Very quick one this! Really interesting paper from Ioannidis et al about citation indices.
I wanted to recreate figure 1, which is:
Closest I could get (code here) is this:
Biggest difference is in NS, where they find all negative correlations, but most of mine are positive.
Source data are Table S1 Data.
Am I doing something wrong? Or is the paper wrong?
UPDATE 9th July 2016
Using Spearman gets us closer but it’s still not quite correct (updated code too)
9th July 2016 at 2:16 am
Probably non normal distribution. Rank correlation maybe more appropriate and gives results as reported in the paper.
If df = the dataframe from the paper
library(dplyr)
df1%select(Field:NC,NS)%>%group_by(Field)%>%summarise(NSpearson=cor(NS,NC),NSspearman=cor(NS,NC,method=”spearman”))%>%gather(Method,value,NSpearson:NSspearman)
ggplot(df1,aes(x=Field,y=value,fill=Field))+facet_grid(Method~.)+geom_bar(stat=”identity”)
Though not familiar with \N field .
9th July 2016 at 2:19 am
The code didn’t come out right:
library(dplyr)
library(ggplot2)
df1%select(Field:NC,NS)%>%group_by(Field)%>%summarise(NSpearson=cor(NS,NC),NSspearman=cor(NS,NC,method=”spearman”))
9th July 2016 at 2:22 pm
Oh autoformatted again! Its the first bit. I will try one last time:
the df1 above (on the first post )should read:
df1=df %>% select(Field:NC,NS)%>%group_by(Field)%>%summarise(NSpearson=cor(NS,NC),NSspearman=cor(NS,NC,method=”spearman”))%>%gather(Method,value,NSpearson:NSspearman)
ggplot(df1,aes(x=Field,y=value,fill=Field))+facet_grid(Method~.)+geom_bar(stat=”identity”)
Hopefully this will be more legible.
Also forgot to mention : library(tidyr).
9th July 2016 at 3:28 pm
Maybe use a gist 🙂
9th July 2016 at 4:01 pm
Ah yes. Thanks! – newbie to commenting that too with code snippet. 🙂
9th July 2016 at 6:36 pm
I added in a Spearman analysis (it isn’t clear from the paper which they used) and it gets us close but not quite….
9th July 2016 at 7:00 pm
Weird. Maybe for one last attempt, try method=”kendall” in cor() for Kendall’s Tau. I recall that method being pretty slow compared to pearson and spearman — so hopefully its not a big dataset.
13th July 2016 at 4:54 pm
I used log-transformed values for np, h, hm, s, sf, sfl for the correlations. Try that.
13th July 2016 at 4:58 pm
log(n+1)?
Which base?
13th July 2016 at 5:30 pm
In(n+1)/ln(max(n)+1) for each of the 6 original indicators puts them each within a [0,1] range. The sum of these gives the composite value.
25th July 2016 at 12:09 pm
Hi Kevin
I tried again with your formula and still couldn’t reproduce the graph:
https://github.com/mw55309/PLOS_citation_graph/blob/master/boyack.R
And
https://raw.githubusercontent.com/mw55309/PLOS_citation_graph/master/boyack_results_pearson.png
Cheers
Mick