Testing Whether Biological Ratio Indices Earn Their Denominators

Two very different fields have independently decided that a ratio beats a raw amount. Plant scientists say crops read the ratio of red to far-red light (R:FR) to decide whether they are shaded, not the total amount of light. Clinicians now argue that a patient's resilience is better captured by muscle volume divided by bone volume, the Sarcopenia Index, than by raw muscle or body mass index. Ratios are attractive because they promise invariance. Bone acts as a ruler that cancels body size, and a spectral ratio cancels lamp brightness. They are also a well-known statistical trap. A ratio used as a predictor silently assumes that numerator and denominator act with equal and opposite strength. If that assumption is false, the ratio can manufacture or hide associations. In lighting experiments the trap has a physical form: adding far-red lamps to lower R:FR also raises the total photon dose.

This note proposes a pre-registered systematic review, paired with reanalysis of individual-level data, that tests one question in both literatures. When a paper credits an outcome to a ratio, does the evidence actually support the ratio, or only one of its parts? Sarcopenia imaging is the main application, because survival and functional-outcome claims there are about to reach clinical cutoffs. Shade-avoidance lighting is the second test domain. Its controlled experiments let us check whether an apparent ratio effect is really a single-band effect. Systematic-review methodology supplies the instrument. It turns "do papers test their ratios?" into a countable, auditable quantity rather than an impression.

Core Claim

A ratio index $M/B$ entered on the log scale is a constrained special case of a model with both parts. For a survival or outcome model with linear predictor

$$\eta = \beta_M \log M + \beta_B \log B + \gamma^\top z,$$

where $M$ is the numerator (muscle volume, or red photon flux), $B$ is the denominator (bone volume, or far-red photon flux), and $z$ holds clinical or environmental covariates, the ratio hypothesis is exactly $\beta_M = -\beta_B$. That constraint can be tested with a one-degree-of-freedom likelihood-ratio test.

We claim three things, each testable. First, almost no published ratio-index claims in either field report this test or the component coefficients. Second, where data allow the test, a meaningful share of ratio effects come from one component, for example muscle volume alone, or far-red dose alone. Third, unequal measurement noise between numerator and denominator inflates ratio noise in predictable ways. One concrete case is segmentation disagreement between annotators, which on single CT slices exceeds algorithm error: 6.12 versus 5.06 cm² in Laousy et al.. The new contribution is not another index. It is a reusable audit protocol and the first quantitative answer to whether these ratios are justified.

Relation to Prior Work

Three papers motivate the work, because their headline results rest on normalization choices that were never tested against components.

Borys et al. segmented whole-body CT in 10,340 cancer patients with the Body and Organ Analysis tool. They defined the Sarcopenia Index as muscle volume over bone volume, reported HR 0.56 (abdomen) and 0.40 (thorax), and showed it outperforms body mass index and the single-slice L3 skeletal muscle index. They also checked bone stability in 250 longitudinal patients. They compared the index against other indices, not against a model containing muscle and bone separately. That missing comparison is exactly what we add.

Teo et al. used nnU-Net, a self-configuring segmentation network, on head and neck CT from stroke patients. They used absolute sternocleidomastoid and temporalis volumes, raising AUC from 0.716 to 0.731. This creates a natural contrast case where the field chose no denominator at all.

In plants, Choi et al. cut pepper time-to-harvest from 170 to 95 days using a 20-hour photoperiod with R:FR of 0.3. Their design changes photoperiod, far-red dose and ratio together. Related spectral work on red-to-blue ratios, by Zhao et al. on stomata and Lu et al. on a reported >500% pepper yield gain, likewise reports effects in ratio terms. That framing is where the photon-dose confound typically lives.

The audit method itself is borrowed. Sav, Demetriou and Liem used a PRISMA-guided audit of 100 recommender-system papers to show that 33% relied on MovieLens and 39% had incomplete dataset links. Dissecting Subjectivity and the "Ground Truth" Illusion used a structured question template covering population, intervention, comparison, outcome and context across 346 papers to show how majority-vote labels hide disagreement. We reuse their machinery of fixed extraction items, a screening audit trail and inter-coder agreement. Our difference is that we target a statistical modelling assumption instead of annotation practice, and we follow the review with reanalysis rather than stopping at prevalence counts.

Schlegel shows that reviews without explicit filtering logic fail reproducibility. This constrains us to pre-register queries and exclusions. Existing narrative syntheses, such as Syed et al. on vertical farming, report efficiency figures without that structure and serve here as the naive reference point.

Research Program

Stage one: the systematic review. We register a protocol (on PROSPERO for the clinical arm, following the practice of Luo et al.) and use two inclusion questions. The clinical question covers studies relating any image-derived muscle measure, absolute or normalized, to a clinical outcome. The plant question covers controlled-environment studies attributing developmental or yield outcomes to a spectral ratio.

Each included paper is double-coded on a fixed set of items:

  • whether components are reported separately;
  • whether the component-versus-ratio test, or anything equivalent, was run;
  • whether the denominator was shown to be stable;
  • whether segmentation or light-measurement reliability was reported, and how;
  • for plants, whether total photon flux or photoperiod was held fixed while the ratio varied.

Agreement between coders is reported as Cohen's kappa.

Stage two: recovering the test from published data. Some papers report component-level effects, or cross-tabulated arms with matched and unmatched photon flux. From these we reconstruct the test approximately. In plants, factorial lighting designs allow a meta-regression of elongation or flowering time on log red flux and log far-red flux separately.

Stage three: individual-level reanalysis. We request patient-level data from the authors of the large imaging cohorts, or run BOA and nnU-Net on any cohort we can access with outcomes. We then fit the unconstrained and constrained models above. To propagate segmentation noise, we use reported inter-annotator variability to simulate how noise in the numerator and in the denominator separately affect the index's estimated hazard ratio.

Baselines

The naive reference point is the field's current practice: accept the ratio index as reported. In the clinical arm, this means taking the Sarcopenia Index hazard ratios of 0.56 and 0.40 from Borys et al. at face value. In the plant arm, it means taking the 44% generation-time reduction attributed to low R:FR in Choi et al. at face value.

The strongest existing method is the Sarcopenia Index as validated against body mass index and the L3 index in that 10,340-patient cohort. It is the most rigorous ratio claim in the file, with external validation and an explicit argument for the denominator. In any reanalysis we will fit the index model and the competing models on identical patients, covariates and follow-up. Discrimination will be scored by C-index with the same cross-validation folds, so that no model gets extra tuning.

For the stroke setting, the reference is the clinical model AUC of 0.716 and the absolute-sternocleidomastoid model AUC of 0.731 from Teo et al.. We add a muscle-over-cervical-bone variant and fit it under the same protocol. For plant ratio-only models, no reported result was found on Lacuna for a component-separated comparison, so the baseline there is the ratio-only meta-regression we fit ourselves.

The key ablation is the unconstrained two-component model. A further single-component model (muscle only, or far-red dose only) isolates which part carries the signal.

"Beating" the index has a specific meaning here. Either the constraint $\beta_M = -\beta_B$ is rejected with a meaningful C-index gain for the unconstrained model, or a single component matches the ratio, which would show the denominator adds nothing. Failing to reject the constraint is also a result: it would give the Sarcopenia Index its first direct justification.

Why not just…?

Why not just reanalyze one big cohort and skip the review? This is the most serious alternative, and for the clinical arm alone a single well-powered reanalysis may answer the ratio question more decisively than any review. We accept that the reanalysis is the sharpest test. The review earns its place in two ways. It measures how many published cutoffs and effect sizes inherit an untested assumption, which one cohort cannot tell us. It also finds the subset of papers whose reported numbers allow approximate testing. If stage-two reconstruction proves impossible across most papers, we will report the review as a prevalence audit only and put resources into stage three.

Why not just use height-normalized indices or body mass index, as clinics do? Height normalization is itself a ratio with the same untested constraint, so it falls inside our audit rather than outside it. Body mass index already loses to volumetric measures in Borys et al. and is included only as a reference row.

Why not a quick narrative review? Sav et al. found field-wide gaps, such as 39% of papers missing dataset links, that informal reading had not exposed. Schlegel shows that unstructured reviews cannot be reproduced. Our coded items and kappa statistics make the prevalence claim checkable.

Why include plants at all? The plant arm is not needed for the clinical conclusion, and we say so. It is included because lighting experiments can physically separate ratio from dose, which observational imaging cannot. That makes the plant literature the cleaner test of whether ratio framing tends to hide single-component effects. If too few factorial designs exist, the arm reduces to a prevalence count.

Decisive Checks

Three outcomes are decisive. The first is the fraction of included imaging studies that report component coefficients. We expect it to be small, and a large fraction would undercut the motivation. The second is whether the one-degree-of-freedom test rejects equal-and-opposite coefficients in reanalyzed cohorts, together with the size of any C-index difference. The third is whether, in matched-flux plant designs, the elongation and flowering effects attributed to R:FR survive once far-red photon flux is entered separately. A null on all three would validate current ratio practice. Any one rejection would show that ratio indices need component-level reporting before clinical cutoffs are fixed.