Median vs Geometric Mean Fluorescence Intensity: Which to Report and When

Two analysts open the same FCS file, gate the same population, and report two different intensities for the same marker. Neither made a mistake. One read the arithmetic mean off the software default; the other switched to median. On a log-scaled fluorescence distribution, those two numbers are not close — and a reviewer who asks “is that the mean, the median, or the geometric mean?” is asking a question that changes the result.

Choosing between median vs geometric mean fluorescence intensity in flow cytometry comes down to the shape of the distribution and how much you trust the tails. Here is how the three statistics behave and a decision path for which one to report.

Median, geometric mean, and arithmetic mean: why they disagree

Fluorescence intensity is measured on a scale that spans several decades, and stained populations are typically log-normal — symmetric once you take the logarithm, heavily right-skewed on the raw scale. That skew is what splits the three central-tendency measures apart.

• Arithmetic mean. The plain average of the raw channel values. On a right-skewed distribution it is dragged toward the high tail, and a handful of bright outliers or aggregates pull it further. It is the least appropriate choice for log-scale fluorescence, yet it is a common software default — which is how it ends up in figures by accident.
• Median. The value with half the events above and half below. It does not care how far out the tail extends, only how many events are on each side, so outliers and aggregates barely move it. It is the same midpoint whether you read it on a linear or a transformed axis.
• Geometric mean. The arithmetic mean computed on the log-transformed values, then transformed back. It is the natural center for log-normal data and sits close to the median for a clean, symmetric-in-log population — but because it is still a mean, it remains more sensitive to outliers and to a skewed shoulder than the median is.

Why median became the default

The median is the de facto standard for reporting fluorescence intensity in peer-reviewed work, for one practical reason: robustness. Real cytometry populations are heterogeneous, carry aggregates, and have variable tails between samples. A statistic that shifts when a few bright doublets sneak past your compensation and gating is a statistic that adds noise to your comparison. The median ignores them. That stability is worth more in most experiments than the geometric mean’s slightly better fit to an idealized log-normal shape.

A decision path

Use the population’s shape and your reporting context to pick:

• Routine marker expression, heterogeneous samples, cross-sample comparison → report the median. It is robust, it is what most readers expect, and it is comparable across samples with different tail behavior.
• Clean, unimodal, log-normal population and a lab convention that uses geometric mean → geometric mean is defensible. State it explicitly. It will track the median closely when the population is well behaved; if it diverges from the median, that divergence is telling you the population is not as clean as you assumed.
• Arithmetic mean on raw fluorescence → avoid unless you have a specific reason. On log-scale data it overstates the center and is the most outlier-sensitive of the three.

The label is part of the number

“MFI” is ambiguous: the M has been read as both mean and median, and the abbreviation hides which one you used. A figure axis that says only “MFI” forces the reader to guess. Report the statistic by name — median fluorescence intensity or geometric mean fluorescence intensity — in the axis label and the methods. This is exactly the kind of metadata that has to travel with the value when you export statistics for a manuscript; the number without its statistic is not reproducible.

A summary table

One place the choice really bites: difference metrics

Anywhere you take a difference or ratio of intensities, the statistic choice compounds. The stain index — the separation between a positive and negative population scaled by spread — is built from intensity values, so reporting it with a mean instead of a median can change which fluorochrome looks brighter. If you are using intensity to compare reagents, keep the statistic consistent across the comparison; the reasoning is the same one behind using stain index to find the right antibody concentration.

The short answer for most experiments: report the median, label it as median, and use the same statistic for every sample you compare. Reach for the geometric mean only when the population is genuinely log-normal and your field expects it — and say so when you do.