Choosing a missing-data threshold

Genotype Data

Data Quality

Choose marker and sample missingness thresholds by diagnosing the data source, matching the downstream analysis, and testing sensitivity.

Author

Alexander M. Sandercock

Published

July 13, 2026

Keywords

missing data, missingness, call rate, SNP filtering, genotype filtering, imputation, GBS

There is no universal missing-data threshold that is correct for every genotype dataset. A useful threshold removes unreliable data without discarding so much information that the downstream analysis becomes weaker or less representative.

First separate two different filters

Missingness is usually evaluated in two directions.

Marker missingness

For each SNP or locus: what proportion of samples lack a usable genotype?

A marker with 30% missingness has calls for roughly 70% of retained samples.

Sample missingness

For each individual: what proportion of retained markers lack a usable genotype?

A sample with unusually high missingness may reflect poor DNA, low coverage, library failure, or assay mismatch.

Do not treat these as the same decision. A few globally poor samples can make many otherwise useful markers appear incomplete, so sample-level diagnostics often deserve attention before final marker filtering.

Why a fixed rule can fail

The same numerical threshold can behave very differently across datasets.

A 20% marker-missingness limit might be:

unnecessarily strict for a sparse sequencing dataset followed by validated imputation;
too permissive for a targeted assay where most loci are expected to genotype nearly every sample;
acceptable for one exploratory population analysis but problematic for a method that requires a nearly complete matrix;
misleading when missingness is concentrated in one family, population, sequencing batch, or taxon.

The central question is not simply “how much is missing?” It is also “why is it missing, and where?”

A defensible workflow

1. Diagnose before filtering

Plot or summarize:

missingness per sample;
missingness per marker;
missingness by sequencing batch, plate, library, family, population, or species;
read depth or genotype confidence against missingness;
the number of samples and markers retained across candidate thresholds.

A bimodal distribution is especially informative. For example, a subset of samples with 70–90% missingness may indicate failed samples rather than a reason to loosen the marker threshold for everyone.

2. Define the downstream objective

Targeted genotyping

High missingness can indicate assay failure or off-target behavior. Investigate the locus or sample before assuming imputation will rescue it.

Population analysis

Missingness that differs among populations can create or distort structure. Check whether filtering changes PCA, distances, differentiation, or clustering.

Genomic prediction

A more permissive marker set may be useful when imputation is accurate, but imputation should be validated under a design that matches deployment. Outcome information from validation samples must not leak into preprocessing.

Linkage or QTL analysis

Genotyping errors and missing parental or progeny calls can disrupt segregation and map construction. The relevant tolerance depends on the population and mapping model.

3. Compare candidate thresholds

Instead of choosing one value by habit, run a small sensitivity analysis. For example, compare several candidate marker filters spanning stricter to more permissive settings.

Illustrative sensitivity grid, not universal recommendations

A project might compare maximum marker missingness of 5%, 10%, 20%, and 50%, then track retained marker count, retained sample count, allele-frequency distribution, imputation performance, and the stability of the biological conclusion. The useful values for a particular dataset may be entirely different.

4. Evaluate the consequence, not only the retained count

For each candidate threshold, ask:

How many markers and samples remain?
Are certain families or populations disproportionately removed?
Does marker density become uneven across the genome?
Do PCA or relationship patterns change materially?
Does prediction performance change under a realistic validation design?
If imputing, how accurate are masked genotypes at the relevant allele frequencies and dosage states?

A threshold is easier to defend when the main conclusion is stable across a reasonable range of settings.

Missingness can be informative

Missing data are not always random. They may correlate with:

sequence divergence from the reference;
restriction-site polymorphism in GBS/RAD-seq;
low DNA quality;
batch effects;
allele-specific assay performance;
paralogous or homoeologous mapping problems;
species or population membership.

That means aggressive filtering can sometimes remove a biologically non-random subset of the dataset. Conversely, retaining systematic missingness can also bias inference.

Special caution for polyploids

In polyploid data, a genotype may be technically non-missing but still highly uncertain. A hard dosage call with weak posterior support is not equivalent to a confident call.

Missing-data filtering should therefore be considered alongside:

read depth;
posterior genotype probability;
allelic bias;
overdispersion;
unexpected ploidy or copy number;
locus behavior consistent with paralogous or homoeologous sequence.

Practical takeaway

Choose a missing-data threshold by combining data diagnostics + downstream requirements + sensitivity analysis. Report both sample- and marker-level rules, and document whether missing genotypes were imputed.

Continue learning

Process MADC genotype data Choose a genomic analysis path