Carpet-dust chemicals as measures of exposure: Implications of variability

Abstract

Background

There is increasing interest in using chemicals measured in carpet dust as indicators of chemical exposures. However, investigators have rarely sampled dust repeatedly from the same households and therefore little is known about the variability of chemical levels that exist within and between households in dust samples.

Results

We analyzed 9 polycyclic aromatic hydrocarbons, 6 polychlorinated biphenyls, and nicotine in 68 carpet-dust samples from 21 households in agricultural communities of Fresno County, California collected from 2003-2005. Chemical concentrations (ng per g dust) ranged from < 2-3,609 for 9 polycyclic aromatic hydrocarbons, from < 1-150 for 6 polychlorinated biphenyls, and from < 20-7,776 for nicotine. We used random-effects models to estimate variance components for concentrations of each of these carpet-dust chemicals and calculated the variance ratio, λ, defined as the ratio of the within-household variance component to the between-household variance component. Subsequently, we used the variance ratios calculated from our data, to illustrate the potential effect of measurement error on the attenuation of odds ratios in hypothetical case-control studies. We found that the median value of the estimated variance ratios was 0.33 (range: 0.13-0.72). Correspondingly, in case-control studies of associations between these carpet-dust chemicals and disease, given the collection of only one measurement per household and a hypothetical odds ratio of 1.5, we expect that the observed odds ratios would range from 1.27 to 1.43. Moreover, for each of the chemicals analyzed, the collection of three repeated dust samples would limit the expected magnitude of odds ratio attenuation to less than 20%.

Conclusions

Our findings suggest that attenuation bias should be relatively modest when using these semi-volatile carpet-dust chemicals as exposure surrogates in epidemiologic studies.

Background

Semi-volatile chemicals can accumulate in carpets over years and decades [1–3], and thus their concentrations in carpet dust could be useful surrogates for long-term indoor exposures in epidemiological studies [2, 4–6]. Moreover, because dust ingestion or inhalation could be responsible for significant chemical exposures in young children [7–9], levels of chemicals in dust may be particularly relevant in studies of childhood diseases.

Although many researchers have measured chemicals in dust [10–12], few have sampled dust repeatedly in the same households [13–16] or characterized the variability of dust measurements within and between households [17, 18]. In two studies that reported variance components of dust levels (of pesticides, lead, and phenanthrene), large variance ratios (i.e., ratio of within-household variance component to between-household variance component, designated here as λ) were observed [17, 18]. Since, the degree of exposure measurement error increases directly with λ, large values of this ratio indicate imprecise exposure classification. In an epidemiological study, exposure misclassification will tend to result in the observation of risk estimates that are smaller than the true risks, a phenomenon referred to as attenuation bias. To employ carpet-dust concentrations as surrogates for chemical exposure with confidence, investigators first need to know how variable these measurements are within a given household, that is, they need some measure of their reliability.

Our objective in this analysis was to quantify the reliability of carpet-dust chemical concentrations as exposure measures for future epidemiological studies. We analyzed 9 polycyclic aromatic hydrocarbons (PAHs), 6 polychlorinated biphenyls (PCBs), and nicotine (as a surrogate for tobacco smoke) in repeated carpet-dust samples. These semi-volatile chemicals are particularly suitable for measurement in household dust because they persist in the indoor environment [10], and their long-term exposures have been associated with health effects [2, 6, 19, 20]. Using random-effects models of repeated carpet-dust measurements, we estimated variance ratios for each of these chemicals. Subsequently, using our variance ratios, we estimated the amount of attenuation bias that would be expected to occur in independent case-control studies that used these carpet-dust chemicals as exposure measures.

Methods

Study households

We obtained dust samples from 21 households in Fresno County, California, from 2003-2005, as part of an investigation to estimate chemical exposures in residences located in agricultural communities. The study protocols were approved by the Institutional Review Boards at Colorado State University and the National Cancer Institute, and we obtained written informed consent from all participating subjects.

Collection of carpet dust

We collected carpet-dust samples using a high-volume surface sampler (HVS3) as previously described [21]. Briefly, the interviewer selected a room on the side of the residence that faced agricultural crops, marked an approximately 4-foot by 6-foot area of a carpet or rug with tape, and vacuumed the surface in 3-inch strips, making four passes back and forth on each strip, until a 10 mL of fine dust had been collected. With few exceptions, all repeated samples we collected from a given household were from the same room. The median number of measurements per household was n = 3 (range of n: 1-7) and the median duration between repeated visits was 5 months (range of 3-15 months).

Laboratory chemical analysis

We analyzed nine PAHs [benzo(a)anthracene, chrysene, benzo(a)pyrene, benzo(b)fluoranthene, benzo(k)fluoranthene, indeno(1,2,3-c,d)pyrene, dibenzo(a,h)anthracene, coronene, and dibenzo(a,e)pyrene], 6 PCBs (PCB 105, PCB 118, PCB 138, PCB 153, PCB 170, and PCB 180), and nicotine in dust samples as previously described [21]. Briefly, we sieved each dust sample using a 100-mesh stainless steel sieve (< 150 μm), extracted 0.5 g of fine dust with either a 1:1 hexane:acetone mixture (PAHs, PCBs) or methylene chloride (nicotine), then cleaned the extract using solid phase extraction (for PAHs and PCBs), and analyzed the concentrated eluate with gas chromatography-mass spectrometry (GC-MS) using p, p-dibromophenyl and d₁₂-benzo(e)pyrene as internal standards for quantitation. PAHs and PCBs were analyzed using an RTx-5 MS column (30 M, 0.25 mm id, 0.25 μm film) with a GC oven temperature programmed from 130-220°C at 2°/min and then 220-330°C at 10°/min. Nicotine was analyzed using a DB-1701 column (30 M, 0.25 mm id, 0.15 μm film) with the GC oven temperature programmed 130-220°C at 2°C/min and then 220-280°C at 10°/min.

Statistical analysis

Since the chemical concentrations were approximately log-normally distributed, we used the natural log-transformed values for all statistical analyses. We assigned all values below the limit of detection a concentration equal to the limit of detection divided by the square root of 2 [22]. We excluded chemicals that had detection rates less than 75% from the random-effects modeling (i.e., PCB 105, PCB 118, and PCB 170).

Random-effects models

To estimate variance components, we used the one-way random-effects model,

$$Y_{ij}=\mathsf{\text{ln }}{X}_{ij}={\mu }_Y+b_i+e_{ij},$$

(1)

for i = 1,2,...,k households and j = 1,2,...,n repeated measurements, where

X _ij = the carpet-dust chemical concentration for the i^th household on the j^th repeated measurement;

Y _ij = the natural log-transform of X _ij ;

μ _Y = the true (logged) mean carpet-dust chemical concentration for the population;

b _i = μ _Yi -μ _Y , and represents the random deviation of the i^th household's true mean (logged) carpet-dust chemical concentration, μ _Yi , from μ _Y ;

e _ij = Y _ij - μ _Yi , and represents the random deviation of the observed (logged) carpet-dust chemical concentration, Y _ij , from μ _Yi for the i^th household on the j^th repeated measurement.

We assumed b _i and e _ij are mutually independent and normally distributed random variables, with means of zero and variances ${\sigma }_{bY}^2$ and ${\sigma }_{wY}^2$, representing the between-household and within-household variances, respectively. These assumptions have been validated using repeated measurements of occupational chemical exposures [23–25].

Using Proc Mixed (SAS v.9.1, Cary, NC) we fit the model described in Equation 1 and estimated variance components (${\widehat{\sigma }}_{bY}^2$,${\widehat{\sigma }}_{wY}^2$, and ${\widehat{\sigma }}_Y^2={\widehat{\sigma }}_{wY}^2+{\widehat{\sigma }}_{bY}^2$) and variance ratios, $\widehat{\lambda }=\frac{{\widehat{\sigma }}_{wY}^2}{{\widehat{\sigma }}_{bY}^2}$. Subsequently, we estimated the expected concentration fold range for 95% of measurements (i.e., the expected ratio of the 97.5^th percentile concentration to the 2.5^th percentile concentration) from a single household $\left[{}_w\widehat{R}{}_{0.95}=\exp \left(3.92\times {\widehat{\sigma }}_{wY}^{}\right)\right]$ and across all households in our study population $\left[{}_b\widehat{R}{}_{0.95}=\exp \left(3.92\times {\widehat{\sigma }}_{bY}^{}\right)\right]$[25].

Estimating attenuation bias

In the context of a case-control study, the following logistic model could be used to assess the risk of disease associated with a particular carpet-dust chemical:

$$\mathsf{\text{Logit}}{Z}_i=\mathsf{\text{ln}}\left(\frac{Z_i}{Z_i-1}\right)={\beta }_0+{\beta }_1\stackrel{-}{Y_i},$$

(2)

where

Z _i = the disease status (1 or 0) of an individual in the i^th household and $\stackrel{-}{Y_i}$ = the (logged) mean carpet-dust chemical concentration for the i^th household.

In this case, the expected value of the estimated logistic regression coefficient, E[${\widehat{\beta }}_1$], is related to the true logistic regression coefficient, β ₁ , by the variance ratio, λ, as follows [26]:

$$\mathsf{\text{E}}\left[{\widehat{\beta }}_1\right]=\frac{{\beta }_1}{1+\frac{\lambda }{n}}.$$

(3)

We define attenuation bias as the normalized difference between the expected value of the estimated logistic regression coefficient and the true logistic regression coefficient:

$$B=\frac{\mathsf{\text{E}}\left[{\widehat{\beta }}_1\right]-{\beta }_1}{{\beta }_1}.$$

(4)

We used Equations 3 and 4 to estimate the amount of attenuation bias that would be expected in case-control studies using carpet-dust chemicals as independent variables in logistic regression analyses. For each chemical, using estimates of the variance ratio,$\widehat{\lambda }$, from the application of the random-effects model (Equation 1), and an assumed true odds ratio of 1.5, we estimated the expected value for ${\widehat{\beta }}_1$, the corresponding expected odds ratio, E[OR], and the expected amount of attenuation bias. It is worth noting that, in Equation 4, the magnitude of the attenuation bias is independent of the true odds ratio. In our calculations we assume that the variance ratio for the case and control populations are the same (i.e., measurement error is assumed to be non-differential).

Investigators can improve the precision of exposure estimates and, thereby, limit attenuation bias by making repeated exposure measurements and finding an average exposure level for each study subject over time. Combining Equations 3 and 4, it is possible to calculate the number of repeated measurements per household, n, that would be necessary to limit attenuation bias to a certain level as follows:

$$n=\frac{\widehat{\lambda }}{\frac{1}{1+B}-1}$$

(5)

Using our variance ratio estimates, we calculated the number of repeated measurements that would be necessary to limit the magnitude of attenuation bias to 20% in a case-control study using these carpet-dust chemicals as measures of exposure.

Results

Chemical concentrations in carpet dust

Our analyses included 21 households with 68 carpet-dust measurements. As shown in Table 1, individual chemical detection rates ranged from 38 to 100% and, as shown in Table 2, individual chemical concentrations ranged from less than the limit of detection to a maximum of 7,776 ng/g. We detected the 9 PAHs in a higher percentage of samples, and at higher median concentrations, than the 6 PCBs. The range in nicotine concentrations was larger than the range in concentrations of either PAHs or PCBs.

Table 1 Limits of detection and frequency of detection for 68 carpet-dust samples

Table 2 Summary statistics for 68 carpet-dust samples

Random-effects models

Table 3 shows the results of the analysis using random-effects models for the 13 chemicals with at least a 75% detection rate. For all models, the between-household variance component was greater than the within-household variance component (i.e., $\widehat{\lambda }$< 1). The median within-household variance component estimate for PAHs was ${\widehat{\sigma }}_{wY}^2$ = 0.38 (interquartile range, IQR: 0.21 - 0.42), for PCBs it was ${\widehat{\sigma }}_{wY}^2$ = 0.41 (IQR: 0.36 - 0.51), and for nicotine it was ${\widehat{\sigma }}_{wY}^2$ = 1.33. For each of the 13 individual chemicals, the within-household variance component ranged from ${\widehat{\sigma }}_{wY}^2$ = 0.16 (coronene) to ${\widehat{\sigma }}_{wY}^2$ = 1.33 (nicotine). Correspondingly, 95% of repeated coronene measurements from a household in our study population would be expected to lie within a 5-fold range versus a 92-fold range for repeated nicotine measurements. The median between-household variance component estimate for PAHs was ${\widehat{\sigma }}_{bY}^2$ = 1.20 (IQR: 1.00 - 1.27), for PCBs it was ${\widehat{\sigma }}_{bY}^2$ = 1.29 (IQR: 1.24 - 1.46), and for nicotine it was ${\widehat{\sigma }}_{bY}^2$ = 1.85. For each of the 13 individual chemicals, the between-household variance component ranged from ${\widehat{\sigma }}_{bY}^2$ = 0.77 [benzo(k)fluoranthene] to ${\widehat{\sigma }}_{bY}^2$ = 1.85 (nicotine). Correspondingly, 95% of the mean benzo(k)fluoranthene concentrations from different households in our study population would be expected to lie within a 31-fold range versus a 207-fold range for mean nicotine levels.

Table 3 Variance parameter estimates from random-effects model regression analyses of repeated measurements of carpet-dust chemicals

Attenuation bias estimates

Table 4 shows the amount of attenuation that would be expected in odds ratios if case-control studies were to use each of the carpet-dust chemicals as independent variables in logistic regression analyses. For each of the 13 chemicals with at least a 75% detection rate, expected bias was calculated using Equations 3 and 4 along with estimates of the variance ratio from Table 3. We found that, by definition, the magnitude of expected bias increased with the estimated variance ratio. For example, for the chemical with the smallest variance ratio [benzo(b)flouranthene, $\widehat{\lambda }$ = 0.13], the expected odds ratio would be 1.43 assuming only one measurement from each household (i.e., n = 1), indicating a -12% bias (true odds ratio = 1.5). However, for the chemical with the highest variance ratio (nicotine, $\widehat{\lambda }$ = 0.72); the expected odds ratio under the same conditions would be 1.27, a -42% bias.

Table 4 Attenuation bias due to measurement error

Figure 1 shows plots of the relationship between the expected odds ratio and the number of repeated measurements per household, using the estimated variance ratios from Table 3 and assuming a true odds ratio of 1.5 for PCB 153, benzo(a)pyrene, and nicotine. For each of the carpet-dust chemicals, Table 4 indicates that the number of repeated measurements necessary to limit attenuation bias to -20% ranged from 1 to 3 measurements per household.

Figure 1

Expected odds ratio attenuation. Odds ratio attenuation in case-control studies that used (logged) carpet-dust chemical concentrations as measures of exposure, given various sampling strategies; for PCB 153 (squares), benzo(a)pyrene (crosses), and nicotine (triangles).

Discussion

Our results can guide epidemiologists in developing sampling strategies for using household dust as a medium for estimating exposures to PAHs, PCBs, or nicotine in their studies. Generally, investigators can improve the precision of their exposure estimates and limit attenuation bias by making repeated exposure measurements on each study subject. However, the analytical advantages of a repeated sampling design must be balanced with the practical concerns of a study's schedule and budget. By evaluating Equation 5 with our own variance ratio estimates, we provide future investigators a blueprint for obtaining precise exposure estimates without unnecessarily inflating study costs. As shown in Table 4, we found that, for each chemical we analyzed in carpet dust, three repeated dust measurements per household would be sufficient to reduce the magnitude of attenuation bias to less than 20%. From a practical standpoint, investigators could resample a particular carpet area as frequently as once a month. However, to observe (and adjust for) seasonal variation that may exist in carpet-dust chemical levels it would appropriate to sample over the course of an entire year. Moreover, for an investigator to estimate exposures that occurred in the distant past, it could be useful to collect samples over an even longer period of time. Indeed, for retrospective exposure assessment, increasing the duration of the dust collection period would enable an investigator to observe (and adjust for) any long-term time trends that may exist in carpet-dust chemical levels.

Moreover, if repeated sampling would not be feasible, Table 4 indicates that for 10 of the 13 chemicals analyzed, the expected magnitude of attenuation bias would still be less than 30%. Notably, nicotine, the most volatile chemical analyzed in our study, had a larger variance ratio than any of the PAHs or PCBs. Based on this observation, it is possible that carpet-dust concentrations of more volatile compounds will be more variable over time.

Our findings are based on a limited sample size (68 dust measurements from 21 households), and our variance ratio estimates are consequently somewhat imprecise (see Table 3). Moreover, our findings are based on dust measurements from only one surface type (carpets) and for only one general class of chemicals (semi-volatiles). However, we are confident that our findings will be externally valid and useful for other investigators measuring these same chemicals in dust. Notably, the dust concentrations of chemicals measured in our study (Table 2) were generally similar to the concentrations reported in recent studies of other households in California with respect to both the medians and the ranges of concentrations [2, 3, 27].

Unfortunately, it is difficult to compare our findings to those from two other studies that repeatedly sampled dust from the same households over time and reported corresponding variance components [17, 18], because these studies published estimates for different chemicals in dust (i.e., pesticides, lead, and phenanthrene). However, our estimated variance ratios (Table 3) were quite similar to those we estimated using unpublished data from Egeghy et al. for several PAHs that were measured in household dust from both studies (Additional file 1). The similarity of variance ratios from two independent populations lends credibility to our findings and suggests that the levels of variability we observed in semi-volatile carpet-dust chemicals may be generalized to other populations.

In using the random-effects model to estimate variance components, we implicitly assume that each household has a true underlying mean dust concentration (for each chemical) that remains constant over the course of the study (i.e., μ _Y + b _i ). As such, we interpret any deviation from a household's true mean level as measurement error or random within-household variability. It is possible that some of the "random" variability that we observed is due to changes in the sources of chemical contamination in the homes, seasonal variations in temperature or ventilation practices, or other unaccounted-for factors that changed during the course of the study. Indeed, since our dust samples were collected over the period of 3 years, it is possible that true mean concentrations of chemicals in household dust did change somewhat over time. Consequently, the long-term timing of our sampling could have artificially inflated the within-household variance component, causing us to overestimate the variance ratios and the associated attenuation bias. Nevertheless, our random-effects model should provide a conservative estimate of the reliability of chemicals measured in carpet dust as measures of exposure.

One limitation of our method for predicting attenuation bias is that we specified that the variance ratios from the case and control populations were the same (i.e., measurement error was defined as non-differential) in Equation 3. In retrospective case-control studies, carpet-dust chemicals will be measured after disease diagnosis. In this scenario, case subjects could be more likely to change their behaviours between diagnosis and dust collection. If cases alter behaviours that result in changes to carpet-dust chemical levels, differential measurement error could occur. In the more complex situation in which case and control populations have differential measurement errors, Equation 3 would be only approximate. We were unable to evaluate whether variance ratios for concentrations of carpet-dust chemicals actually differ for case and control populations.

Conclusions

In summary, we found that estimates of variance ratios of carpet-dust PAHs (0.13 ≤ $\widehat{\lambda }$ ≤ 0.64), PCBs (0.25 ≤ $\widehat{\lambda }$ ≤ 0.37), and nicotine ($\widehat{\lambda }$ = 0.72) were modest for the 21 homes in our study area. Though based on a limited number of measurements (N = 68), our findings suggest that the use of carpet-dust samples as measures of exposure to these 13 chemicals will result in relatively small levels of attenuation bias due to exposure measurement error. Moreover, we have presented a simple guide for investigators to create efficient study designs that will limit bias in future studies that use dust to measure exposures to PAHs, PCBs, or nicotine.