Simpsons (version 0.1.0)

Simpsons: Simpsons

Description

This package detects instances of Simpson's Paradox in datasets of bivariate continuous data. It examines subpopulations in the data, either user-defined or by means of cluster analysis, to test whether a regression at the level of the group is in the opposite direction at the level of subpopulations.

Usage

Simpsons(X, Y, clusterid, clustervars, data, nreps = 5000)

Arguments

X

The first continuous variable to be used in the regression analysis.

Y

The second continuous variable to be used in the regression analysis.

clusterid

If you have a vector describing group membership, such as gender, you can specify it here. This will then be used to test for possible instances of Simpson's Paradox. If left empty, a cluster analysis will attempt to discover clusters in the data. See example 1.

clustervars

By default, the cluster analysis will be carried out on the X and Y variables. If you want to define the clusters on the basis of a different set of variables, such as a questionnaire, you can specify them using this command. See example 3.

data

Describes the data matrix. Should be a dataframe.

nreps

nreps specifies the number of permutations run for each cluster in the permutation significance test. The default is 5000. Each repetition is stored in the matrix 'permutationtest'.

Value

A list of class "Simpson" with the following elements:

Nclusters

Number of clusters estimated in the data (or the number of different groups in the 'clusterid' column defined by the user).

clustersize

the size of each estimated cluster

alldata

the original dataset with the clusterid's appended as a new column

Allbeta

A matrix of beta estimates for each cluster

Allint

A matrix of intercepts for each cluster

permutationtest

The matrix of all permutations. The columns define the clusters, the rows specify the difference in the beta of the group and the beta of that cluster for each iteration, thus generating the null distribution

namex

The first variable used in the analysis

namey

The second variable used in the analysis

pvalues

The p-values for the significance of the regressions

mclustanalysis

Object of class Mclust that contains all mclust results

Details

This package detects instances of Simpson's Paradox in datasets. That is, it tests whether some bivariate relationship found at the level of the whole dataset is consistent (in direction and strength) for possible subpopulations. It examines whether there is evidence for more than one cluster in the data in the data using cluster analysis, either user-defined or by means of cluster analysis. Then, it plots the data, using a different color for every cluster, plots the regression lines for each cluster, and estimates the regression of X on Y for each cluster. Finally, it tests whether the regression at the level of the whole dataset is different from the regression at the level of the subclusters.

Because clusters in the data are part of the whole dataset, and therefore create a dependency, a permutation test is used to test for significant differences. For each cluster, the cluster labels are permuted within the whole dataset, the regression is run within the cluster and the whole dataset, and the difference between these two betas is stored as 1 repetition of the null distribution and stored in the object 'permutationtest'. A regression is considered significantly different from the group if the difference in beta estimate exceeds the lower or upper 2,5 percent of the permuted null distribution. If this is the case, a warning is issued as follows: "Warning: Beta regression estimate in cluster X is significantly different compared to the group!". If the sign of the regression within a cluster is different (positive or negative) than the sign for the group and the beta estimate deviates significantly, a warning states "Sign reversal: Simpson's Paradox! Cluster X is significantly different and in the opposite direction compared to the group!"

References

Kievit, R.A., Frankenhuis, W. E. , Waldorp, L. J. & Borsboom, D. (in preparation). Simpson's Paradox in Psychological Science: A Practical Guide. http://rogierkievit.com/simpsonsparadox.html

Examples

## Not run: 
# #This section contains three examples of the types of analyses you can run
# #using the 'Simpsons' function, illustrating the commmands and the types of #output.
# 
# #Example 1. Here, we want to estimate the relationship between 'Coffee' 
# #and 'Neuroticism', taking into account possible gender differences. 
# #As we have measured gender, we supply this information using the #'clusterid' command. 
# #This means that the function runs the analysis both for 
# #the dataset as a whole and within the two subgroups. 
# #It then checks whether the subgroups deviate significantly 
# #from the regression at the level of the group.
# 
# 	#Simulating 100 males 
# 	coffeem=rnorm(100,100,15)
# 	neuroticismm=(coffeem*.8)+rnorm(100,15,8)
# 	clusterid=rep(1,100)
# 	males=cbind(coffeem,neuroticismm,clusterid)
# 
# 	#Simulating 100 females
# 	coffeef=rnorm(100,100,15)
# 	neuroticismf=160+((coffeef*-.8)+rnorm(100,15,8))
# 	clusterid=rep(2,100)
# 	females=cbind(coffeef,neuroticismf,clusterid)
# 	
# 	data=data.frame(rbind(males,females))
# 	colnames(data) <- c("Coffee","Neuroticism","gender")
# 
# #'normal' data analysis: Plot & regression
# plot(data[,1:2])
# a=lm(data[,1]~data[,2])
# abline(a)
# summary(a) #A normal regression shows no effect
# 
# #Running a Simpsons Paradox analysis, using gender as known clustering #variable
# example1=Simpsons(Coffee,Neuroticism,clusterid=gender, data=data) 
# # Analyze the relationship between coffee and neuroticism for both males 
# # and females. 
# example1
# 
# 
# 
# #example 2. Here we estimate the relationship between 'Coffee' and 'Neuroticism'. 
# #As opposed to example 1, we have not measured any possible clustering #identifiers 
# #such as gender, so we want to estimate whether there is evidence for #clustering based 
# #only on the data we measured: Coffee and Neuroticism.
# 
# #generating data 
# Coffee1=rnorm(100,100,15)
# Neuroticism1=(Coffee1*.8)+rnorm(100,15,8)
# g1=cbind(Coffee1, Neuroticism1)
# Coffee2=rnorm(100,170,15)
# Neuroticism2=(300-(Coffee2*.8)+rnorm(100,15,8))
# g2=cbind(Coffee2, Neuroticism2)
# Coffee3=rnorm(100,140,15)
# Neuroticism3=(200-(Coffee3*.8)+rnorm(100,15,8))
# g3=cbind(Coffee3, Neuroticism3)
# data2=data.frame(rbind(g1,g2,g3))
# colnames(data2) <- c("Coffee","Neuroticism")
# 
# #'normal' data analysis: Plot & regression
# plot(data2)
# b=lm(data2[,1]~data2[,2]) 
# summary(b)
# abline(b)
# 
# # Running the analysis tool identifies three clusters, and warns that the relationship 
# between alcohol and coffee is in the opposite direction in two of the subclusters.
# example2=Simpsons(Coffee,Neuroticism,data=data2) 
# example2
# 
# #example3: 
# 
# #In this final example, we want again want to analyse the relationship
# # between 'Alcohol' and 'Mood'. However, this time 
# #we have reason to believe that responses to a questionnaire 
# #will fall into clusters of response types. Therefore, we want to
# # estimate the clusters in the data on the basis of a different set
# # of variables. In this case, we have simulate three types of responses
# # to a questionnaire of nine questions, with continuous responses 
# #ranging between 1 and 7. We then first estimate the clusters on 
# #the basis of the questionnaire, and then examine the relationship 
# #between 'Alcohol' and 'Mood' based on these detected clusters.
# 
# #group 1
# signal=matrix(rnorm(300,7,1),100,3)
# noise=matrix(rnorm(600,3.5,1),100,6)
# g1=cbind(signal,noise)
# 
# #group 2
# signal=matrix(rnorm(300,1,1),100,3)
# noise=matrix(rnorm(600,3.5,1),100,6)
# g2=cbind(noise, signal)
# 
# #group 3
# signal=matrix(rnorm(300,7,1),100,3)
# noise1=matrix(rnorm(300,3.5,1),100,3)
# noise2=matrix(rnorm(300,3.5,1),100,3)
# g3=cbind(noise1,signal,noise2)
# 
# questionnaire=rbind(g1,g2,g3)
# colnames(questionnaire)=c('q1','q2','q3','q4','q5','q6','q7','q8','q9')
# 
# Alc1=rnorm(100,10,8)
# Mood1=(Alc1*.4)+rnorm(100,3,4)
# A=cbind(Alc1, Mood1)
# Alc2=rnorm(100,15,8)
# Mood2=(Alc2*-.4)+rnorm(100,3,4)
# B=cbind(Alc2,Mood2)
# Alc3=rnorm(100,20,8)
# Mood3=(Alc3*.8)+rnorm(100,3,4)
# C=cbind(Alc3,Mood3)
# data=data.frame(rbind(A,B,C))
# colnames(data) <- c("Alcohol","Mood")
# alldata=cbind(questionnaire,data)
# alldata=as.data.frame(alldata)
# 
# #Run Simpsons Paradox detection algorithm, clustering on the basis of the questionnaire
# example3=Simpsons(Alcohol,Mood,clustervars=c("q1","q2",'q3','q4',
# 'q5','q6','q7','q8','q9'),data=alldata)
# example3 
# ## End(Not run)