BQS Contact Information
Brooke Connor (303) 236-1877
Delicia Beaty (303) 236-1817
Terry Schertz (303) 236-1835

Fax - (303) 236 1880

USGS-WRD BQS
DFC, Bldg 95, MS 401
PO Box 25046
Denver, CO 80225

Long Term-Method Detection Levels (LT-MDL)

**VIEW LTMDL DATA CHARTS**
SELECT A YEAR RANGE
LTMDL data are listed by
(WY data collected - WY LTMDL in effect)
	pre-2005	2009-2010 (spiking starts May 1, 2008
	2006-2007	2010-2011
	2007-2008 (in use)	2011-2012
	2008-2009 (review in process)	2012-2013

Long-Term Method Detection Level Definition

long-term method detection level (LT-MDL) is the lowest detectable concentration of an analyte that can be measured and reported with 99% confidnce that the concentration is greater than zero. At the LT-MDL concentration, one can be confident that the method will differentiate an analyte that is truly present and quantifiable in an environmental sample from lab background-noise or contamination. The LT-MDL limits the chance of a false positive (stating an analyte is present when absent) to one percent or less.

-- show me what an MDL is --

Formula

The LT-MDL is derived from the standard deviation of about 24 low concentration spiked blind sample measurements collected over the course of a year (no less than 3 months), which is multiplied by the student’s t value for n-1 observations. Alternatively, if an analytical method produces a reading for a blank (including negative values), the standard deviation of the blank data (rather than spikes) can be used to determine the LT-MDL. In this case, 24-50 blind blank sample measurements are collected over a year. An LT-MDL is only applicable to analyses that have values approaching zero, and are not arbitrarily scaled (such as temperature or pH).

LT-MDL = s x t

Where:
s = standard deviation of, ideally, 24 samples (no less than 7)
t = student’s t value for n-1 observations

The LT-MDL is set to limit the false positives to less than or equal to one percent. The LT-MDL is always set based on hind-sight. In other words, the previous year ’s data are used to determine the LT-MDL to use this year. This can be frustrating when past events limit current flexibility. However, in an effort to minimize the number of times an LT-MDL might change in its lifetime, the once per year assessment was determined to be optimal for data users.

Setting the LT-MDL too high will over protect against false positives so that more real detections are censored. Setting the LT-MDL too low will result in more environmental sample results released that are due to laboratory background and not from the environmental sample itself. The assumption is that the standard deviation of blank values, times the student ’s t value will result in the concentration equal to about the 99th percentile of the blanks. This assumption relies on a Normal Distribution.

The Correct Spiking Concentration

The purpose of using a spiked sample to calculate the LT-MDL is to overcome the limitation that some methods do not produce signal for blank results. Without knowing the blank population, it is not possible to know where false positives might be more likely to occur. The theory is that a very low concentration spike will have essentially the same distribution as a blank, so very low concentration spike samples can be used to determine the LT-MDL of analyses in the absence of blanks. For information rich analyses (such as mass spectometer methods) the LT-MDL is most often associated with the instrument sensitivity, and not the presence of interferences. More work is being done to investigate better ways to identify the detection level for information rich methods.

You must spike at a concentration where the distribution of results is very similar to the expected blank concentration. The spike concentration should be high enough to be reliably detectable (otherwise you won ’t have any data) and low enough to be able to assume a similar distribution of results as the blank. This is accepted to be about 2 – 5 times the expected LT-MDL.

n alternate process used to assure that the LT-MDL appropriately accounts for blank detections is to use the “Blank-corrected LT-MDL” for methods that produce blank signals. Here the LT-MDL is based on the original goal of setting the censoring level at less-than 1% false positives. To do this, in a set of 100 blank results, the LT-MDL would be set at the 99^th highest ranked value (the second highest). We typically use data sets of around 40 blanks, so setting the LT-MDL equal to the second highest ranked blank value is not as conservative an approach, but an easy process that is easy to visualize and implement.

Setting the LTMDL at the 1% F+ level

Outlier Corrections

utlier detection is employed to keep the LT-MDL as representative as possible. Only the data points that truly represent day-to-day method variability - be they good, bad, or ugly – are kept in the data set. And, the only data that should be removed are those that don’t belong, such as wrong samples, mis-preserved samples, or known sample blunders. The purpose of the LT-MDL data collection process is to assess the total variability of a method, not to limit it.

To remove true outliers, the first line of assurance is to request reanalysis of LT-MDL or Blind Blank samples that fall outside of plus or minus two times the LT-MDL in use. The laboratory will only replace a sample result if the new result is significantly different than the first, and verified.

Then, an outlier test is applied - the Grubbs Test. This is used to eliminate only the most aberrant single data points that individually skew a data set. The Grubbs test will check for instances where a single (or very few) data point(s) are not known to be in error (a rerun produced the same result), and are obviously completely outside of the norm. The reason you would want to apply this filter to the data is because this single aberrant point could increase the LT-MDL so much that 99% of the lab’s customers would be negatively affected by overly-censored data, to save 1% of the customers from a false positive. Granted, this decision is a matter of choice.

There will be data generated under compromised conditions that was later corrected and the data returned to normal. If it can be argued that these results do not represent the normal operational range of the method, it may be appropriate to remove these data points. This practice can get subjective and should be used only in extremely rare instances. An example of when this practice would be acceptable is if a particular lot of sample bottles used only for the LT-MDL test samples (not for environmental samples or lab use) was found to be contaminated and the lot was taken out of use. The data collected during the time the contaminated bottles were used could be manually removed since the variability seen is not representative of the operation of the method, only of the LT-MDL samples.

Data Collection

Twenty-four spiked, and/or between 24-50 blind-blank samples are assessed each year for each analyte to determine if an LT-MDL needs to be updated or can remain the same. The minimum number of spikes and blanks to assess an LT-MDL is 7 or 24 respectively. However, the LT-MDL calculated from a short data set should be considered temporary, and should be updated when a more appropriate number of data points are available.

Results must be returned with enough digits so that the data are not quantized, as shown below.

Calculated LT-MDL values are reviewed with the analysts and supervisors prior to updating or finalizing the LIMS database. In some cases, the LT-MDL calculated value can be overridden by the analysts when the calculated LT-MDL is known to be unachievable or incorrect.

In the case of spiked LT-MDL samples, when variability is so low (not enough samples, quantized data, too high a spike concentration) the analyst may know they can not see the calculated LT-MDL. In these cases, the only option for overriding a calculated LT-MDL is to use the alternate reporting schemes that do not claim to use a statistically derived reporting level. The first is the interim reporting level (IRL), which is some multiplier of the best-guess LT-MDL. The second is the minimum reporting level (MRL) which is a pre-set censoring level and is not associated with any calculation, statistical significance, or an LT-MDL. Any value determined by a process other than the LT-MDL calculation must not be referred to as an LT-MDL.

Updating the LT-MDLs

LT-MDL values are updated yearly (October 1) if they are outside of the 95% confidence limits for the current LT-MDL calculation. Dates of change are noted in the LIMS data base.

LT-MDL values are stored with only one significant digit. Reporting levels are stored with two significant digits.

Reporting Limits

laboratory reporting limit (LRL) is generally equal to twice the yearly determined LT-MDL. The LRL is the less-than value that is reported when nothing is detected in a sample for a particular analyte. The LRL controls false negative error (reporting that the analyte is not detected when present at or greater than the detection level). The probability of falsely reporting a nondetection for a sample that contained an analyte at a concentration equal to or greater than the LRL is predicted to be less than or equal to 1 percent. The concentration of the LRL will be reported with a "less than" (<) remark code for samples in which the analyte was not detected.

The LRL is set at 2 times the LT-MDL

Alternate data reporting schemes beyond the LT-MDL and LRL approach are the method reporting level (MRL) and the interim reporting level (IRL).

The MRL is used as a reporting limit for censoring all values less than the MRL. There is no corresponding detection level value for an MRL because no values are reported below the MRL, even if there is a positive detection from the instrument. The MRL is generally not statistically defined, and often is set at a convenient decadel value such as <0.1 or <10.

The IRL is used when a method does not have at least one year's worth of supporting data to determine an LT-MDL and an LRL. The IRL is based on an estimated detection limit, and is set at least two times that estimation. Data between the IRL and the detection limit may be reported for information rich methods such as mass spectrometry, but detections less than the detection limit are censored in the LIMs and are not available to customers.