Factors Related To The Unemployment Rate: A Statistical Analysis

University of Northern Iowa
UNI ScholarWorks
Honors Program Theses University Honors Program
2009
Factors related to the unemployment rate: A

statistical analysis
Reanna Collins
University of Northern Iowa
Copyright © 2009 Reanna Collins

Follow this and additional works at: http://scholarworks.uni.edu/hpt
Part of the Labor Economics Commons, and the Labor Relations Commons
Let us know how access to this document benefits you
Recommended Citation
Collins, Reanna, "Factors related to the unemployment rate: A statistical analysis" (2009). Honors Program Theses. 9.
http://scholarworks.uni.edu/hpt/9
This Open Access Honors Program Thesis is brought to you for free and open access by the University Honors Program at UNI ScholarWorks. It has
been accepted for inclusion in Honors Program Theses by an authorized administrator of UNI ScholarWorks. For more information, please contact
scholarworks@uni.edu.
FACTORS RELATED TO THE UNEMPLOYMENT RATE 1
FACTORS RELATED TO THE UNITED STATES’ UNEMPLOYMENT RATE:

A STATISTICAL ANALYSIS
Introduction
Current United States economic conditions have everyone searching for ways to solve the
downward spiral that has instigated a crisis in the minds of many people. Rapidly increasing
unemployment rates are key reasons leading to this feeling of emergency. A better
understanding of unemployment trends over the past few decades and factors correlated to the
rate could serve as future indicators of fluctuations in the unemployment rate. A statistical
analysis was conducted to determine if any relationship between the unemployment rate and nine
selected variables existed in order to identify trends in the unemployment rate and to produce a
model that could be used to predict the unemployment rate for a given year. This analysis was
done to examine previously suggested variables associated with the unemployment rate and also
to find if any new variables might be associated with the rate.
Background
Unemployment Rate Details. How the unemployment rate is calculated is essential in
first being able to understand the data that will be presented. The unemployment rate is defined
as the proportion of the nation’s non-institutionalized population sixteen years and older that is
out of work, actively looking for a job, and available for work (Frumkin, 1998). Institutionalized
people confined to facilities such as nursing homes or jails are not included in the rate. U.S.
armed forces members are also excluded so that the measure represents only the civilian labor
force (Frumkin, 1998).
The unemployment rate is computed mathematically as the number of unemployed persons
divided by the civilian labor force and multiplied by 100%. A person is considered employed for
a given month if the person did at least one hour of work during the reference week, the seven
day week that includes the twelfth of that month; this could be as a paid employee or self-
employed. An alternative to the one hour is if the person worked at least fifteen hours at non-
paid jobs in a family business. It also includes people temporarily absent due to vacations,
illness, or other reasons. People with more than one job are only counted once. Citizens of
foreign countries who are temporarily living and working in the United States are also counted as
long as they are not working or living on the premises of an embassy. Unemployed people are
those who were available to work and actively looked for work during the four-week period prior
to the reference week. Merely looking at job advertisements or attending job-training programs
or courses do not count as actively searching for work, but is deemed passively searching.
Actively searching for work includes activities such as contacting an employer for or
participating in an interview, contacting an employment agency, sending out resumes, answering
or placing a job advertisement, or checking a union or professional register. People not pursuing
a job at all, such as students or stay at home mothers, or are passively looking for jobs are not
considered part of the work force and are not included in the unemployment figure at all
(Frumkin, 1998).
The U.S. Bureau of Labor Statistics figures the unemployment rate monthly based on a
survey of a sample of approximately 50,000 households called the Current Population Study.
According to Frumkin (1998), about 47,000 households respond each month. The households
are utilized in a rotation. One-fourth of the households chosen one time are not chosen the next
time, and each household is interviewed for four months and then dropped from the rotation
during the next eight months before being re-included one last time in the survey for another four
month period. Frumkin (1998, p. 209) also states that due to sampling error, a plus or minus of
0.1 percentage point in monthly movement of the unemployment rate is not statistically
significant 66% of the time. A change of 0.2 is considered significant, as well as a consistent
change of 0.1 upward or downward in consecutive months. If raised to 95% of the time, an
unemployment rate would have a range of plus or minus 0.2. This means that monthly
movement would have to change by 0.3 in either direction to be statistically significant. This
helps when trying to analyze whether or not to consider a monthly change in the unemployment
rate measured in this manner as being noteworthy.
The survey method used by the Bureau of Labor Statistics is also said to produce an
upward bias in unemployment statistics (Hoel, Clarkson, & Miller, 1983). Other countries
require people to physically go to an agency to be recorded as unemployed, making their rate
much lower. For example, a 2.5% unemployment rate in the United Kingdom is approximately
equivalent to a 5% rate in the United States. Also, amendments to the food stamp and Aid to
Families with Dependent Children programs in the 1970s facilitated a larger number of people
being considered unemployed. These amendments required recipients of these welfare programs
to register for work. If the number who were forced to register for work to receive benefits but
cannot or do not wish to work is large, then the unemployment rate will increase even though the
number of persons actually seeking work may remain relatively constant. It is hard to determine
the extent to which these registration requirements constitute an over-counting in unemployment
due to those who have no intention to work or are unemployable (Hoel, Clarkson, & Miller,
1983).
The duration of unemployment is another aspect to consider when investigating the
unemployment rate. Duration of unemployment is described as the length of time workers are
unemployed. When the average duration of unemployment rises due to workers remaining
unemployed longer, the unemployment rate can rise even when the number of workers becoming
unemployed per week remains constant. If the average duration increases by a week or two, the
unemployment rate may reflect this by a 0.5% increase (Hoel et al., 1983, p. 193). Some policy
advisors believe an unemployment rate including only those who have been unemployed for a
longer duration, around six months, is more beneficial when considering courses of action in
policies. This idea is due to the fact that most unemployed workers find work within fourteen
weeks during times of recession (Hoel et al., 1983, p. 194). This shows that short term spikes in
unemployment might not be as important, at least in terms of policy making, as focusing on if a
factor like overall duration of unemployment is causing the increased rate.
Factors previously found to be related to the Unemployment Rate. A few factors
considered to affect the unemployment rate are discussed by Frumkin (1998, p. 211-212). He
lists long term birthrate cycles as having an important impact on the labor force, but they lag in
the effect due to the time that must pass for the children to reach working age. Immigration and
death rates are less important according to him. Frumkin (1998) also states that the
unemployment rates for men were 0.5 to 1.5 percentage points lower than those for women from
1950 to the 1980s, when that differential disappeared and later reversed itself in the 1990s.
Teenagers’ unemployment rates were two to three times larger than adults during the entire
period from 1950-1996. Paul Flaim (as cited in Frumkin, 1998, p. 214) conducted a study and
found that the unemployment rate in 1979 would have been 4.4% instead of 5.8%, a difference
of 1.4%, if the labor force composition would have remained the same between 1959 and 1979.
Flaim (Frumkin, 1998) found that most of the rise in unemployment due to demographic factors
over the period resulted from the increased number of teenagers in the population. He also found
that out of the 0.6 percentage point reduction in the unemployment rate in the 1980s, 0.5
percentage point, or 80% of the reduction, was due to the decrease in the proportion of teenagers
in the labor force. This displays a relationship between demographic factors and the
unemployment rate.
The unemployment rate is also assumed to have a connection with wage growth. More
specifically, unemployment has been found to have a negative correlation with wage growth.
The economic expansion ending in the early 2000s was one of the longest and delivered the
lowest unemployment rates in previous decades, yet nominal wage growth remained relatively
contained. Some found the failed acceleration of wages to suggest a breakdown in the historical
relationship between unemployment and wage growth. According to Aaronson and Sullivan
(2000), the relationship has always been a loose one and if the unemployment rates are broken
down amongst states, then the relationship still holds. However, Aaronson and Sullivan (2000)
do recognize that a shift in the relationship between unemployment and wage growth would not
be terribly surprising. This response is due to the many changes in the labor market in recent
years. The general drop in the level of job security, the aging of the work force, higher levels of
education, the growth of temporary services employment, the use of fax machines and the
internet in job search, and even the increase in the prison population could each be changing the
relationship between unemployment and wage growth (Aaronson & Sullivan, 2000). Another
thing that could be affecting the wage growth correlation is the increase in other compensatory
packages other than wages paid to employees. Aaronson and Sullivan (2000) assert that the
fraction of employee compensation paid in the form of wage and salary accruals fell from 92.4%
in 1959 to 83.4% in 1980, then to a minimum of 81.0 % in 1994. Since 1994, the fraction of
compensation paid in the form of wages and salaries has increased again to 83.9% in 1999. All
of these changes in wages and employment could be changing the current connection between
wage growth and unemployment.

Gross domestic product (GDP) has also been found to share a relationship with the
unemployment rate. Zuckerman (2004) claims that unemployment and gross domestic product
are linked, but the connection has been evolving. He reports that usually a 1% decrease in the
GDP results in a reduction of 1.3% to 1.5% in private employment. This trend was broken
during the recession in the early 2000s when a 0.5% decline in GDP was followed by an almost
3% shrinkage of jobs. Zuckerman (2004) also stated that most of the jobs lost in the factory
industry are gone for good, and jobs in the 2003 recovery also shifted from higher to lower
paying positions. The results from this recession/recovery are possibly impacting the current
economic situation and unemployment spike.
Escalated unemployment rates and criminal activity could share a relationship as well.
Moyer (2009) alleges there is a loose relation between unemployment and crime, if there is one
at all. During the Great Depression, the crime rate plummeted, which contradicts the initial
notion that crime spikes during economic hardships. Bruce Weinberg (as cited in Moyer, 2009,
para. 3) states, “People sitting in their houses don't make great targets for crime.” While during
the “Roaring Twenties,” when the economy was booming and people were spending lots of
money, the crime rate was high. Two other spikes in the crime rate later in the century were
during the early 1970s when a boom was occurring and during the early 1990s when there was a
recession taking place. The link tying these two swells in the crime rate has less to do with the
economy, but more to do with the escalation of drug use during these periods, heroin in the
1970s and crack cocaine in the 1990s. However, Weinberg and some colleagues conducted an
analysis on young males with no more than a high school education, because they are the
demographic group that commits most crimes. This examination exposed a direct connection
between unemployment rates and the incidence of property crimes, which include felonies such
as burglary, auto theft and robbery. They also discovered that hard times lead to more domestic
abuse.
On the other hand, murder, a horrendous criminal activity, has never been well linked to
unemployment rates or other standard economic indicators. Rosenfeld and Fornango (as cited in
Moyer, 2009) say this is because of a human emotion factor. They reveal a connection between
murder rates and the Consumer Sentiment Index, a survey of how people view their current
financial situation and how hopeful they are about the future; the lower the index score, the
higher the crime rate. A recent study has also shown that economic stimulus can act as a salve,
because in the 1930s communities that spent more on public works programs had lower crime
rates than the other communities. This helps back the recent governmental economic stimulus
package, which included several public works programs, especially since most Americans’ hopes
were diminishing by the end of the Bush administration.
Another study involving criminal rates and economic downturns was presented by Karen
Heimer and Janet Lauritsen (as cited in Galluzzo, 2009). Their analysis concentrated on
victimization of Latinos and African Americans versus non Latinos and African Americans
during and after periods of economic decline. An association relating economic downturns and
victimization rates among minorities over the period of 1973-2005 was established. This
coincides with Weinberg’s research, where breaking down the society into different
demographics can allow there to be trends where the crime rate rises with downturns in the
economy.
Immigration is another demographically linked variable that has been found to have some
association to the unemployment rate. Employing United States census data covering 1940 to
1980, a study was done to examine the impact of immigration on unemployment and earnings
among racial minorities in the United States (Kposowa, 1995). This study found that increases
in immigration in some periods of U.S. history had significant negative effects on employment
levels among racial minorities in the United States. Specifically, results of the regression
analysis showed that in 1970, a standard deviation change in immigration increased
unemployment among minorities by nearly 14%, while in 1980 unemployment increased by
nearly 10% given a standard deviation increase in immigration. These results, contrary to
Frumkin’s (1998), justify some relationship with the unemployment rate and immigration.
Henry Blodget (2008) discussed the comparison of today’s recession with the Great
Depression in a recent article. He noticed a problem with the argument that today’s problems are
not the same as during the Great Depression, because the unemployment rate spiked to around
25%, while currently the rate is only around 10.2%. Part of the big difference is due to a change
in the method used to calculate the unemployment rate, and if the rate was computed the same
way now as it was during the Great Depression, then the unemployment rate today would be
much higher. In 1933, the definition of the labor force included individuals in the military,
prisoners, and also individuals as young as fourteen (Lieberman, 2008). Even more significant
was the difference in the description of the unemployed, which at the time was not limited to
those actively seeking employment. Many reported that the changes in society also brought
changes to the ruler with which unemployment is measured. More women are encouraged to
enter the workforce now compared to during the 1930s, when several who were in the workforce
were not even counted. All of these disparities make it complicated to compare the two time
periods’ unemployment rates alone without taking into consideration other factors. The
inconsistency in the type of data available between then and now also causes a dilemma when
trying to include data from that time period into a current analysis.
Hypotheses
Before any formal statistical analysis was performed, the reviewed literature and previous
knowledge suggested that each of the chosen factors may correlate with the unemployment rate
in the following manner:
Independent Variable Relationship Type

CPI Inverse
Immigrants Admitted Direct or None
Burglary Direct
Minimum Wage Adjusted for Inflation Inverse
Public Debt Direct
NASDAQ Low Inverse
Trade Balance Direct
Production Workers’ Average Hourly Salary Inverse
Average Gas Price N/A
Earnings
The CPI may be inversely related to the unemployment rate because inflation and the
unemployment rate are usually inversely related in the short run, and inflation influences the CPI
directly. Frumkin (1998) found immigration to be an unimportant factor in the unemployment
rate, while a study done by Kposowa (1995) suggests there is a direct relationship between
immigration and unemployment among minorities. Based on these two findings, there may be
no relationship overall or a possible direct relationship between immigration and the
unemployment rate. A loose direct relationship may be found between the unemployment rate
and the amount of burglaries in a year, because burglary is a main criminal act associated with
unemployment due to the hard times encountered by the unemployed. However, there was
contradictory evidence found by Moyer (2009). The minimum wage and hourly wage of
production workers may have an inverse relationship with the unemployment rate because when
unemployment is low, the economy is generally doing well, which usually leads to higher wages.
Although, Aaronson and Sullivan (2000) said the relationship is a loose one based on the
changes in compensation packages. Public debt might be directly correlated with the
unemployment rate because the unemployment rate is usually a good economic condition
indicator, and if the economy is suffering, then the government might increase the amount it is
borrowing. Also, if the economy is doing poorly, unemployment tends to rise. The NASDAQ
low for the year could be found inversely related with the unemployment rate, because one
would expect the stock market to be down when unemployment is high because the economy is
weak. The United States’ trade balance might be directly related to the unemployment rate due
to a trade surplus being positive (higher) and a trade deficit being negative (lower). A larger
trade deficit is probably associated with higher unemployment because in that case, the United
States is importing more than it is producing and exporting. The average gas price of all gas
types could have either relationship with unemployment if it has any relationship at all. In order
to see whether or not these hypotheses hold any validity, an analysis was performed involving
the unemployment rate and these variables.
Methodology
Only data from 1975-2007 was included in this analysis because of the previously stated
reason of data inconsistencies over long time spans. Nine covariates were chosen to determine if
any have a statistically significant relationship with the unemployment rate. These covariates
include the United States consumer price index (CPI), the total immigrants admitted to the
United States, the amount of burglaries, the minimum wage adjusted for inflation, the public debt,
the NASDAQ low, the United States trade balance, the hourly earnings of production workers,
and the average price of all gas types. Summary statistics of all the variables were first
computed. Next, a linear regression model was found in Statistical Analysis System (SAS) using
multiple model building techniques. The SAS code used for this analysis is listed in Appendix A.
Model assumptions were then verified in order to make the model valid. Furthermore, the data
were searched for damaging points. After the linear regression aspect was complete, more in-
depth correlation checks were performed in order to inspect for problems in the model due to
time correlations; the data were observed each year and consecutive years might affect each
others’ unemployment rates. The results were then summarized and conclusions were produced.
Analysis
Summary Statistics. All of the raw data for the dependent and independent variables for
this analysis are listed in Appendix B. After acquiring the data, summary statistics were
computed, such as means, medians, standard deviations, sample sizes, minimum values, and
maximum values. The summary statistics are provided in the table below:
U.S. Fed. Hourly Avg.

U.S. Total Min. Wage Earnings Price of
U.S. Consumer Immigrants Adjusted U.S. Trade of All Gas
Unemployment Price Admitted for inflation Public Debt NASDAQ Balance Production Types
Rate Index to the U.S. Burglary in 2008 (in Billions) Low (in Millions) Workers (in Cents)
Min 4.00 53.80 385378 2050992 5.53 533.20 60.70 -753283.00 4.73 56.70
Max 9.70 207.30 1826595 3795200 8.80 9007.70 2340.68 12404.00 17.43 274.60
Mean 6.23 132.15 800259 2824952 6.85 3854.44 789.79 -194546.70 10.72 129.04
Median 6.00 136.20 703542 2984434 6.56 3665.30 376.87 -104065.00 10.52 120.50
St. Dev 1.4393 44.9653 326234 535299 0.9412 2601.73 741.86 226361.39 3.6050 50.8084
Sample 33 33 33 33 33 33 33 33 33 33
Size
The dependent variable, the unemployment rate, varied from 4.00%, which occurred in 2000, to
9.70%, which occurred in 1982, and had a mean of 6.23%; this means that the average
unemployment rate between 1975 and 2007 was 6.23% and had a range of 5.70%.
Linear Regression. To assess the trend in the unemployment rate and determine if there
are any factors related to the trend, a linear regression analysis including model building
techniques and assumption verifications were conducted. Linear regression models the
relationship between two or more variables by fitting the line (plane) that best explains the
dependent variable from the explanatory variables. Linear regression does this by finding the
line (plane) that minimizes the sum of the squares of the vertical distances between the actual
points and the fitted line (plane). The procedures used to decide which set of covariates makes
up the best fitting linear regression model include forward selection, backward elimination,
stepwise selection, r-squared values or plot, MSE plot, and Mallow’s C(p) criterion.
The forward selection procedure began by including the most statistically significant
variable. Statistical significance is determined by the p-value, which is the probability of seeing
a result as rare as the one observed in a collection of random data in which the variable had no
effect. The smaller the p-value is, the more significant the result. During the forward selection
process, the variable with the next smallest p-value was added to the model and the regression
was performed again. This iteration continued until none of the variables’ p-values fell beneath a
predetermined cutoff value. A reasonable cutoff value of 0.15 was used in this regression,
meaning that there is only a 15% chance of observing a result at least as rare. The first covariate
inserted into the model was the NASDAQ low with a p-value of less than 0.0001. In the second
step of the forward selection process, the average price of gas was added due to its p-value of
0.0810. Next, the total amount of immigrants admitted to the United States was included in the
model with a p-value of 0.0538. No other variables met the 0.15 significance level for entry into
the model. The resulting model from the forward selection procedure includes the NASDAQ
low, the average price of all gas types, and the total amount of immigrants admitted into the
United States.
Opposite of the forward selection process, the backward elimination procedure was
conducted next. It started with a model including every covariate chosen for evaluation and
eliminated the covariate with the highest p-value in each step until no variables were left with p-
values greater than the cutoff value. Thus, only statistically significant variables were left in the
regression model. The same 0.15 cutoff value was used for this process as was used in the
forward selection. The United States Federal minimum wage adjusted for inflation in the year
2008 was the first variable removed from the model because of its p-value equal to 0.7190. Next,
the hourly earnings of production workers variable was removed from the model with a p-value
of 0.7340. A new regression was conducted each time a variable was discarded, similar to when
a variable was added in the forward selection process. This recalculation is the reason why the
p-value of the second variable removed is higher than the p-value of the first variable removed;
once the first variable was removed, the p-value of the second increased in the new regression
model. Subsequently, public debt in millions was the next variable discarded with a p-value
equal to 0.7388. The burglary count was the next variable eliminated from the model due to its
p-value of 0.5742. Step five in the backward elimination procedure removed the consumer price
index variable with a p-value of 0.4929. Finally, the last variable disposed of was the United
States trade balance with its p-value equal to 0.4370, after which no variables possessed a p-
value greater than 0.15. Identical results to the forward selection process were found by the
backward elimination procedure; the model containing the NASDAQ low, the average price of
all gas types, and the total amount of immigrants admitted into the United States comprised the
end result.
Next, the stepwise selection method was performed. This technique is a variation on
forward selection. This method begins the same as the forward selection process, by starting
with the one variable model with the lowest p-value. The difference is that at each stage in the
process, after a new variable is added, a test is made to check if some previously added variable
can be deleted without significantly increasing the residual sum of squares (RSS), which is the
sum of squares of the residuals and measures the discrepancy between the data and an estimation
model. A cutoff value of 0.15 was used as a limit yet again. The best one variable model found
by the stepwise selection was the model containing the NASDAQ low due to its p-value of
<0.0001. Next, the possible two variable models were considered, and the model containing the
NASDAQ low and the average price of all gas types was determined to be the best. Three
variable models were evaluated after that, and the model with the NASDAQ low, the average
price of all gas types, and the number of immigrants admitted to the United States was deemed
the best. Finally, four variable models were constructed, but none of them met the criteria of
having all covariate p-values below 0.15, so the best three variable model is the model chosen by
the stepwise selection procedure. This is the same model selected by both the forward selection
and backward elimination methods.
Another technique, the r-squared method, is based on a value expressing the squared
correlation of the predicted value from the proposed model to the actual value. This value was
used to judge the goodness of fit of a model. The r-squared value, also referred to as the
coefficient of determination, is a measure of association ranging from 0 to 1; the closer the value
is to 1, the more correlated the predicted value is to the actual value, or the more relevant the
regression model is. R-squared is defined as the ratio of the sum of squares explained by a
𝑆𝑆𝐸
regression model and the total sum of squares around the mean; the equation is (1 − ). It is
𝑆𝑆𝑇
referred to as the proportion of variation explained by the model. This method involves
evaluating the r-squared value for every model and comparing them to find the model with the
highest r-squared value where if another variable was added to the model, then the r-squared
value does not increase significantly. This idea follows parsimony, the concept of using the
model with the smallest amount of variables that gives the most information, in other words, the
model with the highest r-squared value and least amount of variables. The model containing all
covariates will always have the highest r-squared value, but parsimony eliminates the covariates
which do not increase correlation significantly. In order to gauge the r-squared values more
easily, the r-squared values were separated into subsets based on the number of variables in the
model, and only the model with the highest r-squared value within each subset was assessed
further. A graph containing the best fitting model (model with the highest r-squared value) with
one variable, best model with two variables, all the way up to the best model with nine variables,
is displayed below.
R-Squared Plot
0.7
R-squared Value
0.65
0.6
0.55
0.5
1 2 3 4 5 6 7 8 9
Number of Covariates in Model
As shown in the graph, the model with three variables is the elbow in the graph, which means
that it has the highest r-squared value taking into account parsimony; from the best two variable
model to the best three variable model, the r-squared value increases significantly, but from the
best three variable model to the best four variable model, the r-squared value only increases
slightly. The best three variable model turns out to be the same model established by the
previous three methods. This model is the model with covariates including the NASDAQ low,
the number of immigrants admitted to the United States, and the average price of all gas types.
An adjusted r-squared technique using the same principles as the r-squared technique
described previously was also conducted. Adjusted r-squared is a modification of r-squared that
adjusts for the number of explanatory terms in a model. The equation for the adjusted r-squared
𝑆𝑆𝐸 𝑛 −1
is 1 − ( ∗ ). Unlike r-squared, the adjusted r-squared increases only if the new term
𝑆𝑆𝑇 𝑛−𝑝−1
improves the model more than would be expected by chance. The adjusted r-squared value can
be negative, and will always be less than or equal to the corresponding r-squared value. The
graph below displays the adjusted r-squared values for the same subsets as was used before in the
r-squared technique.
Adjusted R-Squared
Plot
0.625
Adjusted R-Squared
0.6
0.575
0.55
0.525
0.5
1 2 3 4 5 6 7 8 9
The model with three covariates is the best fitting model because it is the peak in the graph. This
three covariate model is the same model found by all the previous methods; the model containing
the NASDAQ low, the average price of all gas types, and the number of immigrants admitted to
the United States.
A mean squared error (MSE) plot was also used to decide upon a model. MSE equals
𝑆𝑆𝐸
, where SSE is the sums of squared error, n is the sample size, and p is the number of
𝑛 −𝑝
covariates plus 1, or the number of parameters in the model (covariate slopes plus the intercept).
This approach is opposite of the r-squared method because the smaller the MSE, the better the
model. A smaller MSE means there is less error. Sometimes the model with the absolute
smallest MSE is not the best model to use due to parsimony, so a technique similar to the
technique used in the r-squared method was performed but using the smallest values, not the
largest values. In practice, the square root MSE is usually used to keep the numbers more
manageable. A plot of the smallest square root MSE for each number of covariates is provided
below. Once again, the same model with three covariates is the elbow in the graph.
MSE Plot
Square Root MSE 1
0.98
0.96
0.94
0.92
0.9
1 2 3 4 5 6 7 8 9
Mallows C (p) criterion was another method used to judge the best fitting model. This
method defines a set of reasonable models based on the following criterion: the C (p) for a given
𝑆𝑆𝐸𝑝
model with p covariates is approximately equal to p for the model. C (p) equals
𝑀𝑆𝐸 𝑓𝑢𝑙𝑙
− (𝑛 − 2𝑝), where SSEp is the sums of squared error of the model with p parameters, MSEfull is
the mean squared error for the model with all possible covariates, n is the number of
observations, and p is the number of parameters. The Mallows C (p) for the full model always
𝑆𝑆𝐸 𝑓𝑢𝑙𝑙 𝑆𝑆𝐸 𝑓𝑢𝑙𝑙

equals p because the equation would be − (𝑛 − 2𝑝), and MSEfull = .
𝑀𝑆𝐸 𝑓𝑢𝑙𝑙 𝑛 −𝑝
Accordingly, the full model will always be included in the set defined by the criterion due to the
equation equaling p exactly. Consistent with these guidelines, the model chosen by all other
methods was also included in the set of reasonable models defined by the Mallows C (p)
criterion.
The same model was preferred by all model building techniques. This means that the
model favored by all techniques was decided upon to fit the given data the best. The model
containing the NASDAQ low, the number of immigrants admitted to the United States, and the
average price of all gas types as the independent variables, and the unemployment rate as the
dependent variable was chosen to represent the best relationship between the covariates and the
unemployment rate by this linear regression analysis.
In order to utilize the model found by this linear regression analysis, some model
assumptions had to be checked. These regression model assumptions include: X-values are
measured without error, each observation is equally reliable, residuals are normal with constant
variance, appropriateness of a linear model, and independence of the residuals. The first two
assumptions, X-values are measured without error and each observation is equally reliable,
cannot be tested statistically, but depend on the sample collection methods. The data came from
dependable sources, so the first two assumptions are assumed to hold true.
Normality of the residuals can be verified graphically and formally. Graphically,
normality of the residuals is demonstrated by a normal probability plot of the residuals. A
straight line in the graph suggests normality of the residuals. The plot of the residuals below
illustrates a nearly straight line, so the assumption is valid.
Normal Probability Plot

Legend: * = 1 observations + = 2 observations
2.25+ * ++
| *++++++
| **+*+++
| +**+*+
0.25+ ++*****
| *******
| ***+***
| *++*++
-1.75+ ++*+++
+----+----+----+----+----+----+----+----+----+----+
-2 -1 0 +1 +2
A formal correlation test involving the Re,q value can also be conducted to check for normality.
Re,q is the correlation between the residual and its normal score, the corresponding standard
normal quartile. If the Re,q is less than a given cutoff value from a table of critical values for the
coefficient of correlation, then reject the assumption of normal residuals. Re,q for the preferred
model equals 0.98526, which is greater than the cutoff value of 0.966 for α=0.01 and n=33, so
the assumption of normal residuals was once again confirmed acceptable.
Constant variance of the residuals was the next assumption to verify. This can be done
two ways, graphically by a plot of the predicted values against the residuals or formally by
Breusch-Pagan tests for each covariate. A predicted vs. residuals plot also shows whether a
linear model is appropriate or not. The plot is examined for patterns; a fan-shape indicates there
are non-equivalent variances, while non-linear patterns indicate problems with having a linear
model. The predicted vs. residuals plot below appears to be slightly fan-shaped, so the variances
of the residuals may not be constant. However, there does not appear to be any non-linear
patterns in the graph, so a linear model is not found inappropriate. The plot is displayed below:
Plot of Predicted vs. Residuals

2.5 ˆ
‚
‚
‚ A
2.0 ˆ A Legend:
‚
‚ A = 1 observations
‚ B = 2 observations
1.5 ˆ
‚ A
‚
‚ A
1.0 ˆ A
‚
R ‚
e ‚ A A A
s 0.5 ˆ A A
i ‚
d ‚ A
u ‚ A A
a 0.0 ˆ A A A
l ‚ A A
‚ A A
‚ A A
-0.5 ˆ B A
‚ A A
‚ A A
‚ A
-1.0 ˆ
‚ A
‚
‚ A
-1.5 ˆ
‚ A
Šˆƒƒƒƒƒƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒƒƒƒƒƒˆƒƒƒƒƒƒƒƒƒƒƒƒƒˆƒ
3 4 5 6 7 8
Predicted Value of Unemployment
Breusch-Pagan tests were conducted to further evaluate the variances. In a Breusch-
𝑆𝑆𝑅 𝑋𝑖 /2
Pagan test, ei2 is regressed on Xi for each X. The test statistic equals , where SSRXi is
(𝑆𝑆𝐸/𝑛)2
the sums of squared regression from Xi being regressed on X, SSE is the sums of squared error
from the preferred model, and n is the number of observations. This test statistic follows a Chi-
squared distribution with 1 degree of freedom; reject the assumption of constant variance if the
test statistic is greater than the cutoff value determined by the Chi-squared distribution. The
results from running all three Breusch-Pagan tests with α= 0.05 are summarized in the table
below.
Covariate Test Statistic Cutoff Value Decision

NASDAQ Low 5.321573 3.84 Reject Constant Variance
Gas Price 2.289452 3.84 Accept Constant Variance
Immigrants Admitted 2.188324 3.84 Accept constant Variance
Using this alpha level, results in the assumption of constant variance being rejected for the
NASDAQ low covariate. This helped confirm the results found with the predicted vs. residual
plot. However, if the alpha level is lowered to 0.01, the cutoff value is increased to 6.63 and the
following outcome is produced:
Covariate Test Statistic Cutoff Value Decision

NASDAQ Low 5.321573 6.63 Accept Constant Variance
Gas Price 2.289452 6.63 Accept Constant Variance
Immigrants Admitted 2.188324 6.63 Accept Constant Variance
All of the test statistics are below the new cutoff value. These results show the importance in
choosing an alpha level when validating assumptions. Using the results from the alpha level of
0.01, the variances are considered to be constant for this analysis.
The last assumption to validate is the independence of the residuals, which can be
determined by a runs test. A runs test requires examining the residuals in time sequential order
to determine if any patterns exist. This test is important for this particular model due to the
observations being evenly spaced over time, and due to the question surrounding whether one
year’s value impacts the next year’s value. A lack of independence is depicted by long strings of
positive or negative residuals or by too many alternating signs of residuals. The test statistic
2∗𝑛 1 ∗𝑛 2 1 2∗𝑛 1 ∗𝑛 2 2∗𝑛 1 ∗𝑛 2 − 𝑛 1 −𝑛 2

equation is [ 1 − + ]/ , where r equals the number
𝑛 1 +𝑛 2 2 𝑛 1 +𝑛 2 2 𝑛 1 +𝑛 2 −1
of runs of signs, n1 equals the number of positive residuals, and n2 equals the number of negative
residuals. For the preferred model, r is 12 n1 is 15, n2 is 18 and the test statistic is -1.378429.
The test statistic follows a standard normal probability distribution so the cutoff value for α =
0.05 is ±1.96. Independence of residuals is a valid assumption because the test statistic of -
1.378429 is in between -1.96 and +1.96.
After checking all of the model assumptions, the data were analyzed for individual
damaging points caused by influential observations, large residuals, and high leverages.
Influential observations are determined by judging how much the model changes by removing
that observation; the greater the change the more influence it has. There are three ways to judge
influence, DFITS, Cook’s Distance, and DF Betas. DFITS measures how much Ŷi changes when
the ith observation is removed. The ith observation is influential if its ‫׀‬DFITS‫ ׀‬is greater than
𝑝 4
2* , which equals 2* ≈ 0.696311. According to the DFITS procedure, observations 8, 9
𝑛 33
and 17 are influential. Cook’s Distance is another way to judge influence, and it measures the
influence of the ith observation on all n values, thirty-three in this case. The ith observation is
influential if its Cook’s Distance is greater than the median of the F distribution with degrees of
freedom numerator equal to p (4 in this analysis) and degrees of freedom denominator equal to n
– p (29 in this analysis). As a result, the cutoff value for the Cook’s Distance is approximately
0.890, and no observations were deemed influential. DF Betas measure how much removing the
2
ith observation changes the β parameters. The cutoff value for the DF Betas is ± , which is
𝑛
approximately ±0.348155. Since DF Betas measure the influence on each β, there are influential
values for each β individually, and these influential values are summarized in the table below.
Parameter Influential Observation(s)

β0 17, 25
β1 9, 16, 17
β2 9, 25
β3 9, 17, 25
None of the observations are shared by all the β parameters, so no observations are deemed
influential overall by the DF Betas approach.
Large residuals, which are the distance between the predicted value and the real value,
were investigated. Large residuals can be determined using studentized deleted residuals, which
follow a t-distribution with n – p – 1 (28 for this analysis) degrees of freedom. For this analysis,
the cutoff value is approximately 1.701 with alpha level 0.1. Using this cutoff value, observation
5, 8, and 9 were found to have large residuals.
After large residuals, high leverages were examined. High leverages are points that are
remote in the X-space. From the SAS output, Hat Diag H was used to decide if the value has a
𝑛
high leverage. The value is 2 (0.242424 for this analysis). Points with high leverage are 16,
𝑝
17, 25, 26, 32, and 33 according to this method.
Damaging points are determined by meeting all three criteria: influential, large residual,
high leverage. After looking at the results, no points are damaging because no points meet all
three criteria. Even if there were observations deemed damaging, no observations should ever be
removed unless there is a problem from data error.

One last diagnostic was to check for multicollinearity, which happens when covariates
are highly correlated. There are two ways to test for multicollinearity. One way is called the
variance inflation factor (VIF). If the VIF is greater than 10 for an observation there is a
multicollinearity problem. None of the VIFs were larger than 10, so no problem was found in
this check. Another way to check for multicollinearity is by a condition index. If the index is
larger than 30, then there appears to be a multicollinearity problem. No problems were found in
this check either, so no multicollinearity problems were found for this model.
Time Series: Are there problems with the linear regression model due to undetected
time correlations? The linear regression model established by all of the model building
techniques seems viable since all of the model assumptions can be verified, including constant
variances of the residuals if the alpha level is set to 0.01. However, since the dependent variable
is being measured repeatedly over time, there may be correlation between the measurements for
which the regression analysis fails to account. This would appear as correlation, or lack of
independence, of the errors during the regression analysis. A runs test was previously conducted
to verify independence of the residuals or errors, but since it only takes into account the signs of
the residuals, and not the magnitude, it may fail to detect some correlation between the residuals.
To address this problem, autocorrelation and partial autocorrelation plots were created, also
called ACF and PACF plots. These plots display the correlation between the residuals at
different time lags.
The autocorrelation of a time series Y at lag k is the coefficient of correlation between Yt
and Yt-k. The ACF plot will always equal 1 at lag 0 because this is showing that a residual is
completely correlated with itself. Autocorrelations should be near 0 for all other time-lag
separations. If the residuals are correlated, or non-random, then one or more of the
autocorrelations will be significantly non-zero. A cutoff value, positive and negative, is
displayed on the ACF plot, and all other lags should have a correlation in between the cutoff
value marks to signify no correlation. For this analysis, lag 1 does not fall within the cutoff
value, along with lags 3, 4, 5, 8, and 9. This plot is displayed in Appendix C.
The partial autocorrelation is the amount of correlation between two variables that is not
explained by their mutual correlations with a specified set of other variables. It is the amount of
correlation between a variable and a lag of itself that is not explained by correlations at all lower-
order-lags. The PACF plot has cutoff values marked similar to the ACF plot and the correlation
values should fall within the marks, the closer to 0 the better. In the PACF plot for this analysis,
lags 1, 2, 4, and 6 have correlation values outside the cutoff value range. This plot is displayed
in Appendix D.
Both the ACF and PACF plots suggest that a time series analysis is warranted, so Auto-
Regressive Integrated Moving Average (ARIMA) modeling was performed. ARIMA models are
noted as ARIMA (p,d,q), where the p is the auto-regressive term, d is the differencing term, and
q is the moving average term. Differencing terms are only used for non-stationary processes,
when the ACF plot dies down very slowly. The ACF plot for this analysis does not fall into this
category, so the differencing term was left at 0. The ACF and PACF plots help identify the p
and q components in the ARIMA model. The table below describes the general way to identify
which ARIMA model to use.
Model ACF Plot PACF Plot

ARIMA (p,0,0) Dies down Cuts off after lag p
ARIMA (0,0,q) Cuts off after lag q Dies down
ARIMA (p,0,q) Dies down Dies down
ARIMA (p,d,q) Does not die down Does not die down
The ACF and PACF plots for this analysis do not fit perfectly into any of these
categories, so several ARIMA models were fitted and compared based upon Akaike Information
Criterion (AIC) and the resulting ACF and PACF plots from the ARIMA model. AIC is a
penalized likelihood, and the lower the AIC is, the better the model is. The AIC for model k
equals n*log(SSEk) + 2Pk, where n is the number of residuals, SSEk is the sums of squared error
for model k, and Pk is the number of parameters for model k. SSEk represents the goodness of fit
of the model, and Pk represents a penalty term for how many parameters are used in the model.
As the number of parameters increases, the SSEk will decrease, but the Pk will increase, which is
why it is called a penalized likelihood. ACF and PACF plots are also created in order to make
sure the problem of correlated residuals is resolved with the ARIMA model. The table below
displays the results from the fitted ARIMA models.
Possible Model AIC Comments about

ACF & PACF plots
ARIMA (1,0,0) 70.06815 Both ok
ARIMA (2,0,0) 61.29249 Both ok
ARIMA (3,0,0) 61.02022 Both ok
ARIMA (4,0,0) 54.8522 Both ok
ARIMA (5,0,0) 54.01416 Both ok
ARIMA(6,0,0) 49.28251 Both ok
ARIMA (7,0,0) 49.78799 Both ok
ARIMA (0,0,1) 65.70913 PACF plot lag 4 exceeds cutoff value
ARIMA (2,0,1) 53.84297 Both ok
ARIMA (2,0,2) 55.84295 Both ok
ARIMA (3,0,1) 55.29058 Both ok
The ARIMA model chosen is bolded in the table above, and detailed results of the model
are in appendix E. ARIMA (6,0,0) was decided as the best model because it has the lowest AIC
value and the ACF and PACF plots show no significant correlation at any of the lags. In order to
get perfect results, the ARIMA model and the previously selected linear regression model should
be fitted at the same time in SAS. The signs of the slopes in the new model created by this
process would be expected to remain the same as in the linear regression model found previously.
The main reason for this analysis is to see what variables are correlated with the unemployment
rate and the type of relationship they share, so this process was not deemed necessary because
the relationships established by the linear regression model should not change.
Results
Linear Regression Equation. In general, the linear regression equation is Ŷ= β0 + β1X1
+ β2X2 + β3X3; the subsequent table explains the notations.
Notation Meaning
Ŷ Estimate for the unemployment rate
β0 Intercept of the line
β1 Slope for the NASDAQ low
X1 Value for the NASDAQ low
β2 Slope for the number of immigrants
X2 Value for the number of immigrants
β3 admitted
Slope for to
thethe Unitedprice
average States
of all gas
X3 admitted
Value fortothetheaverage
Unitedprice
Statesof all gas
types
typesconducted model building techniques provides the
The model preferred by the previously
estimates for the intercept and slope coefficients β0, β1, β2, and β3. In linear regression, the size
of the coefficient for each independent variable is equivalent to the size of the effect that variable
has on the dependent variable, and the sign on the coefficient (positive or negative) reflects the
direction of the effect. For the best fitting model, β0 is 6.97312, β1 is -0.00196, β2 is -
0.00000113, and β3 is 0.01328. Following the guidelines stated before, the NASDAQ low has an
inverse relationship with the unemployment rate due to the corresponding slope, β1, being a
negative number. If the NASDAQ low is increased by 1, the unemployment rate would decrease
by 0.00196 on average, which displays how the size of the slope matches the amount of effect
the variable has on the dependent variable, the unemployment rate. The number of immigrants
admitted to the United States also has an inverse relationship with the unemployment rate due to
its negative slope. If the number of immigrants admitted to the United States increased by 1, the
unemployment rate would decrease by 0.00000113 on average. Alternatively, the average price
of all gas types has a direct relationship with the unemployment rate because of its positive slope.
If the average price of all gas types increased by 1, the unemployment rate would increase by
0.01328 on average. All of these individual effects are contingent on holding the other variables
constant while examining it separately.
This linear equation can be used to predict the unemployment rate in a given year.
Plugging in numbers for the average price of all gas types, the number of immigrants admitted to
the United States, and the NASDAQ low will produce an estimate of the unemployment rate for
that year. If a more accurate result is needed, the linear regression model and the ARIMA model
can be fitted together in SAS in the future.
Relationships Established by this Analysis. The table below displays the relationships
believed to be held between the unemployment rate and the chosen independent variables before
this analysis and the relationship types found by the analysis.
Dependent Variable Relationship Type Relationship Type

Believed Before Found After Analysis
Analysis
CPI Inverse Insignificant
Immigrants Admitted Direct or None Inverse
Burglary Direct Insignificant
Minimum Wage Adjusted for Inflation Inverse Insignificant
Public Debt Direct Insignificant
NASDAQ Low Inverse Inverse
Trade Balance Direct Insignificant
Production Workers’ Average Hourly Inverse Insignificant
Earnings Gas Price
Average Not Sure Direct
Several variables were found insignificant by this analysis, some of which have been stated
significant in previous research; these included the CPI, the amount of burglaries, hourly wages,
public debt, and the trade balance. The three variables that were deemed to be significantly
correlated with the unemployment rate include the NASDAQ low, the amount of immigrants
admitted to the United States and the average price of all gas types.
The NASDAQ low was found to have an inverse relationship to the unemployment rate
by this analysis, which is the same as the relationship thought to have existed based on prior
knowledge. An inverse relationship indicates that a downward movement in the NASDAQ low
is associated with an increase in the unemployment rate, and an upward shift in the NASDAQ
low is associated with a decline in the unemployment rate. Logically, this relationship makes
sense because one would naturally think that if the low of the NASDAQ is extremely low one
year, this indicates that the economy is doing poorly so the unemployment rate is probably high.
On the other hand, if the NASDAQ low is relatively high, than the economy is most likely doing
well so the unemployment rate is presumably low. Correlation, not causation is implied by this
relationship; this means that a change in the NASDAQ low from year to year does not
necessarily cause a change in the unemployment rate, and vice versa. The two variables tend to
move opposite of each other; this could mean that one variable causes the change in the other,
another variable all together causes both to change, or that the relationship is completely
coincidental. Although a change in the NASDAQ low does not necessarily cause a change in the
unemployment rate, an inverse relationship does appear to exist according to this analysis, so the
NASDAQ low can be used as an indicator of what change might appear in the unemployment
rate for that year.
Another variable found to be significant by this analysis is the number of immigrants
admitted to the United States yearly. Based on the literature previously reviewed, this variable
was thought to have either no relationship or a direct relationship with the unemployment rate.
Frumkin (1998) did not think any significant relationship existed between immigration and the
unemployment rate, while a study in Kposowa (1995) found that increases in immigration had a
direct relationship with unemployment among minorities throughout some periods of history in
the United States. However, different results were observed in this statistical analysis. An
inverse relationship was established between the number of immigrants admitted to the United
States and the unemployment rate in this analysis. This is contradictory to most people’s first
notion about immigration. A lot of people think that as immigration increases, these immigrants
take jobs from the currently employed and do more work for less money. Although, rationally if
the immigrants come in legally and fill employment openings, they increase the number of
employed and would thus lower the unemployment rate so an inverse relationship could make
sense. Once again this does not mean that immigration causes a change in the unemployment
rate, but rather that the two variables tend to move opposite of each other. This means that the
number of immigrants admitted to the United States each year could be used as an indicator of
that year’s unemployment rate.
Furthermore, the average price of all gas types was observed to have a significant
relationship with the unemployment rate based on this analysis. Before the analysis was
conducted, no specific relationship was thought to exist between the two variables. The analysis
revealed a direct relationship between the average price of all gas types and the unemployment
rate. This means that high gas prices are correlated with high unemployment rates. High
unemployment rates may be associated with high gas prices due to the fact that gas prices effect
a lot of businesses; when gas prices are high business costs increase, which may lead to cuts in
costs elsewhere such as labor reductions. Yet again, this correlation does not imply causation,
which means high gas prices do not necessarily cause the unemployment rate to rise, and vice
versa. This direct relationship can serve as another indicator of the unemployment rate for a
certain year.
Conclusion
The statistical analysis conducted found three variables that may serve as future
indicators of the unemployment rate: the NASDAQ low, the number of immigrants admitted to
the United States, and the average price of all gas types. The NASDAQ low and the number of
immigrants admitted both have an indirect relationship with the unemployment rate, while the
average price of all gas types shares a direct relationship with the unemployment rate. This
means that a higher NASDAQ low and number of immigrants admitted are correlated with a
lower unemployment rate, and that a higher average price of gas is correlated with a higher
unemployment rate. Although these three variables may be used as future indicators of the
unemployment rate, one must be careful of extrapolation. The relationships were found using
data collected from 1975-2007 and can be used to approximate the unemployment rate during
this period, but may not provide correct results when applied to data prior to 1975 and
subsequent to 2007. Even with the risk of extrapolation, linear regression models are reasonable
tools for predictions. Hopefully, the relationships discovered between the significant covariates
and the unemployment rate through this statistical analysis can prove to be helpful in forecasting
future changes in the unemployment rate. More research is left to be done in this field, such as
combining the ARIMA model with the regression model to get a more accurate model, trying
these processes with other variables suspected to have a relationship with the unemployment rate,
and a more in depth study into why these relationships exist. In spite of the future work that can
be done, the relationships found in this statistical analysis can serve as indications of an
increasing unemployment rate and also as indications of an unemployment rate declining back to
its normal state; this could serve useful in the upcoming months as a hint to when the economy
will begin recovering and also in the future to serve as warning signals of potential downturns.
References
Aaronson, D. & Sullivan, D. (2000). Unemployment and wage growth: Recent cross-state
evidence. Chicago Fed: Economic Perspectives, 24. Retrieved from
http://www.chicagofed.org/publications/economicperspectives/2000/2qep4.pdf
Average hours and earnings of U.S. production workers, 1969-2006. (2009). In The world
almanac and book of facts 2008. (p.101). New York: Readers Digest Trade Publishing.
Blodget, H. (2008, December 5). Great depression unemployment didn't hit 25 percent overnight.
Retrieved February 24, 2009, from Yahoo! Finance Web site:
http://finance.yahoo.com/tech-ticker/article/140365/Great-Depression-Unemployment-
Didnt-Hit-25-Percent-Overnight
Bureau of Labor Statistics. Labor statistics from the current population survey. Available from:
http://www.bls.gov/cps/
Chan, J. S. K. & Choy S. T. B. (2008). Analysis of covariance structures in time series. Journal
of Data Science 6, 573-589. Retrieved from http://proj1.sinica.edu.tw/~jds/JDS-432.pdf
Crime in the U.S. (2009). In TIME almanac 2009. (p.692). Chicago: Encyclopedia Britannica
Inc.
Fernandez, G. C. J. (2006). Model selection methods in proc mixed. Retrieved from
http://www.ag.unr.edu/gf/ashs/allmixedwuss2006.pdf
Frumkin, N. (1998). Tracking America’s economy. New York: M.E. Sharpe.
Galluzzo, G. (2009, February 15). Study finds recession associated with increases in minority
victims of crime. Retrieved February 24, 2009, from University of Iowa News Services
Web site: http://www.eurekalert.org/pub_releases/2009-02/uoi-sfr020909.php

Gasoline retail prices, U.S. city average, 1974-2007. (2009). In The world almanac and book of
facts 2008. (p.106). New York: Readers Digest Trade Publishing.
Hoel, A. A., Clarkson, K. W., & Miller, R. L. (1983). Economics sourcebook of government
statistics. Lexington, MA: LexingtonBooks.
Kposowa, A. (1995). The Impact of immigration on unemployment and earnings among racial
minorities in the United States. Ethnic & racial studies, 18(3), 605-628. Retrieved from
http://www.popline.org/docs/1171/250511.html
Kutner, M. H., Nachtsheim, C. J., & Neter, J. (2004). Table B-6. Applied linear regression
models. New York: McGraw-Hill Companies, Inc.
Lieberman, M. (2008, October 28). Jobless rate is a deceptive depression comparison. Retrieved
from http://www.foxbusiness.com/story/markets/economy/jobless-rate-deceptive-
depression-comparison/
Moyer, M. (2009). Stick 'em up. Scientific American, 300(3), 15-16. Retrieved from Academic
Search Elite database.
NASDAQ stock market, 1971-2007. (2009). In The world almanac and book of facts 2008.
(p.60). New York: Readers Digest Trade Publishing.
Public debt of the U.S. (2009). In the world almanac and book of facts 2008. (p.55). New York:
Readers Digest Trade Publishing.
Total immigrants admitted to the US, 1901-2007. (2009). In TIME almanac 2009. (p.649-650).
Chicago: Encyclopedia Britannica Inc.
Trends in U.S. foreign trade, 1970-2006. (2009). In The world almanac and book of facts 2008.
(p.68). New York: Readers Digest Trade Publishing.

U.S. consumer price index, 1913-2007. (2009). In TIME almanac 2009. (p.719-720). Chicago:
Encyclopedia Britannica Inc.
U.S. federal minimum wage rates, 1952-2008. (2009). In TIME almanac 2009. (p.712). Chicago:
Encyclopedia Britannica Inc.
U.S. unemployment rates. (2009). In TIME almanac 2009. (p.715). Chicago: Encyclopedia
Britannica Inc.
Zuckerman, M. (2004, February 9). The case of the missing jobs. U.S. News & World Report,
136(5), 67-68. Retrieved from Middle Search Plus database.

APPENDIX A
SAS Code
Options ls=80;
Data UnemploymentRate;
Input Unemployment CPI Immigrants Burglary MinWage PubDebt NASDAQ Trade HrEarnings
Gas;
Cards;
8.5 53.80 385378 3625300 8.46 533.2 60.70 12404 4.73 56.7
7.7 56.90 499093 3108700 8.76 620.4 78.06 -6082 5.06 59.0
7.1 60.60 458755 3071500 8.22 698.8 93.66 -27246 5.44 62.2
6.1 65.20 589810 3128300 8.80 771.5 99.09 -29763 5.88 65.2
5.8 72.60 394244 3327700 8.65 826.5 117.84 -24565 6.34 88.2
7.1 82.40 524295 3795200 8.15 907.7 124.09 -19407 6.85 122.1
7.6 90.90 595014 3779700 7.98 997.9 170.80 -16172 7.44 135.3
9.7 96.50 533624 3447100 7.52 1142.0 158.92 -24156 7.87 128.1
9.6 99.60 550052 3129900 7.29 1377.2 229.88 -57767 8.20 122.5
7.5 103.9 541811 2984434 6.98 1572.3 223.91 -109072 8.49 119.8
7.2 107.6 568149 3073348 6.74 1823.1 245.82 -121880 8.74 119.6
7.0 109.6 600027 3241410 6.62 2125.3 322.14 -138538 8.93 93.1
6.2 113.6 599889 3236184 6.39 2350.3 288.49 -151684 9.14 95.7
5.5 118.3 641346 3218077 6.13 2602.3 329.00 -114566 9.44 96.3
5.3 124.0 1090172 3168170 5.85 2857.4 376.87 -93141 9.80 106.0
5.6 130.7 1535872 3073909 6.30 3233.3 322.93 -80864 10.20 121.7
6.8 136.2 1826595 3157150 6.76 3665.3 352.85 -31135 10.52 119.6
7.5 140.3 973445 2979884 6.56 4064.6 545.85 -39212 10.77 119.0
6.9 144.5 903916 2834808 6.37 4411.5 645.02 -70311 11.05 117.3
6.1 148.2 803993 2712774 6.21 4692.8 691.23 -98493 11.34 117.4
5.6 152.4 720177 2593784 6.04 5224.8 740.53 -96384 11.65 120.5
5.4 156.9 915560 2506400 6.56 5413.1 978.17 -104065 12.04 128.8
4.9 160.5 797847 2460526 6.95 5526.2 1194.39 -108273 12.51 129.1
4.5 163.0 653206 2332735 6.84 5656.3 1357.09 -166140 13.01 111.5
4.2 166.6 644787 2100739 6.70 5674.2 2193.13 -265090 13.49 122.1
4.0 172.2 841002 2050992 6.48 5807.5 2332.78 -379835 14.02 156.3
4.7 177.1 1058902 2116531 6.30 6228.2 1387.06 -365126 14.54 153.1
5.8 179.9 1059356 2151875 6.20 6783.2 1114.11 -423725 14.97 144.1
6.0 184.0 703542 2154834 6.06 6783.2 1271.47 -496915 15.37 163.8
5.5 188.9 957883 2144446 5.91 7397.1 1752.00 -607730 15.69 192.3
5.1 195.3 1122257 2154126 5.71 7932.7 1904.00 -711567 16.13 233.8
4.6 201.6 1266129 2183746 5.53 8507.0 2020.39 -753283 16.76 263.5
4.6 207.3 1052415 2179140 6.11 2340.7 2340.68 -700258 17.43 274.6
;
Proc reg data=UnemploymentRate;
Model Unemployment= CPI Immigrants Burglary MinWage PubDebt NASDAQ Trade
HrEarnings Gas / selection=forward slentry=0.15;
Proc reg data= UnemploymentRate;

HrEarnings Gas / selection=backward slstay=0.15;

HrEarnings Gas / selection=stepwise slentry=0.15;

HrEarnings Gas / selection=rsquare rmse;

HrEarnings Gas / selection=cp;

HrEarnings Gas / selection=adjrsq;

Model Unemployment= Immigrants NASDAQ Gas;
Output out=next r=resid p=yhat;
Data final;
Set next;
Sqrres = resid**2;
Proc rank normal=blom;

Var resid;
Ranks nscore;
Proc univariate plot data=final;

Var resid;
Proc plot data=final;

Plot resid*yhat;
Proc corr;
Var resid nscore;
Proc glm data=final;

Model sqrres=Immigrants;
Model sqrres=Nasdaq;
Model sqrres=Gas;
Proc reg data=UnemploymentRate;

Model Unemployment=Immigrants Nasdaq Gas / r p influence Collin vif;
Run;
APPENDIX B
Raw Data
US Fed. Hourly Avg

US Total Min. Wage Earnings Price of
US Consumer Immigrants Adjusted for US Trade of All Gas
Unemployment Price Admitted inflation in Public Debt NASDAQ Balance Production Types
Year Rate Index to the US Burglary yr 2008 (in Billions) Low (in millions) Workers (in cents)
1975 8.50 53.8 385,378 3,625,30 8.46 533.2 60.70 12,404 4.73 56.7
1976 7.70 56.9 499,093 3,108,70
0 8.76 620.4 78.06 -6,082 5.06 59.0
1977 7.10 60.6 458,755 3,071,50
0 8.22 698.8 93.66 -27,246 5.44 62.2
1978 6.10 65.2 589,810 3,128,30
0 8.80 771.5 99.09 -29,763 5.88 65.2
1979 5.80 72.6 394,244 3,327,70
0 8.65 826.5 117.84 -24,565 6.34 88.2
1980 7.10 82.4 524,295 3,795,20
0 8.15 907.7 124.09 -19,407 6.85 122.1
1981 7.60 90.9 595,014 3,779,70
0 7.98 997.9 170.80 -16,172 7.44 135.3
1982 9.70 96.5 533,624 3,447,10
0 7.52 1,142.0 158.92 -24,156 7.87 128.1
1983 9.60 99.6 550,052 3,129,90
0 7.29 1,377.2 229.88 -57,767 8.20 122.5
1984 7.50 103.9 541,811 2,984,43
0 6.98 1,572.3 223.91 -109,072 8.49 119.8
1985 7.20 107.6 568,149 3,073,34
4 6.74 1,823.1 245.82 -121,880 8.74 119.6
1986 7.00 109.6 600,027 3,241,41
8 6.62 2,125.3 322.14 -138,538 8.93 93.1
1987 6.20 113.6 599,889 3,236,18
0 6.39 2,350.3 288.49 -151,684 9.14 95.7
1988 5.50 118.3 641,346 3,218,07
4 6.13 2,602.3 329.00 -114,566 9.44 96.3
1989 5.30 124.0 1,090,172 3,168,17
7 5.85 2,857.4 376.87 -93,141 9.80 106.0
1990 5.60 130.7 1,535,872 3,073,90
0 6.30 3,233.3 322.93 -80,864 10.20 121.7
1991 6.80 136.2 1,826,595 3,157,15
9 6.76 3,665.3 352.85 -31,135 10.52 119.6
1992 7.50 140.3 973,445 2,979,88
0 6.56 4,064.6 545.85 -39,212 10.77 119.0
1993 6.90 144.5 903,916 2,834,80
4 6.37 4,411.5 645.02 -70,311 11.05 117.3
1994 6.10 148.2 803,993 2,712,77
8 6.21 4,692.8 691.23 -98,493 11.34 117.4
1995 5.60 152.4 720,177 2,593,78
4 6.04 5,224.8 740.53 -96,384 11.65 120.5
1996 5.40 156.9 915,560 2,506,40
4 6.56 5,413.1 978.17 -104,065 12.04 128.8
1997 4.90 160.5 797,847 2,460,52
0 6.95 5,526.2 1,194.39 -108,273 12.51 129.1
1998 4.50 163.0 653,206 2,332,73
6 6.84 5,656.3 1,357.09 -166,140 13.01 111.5
1999 4.20 166.6 644,787 2,100,73
5 6.70 5,674.2 2,193.13 -265,090 13.49 122.1
2000 4.00 172.2 841,002 2,050,99
9 6.48 5,807.5 2,332.78 -379,835 14.02 156.3
2001 4.70 177.1 1,058,902 2,116,53
2 6.30 6,228.2 1,387.06 -365,126 14.54 153.1
2002 5.80 179.9 1,059,356 2,151,87
1 6.20 6,783.2 1,114.11 -423,725 14.97 144.1
2003 6.00 184.0 703,542 2,154,83
5 6.06 6,783.2 1,271.47 -496,915 15.37 163.8
2004 5.50 188.9 957,883 2,144,44
4 5.91 7,379.1 1,752.00 -607,730 15.69 192.3
2005 5.10 195.3 1,122,257 2,154,12
6 5.71 7,932.7 1,904.00 -711,567 16.13 233.8
2006 4.60 201.6 1,266,129 2,183,74
6 5.53 8,507.0 2,020.39 -753,283 16.76 263.5
2007 4.60 207.3 1,052,415 2,179,14
6 6.11 9,007.7 2,340.68 -700,258 17.43 274.6
0
APPENDIX C
ACF Plot
*** Autocorrelations for series ACFPlotData[["Residuals"]] ***

Call: acf(x = ACFPlotData[["Residuals"]], type = "correlation", plot = F)
Autocorrelation matrix:
lag ACFPlotData
1 0 1.0000
2 1 0.5153
3 2 -0.0882
4 3 -0.4340
5 4 -0.5766
6 5 -0.3939
7 6 -0.1668
8 7 0.1044
9 8 0.3539
10 9 0.3935
11 10 0.3006
12 11 0.0924
13 12 -0.2356
14 13 -0.3183
15 14 -0.2862
16 15 -0.1694
Series : ACFPlotData[["Residuals"]]
1.0
0.5
ACF
0.0
-0.5
0 5 10 15
Lag
APPENDIX D
PACF Plot
*** Partial Correlations for series ACFPlotData[["Residuals"]] ***

Call: acf(x = ACFPlotData[["Residuals"]], type = "partial", plot = F)
Partial Correlation matrix:
lag ACFPlotData
1 1 0.5153
2 2 -0.4815
3 3 -0.2102
4 4 -0.3879
5 5 -0.1060
6 6 -0.4034
7 7 -0.1049
8 8 -0.1423
9 9 -0.0675
10 10 0.0053
11 11 0.0499
12 12 -0.1776
13 13 0.2027
14 14 -0.1505
15 15 0.0781
Series : ACFPlotData[["Residuals"]]
0.4
0.2
Partial ACF
0.0-0.2
-0.4
0 5 10 15
Lag
APPENDIX E
ARIMA Model Chosen
*** ARIMA Model Fitted to Series ACFPlotData[["Residuals"]] ***
Call: arima.mle(x = `ACFPlotData[[\"Residuals\"]]`, model = model, xreg = xreg, max.iter =

nlmin.max.iter, max.fcal = nlmin.max.fcal)
Method: Maximum Likelihood
Model : 6 0 0
Period: 1
Coefficients:
AR : 0.44863 -0.61846 0.06408 -0.52047 0.09813 -0.4211
Variance-Covariance Matrix:
ar(1) ar(2) ar(3) ar(4) ar(5) ar(6)
ar(1) 0.030469345 -0.015085553 0.014788582 -0.001373974 0.009630928 0.003362501
ar(2) -0.015085553 0.037567217 -0.023470306 0.015620472 -0.007774325 0.009630928
ar(3) 0.014788582 -0.023470306 0.041700793 -0.023702884 0.015620472 -0.001373974
ar(4) -0.001373974 0.015620472 -0.023702884 0.041700793 -0.023470306 0.014788582
ar(5) 0.009630928 -0.007774325 0.015620472 -0.023470306 0.037567217 -0.015085553
ar(6) 0.003362501 0.009630928 -0.001373974 0.014788582 -0.015085553 0.030469345
Optimizer has NOT converged

Due to: iteration limit
AIC: 49.28251
ARIMA Model Diagnostics: ACFPlotData[["Residuals"]]

ACF Plot of Residuals
1.0
0.5
ACF
0.0-1.0
0 2 4 6 8 10 12 14
PACF Plot of Residuals

0.4
0.0 0.2
PACF
-0.4
2 4 6 8 10 12 14
ARIMA(6,0,0) Model with Mean 0

Factors Related To The Unemployment Rate: A Statistical Analysis

Uploaded by

Copyright:

Available Formats

Factors Related To The Unemployment Rate: A Statistical Analysis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Factors Related To The Unemployment Rate: A Statistical Analysis

Uploaded by

Copyright:

Available Formats

University of Northern Iowa

Factors related to the unemployment rate: A

Copyright © 2009 Reanna Collins

Let us know how access to this document benefits you

FACTORS RELATED TO THE UNITED STATES’ UNEMPLOYMENT RATE:

to find if any new variables might be associated with the rate.

Unemployment Rate Details. How the unemployment rate is calculated is essential in

force (Frumkin, 1998).

The unemployment rate is computed mathematically as the number of unemployed persons

participating in an interview, contacting an employment agency, sending out resumes, answering

rate measured in this manner as being noteworthy.

require people to physically go to an agency to be recorded as unemployed, making their rate

the extent to which these registration requirements constitute an over-counting in unemployment

The duration of unemployment is another aspect to consider when investigating the

unemployment might not be as important, at least in terms of policy making, as focusing on if a

factor like overall duration of unemployment is causing the increased rate.

Factors previously found to be related to the Unemployment Rate. A few factors

wage growth and unemployment.

economic situation and unemployment spike.

were diminishing by the end of the Bush administration.

analysis showed that in 1970, a standard deviation change in immigration increased

unemployment among minorities by nearly 14%, while in 1980 unemployment increased by

in the following manner:

Independent Variable Relationship Type

directly. Frumkin (1998) found immigration to be an unimportant factor in the unemployment

no relationship overall or a possible direct relationship between immigration and the

the unemployment rate and these variables.

U.S. Fed. Hourly Avg.

and backward elimination methods.

the United States.

𝑆𝑆𝐸 𝑓𝑢𝑙𝑙 𝑆𝑆𝐸 𝑓𝑢𝑙𝑙

unemployment rate by this linear regression analysis.

Normality of the residuals can be verified graphically and formally. Graphically,

normality of the residuals is demonstrated by a normal probability plot of the residuals. A

illustrates a nearly straight line, so the assumption is valid.

Normal Probability Plot

the assumption of normal residuals was once again confirmed acceptable.

Plot of Predicted vs. Residuals

Breusch-Pagan tests were conducted to further evaluate the variances. In a Breusch-

Covariate Test Statistic Cutoff Value Decision

following outcome is produced:

Covariate Test Statistic Cutoff Value Decision

0.01, the variances are considered to be constant for this analysis.

2∗𝑛 1 ∗𝑛 2 1 2∗𝑛 1 ∗𝑛 2 2∗𝑛 1 ∗𝑛 2 − 𝑛 1 −𝑛 2

1.378429 is in between -1.96 and +1.96.

Parameter Influential Observation(s)

influential overall by the DF Betas approach.

5, 8, and 9 were found to have large residuals.

17, 25, 26, 32, and 33 according to this method.

removed unless there is a problem from data error.

different time lags.

The autocorrelation of a time series Y at lag k is the coefficient of correlation between Yt

autocorrelations will be significantly non-zero. A cutoff value, positive and negative, is

value, along with lags 3, 4, 5, 8, and 9. This plot is displayed in Appendix C.

which ARIMA model to use.

Model ACF Plot PACF Plot

displays the results from the fitted ARIMA models.

Possible Model AIC Comments about

Linear Regression Equation. In general, the linear regression equation is Ŷ= β0 + β1X1

* Autocorrelations for series ACFPlotData[["Residuals"]] *

* Partial Correlations for series ACFPlotData[["Residuals"]] *

* ARIMA Model Fitted to Series ACFPlotData[["Residuals"]] *