Sujit K. Ghosh Working Group I: Effects of PM on Mortality; Air Quality and Morbidity Sujit K. Ghosh and http://www.stat.ncsu.edu/people/ghosh/ sujit.ghosh@ncsu.edu Presented at: Statistical Methods and Analysis of Environmental Health Data SAMSI-SAVI Workshop, Piramal Tower Annex, Mumbai, India http://www.tinyurl.com/sami-savi-2016 SAMSI-SAVI Workshop 1 June 3, 2016
Sujit K. Ghosh Current Team • Sujit Ghosh (Co-Leader), NCSU/SAMSI, USA • Brian Reich (Co-Leader), NCSU, USA • Sirajuddin Ahmed, Jamia Milia Islamia University, India • Safraj Shahul Hameed, PHFI, India • Sanjoy Maji, Jamia Milia Islamia University, India • Parul Goel, National Center for Disease Control, India Initial data sets are provided by Sanjoy Maji. SAMSI-SAVI Workshop 2 June 3, 2016
Sujit K. Ghosh Outline • The Delhi Data set • Scientific Questions and Challenges • Methodologies explored • Preliminary Results • Next Steps... SAMSI-SAVI Workshop 3 June 3, 2016
Sujit K. Ghosh The Delhi Data Sanjoy Maji has obtained daily data (FY 2010-2012) on following variables: • Cause specific deaths: accidental, respiratory, circulatory and total number of deaths Available by gender and age group • Air pollutant: SOx, NOx, RSPM Stations: Industrial area: Mayapuri Indl. Area, Shahdara,Shahzada Bagh, Janakpuri Residential area: N.Y. School, Nizamuddin, Pritampura,Siri Fort and Town Hall ( under National Ambient Air Quality Program, monitoring stations are operated twice a week only! ) • Met data: Max temp, Min temp, RH (obtained from Indian Meteorological station at Safdarjung) SAMSI-SAVI Workshop 4 June 3, 2016
Sujit K. Ghosh SAMSI-SAVI Workshop 5 June 3, 2016
Sujit K. Ghosh Scientific Questions and Challenges • It is typical to use quasi-Poisson models to relate number of deaths to pollutant levels after adjusting for other concomitant variables. What appropriate statistical methodologies be used to address the missing values or change of support? • How can we create an Air Quality Index (AQI) by combining several pollutants? • How do we relate the AQI to mortality data and create a warning system (based on futurecasts)? • Is it possible to identify association between cause-specific deaths and air pollutants (after adjusting for other variables) • Do the associations (if any) are of different magnitude for different subgroups of populations and zonal regions? E.g., by gender, age group, industrial vs. residential, etc. SAMSI-SAVI Workshop 6 June 3, 2016
Sujit K. Ghosh Methodologies explored • Create an Air Quality Index (AQI) by combining several (appropriately scaled) pollutants • X j ( s, t ) = Pollutant j measured at site (station) s and time (day) t • ¯ X j ( t ) = � S s =1 X j ( s, t ) /S : Average level of pollutant j measured on time (day) t • However, in most cases X j ( s, t ) are not available for all triplet ( s, t, j ) !! • We thus need an imputation model to compute ¯ X j ( t ) . Explored several models and more needs to be explored. • With imputed versions of each ¯ X j ( t ) we would determine ‘optimal’ weights to create j =1 w j ¯ AQI: λ ( t ) = � J Z j ( t ) where ¯ Z j ( t ) is an appropriately scaled version of ¯ X j ( t ) SAMSI-SAVI Workshop 7 June 3, 2016
Sujit K. Ghosh • Initial explorations included looking at Principal Component Analysis (PCA). Alternatively, we can use Independent Component Analysis (ICA). • However, how do we associate the AQI to health (mortality)? • We use the popular quasi-Poisson models adjusting for other concommitant variables (smooth functions of time and met variavles) and the AQI • A simple version of the model would like like: Y ( t ) Poisson ( µ t ) ∼ µ t = splines ( t, W t ) + g ( λ t ( w )) where W t are available met data and g is an unknown monotone function of the AQI (which depends on unknown weights) • Simultaneous estimation of g and w = ( w 1 , . . . , w J ) is non-trivial. E.g., we plan to explore projection pursuit (or single index) regression methods SAMSI-SAVI Workshop 8 June 3, 2016
Sujit K. Ghosh Preliminary Results • We explored linear regression based methodologies to build initial imputation models • For each pair ( j, t ) we use available values of X j ( s, t ) from several existing monitoring stations and met variable to build the regression models that have ’respectable’ predictive power (e.g., adjusted R 2 > 0 . 5 etc.) SAMSI-SAVI Workshop 9 June 3, 2016
Sujit K. Ghosh PM10 Manual Real time 1500 1000 Concentration 500 0 Mayapuri Indl. Area Pritampura Civili.Lines IHBAS R.K.Puram Station SAMSI-SAVI Workshop 10 June 3, 2016
Sujit K. Ghosh NO2 Manual Real time 600 Concentration 400 200 0 Mayapuri Indl. Area Pritampura Civili.Lines IHBAS R.K.Puram Station SAMSI-SAVI Workshop 11 June 3, 2016
Sujit K. Ghosh SO2 Manual 300 Real time 250 200 Concentration 150 100 50 0 Mayapuri Indl. Area Pritampura Civili.Lines IHBAS R.K.Puram Station SAMSI-SAVI Workshop 12 June 3, 2016
Sujit K. Ghosh Mayapuri Indl. Area Shahdara Shahzada Bagh 500 Concentration Concentration Concentration 600 600 300 200 200 100 0 0 200 600 1000 0 200 600 1000 0 200 600 1000 Day Day Day Janakpuri N.Y. School Nizamuddin 700 1200 Concentration Concentration Concentration 500 400 800 300 200 400 100 0 0 0 200 600 1000 0 200 600 1000 0 200 600 1000 Day Day Day Pritampura Siri Fort Town Hall Obs Obs Obs 600 Impute Impute Impute Concentration Concentration Concentration 400 1000 400 200 200 500 0 0 0 0 200 600 1000 0 200 600 1000 0 200 600 1000 Day Day Day SAMSI-SAVI Workshop 13 June 3, 2016
Sujit K. Ghosh Next Steps... • Explore more sophisticated imputation models to compute AQI. E.g., Robbins, M., Ghosh, S. K. and Habiger, J. (2013). Imputation in High Dimensional Economic Data as Applied to the Agricultural Resource Management Survey, Journal of the American Statistical Association , 108 , 81-95. http://www.tandfonline.com/doi/abs/10.1080/01621459.2012.734158 • Explore more sophisticated quasi-Poisson model based on projection pursuit type regression models. E.g. Lingjarde, O. C. and Liestol, K. (1998) Generalized Projection Pursuit Regression, SIAM Journal of Scientific Computing , 20 , 844-857. http://dx.doi.org/10.1137/S1064827595296574 • Incorporate measure of uncertainty due to imputation SAMSI-SAVI Workshop 14 June 3, 2016
Sujit K. Ghosh • Develop plans for collaborations and training (especially on the use of R and advanced statistical modeling) • E.g., SAMSI uses web-based interface like webex and SAKAI for regular research group meeting • Arrange for short visits for training on an annual basis • E.g., International Indian Statistical Association (IISA) organizes annual conferences in U.S. and India on alternative years • SAMSI will launch a yearlong program on Climate: https://www.samsi.info/programs-and-activities/year-long-research-progra • Create a (secure) Dropbox (or similar web-based sharing tools) to share data, codes and research materials • Other ideas?.... SAMSI-SAVI Workshop 15 June 3, 2016
Sujit K. Ghosh Additional Questions?, Feedback? Interested to join or organize a forum on an emerging area of methodological research? Contact me... Sujit Ghosh, Deputy Director ( ghosh@samsi.info ) Statistical and Applied Mathematical Sciences Institute , NC, USA ( www.samsi.info ) RTP SAMSI-SAVI Workshop 16 June 3, 2016
Recommend
More recommend
