Why the Scale Matters
In a sample-by-feature matrix, entries may all appear as numbers while measuring fundamentally different things. A count may record observed reads, a proportion may express a part of a total, an intensity may be an instrument response, and a score may be calculated from several inputs. The number alone does not determine which comparisons are meaningful. Before comparing entries across samples or omics layers, ask what was measured, how the value was produced, and what its units mean.
This distinction affects interpretation. Subtracting two values always produces an arithmetic difference, but that difference may not correspond to a meaningful change in the measured quantity. Dividing one value by another also produces a number, but a ratio is meaningful only when the scale supports that interpretation and the denominator is appropriate. The aim here is to classify common entries and match each scale to defensible comparisons.
What Each Scale Allows You to Compare
A count has a defined unit, such as reads assigned to a feature or detected peptide-spectrum matches. Differences between counts can describe differences in the number tallied when the counting rule and observation opportunity are comparable. A count ratio can describe a fold difference in tallies, provided the denominators and measurement process make the comparison appropriate. A count is not automatically a direct measure of underlying abundance: observed tallies may also depend on how much material was assayed and how events were detected.
A proportion is dimensionless and bounded. Its difference is often expressed in percentage points: a change from 0.40 to 0.45 is 0.05, or 5 percentage points. A ratio of proportions is a different comparison: \(0.45/0.40=1.125\), so the second proportion is 1.125 times the first. A difference of 0.05 is not the same as a 5% relative increase. At the boundaries, a proportion of zero cannot be used as the denominator of a ratio, and proportions of zero or one can constrain which comparisons are possible.
An intensity can support quantitative comparisons when measurements use a common, interpretable scale. If the signal has been calibrated and zero represents absence of the measured signal, subtraction and ratios may both have useful interpretations. If the instrument scale has an arbitrary offset, a ratio of recorded intensities may be misleading even when the difference is informative. Changing instrument settings, preprocessing, or units can also make raw values incomparable across runs or assays.
A score requires examination of its construction. Some scores are ordinal: they order observations, but the distance between adjacent values need not be equal. Others are designed as quantitative scales with interpretable units, such as standardized scores measured in reference standard deviations. A numerical score is not automatically a probability, a physical quantity, or a measurement with a meaningful zero. Ranking may be justified even when interpreting score differences or ratios is not.
| Entry type | Typical interpretation | Usually defensible | Check before comparing |
|---|---|---|---|
| Count | Number observed under a counting rule | Count differences; sometimes count ratios | Counting rule and observation opportunity |
| Proportion | Part divided by its positive total | Difference in percentage points; sometimes ratio of proportions | Same definition of part and denominator |
| Intensity | Signal on an instrument scale | Differences on a shared calibrated scale; ratios if justified | Calibration, settings, and meaning of zero |
| Derived score | Value generated by a rule or model | Ranking; other comparisons depend on construction | Scale type, reference, and validation |
Worked Examples
Worked Example: Comparing Counts Under a Defined Counting Rule
Suppose a hypothetical targeted assay reports the number of molecules assigned to one feature in two samples. The same counting rule and acquisition protocol are used, and both samples have the same stated input amount. Sample A has 120 assigned molecules and sample B has 180. What comparisons are supported by these values?
The recorded entries are counts, in assigned molecules. The absolute difference is
Under the stated equal-input and comparable-acquisition conditions, sample B has 60 more assigned molecules than sample A. The count ratio is
Thus, the observed count in sample B is 1.5 times the observed count in sample A. This is a comparison of observed tallies under the specified assay conditions, not proof that the underlying biological abundance is 1.5 times as large. If input amounts or detection opportunities differed, these raw counts alone would not establish a fair abundance comparison.
Worked Example: Reporting a Difference Between Proportions
In a hypothetical methylation assay, sample A has 42 methylated reads among 100 reads at a locus. Sample B has 54 methylated reads among 120 reads at the same locus. In each case, the numerator is a part of the total read count, and the locus and classification rule are held fixed.
The proportions are
Both values are between zero and one, as required for proportions. The difference is
The observed proportion is 0.03 higher in sample B, or 3 percentage points higher. The relative ratio is instead
So sample B's proportion is about 1.071 times sample A's proportion, a relative increase of about 7.14% because \((0.45-0.42)/0.42\approx0.0714\). These are distinct descriptions. The denominators matter to the definition of each proportion; the two fractions are meaningful here because both use the same stated locus and methylation classification.
Worked Example: Checking an Intensity Ratio
A hypothetical instrument reports signal in relative fluorescence units (RFU). On the same calibrated scale and under unchanged settings, sample A has intensity 240 RFU and sample B has 360 RFU. Assume for this example that the calibration is linear over this range and that zero RFU represents zero signal.
The difference is
The ratio is
The recorded signal in sample B is 120 RFU higher and 1.5 times the signal in sample A. The stated calibration assumptions support both interpretations on this scale. If the instrument added a constant background offset, the difference might still be interpretable after appropriate calibration, but the ratio of the uncorrected values would not generally equal the ratio of the underlying signals. For example, adding 100 RFU to each reading changes the raw ratio to \(460/340\approx1.353\), even though the difference remains 120 RFU. The example illustrates why a numerical ratio needs scale assumptions.
Worked Example: Interpreting a Derived Score
Suppose a hypothetical algorithm assigns two samples feature scores of 0.82 and 0.64. The algorithm documentation defines these values as ranking scores, but does not state that they are calibrated probabilities or that equal score differences represent equal changes in biology.
The first sample ranks above the second because \(0.82>0.64\). The arithmetic difference is \(0.82-0.64=0.18\) score units, but without further scale validation, that difference should be reported only as a numerical score difference. The ratio is \(0.82/0.64=1.28125\), but it does not establish that the first sample has 1.281 times the biological activity or confidence.
If the documentation instead established that the score was a calibrated probability, then each value could be interpreted as a probability for its defined event. Even then, the difference would be 18 percentage points in predicted probability, not a direct measurement of a biological quantity. The score's definition determines the interpretation.
A Short Justification: Why Proportions Are Bounded
Let \(a\) denote the amount or count in a part and \(t\) the positive total to which that part belongs. By definition, the part cannot be negative or exceed the whole, so \(0\leq a\leq t\), with \(t>0\). Dividing all parts of the inequality by the positive value \(t\) preserves its direction:
Therefore the proportion \(a/t\) lies in \([0,1]\), including the endpoints when the part is empty or equals the entire total. The condition that the numerator is a part of the denominator's total is essential. A quotient of two positive numbers without that relationship need not be a proportion; for instance, \(12/10=1.2\).
A related scale check applies to intensities. If an instrument reading is transformed by adding a constant \(c\), then the difference is unchanged: \((y_2+c)-(y_1+c)=y_2-y_1\). But the ratio generally changes, since \((y_2+c)/(y_1+c)\) need not equal \(y_2/y_1\). This is why ratios depend on a meaningful zero, while differences are invariant to a shared additive offset.
Diagnostics, Alternatives, and Limitations
Before calculating or interpreting a comparison, use the following sequence. It is a measurement audit, not a substitute for an assay-specific analysis.
State whether the entry is a tally, a part-to-total fraction, an instrument signal, or a derived score. Identify its unit or definition.
Verify that the same feature definition, counting or scoring rule, units, and relevant measurement conditions apply to the values being compared.
Use a difference when the scale supports differences; use a ratio only when a meaningful zero and comparable measurement process support multiplicative interpretation. For ordinal scores, ranking may be the justified comparison.
Include units or percentage points as appropriate, and state when the result describes recorded signal or score rather than an unobserved biological quantity.
Warning signs include values called proportions that exceed one, a score described as a probability without calibration, or a fold change computed from an intensity scale with an arbitrary offset. Also investigate whether counts come from comparable observation opportunities and whether intensities share calibration and settings. A calculation can be arithmetically correct yet scientifically misleading if these checks fail.
The main alternative to a raw-scale comparison is often to use a transformation or a model suited to the measurement process. For example, a logarithm can express multiplicative changes as additive differences when its input is positive and the underlying scale supports ratios. A transformation does not repair incomparable measurements, define a score, or create a meaningful zero. The choice should follow the measurement definition rather than convenience.
These categories are useful guides, not guarantees about statistical behavior. Counts are discrete, while intensity values are often recorded on a fine numerical scale; proportions are bounded; scores may be ordinal or quantitative. Each of those facts can matter for later analysis, but scale classification alone does not select a statistical model or establish that samples are exchangeable. It also does not establish causal interpretation.
Common Mistakes
- Calling any quotient a proportion. A proportion's numerator must be a part of the denominator's total; otherwise the quotient can exceed one.
- Confusing percentage points with percent change. Moving from 0.40 to 0.45 is 5 percentage points, but a relative increase of 12.5%.
- Treating raw counts as direct abundance. Counts are observed tallies; input and detection conditions can affect them.
- Assuming intensity ratios are always meaningful. A shared arbitrary offset can change ratios while leaving differences unchanged.
- Interpreting every score as a probability or physical amount. Only its documented construction and validation justify that interpretation.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A hypothetical assay reports 28 positive events among 80 total events. What is the proportion, and why must it lie between zero and one?
- A proportion changes from 0.30 to 0.36. State the difference in percentage points and the relative increase.
- Two calibrated intensities are 50 and 75 units on a shared scale with a meaningful zero. Give their difference and ratio. What assumption supports interpreting the ratio?
- A derived score is documented only as an ordinal ranking. Which comparison is justified: ranking, difference, or ratio?
- Explain why counts of 100 and 200 do not, by themselves, prove that the underlying abundance doubled.
Solutions. (1) The proportion is \(28/80=0.35\). The numerator is a part of the total, so \(0\leq28\leq80\), and division by 80 gives \(0\leq0.35\leq1\). (2) The difference is \(0.36-0.30=0.06\), or 6 percentage points. The relative increase is \(0.06/0.30=0.20\), or 20%. (3) The difference is \(75-50=25\) units, and the ratio is \(75/50=1.5\). The meaningful zero and shared calibrated scale support the ratio interpretation. (4) Ranking is justified; a numerical difference or ratio need not have meaning on an ordinal scale. (5) The tallies may reflect different input amounts or observation and detection opportunities. Without comparable conditions or an appropriate measurement model, the counts alone do not establish a doubling of underlying abundance.