Handling Building Contamination Claims

While “testing” is an intuitive response, there are aspects of mold test data that frequently result in misleading interpretation, as the mathematical nature of the data (what statisticians term “distribution”) is unlike that typically encountered with other environmental/engineering parameters. This can eventually result in unwarranted expense in cleanup and restoration or insufficient response and the associated liability of alleged mold exposure. It is, therefore, essential that the unique characteristics of mold test data and the potential implications for sampling and data interpretation are recognized before mold testing is undertaken.

The Nature of Mold Test Data

Unlike many environmental/engineering parameters that can be tested, there is no standardization in the measurement of mold. In actuality, there is wide variation in sample collection procedures and laboratory analysis. At any given site, the detection and quantification of mold (of which there are potentially thousands of individual species) in environmental samples depends upon the characteristics and selectivity of the particular sampling and analysis employed.

While there have been recent advances in the level of sophistication in detection technology, such as DNA “fingerprinting,” there remains no one methodology that can identify and detect all mold present at a given site. Similarly, there is no general agreement on whether a particular species, group(s) of species, or total mold spore levels are of relevance in a given circumstance. The lack of a clear metric ultimately means there are no fixed numerical, health-based “mold levels,” and only a comparative evaluation can place environmental mold data into a useful context. That is, from a claims perspective, mold test data can sometimes be useful in identifying (or eliminating) sources of mold growth if relative differences in fungal populations are consistent with known or theorized water or moisture sources or pathways. The most fundamental problem with mold test data is therefore not the sophistication of the particular analysis or laboratory but the criteria by which mold test data from different test zones are compared.

Regardless of the methodology chosen, mold test data (whether air or dust) typically exhibits very wide variations in detected fungal types and numerical concentrations. When several samples are collected, many mold species are detected only sporadically. The normal mathematical and statistical approaches, which are more intuitive and almost exclusively utilized by building investigators (frequently environmental/public health consultants), can be misleading.

An actual example of data collected at a building alleged to have been “contaminated” illustrates this point (see p. 48). Data Set A represents numerical concentrations for a particular mold species represented by several short-term air samples taken throughout the day in the outdoor air of the suspect building. Data Set B is from the suspect indoor environment (test zone) collected contemporaneously with the outdoor air (the particular sampling methodology or units of measure are not relevant for purposes of this discussion).

What is immediately apparent is that collecting a small number of the one- to five-minute samples afforded by the most commonly employed sampling devices cannot produce a good representation of either zone. The most commonly utilized approach by building investigators will be to compare the average concentration from each zone and “interpret” the numerical average of 10.13 in the test zone B as indicating “contamination,” compared to the average of 3.94 in the control zone A (outdoor air).

However, directly comparing numerical concentrations is rooted within the standard public environmental-health model (as well as our normal “intuitive” sense of numbers) in which data points with larger values automatically equate to higher contaminant concentrations, and therefore an increased health risk. This model assumes three important conditions: standardized sampling and analysis, fixed numerical standards against which to evaluate data points, and data with known variation. None of these apply to mold-test data.

Further, the validity of a particular criterion in evaluating data—in this case, average numerical concentrations—can be tested using a version of Monte Carlo simulation (bootstrapping), which is a recognized and accepted mathematical analysis method also utilized in catastrophic claims modeling.

A detailed theoretical basis of bootstrapping is beyond this discussion. However, conceptually, it is a computer-based simulation process in which actual field data are used to model the random variability exhibited by the particular parameter in question. The computer selects thousands of random subsets of “resampled” data from the actual field data. The variation that results in the resampling data reflects the random variability in the original field data. Similarly, variation in statistical criteria used to describe differences in the resampling data will also be reflected in the original field data.

Several studies in peer-reviewed technical journals using this analysis to evaluate mold test data, from both “moldy” buildings and those with no history of water damage or mold growth, have demonstrated that using numerical mold levels results in high false positive and negative rates—“clean” building zones (as well as outdoor air) have a high probability of being judged “contaminated,” while buildings that are likely problematic have a high probability of being missed. This also is consistent with the National Academy of Sciences 2004 reference on indoor dampness and mold, which states, “fungal counts alone provide little information about the microbial status of an indoor environment for building assessment.”

Implications for Claims Evaluations

Bootstrap analysis has shown the mathematically and statistically correct criterion for mold test data is significant differences relative to a reference concentration of the two zones of comparison (∆f_d). Actual numerical mold levels are generally irrelevant.

In the example data set (as in most data), the combined median (middle value when all the data are rank ordered) is the best reference value. In the example data sets, the combined median is zero. Both zones exhibit seven values greater than zero (detection), and the two sets of data are statistically no different. Unfortunately, while determination of significant differences in detection frequency is derived from a basic probability formula, the calculations required are lengthy and require a computer due to the number of samples necessary to generate true statistical significance (approximately 15 for each zone of comparison).

As a result, most field investigators will steer data “interpretation” toward using easily calculable numerical mold levels with limited samples, based upon their experiences with other environmental contaminants, environmental and public health training, and imposed budget limitations.

This has important implications for challenges by insurers to building practitioners who may base expert opinions of mold contamination using microbial data. Under rules established to qualify expert scientific testimony (Daubert), a particular test methodology should have a known or determinable error rate. Environmental and public health experts citing fixed numerical levels in mold test data as high, elevated, or in some way unacceptable inevitably must answer the question: What is the criterion?

With no established fixed numerical standards, the error rate for reported numerical “mold levels” cited as problematic cannot be determined, regardless of the type and/or level of sophistication of the mold testing involved.

Similarly, comparative criteria based on numerical concentrations have been shown to be erratic, with high implied error rates given the inability to consistently identify “clean” or “moldy” environments. ASTM Standard E678 (Standard Practice for Evaluation of Scientific or Technical Data) calls for “technical hypotheses and judgmental criteria” to be explained, and for “forensic opinions to be supported by the data.” In the example data set, the probability that mold levels are greater in the indoor air using ∆f_d as the criterion is 0.55. On the generally accepted “scale” used in the scientific community to determine a real difference, the probability should be in the range of at least 0.90, and ∆f_d as the criterion to judge the “mold levels” indicates there is no difference in the test populations. Bootstrap analysis by repeated resampling of the data using ∆f_d generates the same probability. By contrast, while mold levels in the example indoor air are more than twice that of the outdoor air (10.13 vs. 3.94) and superficially appear to indicate contamination, this is, in fact, not supported by the data.

The assessment of buildings with regard to possible mold contamination carries significant public health and economic ramifications, and is often influenced by mold test data in which numerical mold levels are interpreted from the environmental and public health model. This assumes that numerical mold levels can be associated with a potential health hazard and, by implication, denote a contaminated indoor environment. This approach ultimately establishes a context where mere detection of a particular mold type can be presented as “contamination,” as it does not consistently quantify relative differences in mold populations in the true mathematical or statistical sense. The highly charged environment that often accompanies mold claims and the intuitive response to “test” can result in ill-conceived sampling.

Without a valid criterion and sufficient number of samples to allow determination of true, statistically significant differences in mold populations, the data can be manipulated and is ultimately misleading. Similarly, when presented with data as evidence of mold contamination as part of a claim, the insurer should demand that the criteria used to evaluate the data be clearly articulated and that the criteria have been tested/validated in a peer-reviewed publication or industry reference with regard to the ability to characterize an environment as “clean” or “contaminated.”   

DATA SETS

Indoor Zone “B”             9   71    9    36     0    9    0    9    0    0    0    0   0    9    0                      10.13