cat(paste("(C) (cc by-sa) Wouter van Atteveldt, file generated", format(Sys.Date(), format="%B %d %Y")))
Note on the data used in this howto: This data can be downloaded from http://piketty.pse.ens.fr/files/capital21c/en/xls/, but the excel format is a bit difficult to parse at it is meant to be human readable, with multiple header rows etc. For that reason, I've extracted csv files for some interesting tables that I've uploaded to https://github.com/vanatteveldt/learningr/tree/master/data. If you're accessing this tutorial from the githup project, these files should be in your 'data' sub folder automatically.
In this hands-on we continue with the
capital variable created in the transforming data howto.
You can also download this variable from the course pages:
First, let's split our countries into two groups, anglo-saxon countries and european countries (plus Japan):
We can use the
ifelse command here combined with the
anglo = c("U.S.", "U.K.", "Canada", "Australia") capital$Group = ifelse(capital$Country %in% anglo, "anglo", "european") table(capital$Group)
Now, let's see whether capital accumulation is different between these two groups.
t.test(capital$Private ~ capital$Group)
So, according to this test capital accumulation is indeed significantly higher in European countries than in Anglo-Saxon countries. Of course, the data here are not independently distributed since the data in the same year in different countries is related (as are data in subsequent years in the same country, but let's ignore that for the moment) We could also do a paired t-test of average accumulation per year per group by first using the cast command to aggregate the data. Note that we first remove the NA values (for Spain).
library(reshape2) capital = na.omit(capital) pergroup = dcast(capital, Year ~ Group, value.var="Private", fun.aggregate=mean) head(pergroup)
Let's plot the data to have a look at the lines:
plot(pergroup$Year, pergroup$european, type="l", xlab="Year", ylab="Capital accumulation") lines(pergroup$Year, pergroup$anglo, lty=2) legend("top", lty=c(1,2), legend=c("European", "Anglo-Saxon"))
So initially capital is higher in the Anglo-Saxon countries, but the European countries overtake quickly and stay higher.
Now, a paired-sample t-test again shows a significant difference between the two:
t.test(pergroup$anglo, pergroup$european, paired=T)
We can also use a one-way Anova to see whether accumulation differs per country. Let's first do a box-plot to see how different the countries are. Plot by default gives a box plot of a formula with a nominal independeny variable
plot(capital$Private ~ capital$Country)
So, it seems that in fact a lot of countries are quite similar, with some extreme cases of high capital accumulation.
(also, it seems that including Japan in the Europeaqn countries might have been a mistake).
We use the
aov function for this, the
anova function is meant to analyze already fitted models,
as will be shown below.
m = aov(capital$Private ~ capital$Country) summary(m)
So in fact there is a significant difference. We can use
pairwise.t.test to perform
posthoc = pairwise.t.test(capital$Private, capital$Country, p.adj = "bonf") round(posthoc$p.value, 2)
A more generic way of fitting models is using the
aov is a wrapper around
Let's model private capital as a function of country and public capital:
m = lm(Private ~ Country + Public, data=capital) summary(m)
As you can see, R automatically creates dummy values for nominal values, using the first value (U.S. in this case) as reference category. An alternative is to remove the intercept and create a dummy for each country:
m = lm(Private ~ Country + Public - 1, data=capital) summary(m)
- 1 removes the intercept because there is an implicit +1 constant for the intercept in the regression formula)
You can also introduce interaction terms by using either the
: operator (which only creates the interaction term)
* (which creates a full model including the main effects).
To keep the model somewhat parsimonious, let's use the country group rather than the country itself
m1 = lm(Private ~ Group + Public, data=capital) m2 = lm(Private ~ Group + Public + Group:Public, data=capital)
A nice package to display multiple regression results side by side is the
screenreg function from the
library(texreg) screenreg(list(m1, m2))
So, there is a significant interaction effect which displaces the main effect of public wealth.
Comparing and diagnosing models
A relevant question can be whether a model with an interaction effect is in fact a better model than the model without the interaction. This can be investigated with an anova of the model fits of the two models:
m1 = lm(Private ~ Group + Public, data=capital) m2 = lm(Private ~ Group + Public + Group:Public, data=capital) anova(m1, m2)
So, the interaction term is in fact a significant improvement of the model. Apparently, in European countries private capital is accumulated faster in those times that the government goes into depth.
After doing a linear model it is a good idea to do some diagnostics.
We can ask R for a set of standard plots by simply calling
plot on the model fit.
We use the parameter (
mfrow here to put the four plots this produces side by side.
old.settings = par(mfrow=c(2,2)) plot(m) par(old.settings)
See http://www.statmethods.net/stats/rdiagnostics.html for a more exhausitve list of model diagnostics.