10th International Aerosol Conference September 2 - September 7, 2018 America's Center Convention Complex St. Louis, Missouri, USA
Abstract View
Causal Models as a Tool for Analyzing Dependence Structure of Variables in Combustion Aging Data
VILLE LEINONEN, Olli Sippula, Petri Tiitta, Ari Leskinen, Juha Karvanen, Annele Virtanen, Santtu Mikkonen, University of Eastern Finland
Abstract Number: 1139 Working Group: Aerosol Modeling
Abstract We investigated evolvement of small-scale wood combustion emission in environmental aging chamber. On-line measurements of aerosol size distribution and chemical composition and several gas concentrations from the chamber were performed using SMPS, AMS, PTR-MS and several gas analyzers. Total of five primary (POA) and secondary organic aerosol (SOA) factors from AMS data have been identified from OA mass spectra using PMF [1]. Both dark and UV induced aging were studied.
Evolution of emission in aging chambers has been studied by changing different chemical, physical and burning-device related factors that can have an effect on evolvement and SOA potential of emission. Many of these factors are dependent, so changing one of those factors can result in a change in other factors too. Therefore, taking into account the dependence structure of variables is necessary. Otherwise, intended univariate effect of factors are confounded by the changes of other factors, and improper analysis can lead to false conclusion of how intervening of factor changes variable of interest.
Aim of this study is to model evolution of emission in the aging chamber using multivariate statistical model. Furthermore, goal is to determine variables affecting each measured variable. Understanding the structures of impacts between different atmospheric variables is crucial when making decisions about constraints of different emission constituents.
Here we investigate applicability of causal modeling in case of analyzing complex environmental chamber data. We present preliminary results about using causal discovery algorithms and causal models.
Causal discovery algorithms search dependencies between variables by using either correlations observed in data, using fit of statistical model as a criterion, or both. The resulting causal structure is presented as a directed acyclic graph. Because dependencies found using causal discovery algorithms are based on data only, we evaluated meaningfulness of each dependency and removed known false dependencies from causal structure.
Using the obtained causal structure as a starting point, it is possible to estimate the causal relations between the variables and create a causal model. Causal model determines how each variable is influenced by other variables [2]. Model enables us to divide the effect of intervening of a factor to direct and indirect effect and thus make proper conclusions of causes of change in a variable of interest.
By using causal models, we were able to model evolvement of key gases such as NOx and ozone and size distribution and chemical composition of aerosols. We were also able to quantify the effects of components of POA and SOA to O/C-ratio of total organic aerosol. Finding emphasizes the potential of causal discovery algorithms to discover effects from multivariate data. In addition, causal models can estimate effect sizes. This is important when comparing importance of different variables to measured evolvement and comparing effect of intervening value of some variable to other variables.
In general, this study shows that causal models are useful when analyzing multivariate data, where multiple variables affect a variable of interest, and those effects are dependent. Model used in this study can be applied to different fuel types and to different environmental chambers. This enables rapid model development and gaining knowledge about the whole process occurring in the chamber. When the model is constructed, it can be used to predict how change in one emission component affects to others, especially the components of main interest: climate effects and toxicity. Adding new variables, including toxicity of emission is also possible and one of the main research objectives in the future.
References [1]Tiitta et al. Atmos. Chem. Phys., 16, 13251-13269, 2016. [2]Pearl, J., Causality, Cambridge University Press, New York, 2009.