In analytical chemistry, multiple methods are available for sample analysis and applying different techniques to the same sample often give varying results. This discrepancy can also occur with the same technology, such as Near Infrared (NIR) spectroscopy. A notable example is when two samples are sent to ten different laboratories; despite all using NIR spectroscopy, the results can differ significantly. 

Analysis results of the NIRS measurements of the same sample sent to ten different laboratories. Shown are the results for protein content in grass silage and corn silage. The dashed lines show the median value of the ten results. 

To address biases between different methods, it is advisable to adjust NIR calibration applications (predictions) according to your current reference analytical method before use. This practice enhances the comparability of different methods. 

When examining a large dataset of samples with diverse compositions and their corresponding “true” values for a specific analytical parameter, it is common to observe a distinct linear offset between different analysis methods. This phenomenon indicates that, while the results from the two methods may align in terms of trends, the absolute values exhibit a consistent average difference, known as bias. This bias reflects a systematic discrepancy that can affect the interpretation of results, emphasizing the need for calibration and adjustment when comparing outcomes from various analytical techniques. 

Bias correction can be effectively demonstrated using a scatter plot that displays predicted values against reference values. In this graphical representation, the presence of bias becomes readily apparent. The bias value itself is determined by calculating the differences between the results obtained from the NIR spectrometer and those from the reference analytical method. By visualizing these discrepancies, one can easily identify the extent and direction of the bias, facilitating appropriate adjustments to enhance the accuracy and reliability of the analytical results. 

A representative sample set effectively illustrating the bias between predicted values and reference values. This bias manifests as an offset between the trendline, which is derived from a linear regression of the original data, and the angle bisector, which represents the line of perfect correlation.  

After applying the bias correction to the original data, resulting in what is termed “bias-corrected data,” the data points tend to cluster more closely around the line of perfect correlation between the predicted and reference values. 

• The procedures for validating and adjusting your instrument for animal feed applications are outlined in the international standard DIN EN ISO 12099:2018-01.
• According to this standard, at least 20 samples must be measured to achieve statistically significant results. Analyzing these results allows for the evaluation of model performance and the determination of whether adjustments to the model are necessary. When selecting a validation sample set, it is critical to ensure that it is representative of the variety of sample types you will be measuring, as this directly impacts the validity and quality of the adjustments made.
  
  The sample types used in the model description must encompass all potential sample types that will be measured with the instrument and model. 

• It is essential that the sample composition is representative and evenly distributed across the expected range for all predicted parameters. Furthermore, this range should fall within the established limits of the NIR application. 

• Whenever possible, incorporate samples from different regional or seasonal origins, as well as samples from multiple sources. 

• If your application involves many different sample types, it may be necessary to include more than 20 samples to ensure adequate statistical significance and model reliability. 

A retrospective application or modification of bias correction values for previous measurements is not possible. Once a bias value is implemented, the measurements cannot be reverted to their original values. However, you can adjust the bias correction value at any time and promptly reapply it to future measurements, incorporating the updated value immediately. This flexibility allows for continuous improvement in accuracy while maintaining a clear record of changes in the analytical process. 

Every adjustment to the bias is treated as an incremental change. Since biases take effect immediately for future measurements, any modifications refer to the current configuration (and not to the original values without any bias correction). Consequently, we refer to this process as adding a bias rather than editing the existing bias correction.

When configuring bias corrections for animal feed applications, it is crucial to ensure that the reference measurement results are reported on the same basis as the trinamiX predictions. Generally, there are two common ways to express results in feed analytics:

• Percentage “as fed” (which includes moisture content)
• Percentage of “dry matter.”
• It is essential to transform the reference values to match the basis of the trinamiX results, particularly if they are expressed differently. This transformation is outlined in the Application Info and is necessary to maintain consistency and accuracy in the bias correction process.  

For example, a user measuring protein content in wheat samples initially analyzes a set of 25 samples and finds that the trinamiX values are, on average, 5% higher than his reference values. He configures a bias of -5%, which adjusts the original analysis values down by 5%, aligning the results more closely with his reference data. Importantly, all measurements taken prior to this bias configuration remain unchanged. 

After a year, the user switches his reference method and measures new samples, discovering a bias of +7% compared to his results. He then adds a bias of +7% to align with his new reference. This adjustment results in an effective bias of +2% when considering the earlier -5% bias. 

Later, the user refines his bias correction using another sample set and identifies a residual bias of +1%. He adds this to his configuration, resulting in an effective bias of +3%. If he wishes to revert to the original analysis values, he can apply a bias of -3%, which would bring the effective bias back to 0%, effectively restoring the original measurements.