How to Calculate the Limit of Detection (LOD)

The limit of detection (LOD) is calculated by measuring the variability in your blank or low-level samples and then applying a multiplier that accounts for the risk of false positives and false negatives. The most widely used formula is LOD = LoB + 1.645 × SD of a low-concentration sample, where LoB (limit of blank) captures how much noise your instrument generates when nothing is there, and the second term captures how much signal a real but tiny amount of analyte produces.¹ But that single formula glosses over quite a bit. Several competing methods exist, regulatory bodies recommend different ones depending on the field, and the method you choose can meaningfully change the number you report.

What LOD Actually Means

LOD is the lowest concentration of a substance your method can reliably tell apart from “nothing there.” It sits between two related limits that are worth knowing because they frame what LOD does and does not promise. Below LOD is the limit of blank (LoB), the highest reading you expect when testing a sample that truly contains zero analyte. Above LOD is the limit of quantitation (LoQ), the lowest concentration at which you can not only detect the substance but also measure it with acceptable accuracy and precision.¹ In practical terms, a result between LoB and LOD is suspicious but unreliable. A result above LOD tells you the substance is present. A result above LoQ tells you how much is present.

The clinical world makes these distinctions routinely. High-sensitivity cardiac troponin assays, for instance, use LoB, LoD, and LoQ to define how low a troponin concentration the assay can meaningfully report. The LoQ for a given laboratory may or may not match the LoD, depending on how tight the lab’s precision goals are.² If your work only requires you to say “detected” or “not detected,” LOD is the number you need. If you also need to report a concentration, LoQ is your real floor.

The Blank-Based Method

The most common approach starts by measuring blank samples repeatedly. You run a batch of blanks, samples known to contain none of the analyte you are looking for, and record the instrument readings. Even a perfect blank will not read zero every time; electronic noise, reagent impurities, and minor temperature shifts all generate small signals. The average of those readings plus 1.645 times their standard deviation gives you the LoB. That 1.645 multiplier corresponds to a 95% confidence level on one side of the distribution, meaning only about 5% of blank readings would be expected to exceed LoB.¹

Next, you prepare samples spiked with a low but known concentration of the analyte and measure those. LOD equals your LoB plus 1.645 times the standard deviation of those low-concentration sample measurements.¹ The second 1.645 factor controls the false-negative rate: it ensures that when a true low-level sample is present, you will catch it at least 95% of the time. Together, the two steps keep both the false-positive rate and the false-negative rate at roughly 5%.

This is the approach recommended by the Clinical and Laboratory Standards Institute in its EP17 guideline, which is widely followed in clinical chemistry and diagnostic testing.³ It is conceptually straightforward, but it does require you to run a fair number of replicates, typically at least 60 blank measurements and a similar number of low-level spiked samples, to get stable estimates of the standard deviations.

The Signal-to-Noise Ratio Method

In chromatography and spectroscopy, analysts often estimate LOD by comparing the height of the analyte peak to the baseline noise. You inject a low-concentration standard, measure the peak height, measure the noise in a region of the chromatogram where nothing is eluting, and compute the ratio. A signal-to-noise ratio of 3 is the conventional threshold for detection; a ratio of 10 is the conventional threshold for quantitation. This approach is recognized in the United States Pharmacopeia, the European Pharmacopoeia, and other international guidelines.⁴

The appeal is simplicity: you can estimate LOD from a handful of injections. The weakness is subjectivity. Where you draw the boundaries for measuring baseline noise matters, and two analysts looking at the same chromatogram can get different answers. Automated software helps, but different software packages use different algorithms for defining “noise,” which can shift the reported LOD. Comparisons of the signal-to-noise method with the blank-based and calibration-curve methods have shown they can produce meaningfully different LOD values from the same data set.⁵

The Calibration Curve Method

A third common approach uses the regression statistics of a calibration curve. You prepare a series of standards at known concentrations, plot signal against concentration, fit a line, and compute LOD from the residual variability around that line. The formula recommended by ICH (the International Council for Harmonisation of pharmaceuticals) is LOD = 3.3 × σ / S, where σ is the standard deviation of the response (often estimated from the y-intercept residuals of the regression) and S is the slope of the calibration curve.⁶ The related quantitation limit uses a multiplier of 10 instead of 3.3.

This method is popular in pharmaceutical analysis because calibration curves are already part of routine method validation. You do not necessarily need extra blank measurements; you can extract the variability estimate from data you were going to collect anyway. However, how you estimate σ matters enormously. You can take it from the standard deviation of the y-intercepts of several calibration curves, from the residual standard deviation of a single regression, or from replicate measurements at the lowest calibration point. Each of those choices can lead to a different LOD, which is a persistent source of confusion. A tutorial reviewing these options noted that “alternative forms for calculating LOD frequently lead to dissimilar results,” even within the calibration-curve family of methods.⁷

Why Different Methods Give Different Numbers

The reason these three approaches do not converge on the same LOD is that they measure different aspects of variability. The blank-based method captures noise in a real blank matrix. The signal-to-noise method captures instrument noise in a specific region of a chromatogram. The calibration-curve method captures scatter around a fitted line. In a perfectly behaved system, these would be similar. In a real lab with real samples, they often are not.

A comparison of LOD procedures for organochlorine pesticides in water and sediment found that the values from laboratory-fortified-blank experiments were significantly different from those obtained by the signal-to-noise and calibration-curve methods, especially once sample concentration factors were taken into account.⁵ This is not a minor academic footnote. If your reported LOD depends on which formula you picked, the claim “our method detects X down to Y parts per billion” means different things depending on the calculation behind it. When comparing LODs across papers or product specifications, always check which method was used.

How Matrix Effects Change the Picture

LOD is not a fixed property of an instrument. It depends on what else is in your sample. In mass spectrometry, co-eluting compounds from the sample matrix can suppress or enhance the ionization of the analyte, shifting sensitivity up or down in ways that a clean standard curve cannot predict.⁸ In X-ray fluorescence, increasing the matrix concentration from near-zero to 10% ammonium nitrate raised the detection limit for potassium by roughly 17-fold and for strontium by roughly 16-fold.⁹

The practical upshot: an LOD calculated using clean standards or purified water does not automatically apply when you switch to blood, soil, food, or wastewater. If your real samples contain a heavy matrix, you should determine LOD in that matrix, not just in a clean solvent. Otherwise, you may report the method as more sensitive than it actually performs on real-world samples.

The EPA’s Two-Track Approach for Environmental Work

Environmental laboratories in the United States follow a procedure set by the EPA for what it calls the method detection limit (MDL). The current version involves two parallel tracks: one based on the statistical spread of measurements of unloaded blank samples, and another based on lightly spiked samples. The reported MDL is whichever estimate is higher.¹⁰ Taking the maximum of the two estimates is conservative by design: it protects against the situation where blank noise alone underestimates the detection floor because low-level analyte behavior adds variability the blanks cannot capture. If you work in environmental testing, this dual-track MDL is the number auditors and regulators will ask for, even if other LOD formulas might give a lower (and more flattering) result.

Non-Linear Calibration Curves

All three standard methods assume a linear relationship between signal and concentration at low levels. Many analytical systems do not behave that way. Optical sensors for anions, immunoassays, and biosensors often produce sigmoidal or otherwise curved calibration profiles at low concentrations. The traditional IUPAC method for LOD has been adapted for these cases by computing the slope of the calibration curve from the first derivative of a simplified model at low analyte concentration, then feeding that slope into the LOD formula.¹¹ If you are working with a system whose response clearly curves at the low end, using a linear regression to estimate LOD will produce a number that has little connection to the method’s actual sensitivity. Check whether your calibration data are reasonably linear in the concentration range near your expected LOD before committing to a linear formula.

LOD for Counting-Based Methods

Digital PCR and similar molecular techniques work differently from concentration-based assays. Instead of measuring a continuous signal, they partition a sample into thousands of tiny reactions and count how many are positive. The underlying statistics shift from Gaussian to Poisson: the expected variation among replicates follows the pattern of random sampling of discrete events. Droplet digital PCR measurements of HIV DNA have confirmed this, with the coefficient of variation increasing with the template number raised to roughly the 0.5 power, consistent with Poisson behavior.¹²

For these methods, LOD is driven by the number of partitions and the false-positive rate per partition rather than by baseline noise and standard deviation. A Poisson correction adjusts the raw positive count to account for the chance that a single partition received more than one template molecule. Real-time digital PCR platforms use individual amplification curves to filter out false-positive droplets, which tightens the LOD.¹³ In practice, the LOD for a digital PCR assay is often reported as a certain number of copies per reaction or per milliliter of input, established empirically by testing serial dilutions and identifying the lowest level at which the assay reliably distinguishes positive from negative. A LAMP assay for the parasite Entamoeba histolytica, for example, achieved an LOD of one trophozoite when tested with DNA extracted from spiked stool samples.¹⁴

False Positives, False Negatives, and the Error Rates You Are Accepting

Every LOD calculation implicitly accepts a certain rate of wrong answers. The blank-based method with its 1.645 multipliers sets both the false-positive and false-negative rates at about 5%. You can tighten one or both of those by choosing a larger multiplier, say 2.33 for a 1% false-positive rate, but that pushes the LOD higher. Loosening the rates lowers LOD but means more incorrect calls. The framework is the same one used in any binary decision involving noisy measurements: the analyte is either there or it is not, and your measurement will sometimes get the answer wrong in either direction.¹⁵

In high-stakes applications like clinical diagnostics or drinking-water safety, the acceptable error rate is not arbitrary. A false negative on a cardiac troponin test could mean sending a patient home during a heart attack. A false positive on a water-contamination test could shut down a municipal supply unnecessarily. The LOD you calculate should reflect the consequences of each type of error for your specific application, not just the default 5%/5% convention.

Bayesian Alternatives

The classical LOD framework is frequentist: it defines the detection limit in terms of repeated hypothetical experiments and fixed error rates. Bayesian approaches offer a different perspective. Rather than asking “what concentration would be detected 95% of the time in repeated experiments,” a Bayesian detection limit asks “given this particular measurement and everything I know about the measurement uncertainty, how confident am I that the analyte is present?” This framework naturally incorporates all sources of uncertainty, not just the blank noise and low-level sample variability, and it produces a posterior probability rather than a yes/no decision threshold.¹⁶

Bayesian detection limits have been developed most thoroughly in radiation measurement, where background counts are inherently variable and prior information about the sample is often available. They are less common in routine chemical analysis, partly because the math is more involved and partly because regulatory frameworks are built around the classical approach. But if you are working with very low count rates, uncertain backgrounds, or cases where prior information would meaningfully narrow the estimate, the Bayesian framework can give more informative answers than the classical one.

Transferability Across Laboratories

An LOD determined in one lab does not automatically apply in another. Instruments of the same model differ in baseline noise, reagent lots vary, and individual analysts have subtly different sample-handling habits. A multi-laboratory mass spectrometry study found that before normalization, the median coefficient of variation for peptide measurements across sites was about 47%, dropping to about 21% after normalization.¹⁷ That level of inter-site variability propagates directly into LOD. If your blank noise doubles at another site, your LOD roughly doubles too.

This is why method-validation protocols typically require each laboratory to verify its own LOD rather than adopting the number from the method developer’s publication. “Verified” means the lab runs its own set of blanks and low-level samples under its own conditions and confirms that it can achieve the claimed LOD. If it cannot, the lab must report a higher one. Published LODs in papers and instrument brochures represent what is achievable under optimal conditions, not what every lab will achieve in routine practice.

Decades of Terminology Confusion

If you have ever been confused by overlapping terms like detection limit, method detection limit, minimum detectable concentration, sensitivity, lower limit of detection, and instrument detection limit, you are in good company. The analytical chemistry community has wrestled with inconsistent terminology for decades. A historical review of the field noted that there had been “decades of confusion and miscommunication regarding the underlying concepts and terminology,” and that coordinated documents from IUPAC and ISO were developed specifically to provide “a harmonized position on standards and recommendations” for the first time.¹⁸

Despite those harmonization efforts, different regulatory agencies still use slightly different definitions and procedures. The EPA’s MDL is not identical to the CLSI’s LoD, and neither is exactly the same as what ICH calls the detection limit. When reading a reported LOD, pay attention not just to the number but to which protocol produced it. A method claiming an LOD of 0.1 parts per billion under one protocol might report 0.3 parts per billion under another, from identical data, simply because the formulas differ in what variability they capture and what error rates they assume.

Common Mistakes When Reporting LOD

A few errors show up repeatedly in published methods and instrument specifications:

Using too few replicates: Calculating a standard deviation from five or six measurements gives a shaky estimate. Small sample sizes can make LOD look artificially low if you happen to get a few consistent readings, or artificially high if one outlier inflates the spread.
Ignoring the matrix: Reporting an LOD from standards dissolved in pure solvent and applying it to biological or environmental samples overstates the method’s real-world sensitivity, as matrix effects can inflate detection limits by an order of magnitude or more.
Conflating LOD and LoQ: Saying a method “detects down to X” when X is actually the quantitation limit, or vice versa, misrepresents what the method can do at that concentration. Detection means the analyte is there. Quantitation means you can say how much.
Cherry-picking the method: Choosing whichever LOD formula gives the lowest number without disclosing the calculation method makes comparisons between studies meaningless.

LOD is not a fixed stamp of quality. It is a probabilistic statement about how your method performs under specific conditions, using a specific statistical framework, at a specific acceptable error rate. Understanding which calculation method you are using and why gives you a defensible number. Ignoring those details gives you a number that may look good on paper but falls apart in practice.

1
PubMed Central. Limit of blank, limit of detection and limit of quantitation
2
Radcliffe Cardiology. Role of High-sensitivity Cardiac Troponin in Acute Coronary Syndrome
3
PubMed. CLSI EP17-A protocol: a useful tool for better understanding the low end performance of total prostate-specific antigen assays
4
Pharmaceutica Analytica Acta. About estimating the limit of detection by the signal to noise approach
5
PubMed Central. Limit of detection and limit of quantification development procedures for organochlorine pesticides analysis in water and sediment matrices
6
Separation Science. Chromatographic Measurements, Part 5: Determining LOD and LOQ Based on the Calibration Curve
7
PubMed Central. Binding the gap between experiments, statistics, and method comparison: A tutorial for computing limits of detection and quantification in univariate calibration for complex samples
8
PubMed Central. Biological Matrix Effects in Quantitative Tandem Mass Spectrometry-Based Analytical Methods: Advancing Biomonitoring
9
Spectrochimica Acta Part B: Atomic Spectroscopy. Matrix effect on the detection limit and accuracy in total reflection X-ray fluorescence analysis of trace elements in environmental and biological samples – Section: Matrix concentration effect on DL
10
PubMed. Application of the U.S. EPA procedure for determining method detection limits to EDXRF measurement of filter-based aerosol samples
11
PubMed. An IUPAC-based approach to estimate the detection limit in co-extraction-based optical sensors for anions with sigmoidal response calibration curves
12
PLoS ONE. Highly Precise Measurement of HIV DNA by Droplet Digital PCR – Section: Results
13
PubMed Central. Real-time digital polymerase chain reaction (PCR) as a novel technology improves limit of detection for rare allele assays – Section: Results
14
PubMed Central. Loop-mediated isothermal amplification (LAMP) reaction as viable PCR substitute for diagnostic applications
15
PubMed. True and false positive rates in maximum contaminant level tests
16
PubMed. Bayesian decision threshold, detection limit and confidence limits in ionising-radiation measurement
17
Nature Communications. Multi-laboratory assessment of reproducibility, qualitative and quantitative performance of SWATH-mass spectrometry
18
Analytica Chimica Acta. Detection and quantification limits: origins and historical overview