An ADaM dataset is a particular type of analysis dataset that follows the ADaM fundamental principles defined in the ADaM¹² and is compliant to ADaM defined structures or follows as closely as possible to the ADaMIG variable naming and other conventions.¹³ Currently, ADaM has three structures: Subject Level Analysis Dataset (ADSL), Basic Data Structure (BDS), and Occurrence Data Structure (OCCDS). An ADaM dataset contains both source and derived data; it is therefore important to clearly document the variable derivations and how to use them for obtaining the analysis results. ADSL is a required, participant-level dataset that contains participants’ baseline and demographic characteristics, population flags that indicate the participant’s inclusion in different analysis populations, planned and actual treatment variables for each period, and important dates. The BDS datasets contain endpoints and data that vary over time during the course of a study and are organized as one or more records per subject per analysis parameter per analysis timepoint. It is often optimal to have more than one BDS analysis dataset, but not necessarily one dataset per analysis. The BDS datasets are the main data structures used for complex statistical analyses but are not designed to support analysis of incidence of adverse events or other occurrence data. Analysis of such data is supported in the OCCDS. For commonly used analysis methods (eg, analysis of variance or covariance, logistic regression and so on) the BDS implementation is straightforward. A more complex analysis method for time-to-event analyses has its own standardized BDS, ADTTE, that is well developed²¹ and widely used. Although the BDS supports most statistical analyses, it does not support all statistical analyses. For example, it does not support simultaneous analysis of multiple dependent (response/outcome) variables or a correlation analysis across a range of response variables.

In the ADaM design, at a minimum, the analysis datasets should contain the datasets needed for the recreation of specific statistical methods. There is no requirement that every analysis has its own dataset, but rather, a single dataset can support multiple analyses to achieve the optimal number of analysis datasets. Each analysis dataset should contain all the analysis-enabling variables required for performing the statistical analysis it is designed to support (it can even contain supportive variables not needed for the analysis but that are of interest for traceability purposes). This can lead to redundancy, that is, the same data appearing in multiple datasets, but this is necessary for having analysis-ready datasets. Analysis-ready does not mean that the results can be generated in a single statistical procedure, but rather that each of the summary statistics included in the results can be derived with minimal programming effort using standard statistical procedures with the dataset as input.

We briefly describe the fundamental principles governing the structure of BDS in connection to Tidy data principles and discuss the structure of ADLB (analysis datasets for laboratory values) that is used for the ANCOVA-type analyses and ADTTE for time-to-event analyses, as these two datasets, alongside the participant-level ADSL, are the source datasets for ADHCE. Then, following the BDS principles, we construct the ADHCE dataset, which is analysis-ready for multiple analyses (with its metadata traceability describing the source datasets and variables) and provide the minimal steps required to perform these analyses using ADHCE.

The methodology provided here is applicable only for fixed follow-up settings. For settings without fixed follow-up, we explore the challenges associated with the derivation of an analysis dataset that conforms to the BDS principles.

Consider the case of two treatment groups, with active and control treatments, and assume that all participants have the same follow-up and there are no dropouts, meaning all participants were followed for all events of interest until the end of the fixed follow-up. The kidney HCE^4,7 has the following construction: during a fixed follow-up, participants are followed for one of the six dichotomous events in the provided hierarchy described in Table 1.

If a participant experiences death, they are ranked in the category one and the timing of the death is used to determine the ranking within that category, with an earlier death being a worse outcome (a lower rank is assigned). Otherwise, if the participant is alive at the end of the follow-up, then the next event in the hierarchy is considered for ranking this participant and so on. If the participant did not experience any of the six events, then they fall into category seven in which the individual rate of change of glomerular filtration rate (GFR) is used to further rank the participants, with a lower rate of kidney decline being a better outcome (ranked higher).

The time-to-event (TTE) analysis dataset, ADTTE, is an ADaM BDS dataset that includes additional TTE variables designed for survival analyses. The distinguishing feature of survival data is that at the end of the observation period the event of interest may not have occurred for all subjects. The single ADTTE dataset can support multiple survival analyses, for example, Cox proportional hazards regression, Log-rank test and so on. For a given analysis parameter value (PARAM or the short name of the analysis parameter value PARAMCD), ADTTE has one record per subject and the two variables used in all models of survival analyses: the analysis value, AVAL, which shows the timepoint until when the participant was observed for the event of interest and the censoring variable, CNSR, which indicates whether or not the event of interest occurred. The variable ADTTE.AVAL therefore shows either the timing of the occurrence of the event (if CNSR=0) or the length of the fixed follow-up duration for participants without an event (CNSR=1). ADTTE should also include the subject identifier (SUBJID) and the treatment variable showing planned treatment allocation (TRTP) in a randomized, controlled trial. The fixed-follow up duration is stored in Primary Analysis Day (PADY), which is inherited from ADSL, since this variable is a common analysis date for all participants and is needed across multiple datasets. The ADTTE dataset contains the six dichotomous events of interest (Table 1), each having a unique PARAM value.

The BDS for laboratory data, ADLB, has one row per subject per visit per analysis parameter value and contains GFR measurements under a specific analysis parameter, PARAM, and the variables AVISIT, which indicates the timepoint of measurements (categorical variable with visit names); analysis day ADY for the number of days relative to an anchor date (in this case, the date of randomization); the analysis values AVAL, which contain the GFR measurements at each visit; and the BASE variable for the baseline GFR values for each subject. In addition, the individual rate of change of GFR over time can be derived (see the supplementary material)⁷ in ADLB.AVAL corresponding to a new analysis parameter value (PARAM = “Rate of change of GFR”).

An HCE can be analyzed using the methods for ordinal endpoints, for example, rank ANCOVA,²² ordinal logistic regression²³ or win statistics.⁵ We consider the win odds¹⁷ but the same principles can be applied to other win statistics. Based on the hierarchy defined above, based on which each participant in the active group is compared with each participant in the control group using each participant’s clinically most severe outcome. Hence, first we select the clinically most severe outcomes of the participants from the given fixed follow-up duration, then compare participants based on those outcomes. If the participant in the active group has a less severe outcome than the participant in the control group, then this is a “win” for the participant in the active group. Forming all possible comparisons of participants in the active group with participants in the control group, we derive the total number of wins, losses, and ties of the active group. The win odds of the active group against control is formed as the total number of wins (plus half of all ties) divided by the total number of losses (plus the second half of the ties). Win odds greater (less) than 1.0 is indicative of the treatment effect in the active (control) group, while win odds of 1.0 is indicative of no difference between groups.

To visualize HCEs, maraca plots (so named after their visual similarity to the musical instrument) were introduced.⁶ On the maraca plot for a kidney HCE, the x-axis is divided into the seven HCE component categories in severity order from left to right. The six TTE components are visualized with adjoined cumulative Kaplan-Meier plots. For the continuous component, the x-axis corresponds to the annualized rate of change of GFR and a beneficial effect on the continuous component is characterized by a shift to the right. The associated vertical dashed lines show the median values for the annualized rates of changes of GFR among participants without dichotomous outcomes in the two treatment groups. Each participant contributes to the HCE with one event, and the width of each category (dichotomous or continuous outcomes) corresponds to the percentage of that category in the composite. An illustration of analysis results with win odds is provided in Table 2, with the corresponding maraca plot in Figure 1.

The win odds compares every participant in the active group with every participant in the control group (a cartesian product) and hence requires these pair-wise comparisons in a dataset so that the summary of wins/losses/ties is calculated. But a dataset with that structure will not be an ADaM compliant analysis dataset and, in fact, will have a very messy structure according to Tidy principles, since each row will not be an observation, but a combination of observations from two treatment groups. Like BDS principles, the data science community uses Tidy principles,¹⁴ according to which each variable should form a column, each observation should form a row, and each type of observational unit should form a dataset. Any violation of these principles results in messy datasets, for example, if column headers are values, not variable names or if variables are stored in both rows and columns. The Tidy principles are like the BDS principles, but they also describe in detail how these principles can be violated. The use of pair-wise comparisons in a dataset would therefore result in two columns representing the treatment groups and hence having the column names as analysis values (because the treatment group is used as an analysis value), violating another Tidy principle.

Another possible structure for the analysis dataset would be to keep only the number of wins/losses/ties for each participant as a counting response variable. But this would mean having multiple response variables, which is also non-compliant with ADaM principles. Keeping only the wins for each participant plus half the number of ties allows a compliant dataset to be created, but limits analysis to only win odds analysis. For a win ratio analysis, a different definition of the analysis value would be needed to keep only the number of wins without ties. Importantly, different types of analyses, eg maraca visualization or ordinal regression, cannot be performed using these analyses’ values.

We derive an ADaM compliant dataset (see Figure 2), ADHCE, with a single analysis variable that is analysis-ready for multiple analyses. The theoretical justification for this is that the number of wins of a participant can be derived using the rank of the participant in the overall dataset (both treatment groups combined) and the rank of that participant in their own treatment group.^17,18 Therefore, the participant-level ranking from the worst outcome to the most favorable can help to create an analysis value for the win statistics calculation. This methodology is applicable only in the cases of fixed follow-up durations since in case of differing follow-ups between participants comparison issues may arise, known as transitivity issues,^4,24 which would lead to comparisons not being on the participant level (impossibility to rank participants using their outcomes).

To derive AVAL in ADHCE (Table 4), first identify participants with any of dichotomous outcomes by selecting the PARAM value in ADTTE corresponding to this event (for example, selecting ADTTE.PARAM= “All-cause death” and ADTTE.CNSR=0). Then select the most severe event of a participant and the corresponding timing of the event from ADTTE.AVAL. If ADTTE.PADY shows the length of the fixed follow-up, then the algorithm for AVAL for each participant is shown in Box 1

For participants without any dichotomous outcomes, we use the individual rate of change of GFR from ADLB, which can be negative. Regardless of their rate of change, a participant without any outcomes should have a higher AVAL than any other participant in all other categories, as shown in Box 2.

The categorization of AVAL, AVALCAT1, contains the type of the event (presented in Table 1), while AVALCA1N is the numeric order of this categorization. As part of the traceability, we provide an illustration of the ADHCE dataset (Table 3), the metadata of analysis variables (including analysis parameter values) included in ADHCE (Table 4). For full traceability between the results, the analysis datasets and the source datasets the analysis results metadata is presented in Table 5 (for results in Table 2 and Figure 1).

The dataset ADHCE (Table 3) is analysis-ready for win odds analysis and visualization using maraca plots. Win odds in the SAS® software²⁵ (using the procedures freq or npar1way) is provided in the Appendix of Gasparyan et al.¹⁷ For example, using proc freq the win odds can be calculated as follows (caution should be made to select the control group as the reference). See Box 3.

Similarly in the R software, the package hce²⁶ can be used to derive the win odds. This confirms that the analysis dataset ADHCE is analysis-ready for win odds analysis since it is possible to perform the calculations without first having to manipulate the data (Table 5). The package maraca²⁷ in R can be utilized for producing Figure 1 from the dataset ADHCE with minimal programming. The maraca package recognizes the ADHCE data structure as of class “adhce”, meaning that it expects all the variables mentioned in the dataset’s derivation above and hence can effortlessly produce the plot as shown in Box 4.

The maraca plots are ggplot2²⁸ objects and hence allow for customization. The maraca plots have the functionality of also producing an associated analysis dataset that can be used for validating this output.²⁹