# how to map an estimated risk ratio of 0.76 to increased life expectancy

Andrew Gelman's blog contains the following question:

John Ioannidis writes: Assuming the meta-analyzed evidence from cohort studies represents life span–long causal associations, for a baseline life expectancy of 80 years, eating 12 hazelnuts daily (1 oz) would prolong life by 12 years (ie, 1 year per hazelnut)

...With increasing intake (for each daily serving) of ... nuts (RR: 0.76; 95% CI: 0.69, 0.84)... the risk of all-cause mortality decreased . . .

...I assume that hazelnuts count as nuts for this serving size.

The next question is how to map an estimated risk ratio of 0.76 to increased life expectancy. There’s gotta be some standard formula for this

So I'm curious: what is the standard formula for this?

https://statmodeling.stat.columbia.edu/2019/01/26/article-portrays-things-accurately-nutrition-literature-even-worse-shape-thought/#respond

One way to approach this is to use the death rate $$f(j)$$ at a specific age $$j$$ in a specific year, which can be obtained from life tables, in order to predict the life expectancy of a person currently living.

(obviously those death rates will not remain constant and there are many more ways to tackle this problem to get better estimates, but the method suits the purpose of testing the effect of risk ratios on life expectancy)

Then for a person of $$y$$ years old

$$\begin{array}{} P(\text{surival to x years}) &=& \prod_{y\leq j \leq x-1} (1 - f(j))\\ P(\text{death at age = x}) &=& P(\text{surival to x years}) f(x)\\ E(\text{age}) &=& \sum_{0 \leq x < \infty} x P(\text{death at age = x}) \end{array}$$

Example:

Say you use the table 'Life table for the total population: United States, 2003' from the image in that earlier mentioned wikipedia link.

The image below shows the variation of the expected life according to the above formulae. On the x-axis there is a variation in the moment when the RR actually kicks in (Gelman gave an example using 40 years onward).

These results here are far different from the 12 years (but I do not have the numbers of that estimate so clear in order to go into it in more detail). Anyway, I guess that the point from the blogpost was more that the effects should not be considered to add up (which still stands whether or not that 12 years number is correct or not).

# compute
#    - life expextancy
#    - probabiltiy to die at age x
#    - death rate
#    - survival rate
life_expect <- function(base,beginage,rr,rrstart=101) {

# death rate
rel <- rep(1,100)
if (rrstart < 101) {
rel[rrstart:100] <- rr
}
death_rate <- c(base[1:100]*rel, base[101])

# survival rate
survival <- rep(1,101)
for (i in 1:100) {
survival[i+1]  = survival[i]*(1-death_rate[i])
}

# probability to die at age x
p_die <- survival * death_rate

# life expectancy

Elife <- sum(p_die[(beginage+1):101]*c(beginage:100))/
sum(p_die[(beginage+1):101])

list(death_rate = death_rate,
survival = survival,
p_die = p_die,
Elife = Elife)
}

# from  ftp://ftp.cdc.gov/pub/Health_Statistics/NCHS/Publications/NVSR/54_14/Table01.xls
base <- c(0.00686507084137925,0.000468924103840803,0.000337018612082993,0.000253980748012471,0.000193730651433952,0.000177467463768319,0.000160266920016088,0.000146864401608979,0.000132260863615305,0.000117412511687535,0.000108988416427791,0.000117882657537237,0.00015665216302825,0.000233187617725824,0.000339382523440112,0.000459788146727592,0.000576973385719181,0.000684155944043895,0.000768733212499693,0.000831959733234743,0.000894302696081951,0.000954208212234048,0.000989840925560537,0.000996522526309545,0.00098215260061939,0.000959551106572387,0.000942388041116207,0.000935533446389084,0.000946822022702617,0.00097378267030598,0.00100754405484986,0.0010463061900096,0.00109701785072833,0.00116237295935761,0.00124365648706804,0.00133574435463189,0.0014410461391004,0.0015673411143621,0.00171380631074604,0.0018736380419753,0.00203766165711833,0.00220659167333691,0.00238942699716915,0.00259301587170481,0.00281861738406178,0.00306417992710891,0.00332180268908611,0.00358900693685323,0.00386267209667191,0.00414777667611931,0.00445827861595176,0.00479990363846949,0.00516531829562337,0.00555390618653441,0.00597132583819979,0.00642322495833418,0.00692461135042076,0.00749557575640038,0.0081595130519956,0.00892672789984719,0.00982654537395458,0.010830689769232,0.0118723751877809,0.0128914065482476,0.0139080330996353,0.0150030256703387,0.0162668251372316,0.0176990779563976,0.0193202301703282,0.0211079685238627,0.0229501723647085,0.0249040093508705,0.0271512342884117,0.0297841240612845,0.0327533107326732,0.0358306701555879,0.0389873634123265,0.0425026123367764,0.0465565209898809,0.0511997331749049,0.0563354044485466,0.0618372727625818,0.0678564046096954,0.0745037414774353,0.0819753395107449,0.0896822973078052,0.0980311248111167,0.107059411952568,0.116803935241159,0.127299983985204,0.138580592383723,0.150675681864781,0.16361112298441,0.177407732357604,0.192080226605893,0.207636162412373,0.224074899057897,0.241386626061258,0.259551503859515,0.278538968828674,1)

# there are many things that you can do with the above function
# here is an example of computing the life expectancy
# as function of the relative risk rate (of dying)
# and the age when this RR kicks off.

z <- matrix(rep(0,101*101),101)
x <- c(0:100)
y <- seq(0.5,1.5,length.out = 101)
for (i in 1:101) {
for(j in 1:101) {
z[i,j] <- life_expect(base,0,rr = y[j],rrstart = x[i])\$Elife
}
}
min(z)
max(z)

# contour plot
filled.contour(x,y,z,
xlab="age risk starts",ylab="RR",
#levels=c(-500,-400,-300,-200,-100,-10:-1),
color.palette=function(n) {hsv(c(seq(0.15,0.7,length.out=n),0),
c(seq(0.7,0.2,length.out=n),0),
c(seq(1,0.7,length.out=n),0.9))},
levels=70:85,
plot.axes= c({
title("life expectancy for someone who is currently 0 years")
axis(1)
axis(2)
},""),
xlim=range(x)+c(-0.0,0.0),
ylim=range(y)+c(-0.0,0.0)
)


The following answer was posted on Gelman's blog (the source of the original question), and seems to provide a good rough approximation:

Joshua R Goldstein says: January 29, 2019 at 11:35 am

There’s a nice literature on life table “entropy” that has analytic expressions on how to convert a uniform change in mortality by age into a change in life expectancy at birth.

The basic result is that the proportional change in life expectancy is currently equal to about .1 to .2 times the change in mortality rates. So if some treatment causes mortality to drop by 10% at all ages, life expectancy at birth will increase by 1-2%.

For those interested, here is our paper that reviews this classic result with some extensions to slowing the rate of aging: Goldstein, Joshua R., and Thomas Cassidy. “How slowing senescence translates into longer life expectancy.” Population studies 66.1 (2012): 29-37.