Calculate within and between variances and confidence intervals in R
https://stackoverflow.com/questions/1401894
-
05-07-2019 - |
italiano
english
français
española
中国
日本の
العربية
Deutsch
한국어
Português
RussianQuestion
I need to calculate the within and between run variances from some data as part of developing a new analytical chemistry method. I also need confidence intervals from this data using the R language
I assume I need to use anova or something ?
My data is like
> variance
Run Rep Value
1 1 1 9.85
2 1 2 9.95
3 1 3 10.00
4 2 1 9.90
5 2 2 8.80
6 2 3 9.50
7 3 1 11.20
8 3 2 11.10
9 3 3 9.80
10 4 1 9.70
11 4 2 10.10
12 4 3 10.00
Solution 4
I've been looking at a similar problem. I've found reference to caluclating confidence intervals by Burdick and Graybill (Burdick, R. and Graybill, F. 1992, Confidence Intervals on variance components, CRC Press)
Using some code I've been trying I get these values
> kiaraov = aov(Value~Run+Error(Run),data=kiar)
> summary(kiaraov)
Error: Run
Df Sum Sq Mean Sq
Run 3 2.57583 0.85861
Error: Within
Df Sum Sq Mean Sq F value Pr(>F)
Residuals 8 1.93833 0.24229
> confint = 95
> a = (1-(confint/100))/2
> grandmean = as.vector(kiaraov$"(Intercept)"[[1]][1]) # Grand Mean (I think)
> within = summary(kiaraov)$"Error: Within"[[1]]$"Mean Sq" # S2^2Mean Square Value for Within Run
> dfRun = summary(kiaraov)$"Error: Run"[[1]]$"Df"
> dfWithin = summary(kiaraov)$"Error: Within"[[1]]$"Df"
> Run = summary(kiaraov)$"Error: Run"[[1]]$"Mean Sq" # S1^2Mean Square for between Run
> between = (Run-within)/((dfWithin/(dfRun+1))+1) # (S1^2-S2^2)/J
> total = between+within
> between # Between Run Variance
[1] 0.2054398
> within # Within Run Variance
[1] 0.2422917
> total # Total Variance
[1] 0.4477315
> betweenCV = sqrt(between)/grandmean * 100 # Between Run CV%
> withinCV = sqrt(within)/grandmean * 100 # Within Run CV%
> totalCV = sqrt(total)/grandmean * 100 # Total CV%
> #within confidence intervals
> withinLCB = within/qf(1-a,8,Inf) # Within LCB
> withinUCB = within/qf(a,8,Inf) # Within UCB
> #Between Confidence Intervals
> n1 = dfRun
> n2 = dfWithin
> G1 = 1-(1/qf(1-a,n1,Inf)) # According to Burdick and Graybill this should be a
> G2 = 1-(1/qf(1-a,n2,Inf))
> H1 = (1/qf(a,n1,Inf))-1 # and this should be 1-a, but my results don't agree
> H2 = (1/qf(a,n2,Inf))-1
> G12 = ((qf(1-a,n1,n2)-1)^2-(G1^2*qf(1-a,n1,n2)^2)-(H2^2))/qf(1-a,n1,n2) # again, should be a, not 1-a
> H12 = ((1-qf(a,n1,n2))^2-H1^2*qf(a,n1,n2)^2-G2^2)/qf(a,n1,n2) # again, should be 1-a, not a
> Vu = H1^2*Run^2+G2^2*within^2+H12*Run*within
> Vl = G1^2*Run^2+H2^2*within^2+G12*within*Run
> betweenLCB = (Run-within-sqrt(Vl))/J # Betwen LCB
> betweenUCB = (Run-within+sqrt(Vu))/J # Between UCB
> #Total Confidence Intervals
> y = (Run+(J-1)*within)/J
> totalLCB = y-(sqrt(G1^2*Run^2+G2^2*(J-1)^2*within^2)/J) # Total LCB
> totalUCB = y+(sqrt(H1^2*Run^2+H2^2*(J-1)^2*within^2)/J) # Total UCB
> result = data.frame(Name=c("within", "between", "total"),CV=c(withinCV,betweenCV,totalCV),LCB=c(sqrt(withinLCB)/grandmean*100,sqrt(betweenLCB)/grandmean*100,sqrt(totalLCB)/grandmean*100),UCB=c(sqrt(withinUCB)/grandmean*100,sqrt(betweenUCB)/grandmean*100,sqrt(totalUCB)/grandmean*100))
> result
Name CV LCB UCB
1 within 4.926418 3.327584 9.43789
2 between 4.536327 NaN 19.73568
3 total 6.696855 4.846030 20.42647
Here the lower confidence interval for between run CV is less than zero, so reported as NaN.
I'd love to have a better way to do this. If I get time I might try to create a function to do this.
Paul.
--
Edit: I did eventually write a function, here it is (caveat emptor)
#' avar Function
#'
#' Calculate thewithin, between and total %CV of a dataset by ANOVA, and the
#' associated confidence intervals
#'
#' @param dataf - The data frame to use, in long format
#' @param afactor Character string representing the column in dataf that contains the factor
#' @param aresponse Charactyer string representing the column in dataf that contains the response value
#' @param aconfidence What Confidence limits to use, default = 95%
#' @param digits Significant Digits to report to, default = 3
#' @param debug Boolean, Should debug messages be displayed, default=FALSE
#' @returnType dataframe containing the Mean, Within, Between and Total %CV and LCB and UCB for each
#' @return
#' @author Paul Hurley
#' @export
#' @examples
#' #Using the BGBottles data from Burdick and Graybill Page 62
#' assayvar(dataf=BGBottles, afactor="Machine", aresponse="weight")
avar<-function(dataf, afactor, aresponse, aconfidence=95, digits=3, debug=FALSE){
dataf<-subset(dataf,!is.na(with(dataf,get(aresponse))))
nmissing<-function(x) sum(!is.na(x))
n<-nrow(subset(dataf,is.numeric(with(dataf,get(aresponse)))))
datadesc<-ddply(dataf, afactor, colwise(nmissing,aresponse))
I<-nrow(datadesc)
if(debug){print(datadesc)}
if(min(datadesc[,2])==max(datadesc[,2])){
balance<-TRUE
J<-min(datadesc[,2])
if(debug){message(paste("Dataset is balanced, J=",J,"I is ",I,sep=""))}
} else {
balance<-FALSE
Jh<-I/(sum(1/datadesc[,2], na.rm = TRUE))
J<-Jh
m<-min(datadesc[,2])
M<-max(datadesc[,2])
if(debug){message(paste("Dataset is unbalanced, like me, I is ",I,sep=""))}
if(debug){message(paste("Jh is ",Jh, ", m is ",m, ", M is ",M, sep=""))}
}
if(debug){message(paste("Call afactor=",afactor,", aresponse=",aresponse,sep=""))}
formulatext<-paste(as.character(aresponse)," ~ 1 + Error(",as.character(afactor),")",sep="")
if(debug){message(paste("formula text is ",formulatext,sep=""))}
aovformula<-formula(formulatext)
if(debug){message(paste("Formula is ",as.character(aovformula),sep=""))}
assayaov<-aov(formula=aovformula,data=dataf)
if(debug){
print(assayaov)
print(summary(assayaov))
}
a<-1-((1-(aconfidence/100))/2)
if(debug){message(paste("confidence is ",aconfidence,", alpha is ",a,sep=""))}
grandmean<-as.vector(assayaov$"(Intercept)"[[1]][1]) # Grand Mean (I think)
if(debug){message(paste("n is",n,sep=""))}
#This line commented out, seems to choke with an aov object built from an external formula
#grandmean<-as.vector(model.tables(assayaov,type="means")[[1]]$`Grand mean`) # Grand Mean (I think)
within<-summary(assayaov)[[2]][[1]]$"Mean Sq" # d2e, S2^2 Mean Square Value for Within Machine = 0.1819
dfRun<-summary(assayaov)[[1]][[1]]$"Df" # DF for within = 3
dfWithin<-summary(assayaov)[[2]][[1]]$"Df" # DF for within = 8
Run<-summary(assayaov)[[1]][[1]]$"Mean Sq" # S1^2Mean Square for Machine
if(debug){message(paste("mean square for Run ?",Run,sep=""))}
#Was between<-(Run-within)/((dfWithin/(dfRun+1))+1) but my comment suggests this should be just J, so I'll use J !
between<-(Run-within)/J # d2a (S1^2-S2^2)/J
if(debug){message(paste("S1^2 mean square machine is ",Run,", S2^2 mean square within is ",within))}
total<-between+within
between # Between Run Variance
within # Within Run Variance
total # Total Variance
if(debug){message(paste("between is ",between,", within is ",within,", Total is ",total,sep=""))}
betweenCV<-sqrt(between)/grandmean * 100 # Between Run CV%
withinCV<-sqrt(within)/grandmean * 100 # Within Run CV%
totalCV<-sqrt(total)/grandmean * 100 # Total CV%
n1<-dfRun
n2<-dfWithin
if(debug){message(paste("n1 is ",n1,", n2 is ",n2,sep=""))}
#within confidence intervals
if(balance){
withinLCB<-within/qf(a,n2,Inf) # Within LCB
withinUCB<-within/qf(1-a,n2,Inf) # Within UCB
} else {
withinLCB<-within/qf(a,n2,Inf) # Within LCB
withinUCB<-within/qf(1-a,n2,Inf) # Within UCB
}
#Mean Confidence Intervals
if(debug){message(paste(grandmean,"+/-(sqrt(",Run,"/",n,")*qt(",a,",df=",I-1,"))",sep=""))}
meanLCB<-grandmean+(sqrt(Run/n)*qt(1-a,df=I-1)) # wrong
meanUCB<-grandmean-(sqrt(Run/n)*qt(1-a,df=I-1)) # wrong
if(debug){message(paste("Grandmean is ",grandmean,", meanLCB = ",meanLCB,", meanUCB = ",meanUCB,aresponse,sep=""))}
if(debug){print(summary(assayaov))}
#Between Confidence Intervals
G1<-1-(1/qf(a,n1,Inf))
G2<-1-(1/qf(a,n2,Inf))
H1<-(1/qf(1-a,n1,Inf))-1
H2<-(1/qf(1-a,n2,Inf))-1
G12<-((qf(a,n1,n2)-1)^2-(G1^2*qf(a,n1,n2)^2)-(H2^2))/qf(a,n1,n2)
H12<-((1-qf(1-a,n1,n2))^2-H1^2*qf(1-a,n1,n2)^2-G2^2)/qf(1-a,n1,n2)
if(debug){message(paste("G1 is ",G1,", G2 is ",G2,sep=""))
message(paste("H1 is ",H1,", H2 is ",H2,sep=""))
message(paste("G12 is ",G12,", H12 is ",H12,sep=""))
}
if(balance){
Vu<-H1^2*Run^2+G2^2*within^2+H12*Run*within
Vl<-G1^2*Run^2+H2^2*within^2+G12*within*Run
betweenLCB<-(Run-within-sqrt(Vl))/J # Betwen LCB
betweenUCB<-(Run-within+sqrt(Vu))/J # Between UCB
} else {
#Burdick and Graybill seem to suggest calculating anova of mean values to find n1S12u/Jh
meandataf<-ddply(.data=dataf,.variable=afactor, .fun=function(df){mean(with(df, get(aresponse)), na.rm=TRUE)})
meandataaov<-aov(formula(paste("V1~",afactor,sep="")), data=meandataf)
sumsquare<-summary(meandataaov)[[1]]$`Sum Sq`
#so maybe S12u is just that bit ?
Runu<-(sumsquare*Jh)/n1
if(debug){message(paste("n1S12u/Jh is ",sumsquare,", so S12u is ",Runu,sep=""))}
Vu<-H1^2*Runu^2+G2^2*within^2+H12*Runu*within
Vl<-G1^2*Runu^2+H2^2*within^2+G12*within*Runu
betweenLCB<-(Runu-within-sqrt(Vl))/Jh # Betwen LCB
betweenUCB<-(Runu-within+sqrt(Vu))/Jh # Between UCB
if(debug){message(paste("betweenLCB is ",betweenLCB,", between UCB is ",betweenUCB,sep=""))}
}
#Total Confidence Intervals
if(balance){
y<-(Run+(J-1)*within)/J
if(debug){message(paste("y is ",y,sep=""))}
totalLCB<-y-(sqrt(G1^2*Run^2+G2^2*(J-1)^2*within^2)/J) # Total LCB
totalUCB<-y+(sqrt(H1^2*Run^2+H2^2*(J-1)^2*within^2)/J) # Total UCB
} else {
y<-(Runu+(Jh-1)*within)/Jh
if(debug){message(paste("y is ",y,sep=""))}
totalLCB<-y-(sqrt(G1^2*Runu^2+G2^2*(Jh-1)^2*within^2)/Jh) # Total LCB
totalUCB<-y+(sqrt(H1^2*Runu^2+H2^2*(Jh-1)^2*within^2)/Jh) # Total UCB
}
if(debug){message(paste("totalLCB is ",totalLCB,", total UCB is ",totalUCB,sep=""))}
# result<-data.frame(Name=c("within", "between", "total"),CV=c(withinCV,betweenCV,totalCV),
# LCB=c(sqrt(withinLCB)/grandmean*100,sqrt(betweenLCB)/grandmean*100,sqrt(totalLCB)/grandmean*100),
# UCB=c(sqrt(withinUCB)/grandmean*100,sqrt(betweenUCB)/grandmean*100,sqrt(totalUCB)/grandmean*100))
result<-data.frame(Mean=grandmean,MeanLCB=meanLCB, MeanUCB=meanUCB, Within=withinCV,WithinLCB=sqrt(withinLCB)/grandmean*100, WithinUCB=sqrt(withinUCB)/grandmean*100,
Between=betweenCV, BetweenLCB=sqrt(betweenLCB)/grandmean*100, BetweenUCB=sqrt(betweenUCB)/grandmean*100,
Total=totalCV, TotalLCB=sqrt(totalLCB)/grandmean*100, TotalUCB=sqrt(totalUCB)/grandmean*100)
if(!digits=="NA"){
result$Mean<-signif(result$Mean,digits=digits)
result$MeanLCB<-signif(result$MeanLCB,digits=digits)
result$MeanUCB<-signif(result$MeanUCB,digits=digits)
result$Within<-signif(result$Within,digits=digits)
result$WithinLCB<-signif(result$WithinLCB,digits=digits)
result$WithinUCB<-signif(result$WithinUCB,digits=digits)
result$Between<-signif(result$Between,digits=digits)
result$BetweenLCB<-signif(result$BetweenLCB,digits=digits)
result$BetweenUCB<-signif(result$BetweenUCB,digits=digits)
result$Total<-signif(result$Total,digits=digits)
result$TotalLCB<-signif(result$TotalLCB,digits=digits)
result$TotalUCB<-signif(result$TotalUCB,digits=digits)
}
return(result)
}
assayvar<-function(adata, aresponse, afactor, anominal, aconfidence=95, digits=3, debug=FALSE){
result<-ddply(adata,anominal,function(df){
resul<-avar(dataf=df,afactor=afactor,aresponse=aresponse,aconfidence=aconfidence, digits=digits, debug=debug)
resul$n<-nrow(subset(df, !is.na(with(df, get(aresponse)))))
return(resul)
})
return(result)
}
OTHER TIPS
You have four groups of three observations:
> run1 = c(9.85, 9.95, 10.00)
> run2 = c(9.90, 8.80, 9.50)
> run3 = c(11.20, 11.10, 9.80)
> run4 = c(9.70, 10.10, 10.00)
> runs = c(run1, run2, run3, run4)
> runs
[1] 9.85 9.95 10.00 9.90 8.80 9.50 11.20 11.10 9.80 9.70 10.10 10.00
Make some labels:
> n = rep(3, 4)
> group = rep(1:4, n)
> group
[1] 1 1 1 2 2 2 3 3 3 4 4 4
Calculate within-run stats:
> withinRunStats = function(x) c(sum = sum(x), mean = mean(x), var = var(x), n = length(x))
> tapply(runs, group, withinRunStats)
$`1`
sum mean var n
29.800000000 9.933333333 0.005833333 3.000000000
$`2`
sum mean var n
28.20 9.40 0.31 3.00
$`3`
sum mean var n
32.10 10.70 0.61 3.00
$`4`
sum mean var n
29.80000000 9.93333333 0.04333333 3.00000000
You can do some ANOVA here:
> data = data.frame(y = runs, group = factor(group))
> data
y group
1 9.85 1
2 9.95 1
3 10.00 1
4 9.90 2
5 8.80 2
6 9.50 2
7 11.20 3
8 11.10 3
9 9.80 3
10 9.70 4
11 10.10 4
12 10.00 4
> fit = lm(runs ~ group, data)
> fit
Call:
lm(formula = runs ~ group, data = data)
Coefficients:
(Intercept) group2 group3 group4
9.933e+00 -5.333e-01 7.667e-01 -2.448e-15
> anova(fit)
Analysis of Variance Table
Response: runs
Df Sum Sq Mean Sq F value Pr(>F)
group 3 2.57583 0.85861 3.5437 0.06769 .
Residuals 8 1.93833 0.24229
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
> degreesOfFreedom = anova(fit)[, "Df"]
> names(degreesOfFreedom) = c("treatment", "error")
> degreesOfFreedom
treatment error
3 8
Error or within-group variance:
> anova(fit)["Residuals", "Mean Sq"]
[1] 0.2422917
Treatment or between-group variance:
> anova(fit)["group", "Mean Sq"]
[1] 0.8586111
This should give you enough confidence to do confidence intervals.
If you want to apply a function (such as var) across a factor such as Run or Rep, you can use tapply:
> with(variance, tapply(Value, Run, var))
1 2 3 4
0.005833333 0.310000000 0.610000000 0.043333333
> with(variance, tapply(Value, Rep, var))
1 2 3
0.48562500 0.88729167 0.05583333
I'm going to take a crack at this when I have more time, but meanwhile, here's the dput() for Kiar's data structure:
structure(list(Run = c(1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4), Rep = c(1,
2, 3, 1, 2, 3, 1, 2, 3, 1, 2, 3), Value = c(9.85, 9.95, 10, 9.9,
8.8, 9.5, 11.2, 11.1, 9.8, 9.7, 10.1, 10)), .Names = c("Run",
"Rep", "Value"), row.names = c(NA, -12L), class = "data.frame")
... in case you'd like to take a quick shot at it.
Licensed under: CC-BY-SA with attribution
Not affiliated with StackOverflow
