Hierarchical, ordinal and ranked composite outcomes in clinical trials, and the place of family priorities in neonatal outcome selection

1 Introduction

The dominant primary endpoint in trials of extremely preterm infants is a binary composite, usually “death or severe neurodevelopmental impairment” assessed at 18–24 months corrected age. Its appeal is practical: it raises event rates, so trials need fewer infants, and it sidesteps the competing-risk problem that early death creates for any outcome measured later.

Three properties limit it. First, it is non-ordinal: an infant who dies at three days and an infant who survives with moderate impairment contribute identically to the endpoint. Second, it is insensitive to gradation — a shift from severe to mild impairment, which matters greatly to families, is invisible if both fall on the same side of the dichotomy. Third, it is narrow: the outcomes it contains were selected by clinicians and researchers, and several domains that parents identify as critical do not appear in it.

Statistical methods that avoid the first two problems are well established outside neonatology. Desirability of outcome ranking (DOOR) assigns each participant a single ordinal rank across a hierarchy of outcomes. The win ratio, generalised pairwise comparisons and the Finkelstein–Schoenfeld procedure compare participants pairwise, resolving each comparison at the most severe outcome on which the pair differs. These methods retain the ordering that a binary composite destroys, and they permit outcomes of unequal severity to be combined without assuming equal weight.

Whether neonatology has adopted them, and whether the hierarchies used would reflect what families value, are open questions. This review addresses both.

Review questions

1. Which hierarchical, ordinal and ranked composite outcome methods are used in clinical trials, in what volume, and in which clinical domains?
2. To what extent have these methods been applied in neonatal and perinatal research?
3. How do outcome priorities elicited directly from families compare with those embedded in clinician-derived trial endpoints and core outcome sets?

2 Methods

2.1 Protocol

A protocol was written and fixed before screening began (scoping_protocol.md), specifying the review question in PCC terms, eligibility criteria, search blocks and charting fields. Deviations from it are reported in §2.7.

2.2 Eligibility criteria

Population/Concept/Context (PCC). Population: participants in clinical trials or outcome-development studies, with a focus on neonatal and preterm populations. Concept: hierarchical, ordinal or ranked composite outcome measures; neonatal core outcome sets; family- or parent-derived outcome priorities. Context: any clinical setting, any country, human studies.

Included if a record: (a) developed, evaluated or applied a named hierarchical, ordinal or ranked composite outcome method as a trial endpoint (INC_METHOD); or (b) did so in a neonatal or perinatal population (INC_NEO); or (c) developed or reported a neonatal core outcome set (INC_COS); or (d) elicited or reported outcome priorities from parents, families or former patients in a neonatal context (INC_FAM).

Excluded if a record: used only a conventional binary composite endpoint without hierarchical treatment (EX_COMPOSITE); concerned ordinal measurement scales rather than ordinal endpoints (EX_SCALE); was off-topic (EX_TOPIC); was an editorial, comment or letter (EX_NOORIG); was non-human (EX_NONHUMAN); or had no retrievable abstract (EX_NOABSTRACT).

No language or date restriction was applied.

2.3 Information sources and search

PubMed (via E-utilities) and Europe PMC (via REST) were searched to August 2026. Thirteen PubMed queries and four Europe PMC queries were run across four conceptual blocks. Europe PMC queries were restricted to title and abstract fields (TITLE_ABS) to match the field scope of the PubMed strategy. All queries and per-query yields are reported verbatim in search_strategy.csv and summarised in Appendix A.

2.4 Selection of sources of evidence

Records were de-duplicated on a composite key applied in sequence: PubMed identifier, then normalised title, then DOI. Screening was conducted on title and abstract against the criteria in §2.2, using a combination of large-language-model-assisted classification against the written rubric, a validated high-recall keyword filter, and a text classifier trained on the decisions already made. Records the classifier flagged as uncertain were adjudicated by direct reading. Every record coded as both neonatal and naming a method was read individually.

2.5 Data charting

A charting form was fixed in the protocol before screening. Each included record was charted for: named method (10 categories), clinical domain (8 categories), study design (5 categories), publication year, journal and evidence strand. Domain and design assignment required the signal to appear in the title, or at least three times in the abstract, to reduce incidental matches.

2.6 Synthesis of results

Results are presented as frequency counts by method, domain, design and year, and as a qualitative divergence map (§3.5) contrasting family-derived and clinician-derived positions for 20 outcome domains. Consistent with scoping review methodology, no risk-of-bias appraisal was performed and no effect estimates were pooled.

2.7 Deviations from protocol

Two deviations are recorded. First, Europe PMC was initially omitted from the executed search and was added after the omission was identified during flow-diagram construction; the review reports the complete two-database search. Second, the first Europe PMC pass ran unbounded across full text for two blocks, returning implausible yields; those blocks were re-run restricted to title and abstract, consistent with the protocol's stated field scope.

3 Results

3.1 Selection of sources of evidence

Of 5,476 records identified (4,406 PubMed, 1,070 Europe PMC), 888 were duplicates or cross-database overlap and 571 fell beyond per-query retrieval caps and were not retrieved. Of 4,017 unique records, 137 were removed before screening (no retrievable abstract, n = 103; editorial, comment or letter, n = 34), leaving 3,880 screened and 996 included: 647 method records, 202 family-priority records, 118 core-outcome-set records and 29 neonatal method applications.

3.2 Characteristics of sources of evidence

Included records span 1994 to 2026. By record type, the largest group was trial applications and reports (n = 284), followed by methodological and statistical papers (n = 205), consensus and Delphi studies (n = 173), reviews (n = 153) and qualitative or interview studies (n = 106).

3.3 Hierarchical and ordinal methods: volume, growth and distribution

Figure 2 summarises the methods landscape.

Uptake is recent and steep: records naming a hierarchical or ordinal method rose from 13 in 2018 to 143 in 2025. The win ratio dominates, reflecting its adoption in cardiovascular outcome trials.

The concentration by clinical domain is marked. Cardiovascular research contributes 144 records naming a method, infectious disease 114, and neurology/stroke 80.

3.4 Neonatal and perinatal applications

This is the review's central finding. Against 267 included records that are neonatal or perinatal in focus — a substantial literature, dominated by core outcome sets and parent-priority work — only 9 both concern a neonatal/perinatal population and name a hierarchical or ordinal endpoint method.

Three features of this set are worth noting. Several are cardiac-surgical trials in infants rather than trials in preterm neonates, reflecting the transfer of the method from adult cardiovascular research along disciplinary rather than population lines. The clearest neonatal DOOR application is a post hoc re-analysis of an existing trial, not a trial designed around a DOOR endpoint. And five of the nine appeared in 2026, suggesting adoption may now be beginning.

The percentage columns of Table 2 make the same point method by method: the neonatal share of the win ratio literature is 0.9%, of hierarchical composite endpoints 0.5%, and of DOOR 2.3%. Q-TWiST, generalised pairwise comparisons, sliding dichotomy and the Finkelstein–Schoenfeld procedure have no neonatal application in this corpus at all.

3.5 Divergence between family-derived and clinician-derived priorities

Ten studies in the corpus elicited outcome priorities from parents, families or former patients, several comparing them directly against professionals. Each was verified against its PubMed record before charting.

Four patterns emerge.

4 Discussion

4.1 Summary of evidence

Three findings stand out. Hierarchical and ordinal endpoint methods have grown roughly tenfold since 2018 and are now routine in cardiovascular trials. Neonatology has almost entirely missed this development: 9 applications in a corpus of 996 records, several of them in infant cardiac surgery rather than preterm neonatology, and the clearest preterm DOOR example a post hoc re-analysis. And the neonatal outcome literature that is large — core outcome sets and family-priority research — has developed on a separate track, producing a well-characterised account of what families value that has not been connected to the statistical machinery capable of expressing it.

The two gaps compound. Hierarchical methods need an explicit severity ordering, which is exactly what the family-priority literature supplies. The family-priority literature identifies outcomes that cannot be accommodated in a binary composite, which is exactly what hierarchical methods can absorb. Neither field has taken up the other's contribution.

4.2 Implications

For trialists, the practical implication is that a neonatal DOOR endpoint is constructible now: the methodology is mature, published in high-profile trials, and supported by software. What is missing is an agreed hierarchy.

For hierarchy construction, the divergence map indicates that simply ordering existing endpoints by clinical severity would reproduce the professional valuation the evidence contests. Two specific corrections follow: the ordering of death against severe impairment should not be assumed, because the elicitation evidence points the opposite way from the standard composite; and any hierarchy claiming to integrate family priorities must add tiers for domains that currently have no endpoint, rather than reweighting the tiers that already exist.

The companion concept paper to this review develops such a hierarchy and evaluates its statistical behaviour by simulation.

4.3 Limitations

Screening was not dual independent human screening. Records were screened using model-assisted classification against a written rubric, a validated keyword filter, and a classifier trained on those decisions, with uncertain records adjudicated by direct reading. This is reproducible and the full decision log is published, but it is not the dual-reviewer standard, and misclassification at the margins is likely.

Charting is based on titles and abstracts, not full texts. Method, domain and design were assigned by pattern matching over title and abstract. This is appropriate for the descriptive mapping a scoping review performs, but counts for the large domains should be read as indicative magnitudes rather than exact censuses. The neonatal application count, being small enough to verify by hand, was read individually and is exact; one record initially flagged was removed on reading because it matched on “prioritised outcomes” in a Delphi-methods trial rather than a hierarchical endpoint.

Two databases only. Embase, CINAHL, PsycINFO and trial registries were not searched. Trial registries in particular would likely identify in-progress neonatal applications not yet published, and their omission means the 9 applications reported here are a lower bound on current activity.

Retrieval caps. 571 records fell beyond per-query retrieval caps and were not retrieved. These were distributed across the highest-yield queries and are unlikely to be enriched for neonatal applications, but they were not screened.

The divergence map is an interpretive synthesis. The ten anchor studies used different elicitation methods — utility measurement, Delphi, discrete choice experiment, qualitative interview — and report on no common scale. Tier positions in Table 6 and Figure 3 are the reviewer's reading of what each study reports, traceable to a named source but not a pooled quantitative estimate.

No quality appraisal. Consistent with scoping review methodology, included sources were not appraised for risk of bias, so the evidence underlying the divergence map varies in strength.

4.4 Conclusions

Hierarchical outcome methodology and neonatal family-priority research are both mature, both growing, and almost entirely disconnected. Nine records in this corpus of 996 sit at their intersection. The evidence needed to build a family-integrated hierarchical endpoint for preterm trials already exists in the literature — it has simply never been assembled into one.

Appendix A. Search strategy

Searches were executed to August 2026. Full query strings are in search_strategy.csv.

Appendix B. Data availability

All review data are published as machine-readable files: screening_log.csv (every record with decision and reason, including duplicates), included_studies.csv (996 included records), charted_evidence.csv (charting form output), divergence_table.csv and divergence_sources.csv (§3.5), prisma_flow_counts.csv, neonatal_applications.csv (Table 4) and search_strategy.csv.