_{1i}\).
• Leaving the observation order implicit. A tuple such as \((L_{0i},L_{1i})\) is interpreted in visit order; rearranging entries changes what the tuple represents.
• Treating the outcome as a treatment or covariate-history entry. Here \(Y_i\) is a separate final outcome, recorded after the two decisions.

The notation also has limits. If visits occur at irregular times, the subscript \(1\) alone may not record the elapsed time; the schedule or actual time variable must be supplied. If participants have different numbers of treatment decisions, two-visit notation does not describe every record without modification. And if a measurement's timing relative to treatment is unclear, no choice of subscripts can resolve that ambiguity: the study's measurement schedule must be clarified first.