![Page 1: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/1.jpg)
Using Principal Components Analysis to construct a wealth
index
Laura HoweJames Hargreaves, Bianca De Stavola,
Sharon Huttly
![Page 2: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/2.jpg)
Wealth Index
![Page 3: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/3.jpg)
Principal Components Analysis
� Data reduction technique
� From set of correlated variables, PCA extracts a set of uncorrelated ‘principal components’
� Each principal component is a weighted linear combination of the original variables
![Page 4: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/4.jpg)
Principal Components Analysis
i.e. if we have n correlated variables X1 - Xn
each principal component is the sum of each variable multiplied by its weight (the weight for each variable is different in each principal component)
PCi=a1X1 + a2X2 + … + anXn
![Page 5: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/5.jpg)
Principal Components Analysis
� If we use 10 variables in PCA, we get 10 ‘principal components’
� The components are ordered so that the first principal component (PC1)explains the largest amount of variation in the data
� We assume that this first principal component represents wealth/SEP
![Page 6: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/6.jpg)
Principal Components Analysis
� Assumption:
The most important reason households have different values of the indicators we have put in the PCA is their wealth/SEP
![Page 7: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/7.jpg)
Issues in using PCA
1. Obscure
2. 1st principal component often explains a low proportion of the total variance
3. Designed for continuous data
![Page 8: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/8.jpg)
PCA with discrete data
� Dummy variables most common approach
� ‘Confuses’ PCA:
Variation arises both from the underlying concept of wealth and from the dependence between dummies for categorical variables
![Page 9: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/9.jpg)
Using dummy variables
� General practice involves omission of the dummy for the baseline category
� In PCA, a dummy is often used for ALL categories(Filmer & Pritchett, Vyas & Kumaranayake, DHS report)
� Linear dependence when all dummies are included; affects weights
![Page 10: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/10.jpg)
Alternatives to dummy variables
1. Treat categorical variables as continuous REQUIRES ORDINAL VARIABLES
2. Multiple Correspondence Analysis (MCA)
![Page 11: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/11.jpg)
Kolenikov & Angeles, 2004:
� Large simulation study
� Treating ordinal variables as continuous is ‘better’ than using dummy variables
![Page 12: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/12.jpg)
Analysis
Aim:
� Compare the agreement of indices constructed using different methods for dealing with categorical variables
![Page 13: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/13.jpg)
Methods
� Dataset: Malawi IHS2 2004/5, N=11,280
� Wealth index as in DHS of same year
� Agreement measured by classification into quintiles
![Page 14: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/14.jpg)
Methods
� 5 wealth indices:
1. PCA: Include dummies for all categories
2. PCA: Exclude lowest SEP category dummy
3. PCA: Exclude lowest frequency dummy
4. PCA: Treat ordinal variables as continuous
5. MCA
![Page 15: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/15.jpg)
-87 %75 %44 %76 %MCA
-67 %42 %67 %PCA: treat categorical variables as continuous
-60 %97 %PCA: exclude lowest frequency
dummy
-61 %PCA: exclude lowest SEP
dummy
-PCA: dummy for each category
MCAPCA: treat categorical variables as continuous
PCA: exclude lowest
frequency dummy
PCA: exclude
lowest SEP dummy
PCA: dummy for
each category
![Page 16: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/16.jpg)
-87 %75 %44 %76 %MCA
-67 %42 %67 %PCA: treat categorical variables as continuous
-60 %97 %PCA: exclude lowest frequency
dummy
-61 %PCA: exclude lowest SEP
dummy
-PCA: dummy for each category
MCAPCA: treat categorical variables as continuous
PCA: exclude lowest
frequency dummy
PCA: exclude
lowest SEP dummy
PCA: dummy for
each category
![Page 17: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/17.jpg)
-87 %75 %44 %76 %MCA
-67 %42 %67 %PCA: treat categorical variables as continuous
-60 %97 %PCA: exclude lowest frequency
dummy
-61 %PCA: exclude lowest SEP
dummy
-PCA: dummy for each category
MCAPCA: treat categorical variables as continuous
PCA: exclude lowest
frequency dummy
PCA: exclude
lowest SEP dummy
PCA: dummy for
each category
![Page 18: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/18.jpg)
Weights from PCA
� Example 1: floor material
![Page 19: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/19.jpg)
-0.7160Mud floor
0.0209Other floor
0.0478Tile floor
0.1823Sand floor
0.6719Cement floor
Index
including all
dummies
Including all dummies can result in counter-intuitive weights
![Page 20: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/20.jpg)
Weights from PCA
� Example 2: toilet facility
![Page 21: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/21.jpg)
-0.10410-0.0923None/other facility
(lowest SEP)
-0.0659-0.0631-0.0611Latrine no roof
Index
excluding
lowest
frequency
dummy
Index
excluding
lowest SEP
category
Index
including all
dummies
Excluding lowest SEP dummy can result in counter-intuitive weights
![Page 22: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/22.jpg)
� Example 3: drinking water source
![Page 23: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/23.jpg)
0.18730.16630.1630Piped outside of dwelling
00.31320.2762Piped into dwelling
(lowest frequency)
Index
excluding
lowest
frequency
dummy
Index
excluding
lowest SEP
category
Index
including all
dummies
Excluding lowest frequency dummy can result in counter-intuitive weights
![Page 24: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/24.jpg)
� So different methods can alter the order of categories
� Weights from PCA should ‘make sense’
![Page 25: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/25.jpg)
What to do?
� All methods of using dummy variables can be flawed…
� Assign an order to the categories if possible, and treat as continuous
� If not, MCA
![Page 26: Using Principal Components Analysis to construct a wealth ...epi2008.com.br/apresentacoes/CC_24_09_Tarde_PDF/Laura Howe.pdfUsing Principal Components Analysis to construct a wealth](https://reader033.vdocuments.mx/reader033/viewer/2022041811/5e58168ab37b6a57f83ef616/html5/thumbnails/26.jpg)
Acknowledgements
� Sharon Huttly
� James Hargreaves
� Bianca De Stavola
� Paul Clarke