Why Diagnose Missingness Before Doing Anything Else?
A feature table may look like a rectangular array, but some cells can be unobserved: a peptide may not have been quantified in a sample, or a metabolite peak may be absent from the recorded table. The locations of those gaps matter. Missingness concentrated in a few samples can point to a different issue than missingness concentrated in a few features, while gaps that cluster in one assay deserve attention even when another assay is nearly complete.
The goal here is diagnosis, not repair. Summaries and maps show where the gaps occur and help formulate questions about assay behavior and data handling. They do not, by themselves, identify why a value is missing or justify replacing it with an estimate.
The word eligible is essential. A denominator should include only cells that were meant to be measured and are within the scope of the table being summarized. A sample not run in a particular assay, or a feature not included in a particular assay panel, is not automatically a missing measurement in that assay. Define the table and its eligible cells before counting. Also use the dataset’s conventions consistently: a recorded zero is not a missing value unless the data documentation explicitly defines it as a missing-value code.
Rates and Their Denominators
For a complete \(n\)-by-\(p\) table of eligible cells, the sample-wise missingness rate for sample \(i\), the feature-wise rate for feature \(j\), and the overall rate are:
Rates can be reported as proportions or percentages, but state which. The denominator makes a rate interpretable: \(m_i=0.25\) means that one quarter of the eligible feature entries for sample \(i\) are missing. If eligibility varies by row or column, replace \(p\) or \(n\) with the number of eligible cells for that sample or feature, and exclude ineligible cells from the overall denominator. A rate with a zero denominator is undefined, not zero.
The overall rate is a cell-weighted summary. With a complete table, it equals the mean of the sample-wise rates and also the mean of the feature-wise rates:
For example, each sample contributes \(p\) cells, so the total number of missing cells is \(\sum_i p m_i\). Dividing by the total \(np\) cells gives \(m_{\mathrm{all}}=\sum_i m_i/n\). Counting by columns gives the same total missing cells, \(\sum_j n f_j\), and hence \(m_{\mathrm{all}}=\sum_j f_j/p\). If eligibility differs across rows or columns, these simple equal-weight averages need not hold; calculate the overall rate directly from eligible cells.
Worked Example: Locate Gaps in a Small Feature Table
Consider this hypothetical table for four samples and five features. “O” means observed and “M” means missing. Assume every displayed cell was eligible for measurement.
| Sample | F1 | F2 | F3 | F4 | F5 |
|---|---|---|---|---|---|
| S1 | O | O | M | O | M |
| S2 | O | M | M | O | O |
| S3 | O | O | O | O | O |
| S4 | M | M | M | O | O |
For S1, two of five entries are missing, so \(m_1=2/5=0.400\), or 40.0%. The sample-wise rates for S2, S3, and S4 are \(2/5=40.0\%\), \(0/5=0\%\), and \(3/5=60.0\%\), respectively. S4 has the highest sample-wise rate in this table.
Feature F3 is missing for S1, S2, and S4, so its feature-wise rate is \(f_3=3/4=0.750\), or 75.0%. For comparison, F1 is missing once, \(f_1=1/4=25.0\%\); F2 is missing twice, \(f_2=2/4=50.0\%\); F4 is never missing; and F5 is missing once. There are seven missing cells out of \(4\times5=20\), so the overall rate is \(7/20=0.350\), or 35.0%. The row counts \(2+2+0+3\) and the column counts \(1+2+3+0+1\) both give seven.
The pattern is more informative than the overall percentage alone: the gaps are concentrated in F3 and S4 is comparatively incomplete. This description does not explain why. In particular, it does not establish that F3 is biologically absent or that S4 failed for a technical reason.
Reading Missingness Maps
A missingness map displays the observed/missing status of many cells at once. A simple binary map uses one mark or color for observed entries and another for missing entries. A useful map retains sample and feature labels, or makes clear how rows and columns correspond to those labels. Sorting samples or features by their missingness rates can make blocks easier to see, but the original identifiers and ordering must remain recoverable.
Look for patterns at more than one scale: isolated gaps, a feature missing in many samples, a sample missing many features, a contiguous block, or a pattern shared across a subset of samples. Blocks may reflect sample batches, assay runs, or groups of features, but the map alone does not tell which explanation is correct. If the layout groups samples or features, check whether that grouping was chosen independently of the missingness pattern; a visually striking arrangement can be created by sorting on the very pattern being interpreted.
A map is most useful alongside the rates. Rates quantify how much is missing, while the map shows how those gaps are arranged. Two tables can have the same overall rate but very different structures: one can have evenly scattered gaps, and another can have a fully missing feature or a single highly incomplete sample. The distinction can affect later analyses, so an overall percentage should not stand in for inspecting the pattern.
Worked Example: Compare Missingness Across Two Assays
Suppose a hypothetical study has four samples measured in both an RNA assay with five eligible features per sample and a protein assay with three eligible features per sample. The RNA table has sample-wise missing counts of 0, 1, 1, and 2. The protein table has counts of 1, 2, 2, and 1.
The RNA assay has \(0+1+1+2=4\) missing cells among \(4\times5=20\) eligible cells. Its overall missingness rate is \(4/20=0.200\), or 20.0%. The protein assay has \(1+2+2+1=6\) missing cells among \(4\times3=12\) eligible cells, giving \(6/12=0.500\), or 50.0%.
On these denominators, the protein table has the higher fraction of missing entries. The raw counts, four versus six, do not make that comparison fairly because the assays have different numbers of eligible cells. Nor does the rate comparison establish that the protein assay performed worse: the assays measure different features and may use different measurement processes. The result is a descriptive signal to investigate, not a causal conclusion about assay quality.
Checking for a Possible Detection-Limit Pattern
Missingness summaries can be paired with the observed measurement values. A detection-limit effect is plausible when values are more likely to be unobserved at the low end of an assay’s measurement range. The earlier material on biological zeros, detection limits, and censored signals distinguishes these cases conceptually; here the practical diagnostic is to compare missingness locations with observed values and assay documentation. A missingness map alone has no measurement magnitudes, so it cannot show whether gaps occur near a low-intensity boundary.
Worked Example: Inspect Values Near a Documented Assay Floor
In a hypothetical intensity assay, a feature has measurements for six samples. The instrument documentation gives a detection threshold of 10 arbitrary intensity units. The recorded values are missing, missing, 11, 12, 18, and 25.
The feature-wise missingness rate is \(2/6=0.333\), or 33.3%. Both observed values closest to the documented threshold, 11 and 12 units, are just above it; the other observed values are higher. This pattern is consistent with a possible low-end detection effect, but it does not prove that the two missing entries were below 10 units. Other causes, such as sample handling or processing, could also produce missing values. The units matter: 11 and 12 are arbitrary intensity units, not probabilities or percentages.
For comparison, imagine a second feature with two missing entries out of six but observed values of 40, 55, 70, and 80 units. It has the same missingness rate, \(2/6=33.3\%\), but the observed values do not cluster near the stated threshold. This comparison shows why a rate alone cannot diagnose a detection-limit mechanism. The threshold, feature-level values, and relevant assay records are needed to assess plausibility.
Assumptions, Diagnostics, and Alternatives
These calculations are straightforward, but their interpretation depends on careful data preparation. Before counting, confirm that missing-value codes have been decoded correctly, that zeros have not been silently converted to missingness, and that the table contains the intended sample-by-feature orientation. Verify which entries were eligible for measurement, especially when assays cover different feature sets or samples. For each reported rate, retain the numerator and denominator so a reviewer can reproduce it.
Useful checks include confirming that each sample or feature has the expected number of eligible cells; checking that counts are between zero and their denominators; and recomputing the total number of missing cells by rows and columns. Inspecting the distribution of sample-wise and feature-wise rates can reveal outliers hidden by an overall mean. In a map, compare labels with metadata and assay records rather than inferring technical causes from visual blocks.
| Diagnostic | What it shows well | What it cannot establish alone |
|---|---|---|
| Overall rate | Fraction of eligible cells missing | Whether gaps are concentrated in particular samples or features |
| Sample-wise rates | Samples with unusually many missing entries | Why those samples have gaps |
| Feature-wise rates | Features missing across many samples | Whether missingness reflects low abundance, assay design, or another cause |
| Missingness map | Clusters, blocks, and shared gap patterns | Measurement values near a detection threshold or a causal mechanism |
| Map paired with values and assay records | Whether gaps coincide with low observed values or known processing conditions | Proof of the unobserved values or a definitive explanation without further evidence |
These tools answer complementary questions. A single overall rate is compact but can hide concentration. Sample-wise and feature-wise rates locate concentration numerically but discard the arrangement of the remaining gaps. A map preserves arrangement, yet can be difficult to read for a very large table and does not provide a concise summary. Comparing assays adds context, but is meaningful only after checking their eligible cells and measurement scopes. Neither substituting values nor restricting to complete entries is a diagnostic substitute for documenting the original missingness pattern; those are separate analysis decisions.
A further limitation is that descriptive patterns do not classify the missingness mechanism. A high feature-wise rate does not show whether missingness depends on its unobserved value, and a block across samples does not identify whether a batch, biological subgroup, or data-processing step produced it. The summaries can motivate targeted checks, but they are not a test of a missingness assumption and do not reveal the values that were not recorded.
Common Mistakes
- Using the wrong denominator. Dividing by all cells in a combined table when some were never eligible for measurement changes the meaning of the rate. Define eligibility and use the number of eligible cells.
- Calling an observed zero missing. Zero and missing are distinct unless documentation explicitly encodes zero as a missing-value code. Check the assay’s measurement convention before recoding.
- Reporting only the overall rate. The same overall rate can arise from evenly scattered gaps or severe concentration in one sample or feature. Report sample-wise and feature-wise summaries and inspect the map.
- Treating a pattern as its explanation. A block of missing entries may suggest a shared process, but it does not establish a batch effect or a biological cause. Use metadata and assay documentation to investigate.
- Claiming a detection limit from missingness alone. A map shows where values are absent, not their unobserved magnitudes. Compare observed values with a documented threshold and keep the conclusion provisional.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A table has 10 samples and 8 eligible features. One sample has 3 missing entries. What is its sample-wise missingness rate?
- For one feature, 4 of 12 eligible samples are missing. State the feature-wise rate and its interpretation.
- A complete table has 5 samples and 4 features, with 6 missing cells. What is the overall missingness rate?
- Two assays have missingness rates of 20% and 45%. What can and cannot be concluded from this comparison?
- A feature has several missing values, while its observed values are close to a documented low detection threshold. What is a defensible interpretation, and what remains unknown?
Solutions. (1) \(3/8=0.375\), or 37.5% of that sample’s eligible feature entries. (2) \(4/12=0.333\), or 33.3%; one third of the eligible samples lack an observed value for that feature. (3) There are \(5\times4=20\) eligible cells, so \(6/20=0.300\), or 30.0%. (4) The second assay has a higher fraction of eligible cells missing in the tables being compared. The rates alone do not show that its measurement process is worse or explain the difference; assay scope and eligibility also matter. (5) A low-end detection effect is plausible and should be checked against assay documentation and other records. The missing values’ actual magnitudes and the cause of their absence remain unknown.