Vous êtes sur la page 1sur 24

4/26/2016

Factors Related to the


Poverty Rate in Ohio:
A Multilevel Analysis

Yiqing Yang
UNIVERSITY OF CINCINNATI

Yang 1

Contents
1.

Introduction ..................................................................................................................................... 2
a)

Background............................................................................................................................... 2

b)

Problem Statement................................................................................................................ 3

c)

Goal ................................................................................................................................................. 3

2.

Literature Review ........................................................................................................................... 3

3.

Study Area and Data ..................................................................................................................... 6

4.

Methodology .................................................................................................................................... 7

5.

Results and Analysis ..................................................................................................................... 8


Model 1: Null Model ........................................................................................................................... 8
Model 2: Add Census Tract Level Variables .............................................................................. 9
Model 3: Add County Level Variables to Predict Intercept ............................................... 12
Model 4: Add County Level Variables to Predict the Slope of Education ..................... 14
Model 5: Add County Level Variables to Predict the Slope of Unemployment .......... 16
Model 6: Add County Level Variables to Predict the Slope of Black .............................. 18
Best Model .......................................................................................................................................... 19

6.

Discussion and Conclusion ........................................................................................................ 21

7.

References ...................................................................................................................................... 23

Yang 2

1.

Introduction

a) Background
Poverty is an exhausting status and serious social problem that not only lead to a
low life quality, but also result in a high daily stress. Following the Office of
Management and Budget's (OMB) Statistical Policy Directive 14, the Census Bureau uses
a set of money income thresholds that vary by family size and composition to determine
who is in poverty. If a family's total income is less than the family's threshold, then that
family and every individual in it is considered in poverty. The official poverty thresholds
do not vary geographically, but they are updated for inflation using Consumer Price
Index (CPI-U). (U.S. Census Bureau).
Although the U.S. is a rich country with strong economic strength, and 34% of all
billionaires are Americans, according to the Organisation for Economic Co-operation and
Development (OECD), the U.S. poverty rate is the highest in the developed world. In
2014, the official poverty rate was 14.8 percent, and there were 46.7 million people in
poverty. Considering these statistical result, poverty is a serious social problem in the
U.S. and deserves attention from society and government. In order to have a better
understanding of poverty and find solutions to control the poverty rate, it is necessary
to analyze the factors which are related to the poverty rate. There are many scholars
use different kinds of models to analyze the poverty rate. This project will build
multilevel model to evaluate both the census tract level and county level factors.

Yang 3

b) Problem Statement
First, according to the UN Poverty Index, the U.S. has 20% of its population without
functional literacy skills. In general, a high educated person is more competitive when
finding a job and is more likely to have a high income. Second, there is no doubt that
economic recession will lead to bankruptcy, unemployment, and poverty. Third,
different racial status may result in different social status as well as economic status.
Forth, females who are divorced or widowed are in a disadvantage group, they struggle
to live without their husbands and might more likely to have a problem in money.
After considering these factors, this project will focus on the poverty rate in the
state of Ohio 2014, and build a multilevel model based on census tract level (level 1) as
well as county level (level 2) to investigate the relationship between poverty rate and
factors related to education, economy, race, and disadvantaged female.
c) Goal
Poverty can lead to health and mental problems. The objective of this project is to
find the influencing factors that are related to poverty rate, raise peoples awareness
and help government to find solutions to reduce poverty rate. In addition, this paper
can also shows an application of multilevel modeling, which is a useful method when
there is a significant variance between groups.
2.

Literature Review
The analysis of poverty have been subject to numerous studies through diverse

data and method.

Yang 4

Education is a key factor that relate to ones job and income. Bilenkisi, Gungor
and Tapsin (2013) investigated the relation between education and poverty through the
logistic regression method. In this study, they used data that were obtained from the
Income and Living Conditions Survey (SILC), divided the household heads education
levels into seven categories, and controlled the effects of other variables like age,
location, gender and so on. The result shows that the possibility of a household to be
poor will decrease if the education level of the household head increase. Furthermore,
Janjua and Kamal (2011) found out that secondary education emerged as the main
contributor to poverty alleviation while per capita income growth just played a
moderate role.
People tend to be poor if they are unemployed for a long time. Xue and Zhong
(2003) analyzed how unemployment relate to poverty rate in China, benefiting from the
1999 survey data of the Institute of Economics, Chinese Academy of Social Sciences
(CASS). One of the key point of this paper is that it used the population census data,
the survey data, and the registered unemployment data, to calculate an adjusted
unemployment, which will make the result more concrete. The result showed that the
unemployment rate was a major cause of urban poverty. In addition, Elizabeth
T.Powers (1995) also demonstrated that inflation and unemployment have strong
influence on poverty rate.
Anwar Ali Shah G.Syed et al. (2011) examined social and economic factors that
are responsible for the poverty in rural Sindh based on 2500 households data from four
districts. They evaluated the education, shelter, health, income, occupational status,

Yang 5

and government expenditures, and found out that government expenditures on road
and on education have significant impact on poverty rate. However, they just used
some graphs to analyze the variables and there is not a convincing model to
demonstrate the result.
Multilevel modeling is an advanced mothed. The goal of multilevel analysis is to
account for variance in a dependent variable that is measured at the lowest level of
analysis by considering information from all levels of analysis (Steenbergen and Jones
2002). Brady and Burroway (2012) did a multilevel analysis based on individual level
and country level data to exam the factors related to single mother poverty in
18affluent democracies because single mother is a disadvantage group and deserved to
be researched. The result showed that single mothers are more likely to be poor, and
well-educated, multiple earners, universal welfare will decrease the poverty. In order to
decrease risk of making a type I error, Kim, Yongwoo Lee, and Yu-jeong Lee (2010)
built a hierarchical linear models to analyze the factors related to poverty rate and
found out that the individual level character like younger age, being female, low levels
of education, labor market participation, and not being married will led to a high
poverty rate, and only the public social expenditures account for poverty rate as a state
level factor.

Yang 6

3.

Study Area and Data


The study area I chose for
this project is the state of Ohio. Ohio
is located in the Midwestern part of
the United State, with 88 counties
and 2952 census tracts in total. The
poverty rate in Ohio 2014 is 15.6%,
which ranked as 19th among the 51

Table. The Change of Poverty Rate (Source: U.S. Census Bureau)

states in the United State.

The change of Ohio poverty rate is similar to the change of U.S. poverty rate,
which is worth to be analyzed. In addition, although the poverty rate in Ohio is reducing
in recent years, the change is not significant enough, and a research on this field can
help people have a better understanding in the potential factors related to poverty rate.
The dependent variable is poverty rate in the state of Ohio 2014 by census tract.
The census tract level (level 1) independent variables and county level (level 2)
independent variables are shown as below.
Level 1 Variable
education
black
unemp
single_female
Level 2 Variable
house_c

Explanation
percentage of population 25 years and over who have
bachelors degree
percentage of black people
unemployment rate for population 16 years and over
percentage of female who are older than 15 and were
divorced or widowed
Explanation
Predict the intercept
median value of an owner-occupied home
Predict the slope of education

Yang 7

priv_c
uni_c

number of private schools


number of universities
percent of age 16 years or older who were employed in
science_c
the Professional, Scientific, and Technical Services
industry
percent of age 16 years or older who were employed in
agri_c
the Agriculture, Forestry, Fishing and Hunting industry
Predict the slope of black rate
fore_born_c
percent of population who are not U.S. citizens at birth
Predict the slope of unemployment rate
bankurptcy_c
number of business bankruptcy filings
bigfirm_c
percent of jobs at firms with 500 or more employees
All the variables are statistical data in 2014. The census tract level variables are
collected from U.S. Census Bureau and the county level variables are gathered from
policy map, which is an online data and mapping tool. This two resources are both
authoritative and reliable.
4.

Methodology
A major assumption of single-level, ordinary least squares (OLS) models is that

the observations are independent from one another. However, the independent
assumption will be violated in many cases because of some cross-level effects.
Multilevel modeling is an efficient regression method that can do micro and macro
models analyzed simultaneously and partition residual variance into between-group
(level-2) and within-group (level-1) components to avoid ecological fallacy and atomistic
fallacy. For poverty rate, some county level variables like the number of schools will
affect the effect of some census tract level variables on poverty rate like education
background. So that in this project, a multilevel model will be conducted to explain the
poverty rate in Ohio based on both census tract level and county level variables.

Yang 8

First, the raw data are being processed, and the census tracts that contain some
missing data are deleted. After this step, the 2952 census tracts are reduced to 2939.
Second, a null model that include zero independent variable is generated in HLM as a
baseline model to see the total variance and deviance statistics. Third, in order to better
explain the variance and illustrates the relationship between poverty rate and
independent variables in a numeric way, level 1 and level 2 variables began to be added
into the multilevel model step by step, and a best model will be selected to predict the
poverty rate in Ohio based on the chosen independent variables. Forth, a map will be
drawn by ArcMap to display the real poverty rate, predicted poverty rate, and the
residuals to check the fitness of the model. At last, a conclusion can be drawn on the
factors that relate to poverty rate, and some suggestions to reduce the poverty rate will
be given.
5.

Results and Analysis

Model 1: Null Model


First, a null model is generated which only include the dependent variable
poverty rate and no independent variable is included. The results are shown as below:
Level-1 Model POVERTYij = 0j + rij
Level-2 Model 0j = 00 + u0j

Yang 9

1) G00, which is interpreted as the average poverty rate, is statistically different from
zero and is equal to 16.83
2) Census tract (Level-1) variance is 216.06 and county (level-2) variance is 17.43. Hence,
total variability is 233.49 =216.06(within group) + 17.43(between group)
3) Variation within counties is 0.92 = 216.06/233.49. Intraclass correlations coefficient
(ICC) is 0.08 = 17.43/233.49. These numbers means that 8% of the total variability
in poverty rate can be attributed to the counties, and 92% is within counties.
4) Considering there is a significant variation of the average poverty rate in different
counties with a p-value <0.001, we cannot ignore the difference between counties.
The model can be improved. We add level-1 and level-2 predictors to reduce R and
explain U0.
Model 2: Add Census Tract Level Variables
In general, black people, divorced single female, low educated people, and
unemployment people are more likely to be poor. So that these factors are used as
census tract level variable to predict poverty rate.

Yang 10

From the table above, we can notice that the p-value of percentage of female
who are older than 15 and were divorced or widowed (single_f) is 0.166, which means
that although these people seems to have to struggle for life without husband, it do not
has statistically significant relationship with poverty rate. Then, a model that delete this
insignificant variable is built and the result is shown as below.
Level-1 Model
POVERTYij = 0j + 1j*(EDUCATIOij) + 2j*(BLACKij) + 3j*(UNEMPij) + rij
Level-2 Model
0j = 00 + u0j

1j = 10 + u1j

2j = 20 + u2j

3j = 30 + u3

Yang 11

1) G00, which is interpreted as the average poverty rate, is statistically different from
zero and is equal to 9.9773.
2) G10, which represent the average extent or slope that education can influence the
poverty rate, equals -0.3554 with a p-value<0.001. This means that education is
negatively and significantly related to the poverty rate within counties. The poverty
rate will decrease 0.3554 for each one-unit increase in the percentage of population
25 years and over who have bachelors degree (education).
3) G20, which represent the average extent or slope that black percentage can influence
the poverty rate, equals 0.4792 with a p-value<0.001. This means that black is
positively and significantly related to the poverty rate within counties. The poverty
rate will increase 0.4792 for each one-unit increase in the percentage of black people
(black).
4) G30, which represent the average extent or slope that unemployment rate can
influence the poverty rate, equals 0.9455 with a p-value<0.001. This means that
unemployment is positively and significantly related to the poverty rate within counties.
The poverty rate will increase 0.9455 for each one-unit increase in the unemployment
rate for population 16 years and over (unemp)
5) Census

tract

(level-1)

variance

is

81.9807,

county

(level-2)

variance

is

28.2592=27.9536+0.0918+0.1091+0.1047, and total variance is 110.2399. After


including education, black rate, and unemployment rate as predictors of poverty rate,
within county variance is reduced by 86.92% = (216.06 28.2592) / 216.06.
6) Variation within counties is 0.74 = 81.9807/110.2399. Intraclass correlations

Yang 12

coefficient (ICC) is 0.2563 = 28.2592/110.2399. These numbers means that 26% of


the total variability in poverty rate can be attributed to the counties, and 74% is within
counties.
7) Considering the within county variance is reduced by 86.92%, which is a great change,
model 2 is better than model 1. However, variation in different counties are still
statistically significant, and it can be explained by incorporating level 2 variables into
the model.
Model 3: Add County Level Variables to Predict Intercept
In order to explain the variance in the intercept between counties, median value
of an owner-occupied home is used as a level-2 predictor because we can assume that
a region that has high house price is a relatively rich region and has lower poverty rate.
Level-1 Model
POVERTYij = 0j + 1j*(EDUCATIOij) + 2j*(BLACKij) + 3j*(UNEMPij) + rij
Level-2 Model
0j = 00 + 01*(HOUSE_Cj) + u0j
1j = 10 + u1j 2j = 20 + u2j 3j = 30 + u3j

Yang 13

1) The average poverty rate G00 is 12.9692 and it is statistically significant.


2) The effect of house price on poverty rate is negative and statistically significant (G01).
The poverty rate will decrease 0.000025 for each one-unit increase in median value
of an owner-occupied home (house_c).
3) The poverty rate will decrease 0.3398 for each one-unit increase in education. The
poverty rate will increase 0.4855 for each one-unit increase in black percentage. The
poverty rate will increase 0.3398 for each one-unit increase in unemployment rate.
4) Census

tract

(level-1)

variance

is

81.9578,

county

(level-2)

variance

is

24.6813=24.386+0.0914+0.1105+0.0934, and total variance is 106.6391. Intraclass


correlations coefficient (ICC) is 0.2331.
5) After including median house value as a level-2 predictor for intercept, 13% =
(27.9536-24.386)/27.9536 of the variance for mean poverty rate between counties is
explained.
6) A large percentage of the variance between groups can be explained by median house
value, and the total variability is decreased, so that median house value is a good
predictor between counties.

Yang 14

Model 4: Add County Level Variables to Predict the Slope of Education


In order to explain the variance in the slope of education between counties, four
county level predictors are added in to model to check whether they have significant
effect. This four variables are: number of private schools (priv_c), number of
universities (uni_c), percent of age 16 years or older who were employed in the
professional, scientific, and technical services industry (science_c), and who were
employed in the agriculture, forestry, fishing and hunting industry (agri_c).

From the table above, we can notice that only science_c has a p-value < 0.05
and is statistically significant. Then, a model that delete all the insignificant variables is
built and the result is shown as below.
Level-1 Model
POVERTYij = 0j + 1j*(EDUCATIOij) + 2j*(BLACKij) + 3j*(UNEMPij) + rij
Level-2 Model
0j = 00 + 01*(HOUSE_Cj) + u0j
2j = 20 + u2j

3j = 30 + u3j

1j = 10 + 11*(SCIENCEj) + u1

Yang 15

1) The percent of age 16 years or older who were employed in the professional, scientific,
and technical services industry (science_c) has a negative 0.0446 impact on the effect
of education. Every one unit increase of the science will result in 0.0446 decrease in
the effect of education on poverty rate.
2) Census

tract

(level-1)

variance

is

81.6827,

county

(level-2)

variance

is

24.8226=24.4854+0.0908+0.138+0.1084, and total variance is 106.5053. Intraclass


correlations coefficient (ICC) is 0.2314.
3) After including the percentage of people who have a technical or scientific work as a
level-2 predictor for the slope of education, 7% = (0.0914-0.0908)/27.9536 of the
variance for slope of education between counties is explained.
4) A large percentage of the variance for slope of education between groups can be
explained by percentage of people who have a technical or scientific work, and the
total variability is decreased, so that it is a good predictor between counties.

Yang 16

Model 5: Add County Level Variables to Predict the Slope of Unemployment


In order to explain the variance in the slope of unemployment rate between
counties, two county level predictors are added in to model to check whether they have
significant effect. This two variables are: number of business bankruptcy filings
(bankurptcy_c), and percent of jobs at firms with 500 or more employees (bigfirm_c).

From the table above, we can notice that only bankurpt has a p-value < 0.05
and is statistically significant. Then, a model that delete the insignificant variable is built
and the result is shown as below.
Level-1 Model
POVERTYij = 0j + 1j*(EDUCATIOij) + 2j*(BLACKij) + 3j*(UNEMPij) + rij
Level-2 Model
0j = 00 + 01*(HOUSE_Cj) + u0j

1j = 10 + 11*(SCIENCEj) + u1j

2j = 20 + u2j

3j = 30 + 31*(BANKURPTj) + u3j

Yang 17

1) The number of business bankruptcy filings (bankrupt) has a positive 0.0029 impact
on the effect of unemployment rate. Every one unit increase of the bankrupt will result
in 0.0029 increase in the effect of unemployment rate on poverty rate.
2) Census

tract

(level-1)

variance

is

81.6444,

22.0384=21.6833+0.0812+0.1896+0.0843,

and

county
total

(level-2)
variance

variance
is

is

103.6828.

Intraclass correlations coefficient (ICC) is 0.2126.


3) After including the number of business bankruptcy filings as a level-2 predictor for the
slope of unemployment rate, 22% = (0.1084-0.0843)/0.1084 of the variance for slope
of unemployment between counties is explained.
4) A large percentage of the variance for slope of unemployment between groups can
be explained by percentage of people who have a technical or scientific work, and the
total variability is decreased, so that it is a good predictor between counties.

Yang 18

Model 6: Add County Level Variables to Predict the Slope of Black


In order to explain the variance in the slope of black people rate between counties,
the percent of population who are not U.S. citizens at birth is added into the model as a
county level predictors.
Level-1 Model
POVERTYij = 0j + 1j*(EDUCATIOij) + 2j*(BLACKij) + 3j*(UNEMPij) + rij
Level-2 Model
0j = 00 + 01*(HOUSE_Cj) + u0j

1j = 10 + 11*(SCIENCEj) + u1j

2j = 20 + 21*(FORE_BORj) + u2j

3j = 30 + 31*(BANKURPTj) + u3j

1) The percent of population who are not U.S. citizens at birth (fore_born) has a negative
0.1026 impact on the effect of black rate. Every one unit increase of the fore_born
will result in 0.1026 decrease in the effect of black rate on poverty rate.
2) Census

tract

(level-1)

variance

is

81.6015,

county

(level-2)

variance

is

Yang 19

25.01=24.6908+0.0878+0.1533+0.0781, and total variance is 106.605. Intraclass


correlations coefficient (ICC) is 0.2345.
3) After including the percent of population who are not U.S. citizens at birth as a level2 predictor for the slope of black rate, 19% = (0.1896-0.1533)/0.1896 of the variance
for slope of black rate between counties is explained.
4) A large percentage of the variance for slope of black between groups can be explained
by percentage of people who have a technical or scientific work, and the total
variability is decreased, so that it is a good predictor between counties.
Best Model
Total
variance
233.49
110.2399

ICC

Deviance

Improvement

Model 1 null model


0.08
24222.4
/
Model 2
0.2563 21570.83 explain 86.92% variance within
add level 1 variable
county
Model 3
106.6391
0.2331 21488.34 explain 13% variance for
level 2 to explain intercept
intercept between county
Model 4
106.5053
0.2314 21484.22 explain 7% variance for
level 2 to explain education
education slop between county
Model 5
103.6828
0.2126 21493.94 explain 22% variance for
level 2 to explain unemploy
unemploy slop between county
Model 6
106.605
0.2345 21486.06 explain 19% variance for black
level 2 to explain black
slop between county
The table above shows the improvement through the 6 models. Model 4 has the
lowest deviance, model 5 has the lowest ICC, and model 6 has most significant
variables. In order to compare this 3 model, mean square error (MSE) is calculated.
Model 4
Model 5
Model 6
MSE
213.3712
268.0742
129.9742
From the table above, model 6 is the best one, and will be used to predict the
poverty rate in Ohio. The model is POVERTYij = 15.3992 0.000044*HOUSE_Cj
0.5254*EDUCATIOij + 0.0384*SCIENCEj*EDUCATIOij + 0.8066*BLACKij

Yang 20

0.1026*FORE_BORj*BLACKij + 0.8866*UNEMPij +0.0045*BANKURPTj*UNEMPij + u0j


+ u1j*EDUCATIOij + u2j*BLACKij + u3j*UNEMPij + rij

The overall distribution of real poverty


and predicted poverty is similar with several
differences.
The residual map show that the model
suit the data well in most regions, but for
some tracts especially those have high
poverty rate, the residual is a little high,
which means there might be some other
factors that lead to this high poverty rate.

Yang 21

6.

Discussion and Conclusion

After adding the independent variables step by step, conclusions can be drawn as below.
For census tract level predictor
1) Education has a negative impact on poverty rate, which means the census tracks that
have a high percentage of population 25 years and over who have bachelors degree
are more likely to have a low poverty rate. This is understandable because a high
educated person is more likely to find a good job and get a content salary.
2) The percentage of black people and unemployment rate for population 16 years and
over have positive impact on poverty rate, which means the census tracks that have
a high proportion of black people and unemployed people also are more likely to have
a high proportion of poor people.
3) The percentage of female who are older than 15 and were divorced or widowed do
not have a significant relationship with poverty rate although this groups of people
seems to live tougher.
For county level predictor used to adjust the intercept
1) Median value of an owner-occupied home has a negative impact on poverty rate,
which means that counties that have low value of house are more likely to have a
high poverty rate. This can be explained because a poor region will have a low house
value to make sure people live there have ability to afford the house and people who
live in rich region are more willing to buy more house which will raise the house price.
For county level predictor used to adjust the slope of education
1) Counties that have a higher percentage of people who have a technical or scientific

Yang 22

work will have weaken the effect of education background on poverty rate. This result
is beyond expectation because first it is assumed that technical jobs have a high
require of education so that education might count more. However, this result can be
interpreted as if a county has a high proportion of technical jobs which might have
higher payment, the poverty rate will be lower regardless of education background.
2) The percentage of people who work in agriculture, forestry fishing and hunting
industry, the number of private schools and universities do not have significant impact
on the slope of education.
For county level predictor used to adjust the slope of unemployment rate
1) The effect of unemployment rate will be stronger in counties that have a high number
of business bankruptcy filings, which demonstrates the
2) The percent of jobs at firms with 500 or more employees do not has significant impact
on the slope of unemployment rate.
For county level predictor used to adjust the slope of black
1) The effect of percentage of black people on poverty rate will be stronger in counties
that have a high percent of population who are not U.S. citizens at birth. This result
shows the racial discrimination to some extent.
Overall, in this case, education, economic, and race factors have significant
relationship with poverty rate in Ohio 2014, while the disadvantage female do not has
significant contribution. For the recommendations to reduce the poverty rate,
government should appeal people to enhance their education background because
knowledge is strength. In addition, considering that economic recession will led to

Yang 23

bankruptcy and unemployment, economic should be controlled. Last, all people should
be given equal opportunity to study as well as work no matter what their race are.
7.

References

Bilenkisi, Fikret, Mahmut Sami Gungor, and Gulcin Tapsin. "The Impact of Household
Heads' Education Levels on the Poverty Risk: The Evidence From
Turkey." Educational Sciences: Theory & Practice 15.2 (2015): 337-348.
Brady, David, and Rebekah Burroway. "Targeting, Universalism, and SingleMother Poverty: A Multilevel Analysis across 18 Affluent
Democracies." Demography49.2 (2012): 719-46.
Janjua, P. Z., & Kamal, U. A. (2011). The role of education and income in poverty
alleviation: A cross-country analysis. The Lahore Journal of Economics, 16(1),
143-172.
Kim, Kyo-seong, Yongwoo Lee, and Yu-jeong Lee. A Multilevel Analysis of Factors
Related to Poverty in Welfare States. Social Indicators Research 99.3 (2010):
391404.
Steenbergen, Marco R., and Bradford S. Jones. Modeling Multilevel Data
Structures. American Journal of Political Science 46.1 (2002): 218237.
Powers, Elizabeth T. "Inflation, Unemployment, and Poverty Revisited." Economic

Review - Federal Reserve Bank of Cleveland 31.3 (1995): 2.


Xue, Jinjun, and Wei Zhong. "Unemployment, Poverty and Income Disparity in Urban
China." Asian Economic Journal (2003): 383-405.

Vous aimerez peut-être aussi