shows summary statistics for the particulate and gaseous underground area air measurements. There were 179 area basket samples, resulting in 2336 observations of the 14 agents. There were generally few missing observations (<5%), except for SMD. For most agents, the percent of measurements less than or equal to LOD was <10%, except for SEC, NO, NO_{2}, CO, and SCD.

| **Table 1.** Descriptive statistics of underground area air measurements for DE components in seven non-metal mining facilities |

Levels of REC varied between (GSD

_{bf} = 2.6) and within (GSD

_{wf} = 3.4) underground operations of the facilities (), with facility-wide geometric averages for REC ranging from 21 μg m

^{−3} in Facility I to 459 μg m

^{−3} in Facility A (). ROC measurements showed less variation both between and within facilities (e.g. GSD

_{bf} = 1.3 and GSD

_{wf} = 1.8). The between- and within-facility variances for NO and CO were more similar to those for REC than for ROC; however, CO

_{2} showed less total variation (GSD

=

1.5), while NO

_{2} showed more variation overall, especially between facilities (GSD

=

7.5 and GSD

_{bf} = 5.1, respectively). Particulate measurements (TD, RD, SMD, and SCD) varied somewhat less between facilities than did the REC measurements. In general, gaseous and particulate concentrations were highest in the limestone facility (A) and lowest in the trona facilities (G–I) ().

| **Table 2.**Geometric means, GSDs, and numbers of measurements for underground area air measurements of selected DE components, by mining facility |

displays scatter plots of the relationship of REC with TEC and with SEC. Ln(REC) and Ln(TEC) were strongly correlated (

*r*_{P}=

0.92). REC and TEC showed strong clustering around the line of identity (

*x*=

*y*) (, left panel) and no apparent deviation from this line for the individual facilities (median TEC/REC ratio: 1.01). The correlation between Ln(SEC) and Ln(REC) was similar to that for TEC (

*r*_{P}=

0.94), although SEC measurements were on average lower than the REC measurements (median SEC/REC ratio: 0.77) (, right panel).

shows Pearson correlations on the log scale of REC with the other DE components. Spearman correlations were similar (data not shown). Correlations of REC with the non-EC agents were weaker than for TEC and SEC but were still high for ROC, NO, CO_{2}, SMD, and SCD (range: 0.66–0.79). Moderate to weak correlations were found with the remaining agents (range: 0.22–0.56). Correlations between REC and the other agents tended to be highest in the limestone, potash, and salt facilities (where REC levels were higher) and lowest in the trona facilities (where REC levels were lowest).

| **Table 3.**Pearson correlations (*r*_{p}) between the natural logarithm of underground area air measurement results for REC and other DE components, overall and by mining facility |

Factor analysis using log-transformed concentrations of 132 complete sets of DE component measurements showed that the EC fractions, ROC, the gases, and SCD loaded heavily on the first factor, which we call ‘Diesel exhaust’ for convenience (). This is the factor that explained the most of the variance (i.e. 39%). TOC and two dust components (TD and RD) loaded heavily on the second factor, suggesting an additional ‘Mine dust’ factor that accounted for 21% of the variance. ROC and SOC loaded most heavily on the third factor, suggesting a separate ‘Organic carbon’ component that accounted for 12% of the variance.

| **Table 4.**Results of factor analysis of multiple correlations with Varimax rotation among components of DE (transformed by the natural logarithm)^{a}^{b} |

Results of the regression of Ln(REC) on Ln(CO) for all facilities combined are displayed in (upper left panel). For Ln(CO), the regression model showed a linear association with Ln(REC) with an overall slope estimate of 0.47 [95% confidence intervals (95% CI) 0.31–0.63] (AIC

=

586.6). However, there was significant heterogeneity in this relation between facilities (, remaining panels). A mixed-effect model allowing for facility-specific intercepts (as fixed effects) and slopes (as random effects) fitted the data significantly better (AIC

= 516.8). In this model, the mean slope for all facilities was 0.58 (95% CI 0.22–0.94) with facility-specific slopes ranging from 0.13 for Facility I to 1.17 for Facility E ().

For Ln(NO

_{2}), the model including data from all facilities showed a linear association with Ln(REC), with an overall slope estimate of 0.40 (95% CI 0.30–0.49) (AIC

=

600.6) (). Similarly to CO, a mixed-effect model with a random slope fitted the data significantly better (AIC

= 562.6). The mean slope was 0.44 (95% CI 0.13–0.75) and facility-specific slopes ranged from 0.16 for Facility A to 1.04 for Facility E ().

Because REC and CO were associated overall and because CO was the DE component used in the back-extrapolation of REC, we further investigated whether the REC:CO relationship was approximately linear over the full range of measurement concentrations. Non-parametric regression analyses allowing for facility-specific intercepts using GAM showed that the association of REC with CO was essentially linear in log–log space (data not shown).