Limit of detection (LOD) is calculated separately for each Olink assay and sample plate. The LOD is based on the background, estimated from negative controls included on every plate, plus three standard deviations. The standard deviation is assay specific and estimated during product validation for every panel.
For studies including more than one plate per panel, the maximum observed LOD for each assay is selected as study LOD. Consequently, all plates included in the study receive the same assay specific LOD. The estimated LOD is a conservative measurement especially in large multiplate studies where there is high probability that observed data is in fact above the true background signal.
Consider excluding assays with low detection from analysis
Olink recommends that assays with a large proportion of samples below LOD is excluded from the analysis. The limit for exclusion should be decided on a study basis and consider design including sample size and experimental variables. Suitable exclusion limits may be in the range of less than 25-50% of samples above LOD.
Characteristics of data below LOD
As with all affinity based assays, data from Olink’s platform have a S-curve (sigmoid) relationship with the true protein concentration in a sample. Data below LOD have a higher risk to be in the non-linear phase of the S-curve meaning that 1 NPX difference may not correspond to 2x protein concentration in this region. This may bias estimates including data below LOD and should be considered when interpreting any results that are based on data below LOD.
Strategies for handling data below LOD in data analysis
Several strategies exist for handling data below LOD that varies in complexity. Olink delivers data below LOD to allow researches to choose the strategy that is best for their study and interpret results with the complexity of data below LOD in mind. Some examples of strategies include:
- Replace data below LOD: It is common to replace data below LOD with a specific value. This will left-censor the data which creates a skewed distribution. Estimates of, for example, mean will be biased and parametric statistical tests may have lower statistical power. Common values to use for replacement is the value for LOD or LOD/√2. The latter have been reported in literature to give less biased estimate of means.
- Use actual data below LOD: As data below LOD may be non-linear, estimates of for example mean may be biased. However, especially in large multiplate studies LOD is a conservative measurement. Using actual data may increase the statistical power and give a less skewed distribution compared to replace data below LOD with a value.
- Impute data below LOD: A more complex approach for handling data below LOD is to impute the true value. Several methods for imputation exist and includes maximum-likelihood estimation (MLE) of the distribution below LOD.
- Set data below LOD to missing: Olink do not recommend that data below LOD is excluded from analysis as the most distinct biomarkers may have a low concentration under specific conditions.