User Tools

Site Tools


sampling

Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Both sides previous revision Previous revision
sampling [2018/03/13 16:48]
hkimscil [Sample statistics]
sampling [2018/03/13 16:49] (current)
hkimscil [Sample statistics]
Line 66: Line 66:
  
 var_ <- new.env() var_ <- new.env()
-n<​-20 ​           ## Sample n individuals at a time +n<​-20 ​       ## Sample n individuals at a time 
-p_mean<​-0 ​       ## Population mean +p_mean<​-0 ​   ## Population mean 
-p_sd<​-1 ​           ## Population standard deviation +p_sd<​-1 ​     ## Population standard deviation 
-N<​-500 ​           ## Number of times the experiment (sampling) is replicated+N<​-500 ​      ​## Number of times the experiment (sampling) is replicated
  
 pdf('​SE.pdf'​) pdf('​SE.pdf'​)
  
-for(i in 1:N)                                ## do the experiment N times+for(i in 1:N)     ​## do the experiment N times
 { {
-smp<​-rnorm(n,​p_mean,​p_sd) ​                ​## sample n data points from the population+smp<​-rnorm(n,​p_mean,​p_sd) ​   ## sample n data points from the population
  
-var_$x_bar<​-c(var_$x_bar,​mean(smp)) ​        ​## keep track of the mean (x_bar) from each sample+var_$x_bar<​-c(var_$x_bar,​mean(smp)) ​    ​## keep track of the mean (x_bar) from each sample 
 + 
 +hist(var_$x_bar,​probability=TRUE,​col="​red",​xlim=c(-4,​4),​xlab="​x / x_bar",​main="",​ylim=c(0,​2.2)) ​  
 +# Plot a histogram of x_bar values
  
-hist(var_$x_bar,​probability=TRUE,​col="​red",​xlim=c(-4,​4),​xlab="​x / x_bar",​main="",​ylim=c(0,​2.2)) ​ # Plot a histogram of x_bar values 
 points(mean(smp),​0,​pch=19,​cex=1.5,​col='​black'​) points(mean(smp),​0,​pch=19,​cex=1.5,​col='​black'​)
 curve(dnorm(x,​p_mean,​p_sd/​sqrt(n)),​lwd=3,​add=TRUE) curve(dnorm(x,​p_mean,​p_sd/​sqrt(n)),​lwd=3,​add=TRUE)
Line 86: Line 88:
 text(2.5,​1.5,​labels=paste('​standard deviation of\nsample means = ',​round(sd(var_$x_bar),​2),​sep=''​) ) text(2.5,​1.5,​labels=paste('​standard deviation of\nsample means = ',​round(sd(var_$x_bar),​2),​sep=''​) )
  
-curve(dnorm(x,​p_mean,​p_sd),​main="",​ylab="",​xlim=c(-4,​4),​xlab="​X",​col="​blue",​lwd=3,​add=TRUE) ## Plot the sample+curve(dnorm(x,​p_mean,​p_sd),​main="",​ylab="",​xlim=c(-4,​4),​xlab="​X",​col="​blue",​lwd=3,​add=TRUE) ​ 
 +## Plot the sample
  
 text(2.5,​0.5,​labels=paste('#​ of means drawn = ',​i,​sep=''​)) text(2.5,​0.5,​labels=paste('#​ of means drawn = ',​i,​sep=''​))
Line 100: Line 103:
 dev.off() dev.off()
 </​code>​ </​code>​
- +{{SE.pdf}} ​
  
   * Variation See, [[:​Variance]]:​ 225.0584138 (15^2)   * Variation See, [[:​Variance]]:​ 225.0584138 (15^2)
sampling.txt · Last modified: 2018/03/13 16:49 by hkimscil