A Bayesian hierarchical modeling approach can improve measurement accuracy of microcystin concentrations

Importance of the topic

The accurate measurement of microcystins (MC) in freshwater is critical for protecting public health, guiding water-management decisions, and avoiding costly false alarms during cyanobacterial harmful algal blooms (HABs). Routine monitoring programs rely heavily on calibration-based assays such as ELISA, but small calibration sample sizes and variability between tests and operators can produce large estimation uncertainty. Improving calibration precision and stabilizing concentration estimates thus has immediate practical value for environmental monitoring and risk management.

Objectives and study overview

This study evaluates a Bayesian hierarchical modeling (BHM) framework, combined with a sequential updating algorithm, to reduce calibration and measurement uncertainty for ELISA-derived microcystin concentrations. The approach was applied to 214 ELISA tests (NOAA GLERL Lake Erie HAB monitoring, 2012–2021) to compare conventional inverse-function calibration approaches with Bayesian and hierarchical shrinkage estimators. The practical goal was to provide a method that (1) improves accuracy and consistency of MC estimates, (2) can be implemented without changing laboratory workflows, and (3) can be updated easily as new data are generated.

Methods

The original ELISA procedure provided six standards (0–5.00 μg/L) with two replicates each and a quality-control sample at 0.75 μg/L. Calibration modeling followed two regimes: a log–log linear model used for early tests and, from mid-2016 onward, a four-parameter logistic (4PL) nonlinear model considered more appropriate for ELISA response curves. Key statistical and computational steps were:

• Log-transform of concentrations to stabilize errors.
• Comparison of the conventional inverse-function estimator fitted with five pooled relative-absorbance points (IFE5) versus fitting using all 12 standard replicate observations (IFE12).
• Bayesian single-test estimator (non-hierarchical Bayes) using weakly informative priors to combine fitting and inverse estimation.
• BHM variants that implement shrinkage: BHM1 (sharing information within each test across unknown samples), BHM2 (sharing calibration-curve parameter information across tests using sequential updating), and BHM3 (sharing both within- and across-test information).
• Sequential updating: posterior summaries of hyperparameters (mean and among-test variance of calibration coefficients) obtained from an initial set of tests were converted into conjugate prior parameters and used as informative priors for subsequent tests; this avoids re-fitting a large pooled hierarchical model as new tests arrive.
• MCMC-based inference implemented in Stan called from R; posterior predictive checks and Monte Carlo simulation (5,000 draws) were used to quantify accuracy for the kit quality-control (QC) sample (true value 0.75 μg/L).

Used instrumentation

• ELISA kits (Abraxis) used by NOAA GLERL for field monitoring of Lake Erie.
• Absorbance microplate measurements aggregated per-sample from duplicate wells.
• Computing: R and Stan for Bayesian model implementation and MCMC sampling.

Main results and discussion

Summary of the primary quantitative findings (accuracy reported as median absolute deviation from the known QC concentration 0.75 μg/L):

• IFE5 (five pooled standards): 0.261 μg/L — largest error and variance.
• IFE12 (twelve standard replicate observations): 0.145 μg/L — significant improvement over IFE5 (F-test p = 1.176e−12 for variance of accuracies).
• Bayes (single-test Bayesian estimator): 0.127 μg/L — improved by using weakly informative priors that limit extreme estimates.
• BHM1 (within-test shrinkage): 0.114 μg/L.
• BHM2 (across-test shrinkage with sequential updating): 0.121 μg/L.
• BHM3 (within- and across-test hierarchical model with sequential updating): 0.109 μg/L — best-performing approach overall.

Key interpretations:

• Using all 12 measured standard observations (rather than pooling into five points) materially improves calibration stability; this is an immediate practical recommendation.
• Shrinkage-based Bayesian methods consistently outperform inverse-function estimators in both accuracy and variance, because they regularize extreme curve fits and leverage information across samples and tests.
• Within-test shrinkage (BHM1) provided substantial benefit because simultaneous estimation of multiple unknown samples enables Stein-type improvements in point estimates; across-test pooling (BHM2) provided additional guardrails but its benefit was limited when between-test heterogeneity of calibration curves is large, as observed here.
• BHM3 yields the best median accuracy; however, when across-test variance is high the across-test component contributes less weight and results are dominated by the within-test level.
• Sequential updating makes BHM operationally feasible: posterior hyperparameters from a small initial set of tests are summarized into conjugate priors and then used for single-test analysis, avoiding repeated pooling of large historical datasets and heavy computation.

Benefits and practical applications

• Improved point accuracy and reduced estimation variance for microcystin measurements, translating to more reliable data for public-health decisions and resource management.
• No required changes to existing laboratory ELISA protocols or plate layouts — implementation is purely computational, so adoption cost is negligible.
• Sequential updating enables real-time incorporation of new information and efficient per-test analysis, which fits existing lab throughput and reporting timelines.
• Methods and code (R + Stan) can be wrapped in user-friendly software (for example, a Shiny app) to allow non-statistical laboratory personnel to run the approach.
• Reduces false positives and extreme estimates that could trigger unnecessary advisories or public alarms, and can reduce missed events by stabilizing estimates near management thresholds.

Future trends and potential applications

• Automation and user interfaces: packaging the sequential-BHM pipeline in a GUI (Shiny or similar) will facilitate uptake in routine monitoring labs.
• Dynamic discounting of older data: applying time-dependent inflation of hyperparameter variances or explicit discount factors to downweight outdated calibration information as assay kits and operators change.
• Extension to other calibration-based assays: the approach is applicable beyond ELISA (e.g., ligand-binding assays, immunoassays, and other analytical platforms using standard curves) where small per-run standard sizes create estimation problems.
• Hybrid models combining mechanistic assay knowledge (e.g., variability sources tied to kit lots or operator) with hierarchical structure could improve across-test pooling by explaining heterogeneity instead of shrinking toward a global mean.
• Regulatory and operational adoption: improved uncertainty quantification could be integrated into decision frameworks, e.g., flagging measurements with high posterior uncertainty for reanalysis before issuing public advisories.

Limitations

• When among-test heterogeneity is large, across-test pooling provides limited benefit unless heterogeneity sources are modeled explicitly.
• Posterior summaries used as conjugate priors approximate the joint posterior; in highly non-normal or multimodal cases this approximation could be suboptimal.
• Implementation requires computational infrastructure and basic statistical workflow changes (R/Stan or pre-built applications), which may present an initial barrier for some laboratories.

Conclusions

A Bayesian hierarchical framework with sequential updating substantially improves ELISA-derived microcystin measurement accuracy compared with traditional inverse-function calibration, especially when using all replicate standard observations. Shrinkage estimators (within-test and two-level hierarchical models) reduce both bias and variability and can be implemented without altering laboratory protocols. Sequential updating makes BHM practical for routine monitoring, and automation will facilitate adoption by monitoring programs to improve public health decision-making during HAB events.

References

The following references are representative of sources used and cited in the analysis:

• Box G.E.P., Tiao G.C., 1973. Bayesian Inference in Statistical Analysis. Addison-Wesley.
• Efron B., 1996. Empirical Bayes methods for combining likelihoods. Journal of the American Statistical Association 91, 538–550.
• Gelman A., Carlin J.B., Stern H.S., Dunson D.B., Vehtari A., Rubin D.B., 2014. Bayesian Data Analysis (3rd ed.). CRC Press.
• Gelman A., Chew G.L., Shnaidman M., 2004. Bayesian analysis of serial dilution assays. Biometrics 60, 407–417.
• Findlay J.W., Dillard R.F., 2007. Appropriate calibration curve fitting in ligand binding assays. AAPS Journal.
• Klauenberg K., Walzel M., Ebert B., Elster C., 2015. Informative prior distributions for ELISA analyses. Biostatistics 16, 454–464.
• Miller J.N., Miller J.C., 2010. Statistics and Chemometrics for Analytical Chemistry (6th ed.). Pearson.
• Nummer S.A., Weeden A.J., Shaw C., et al., 2018. Updating the ELISA standard curve fitting process to reduce uncertainty in estimated microcystin concentrations. MethodsX 5, 304–311.
• Qian S.S., Chaffin J.D., DuFour M.R., et al., 2015. Quantifying and reducing uncertainty in estimated microcystin concentrations from the ELISA method. Environmental Science & Technology 49, 14221–14229.
• West M., Harrison J., 1997. Bayesian Forecasting and Dynamic Models (2nd ed.). Springer.