Open markotainio opened 6 years ago
Functional form of IER is following:
Where: C as PM2.5 exposure concentration (in ITHIM intake concentration); C0 as theoretical minimum risk exposure level; 𝛽 as ratio of the 𝑅𝑅 at low to high PM2.5 exposure concentrations; 𝛿 as the power of PM2.5 exposure concentration.
The R model to estimate these parameters are included in Cohen et al. 2017 appendix. This gives us two options:
I would prefer to start with second option.
I got reply from Burnett.
Newest GBD2016 parameters are here: https://cloud.ihme.washington.edu/index.php/s/puzbu28QteEHTmS
there are 4 parameters, TMREL, alpha, beta, delta with the model IER(PM2.5)=1+alpha(1-exp(-beta(PM2.5-TMREL)^delta))
Here is also R-code to explore the results:
setwd ("C:/Pooling/IER_CURVES_2015") dataIER<- read.csv("parameter_draws.csv ", header = T) PM=seq(0, 500, 1 ) RR=matrix(0, length(PM), 1000) LCL=matrix(0, length(PM), 1) UCL=matrix(0, length(PM), 1) MEAN=matrix(0, length(PM), 1) alpha=dataIER[,2] beta=dataIER[,3] gamma=dataIER[,4] tmrel=dataIER[,5] for (i in 1:length(PM)) { for (j in 1:1000) { if (PM[i]<tmrel[j]) {RR[i,j]=1} if (PM[i]>=tmrel[j]) { RR[i,j]=(1+alpha[j](1-exp(-beta[j](PM[i]-tmrel[j])^gamma[j]))) } } MEAN[i]=mean(RR[i,]) UCL[i]=quantile(RR[i,], 0.975) LCL[i]=quantile(RR[i,], 0.025) } plot(PM, UCL, type="l", lty=1, col="lightgrey", ylab="Relative Risk", xlab=expression(paste("PM"[2.5], " - ", mu, "g/m"^3))) polygon(x=c(PM, rev(PM)), y=c(LCL, rev(UCL)), col="lightgrey", border=NA, lty=2) lines(PM, MEAN, lwd=4, col="blue") abline(1,0)
Is the file _parameterdraws.csv available? I see it now on NextCloud.
File is available behind this link: https://cloud.ihme.washington.edu/index.php/s/puzbu28QteEHTmS
But we will also add these to Google drive.
Yes, thank you. I see it now.
@markotainio Do you know how _agecode is defined in _parameterdraws.csv? Thanks.
I have posted an R markdown script on Slack here. In it, spline functions are created to fit the mean RR across a range of PM 2.5 values. One of these functions is then used in an example population attributable fraction calculation. The R markdown script can be converted to html with knit2html("PM25.Rmd")
. Please post comments on Slack. Thanks.
@syounkin About age: Assuming that age-groups are same as in Burnett et al. 2014 (https://ehp.niehs.nih.gov/1307049/), age 25 refers to age group 25-29, age 30 refers to age group 30-34 etc. Last age group is 95+. Age 99 means that same RR is used for all age groups. With exception of ALRI these RR's are estimated for age groups 25+ so for age group 20-24 effect is assumed to be zero. However, I think we can use 25-29 data also for that age group.
From the ITHIM-R meeting reminding that the tool would need to have a possibility to use different RRs if local decision makers have preferred method (expressed by Neil). Implementation of this should be doable.
I propose that future methods that we write for PM2.5 include an argument for the relative risk function, RR
. If RR
is omitted by the user, then a default function is used, say one of the spline functions we've already computed here.
Here is some sample code. CRA
stands for comparative risk assessment.
CRA.PM25 <- function( pm25.baseline, pm25.scenario, RR ){
# RR is a function.
# Include some code here.
# We need to assign a distribution to PM25.
return(PAF)
}
# for default RR function
CRA.PM25( pm25.baseline, pm25.scenario )
# for user-specified RR function
CRA.PM25( pm25.baseline, pm25.scenario, RR = function(PM) exp(-PM) )
Something like that. If we limit user defined options to log-linear RRs with possible threshold value, then user would need to provide just two numbers (RR and threshold), and model would do rest.
Something like this?
logLinearPM <- function(PM, alpha, beta, gamma){
logRR <- ifelse( PM < gamma, alpha*log(PM) + beta, alpha*log(gamma) + beta )
RR <- exp(logRR)
return(RR)
}
Don't we need two parameters for the log-linear relationship (alpha, beta) and one for the threshold (gamma)?
Typical form for log-linear DRF for PM2.5 has equation:
RR = EXP(LN(RR)*(PM2.5concentration/PM2.5concentrationRR))
For example, based on Heroux et al. 2015 paper, which defines WHO recommended RR for all-cause mortality, the equations is:
RR = EXP(LN(1.062)*(PM2.5concentration/10))
So here parameters needed are: RR = 1.062 PM2.5 concentration for which the RR has been define = 10 Possible threshold value under which RR is 1 = Here none.
So in fact we need 1-3 parameters, if we restrict the RRs for log-linear models (which is the most common way). @syounkin
This issue is about discussion how we transfer air pollution concentration for RRs. With air pollution we mean at this stage fine particulate matter (PM2.5).
This divides to two parts:
For disease specific RRs we use the method from IHME Global Burden of Disease study (so called Integrated Risk Functions, IRFs). The methodology has been presented in Burnett et al. 2014 article. http://dx.doi.org/10.1289/ehp.1307049.
Diseases for adults are:
Methodology to apply these values is in Cohen et al. 2017 paper. http://dx.doi.org/10.1016/ S0140-6736(17)30505-6
Key equation is in page 1910 in Cohen et al. However, that paper wont include parameters needed to do the calculations. This we need to ask from Rick Burnett directly.
For all-cause mortality we will use similar method after the new study describing the method and results will be published. This should happen in 2018.