Case studies of vector generalized additive models in environmental health and combustion research

✍ Scribed by James E. Dunn

• Publisher: John Wiley and Sons
• Year: 1999
• Tongue: English
• Weight: 188 KB
• Volume: 10
• Category: Article
• ISSN: 1180-4009
• DOI: 10.1002/(sici)1099-095x(199909/10)10:5<531::aid-env357>3.0.co;2-m

✦ Synopsis

Two case studies are presented in which vector-valued, generalized additive models (VGAM) have played pivotal roles. In Case 1, each subject was sequentially exposed via face mask to ®ve concentrations of butanol, with 60 s allowed between exposures to record sensations on six response scales (overall and odor intensity, eye, nose and throat irritation, and pleasantness). This was repeated immediately for a second replication, and then repeated twice again using ®ve concentrations of pentanone. Preliminary data analysis indicated both a lack of trend in response with increasing exposure concentration and poor reproducibility between replications. Both problems were resolved by introducing two covariates, namely optimally scaled, monotone transforms of the previously recorded overall and odor intensity scores. This suggested that either (1) the mechanical system was not zeroing between exposure levels, or (more likely) (2) subjects retained either a residual of the actual stimulus or sensations of the stimulus from the previous exposure. This case illustrates that optimal transforms need not always be identi®ed in terms of known functions. Case 2 is based on molecular structures of dioxin and furan, each of which admits the presence of 1±8 chlorine atoms. Measurements of these 16 homologues were reported for 24 test runs when burning refuse derived fuel (RDF) under varying conditions of HCl, coal, and lime injection. VGAM results, based on log-ratios of homologue composition, suggested that the eects of [SO 2 ] and [SO 2 ]/[HCl] were implemented through Michaelis±Menten kinetics. All-possible-subsets selection among VGAMs was crucial to model selection, and is illustrated. Empirical transforms were sucient for simulations under alternative test conditions.