Why the Format of Interim Data Matters
An interim data view is not just a smaller version of the final trial database. Its format determines which relationships can be examined and which distinctions have already been erased. A table with one row per participant can show how treatment assignment, outcome, and assessment time come together for each person. An aggregate table can efficiently report counts or summaries, but it may conceal the individual patterns that produced them.
The distinction matters even when two data views report the same treatment-arm means. They may differ in whether they identify each participant's assigned group, preserve repeated measurements, distinguish observed outcomes from missing ones, or show when outcomes were recorded. To compare formats responsibly, first ask what the unit of each row is, what is summarized, and what information is no longer recoverable.
What Each Format Exposes
A participant-level view makes the joint structure of the available records visible. Depending on the trial and the authorized data view, fields may include a participant identifier, randomization code, baseline characteristics, outcome measurements, visit or event dates, and status indicators such as whether an assessment is available. A code identifies an assigned group only to the extent that the code key is available to the person viewing the data. A participant-level file does not automatically disclose treatment identity.
An aggregate view displays summaries for a specified group. The group might be all participants combined, or participants grouped by a treatment code; the data may or may not reveal the treatment identity behind that code. Summaries can include group sizes, outcome counts, means, or measures of variability. However, the usefulness of an aggregate depends on exactly how it was constructed. A mean among observed outcomes is different from a mean among all randomized participants if some outcomes are missing.
| Question | Participant-level records | Aggregate summaries |
|---|---|---|
| Who contributed each outcome? | Can be traced to a participant identifier. | Usually not identifiable from a group summary. |
| Which assigned group? | May show a randomization code for each participant. | May summarize by code, or pool groups; identity is not guaranteed. |
| When was the outcome recorded? | Can preserve participant-specific dates or visit times if included. | May show a common time window, but often not individual timing. |
| What denominator was used? | Record status can help identify who has an available outcome. | Must be explicitly reported; a count alone may not define it. |
| Can individual patterns be examined? | Yes, for fields retained and accessible. | Only to the extent the summaries preserve those patterns. |
Neither format is inherently blinded or unblinded. Participant-level records can carry coded assignments without revealing treatment names; aggregate summaries can also be shown by coded arm. Conversely, pooled aggregates can omit group distinctions, while a participant-level file with an accessible code key can reveal them. The format describes the granularity of information, not by itself who knows treatment identity.
Worked Example: From Coded Participant Records to Arm Summaries
Suppose a hypothetical trial has eight participants with an outcome measured in points at a prespecified interim visit. Codes K and L denote randomized groups, but the treatment names are not available in this view. All eight participants have an outcome at that visit.
| Participant | Code | Outcome (points) | Assessment day |
|---|---|---|---|
| P01 | K | 8 | 28 |
| P02 | K | 10 | 29 |
| P03 | K | 9 | 28 |
| P04 | K | 7 | 30 |
| P05 | L | 11 | 28 |
| P06 | L | 12 | 29 |
| P07 | L | 8 | 28 |
| P08 | L | 13 | 30 |
For code K, the number of participants is \(n_K=4\), and the mean outcome is
For code L, \(n_L=4\), and
An aggregate view could therefore report four observed outcomes in each coded group and means of 8.5 and 11 points. That summary preserves group sizes and mean outcomes, but it no longer identifies which participant had which value or which assessment occurred on day 29 rather than day 30. The difference between these coded-group means is 2.5 points; by itself, it does not identify a causal treatment effect or reveal treatment names.
Aggregation Preserves Some Quantities, Not Every Relationship
An aggregate can answer questions about the particular summaries it retains. For example, a total and a group size determine a group mean. But aggregation is generally a many-to-one transformation: multiple different participant-level data sets can produce the same summary. Once records have been combined, there is usually no unique way to reconstruct which outcomes belonged to which people or how measurements were paired over time.
A simple derivation shows why marginal means do not determine the variability of within-participant changes. Let \(B_i\) and \(F_i\) be a participant's baseline and follow-up measurements, and define the change \(D_i=F_i-B_i\). The mean change is determined by the two means:
But the sample variance of the changes depends on the pairing of each follow-up value with its own baseline value. Equivalently, its calculation includes the paired products \(B_iF_i\), which cannot generally be recovered from separate baseline and follow-up means. Thus, summaries can preserve a mean change while losing information needed to describe the distribution of individual changes.
Worked Example: Identical Marginal Means, Different Change Patterns
Consider two hypothetical participants with baseline values of 10 and 20 points. In one possible data set, their follow-up values are 12 and 22. The baseline mean is 15 points, the follow-up mean is 17 points, and the mean change is \(17-15=2\) points. The individual changes are 2 and 2 points, so their sample standard deviation is 0 points.
Now pair the same two follow-up values differently: participant 1 has 22 and participant 2 has 12. The baseline and follow-up means are still 15 and 17 points, so the mean change remains 2 points. But the individual changes are now 12 and \(-8\) points. Their sample mean is 2 points, and their sample standard deviation is
The two data sets have the same baseline mean and the same follow-up mean, but radically different participant-level change patterns. Separate aggregate means recover the mean change here, but not the variability of paired changes. A summary that reports only those means cannot support every analysis that the participant-level records could support.
Outcome Timing and Denominators Must Be Explicit
Interim summaries are especially easy to misread when follow-up is incomplete. A count of events needs a defined group and observation window. A proportion needs both a numerator and a denominator, and the denominator should state whether it includes all randomized participants, only participants with observed outcomes, or some other specified set. Timing also matters: an event count accumulated through a common calendar date can reflect different follow-up durations for participants enrolled at different times.
Participant-level records can preserve the dates and statuses needed to examine such differences, if those fields are collected and included. An aggregate can still be informative, but it should report what the counts refer to: for example, the number randomized, the number with outcome data, the event definition, and the period over which outcomes were observed. A single rate without these details can hide important differences in data availability.
Worked Example: Equal Observed Rates, Unequal Follow-Up Counts
In a hypothetical interim review, code K has 50 randomized participants, 30 with an observed binary outcome by the review date, and 6 observed events. Code L has 50 randomized participants, 40 with an observed outcome, and 8 observed events. Among those with observed outcomes, the event proportions are
Both observed-outcome proportions are 20%. However, the view also shows that outcome availability differs: 30 of 50 participants in K and 40 of 50 in L have observed outcomes. If someone instead divides by all randomized participants, the calculations are \(6/50=0.12\) and \(8/50=0.16\), or 12% and 16%. Those are not the same estimand as the proportions among observed outcomes, and treating everyone without an observed outcome as a non-event would require justification that the data given do not provide.
The aggregate counts reveal the observed proportions and the different numbers observed, but not which participants remain without outcomes or how long each has been followed. A participant-level view with relevant status and timing fields could reveal those details. Neither calculation alone establishes why outcomes are unavailable or whether the observed participants represent the intended analysis population.
How to Audit an Interim Data View
Before using either format, specify the question the view is meant to answer. Then check whether its fields or summaries align with that question. A useful audit separates what is visible from what would have to be inferred or obtained elsewhere.
Determine whether each row represents a participant or a group summary. State whether groups are pooled or separated by assignment code, and whether treatment identities are visible.
Check the outcome variable, units, event definition, and assessment time or window. A label such as “response” is not enough unless the response rule is specified.
For each summary, identify the number randomized, the number with available outcomes, and the rule used to include records. Distinguish missing outcomes from observed non-events.
Use participant-level detail when the question depends on individual pairing, timing, or status. Use aggregates when the target is a reported summary and the grouping and construction are sufficiently clear.
Common Mistakes, Alternatives, and Limitations
- Assuming an aggregate is blinded. A summary by coded arm still distinguishes groups, while a participant-level file can use masked codes. Data granularity and treatment-identity access are separate properties.
- Comparing percentages with unclear denominators. A rate among participants with observed outcomes cannot be treated as a rate among all randomized participants without an explicit rule and justification.
- Assuming group means describe individual patterns. Identical means can come from records with very different variation, pairing, and timing, as the paired-change example shows.
- Treating a missing record as an observed zero or non-event. The format must distinguish an unavailable outcome from an observed outcome equal to zero or an observed absence of an event.
- Assuming participant-level data are complete. Individual rows help only for fields actually collected and available. Missing dates, ambiguous outcome definitions, or unrecorded status cannot be repaired merely by retaining one row per participant.
The main alternative to sharing participant-level records is to share aggregates tailored to a prespecified question. Aggregates can reduce detail and make a view easier to interpret, but they restrict later checks to what was summarized. Participant-level records preserve more structure for analysis and auditing, but require careful control of access and interpretation; they do not automatically clarify treatment identity or remove bias from incomplete observations. Neither format is universally preferable. The right comparison is between the information needed for the intended task and the information the proposed view actually exposes.
Check Your Understanding
Answer each question, then compare with the short solutions below.
- An aggregate report gives a mean outcome of 6.4 points for code M. Name two additional details needed to interpret it clearly.
- Can separate baseline and follow-up means determine the sample standard deviation of participant-level changes? Explain.
- A report gives 9 events among 45 participants with observed outcomes. What is the observed event proportion, and what does it not establish about all randomized participants?
- Does a participant-level file necessarily reveal treatment identity? Explain the role of coded assignment.
Solutions. (1) State the group and its size, the outcome definition and units, the assessment time or window, and how participants with missing outcomes were handled; any two of these address essential interpretive gaps. (2) No. The means determine the mean change, but the standard deviation of individual changes depends on how each participant's baseline and follow-up values are paired. (3) The observed proportion is \(9/45=0.20\), or 20%; it does not establish the event proportion among all randomized participants if some outcomes are unavailable. (4) No. A file may show a randomization code without revealing the treatment names; whether identity is disclosed depends on access to the code key and other information.