[Type text] [Type text] [Type text]

Project 2

Group 7
Multivariate Assessment for

Life Expectancy
Robert Lesperance, Megan Withers,
Nick Cellitti, and Amanda Fisher
 Multivariate Assessment for Life Expectancy

Table of Contents
Table of Figures ............................................................................................................................................. 2

Executive Summary....................................................................................................................................... 3

Introduction .................................................................................................................................................. 4

Multiple Regression Analysis ........................................................................................................................ 6

Overall Fit ...................................................................................................................................................... 7

Hypothesis Tests ........................................................................................................................................... 9

Binary Significance Test .............................................................................................................................. 11

Tests for Multicollinearity ........................................................................................................................... 13

VIF’s ......................................................................................................................................................... 13

Correlation Matrix................................................................................................................................... 14

Error Assumption Testing ........................................................................................................................... 15

Errors Are Normally Distributed ............................................................................................................. 15

Tests for Non-constant variance ............................................................................................................. 17

Autocorrelation Tests ............................................................................................................................. 18

Unusual Observations ................................................................................................................................. 19

Conclusion ................................................................................................................................................... 20

Bibliography ................................................................................................................................................ 22

Appendix A .................................................................................................................................................. 23

Appendix B .................................................................................................................................................. 24

1
 Multivariate Assessment for Life Expectancy

Table of Figures

Figure 1: Multiple Regression Table.............................................................................................................. 6

Figure 2: Overall Fit Test ............................................................................................................................... 7

Figure 3: Evan’s and Doane’s Rule ................................................................................................................ 8

Figure 4: Prediction Model ........................................................................................................................... 8

Figure 5: Intercept Test ................................................................................................................................. 9

Figure 6: Smoking Adults Test ....................................................................................................................... 9

Figure 7: Cancer Death Rate Test................................................................................................................ 10

Figure 8: Average Income Test.................................................................................................................... 10

Figure 9: Obesity/ Overweight Test ............................................................................................................ 11

Figure 10: North South Test ........................................................................................................................ 11

Figure 11: VIF Evaluation and Independent Variable Adjustment.............................................................. 13

Figure 12: Correlation Matrix...................................................................................................................... 14

Figure 13: Adjusted Correlation Matrix ...................................................................................................... 15

Figure 14: Histogram Plot ........................................................................................................................... 15

Figure 15: Normal Probability Plot of Residuals ......................................................................................... 16

Figure 16: Residuals by Predicted ............................................................................................................... 17

Figure 17: Residuals by Observation ........................................................................................................... 18

Figure 18: Test for Autocorrelation ............................................................................................................ 19

2
 Multivariate Assessment for Life Expectancy

Executive Summary

This report was created on behalf of the actuary division in the S.C. insurance because of the

new fiscal year predictions for next year. The statistical analysis found in this report depicts the

relationship between average income, obesity, cancer deaths, smoking, and life expectancy between all

50 states in the United States. The data came from the Kaiser Family Foundation and the US Census

Bureau. The data was analyzed using multivariate regression analysis and hypothesis testing to

determine the overall effect that the independent variables have on the dependent variable, life

expectancy. By utilizing Mega-Stat, S.C. Insurance Agency was able to calculate the coefficient

correlation which showed that our linear equation was significant and had a strong positive correlation

(0.925). The correlation of determination analysis concludes that 85.6% of the variability in life

expectancy is explained by the independent variables. 14.4% are unexplained by the variables in this

model.

The overall fit of the model was found significant at the 10 percent significance level; while adult

obesity/overweight rate and average income rate were not found significant at the 10 percent

significance levels. Further testing revealed that multicollinearity did exist, and if the smoking adult

variable is removed, the overall significance for independent variables increases and all remaining

variables become significant. There were no violations of assumptions. There were 3 outliers in the data;

Utah, Minnesota, and Maryland. States that have a lower life expectancy and have a lower significance

to all of the four variables will in return give S.C. insurance the most profit. On the other side, if the

variables for life expectancy show high and have a strong correlation to all four variables, S.C. insurance

will lose money. Future testing on genetics, life style, and extracurricular activities should be conducted

to see how they impact life expectancies and future profits.

3
 Multivariate Assessment for Life Expectancy

Introduction

Every year S.C. Insurance Agency looks for ways to increase profit margins and better their

company as a whole. One of the easiest ways to do this is to determine which states will help S.C.

Insurance achieve the highest possible total profit and then to focus marketing in these specific states.

In this department of S.C. Insurance Agency is the actuary division. Due to advances in medicine and

new technological procedures, insurance payouts have become higher and higher. Clients are requesting

more insurance money to cover a larger portion of their expenses incurred for medical procedures and

tests. The primary objective of this department, as company statistician’s, is to forecast the best possible

insurance practices that will benefit S.C. Insurance financially, while still providing adequate health

insurance to potential clients. The division will be separated into four different groups, each testing a

different independent variable to determine which state has the shortest lifespan and which factors

affect lifespan most. Once each report on the four independent variables have been composed, they will

be compiled into a larger multivariable regression analysis.

This report will focus on the impact that cancer per 100,000, average income per state, rate of

obesity/ overweight adults, rate of smoking adults, and geographic location (Northern part of U.S. or

Southern part of U.S.) have on average life expectancy. Based on the Kaiser Family Foundation finding's1,

the average life expectancy (in years) is determined by each state's own observations (State Health

Facts, 2014).The percentage of adults (18 years or older) that smokes or are obese/ overweight are

calculated per state as well (State Health Facts, 2014). The target population for samples in 2010

consists of persons living in households who have a working cellphone, are aged 18 and older, and

received 90 percent or more calls on cellphones. Samples are chosen randomly for the 50 states from a

set area, determined by area code and response to phone calls (Behavioral Risk Factor Surveillance

System, 2013). Average income per state is based on the real median household income, and the

1
Statistics are taken from the ‘CDC’- the Center for Disease and Control

4
 Multivariate Assessment for Life Expectancy

householder and any individuals 15 years or older in that household, even if there is no relation to the

homeowner (Noss, 2012). The data provided from the site is based on sample approximates and closely

represent the entire household and group quarter populations for all states and recorded by the U.S.

(Christie, 2009) Census Bureau (Noss, 2012). Finally, cancer death rates are age-adjusted rates per

100,000 standard populations (Number of Cancer Deaths per 100,000 Population, 2013). Again, Kaiser

Family Foundation accumulated these findings recorded by the CDC2, and provided the figures for the

2010 year findings.

The correlation will be analyzed to determine if the regression analysis is significant and useful

for S.C. Insurance Agency’s goal of maximizing profits. The report will begin with a multivariate

regression analysis and then a series of hypothesis tests that will help conclude whether or not each

independent variable is in fact significant when compared to life expectancy. This report will also include

a multicollinearity test, normality test, non-variance test, and an autocorrelation test to evaluate error

potential in the data set. All of the testing and computational analysis will be completed in Microsoft

Excel with the use of Mega-Stat.

2
‘CDC’- Center for Disease and Control

5
 Multivariate Assessment for Life Expectancy

Multiple Regression Analysis

Figure 1: Multiple Regression Table

The multiple regression analysis was performed using Mega-Stat to analyze the multiple data

sets. The original data can be referenced in Appendix A. The summary of the analysis, provided by

Mega-Stat, can be found in Figure 1. The regression equation for the model is: Life Expectancy = 91.1658

- Smoking Adults (20.7315) - Cancer Death Rate (0.0311) + Average Income (0.00001391) - Adult

Obesity/Overweight Rate (5.0349) - North or South (1.1314). The correlation coefficient (r) = 0.925

concludes that the data set has a strong positive linear correlation. The regression analysis shows that if

all independent factors are zero, the average life expectancy would still be 91.1658. This value is not

completely accurate due to other factors that may have a negative impact on average life expectancy.

Some of this can be explained by genetics, life style, and extracurricular activities. The coefficient of

6
 Multivariate Assessment for Life Expectancy

determination (R2) shows the linear relationship between the independent variables and dependent

variable. The coefficient of determination (R2) is equal to 0.856 and indicates that 85.6% of the

variability in Life Expectancy can be contributed to smoking adults, cancer death rates, average income,

and obesity. The adjusted R2 value (0.840) penalizes the regression data for adding unnecessary

predictor variables that do not prove to be helpful for the overall model. Due to the fact that the R2

value (0.856) is not significantly different than the adjusted R2 value (0.840), this demonstrates that the

model does not contain any useless predictors.

Overall Fit

Figure 2: Overall Fit Test

To test overall goodness of fit for the multiple regression, a hypothesis test was performed

comparing F calculated values to F critical values. The F calculated value (52.50) is determined by taking

the mean squares of the explained and dividing by the mean squares of the error. Based on Figure 2, the

analysis declares that there is a goodness of fit, and at least one of the predictors in the regression

model is significant, because 52.50 is greater than Fisher’s critical value (1.982752).

7
 Multivariate Assessment for Life Expectancy

Figure 3: Evan’s and Doane’s Rule

Using Evan’s rule, the determination of reliable fit is assessed for the model. This verifies

whether R2 is reliable based on the sample size and number of predictors. The basis of Evan’s rule states

that at least 10 observations per predictor are necessary in order to eliminate over fitting of the model

whereas Doane’s rule is more relaxed and allows for only 5 observations for predictors, see Figure 3. The

model is utilized to limit the number of predictors based on sample size. The data provided in the

regression analysis shows that the model surpasses the minimum requirement for observations which

means the coefficient of determination (R2) is reliable and that there is no overfitting in this model.

Figure 4: Prediction Model

intercept smoking cancer income obesity $50

The coefficient of determination (0.856) indicates that 85.6% of the correlation between

independent and dependent variables can be explained, and 14.4% of the correlation remains

unexplained by the independent variables. The predictor values are utilized to determine the impact of

each independent variable on the dependent variable. For this model, we chose a predictor value of 50

to multiply each slope (β1) value by. In each prediction test, we chose one specific independent variable

to set as zero to clearly show its impact on the entire model. The table shown in Figure 4 shows the

prediction model. The prediction model portrays the regression equation with a predictor value of 50

inputted in each independent variable except 1.

8
 Multivariate Assessment for Life Expectancy

Hypothesis Tests

Figure 5: Intercept Test

The intercept test shows that this model is significant at 10% level because the tcalc= 21.621 >

tcrit= 1.68023. The values for tcalc come directly from the regression output table from the regression

analysis. This tells S.C. Insurance that at 0 additional factors there would still be an average life

expectancy of 91.2 years and that the independent variables do in fact influence the dependent variable

life expectancy.

Figure 6: Smoking Adults Test

9
 Multivariate Assessment for Life Expectancy

The hypothesis test in Figure 6 shows that smoking adults is significant at the 10% significance

level. Smoking is in fact a significant variable for S.C. insurance to compare with Life Expectancy.

Figure 7: Cancer Death Rate Test

The hypothesis test in Figure 7 shows that cancer death rate is significant at the 10% significance

level. Cancer rate is not a significant variable for S.C. insurance to compare with Life Expectancy.

Figure 8: Average Income Test

The hypothesis test in Figure 8 shows that average income is significant at the 10% significance

level. Average Income is a significant variable for S.C. insurance to compare with Life Expectancy.

10
 Multivariate Assessment for Life Expectancy

Figure 9: Obesity/ Overweight Test

The hypothesis test in Figure 9 shows that obesity is not significant at the 10% significance level.

Obesity rate is not a significant variable for S.C. insurance to compare with Life Expectancy.

Binary Significance Test

Figure 10: North South Test

The binary significance test enables S.C. insurance to capture the effects of non-quantitative

variables, in this model it was location. The categorical predictor is chosen to be North=0 and South=1.

South is considered significant, based on the test in Figure 10, as there is a correlation between living in

warmer climates versus colder climates in the United States and the effects on Life Expectancy. The

11
 Multivariate Assessment for Life Expectancy

division of the United States was done unscientifically, based on estimation of which states lean toward

warmer climates and those that generally are colder. The states classified as warmer can be found in

Appendix A. The binary variable in this case will shift the intercept, but will not affect the slope of the

multi-regression model.

South = 1 (if the state is a southern state)

Not South (North) = 0 (otherwise)

The binary predictors are important because they enable us to view the effects of categorical

data (such as state location). If the binary coefficient is found to be significantly different from zero then

we can conclude that the binary predictor is a significant predictor for y. The new regression equation

when including the binary variable is Life Expectancy = 91.1658 - Smoking Adults (20.7315) - Cancer

Death Rate (0.0311) + Average Income (0.00001391) - Adult Obesity/Overweight Rate (5.0349) - North or

South (1.1314). The discrepancy between equations, depicted below, show the shift in the graph (

intercept changes from 91.1658 to 90.0344) that is supposed to occur due to the addition of the binary

variable.

X1 = 0, then y = 91.1658 - (20.7315)smoking adults - (0.0311)cancer death rates + (0.00001391)average

X1 = 1, then y = 91.1658 - (20.7315)smoking adults - (0.0311)cancer death rates + (0.00001391)average

12
 Multivariate Assessment for Life Expectancy

Tests for Multicollinearity

VIF’s

Multicollinearity is when there are strong correlations between independent variables, which

increases the standard errors of the coefficients, and may find that the values would otherwise be

significant if Multicollinearity did not exist in the model. Looking at the variance inflation factor’s(VIF’s),

found in Figure 1, there are multiple values highlighted in blue. Since no values equal 1, meaning no

collinearity, and some values are greater than 5, the determination is that multicollinearity exists (Doane

& Seward, 2013). In this instance, the value with the highest VIF is smoking adults; to correct for

Multicollinearity the simplest option is to remove the independent variable and rerun the correlation

regression analysis.

Figure 11: VIF Evaluation and Independent Variable Adjustment

13
 Multivariate Assessment for Life Expectancy

Figure 11 demonstrates the changes when the independent variable Smoking Adult rate is

removed from the data set. As indicated in the far right column of the table, the VIF is adjusted and all

VIF’s are reduced by removing the variable (no blue highlights). Also, all variables are now showing as

significant (highlighted in yellow).

Correlation Matrix

Figure 12: Correlation Matrix

Another way to test for multicollinearity is the correlation matrix; the values highlighted in

yellow, shown in Figure 12, demonstrate high collinearity between variables. This indicates that there is

strong relationships between independent variables and could be affecting the data set in a positive

way. This also could affect the significance of other variables, had certain variables not be introduced to

this data set. According to Klein’s rule, we should not worry about the stability of the regression

because the pairwise predictor correlation values do not exceed R value (0.925). This tells us that the

confidence intervals and t-tests may not be affected.

14
 Multivariate Assessment for Life Expectancy

Figure 13: Adjusted Correlation Matrix

In Figure 13, the independent variable Smoking Adult rate is removed from the regression

model. By removing the variable, it is apparent that correlation rate between independent variables has

improved, and that multicollinearity did exist.

Error Assumption Testing

Errors Are Normally Distributed

Figure 14: Histogram Plot

15
 Multivariate Assessment for Life Expectancy

The analysis has three major limitations that result in potential errors for the overall model is

called residuals. These limitations are all assumptions that are made: the errors are normally distributed,

the errors have constant variance, and the errors are dependent of one another (Doane & Seward,

2013). Figure 14 represents a histogram in relation to our residuals. This histogram was constructed

using the standardized residuals. The histogram shows that the data, shown in Appendix B, does not

contain any major outliers; all values fall within -1.5 to +1.5 with a normal bell curve distribution.

Figure 15: Normal Probability Plot of Residuals

The first limitation (normally distributed) is also tested in Figure 15. The figure shows that the

regression fit is highly linear with possible unusual observations. The coefficient of determination

depicts whether Figure 15 passes the normality of errors assumption. The large R2 value (0.9898)

indicates a near perfect linear regression line (98.98% of the errors are explained), which means it

passes the normality of errors assumption.

16
 Multivariate Assessment for Life Expectancy

Tests for Non-constant variance

Figure 16: Residuals by Predicted

The second assumption that needs to be tested is whether or not the errors have constant

variances. If the variance is constant among the errors then the model would be considered

homoscedastic, which is the ideal condition. If the variances are not constant then the errors are

considered heteroscedastic; this indicates bias and makes the values inefficient, potentially overstating t

values. The residual plot in Figure 16 is used to test for homoscedasticity. Since the residual plot does

not fan-out or funnel-in, it is passes the homoscedasticity test, and the second assumption is met.

17
 Multivariate Assessment for Life Expectancy

Autocorrelation Tests

Figure 17: Residuals by Observation

The final assumption for residual tests is that the errors are independent from one another.

Figure 17 will be used to test assumption 3: autocorrelated errors. If the errors are in fact independent

from one another, then the data is considered to be non-autocorrelated. If the errors are

autocorrelated, this means that estimated variances are biased that affects the confidence intervals. The

Durbin-Watson test will be used to determine autocorrelation in the data set. As a general rule the

residual plot should cross the mean at zero (n/2) times or 50/2 = 25. If the centerline crossing is more or

less than 25, the data is considered negative autocorrelated and positive autocorrelated respectively.

There are 26 values above and 24 values below the 0.00 axis. This is indicative of a slight negative

autocorrelation, but is near enough to be call un-correlated. No autocorrelation exists for this test.

18
 Multivariate Assessment for Life Expectancy

Figure 18: Test for Autocorrelation

The Durbin-Watson test statistic seen in Figure 18 is determined from the regression analysis is

2.68. This value suggests a slight negative autocorrelation because the Durbin-Watson is greater than 2,

but not significantly close to 4. This is confirmed when compared to the centerline crossings that should

be (50/2) = 25. The residual plot actually has 26 centerline crossings and more than (n/2) centerline

crossings suggests a negative autocorrelation. Based on the Durbin-Watson test, which has slightly more

validity than the runs test, it can be concluded that no autocorrelation exists for this model.

Unusual Observations

The last step to regression analysis is to evaluate the unusual residuals. This is done with the use

of a leverage test. If the leverage is beyond = 0.10, then there is indication for high leverage.

A list of all the leverages for the data set can be found in Appendix B. This data set shows that Colorado

and Hawaii have a very high level (5/n) = 0.1 and North Dakota and Vermont still have high leverage but

are not as significant as Colorado and Hawaii. These four data points are considered unusual

observations and may be potential outliers.

19
 Multivariate Assessment for Life Expectancy

Another method for testing unusual residuals is evaluating the studentized residual in Appendix

B. The studentized residual values help reveal extreme outliers, outside the -2 to 2 range. For this data,

there are three states that fall below -2 range: Maryland, Minnesota, and Utah. Utah and Minnesota

have very high life expectancies, where Maryland has a moderate life expectancy. Utah and Maryland

both have unusual dependent variables; Utah has low smoking rate and low obesity rate, while

Maryland has the highest average income level. The possible explanation for Utah’s low obesity and

smoking rates can be contributed to religious groups located in Utah (Cannon, 2010). The religious

beliefs established there have apparently affected the life style, improving logistics. Minnesota also has

a very high life expectancy; a contribution could be state wide bans on smoking and unhealthy living

practices (McCarthy, 2014). Maryland has a high average income due to cost of living in Maryland; the

cost of living is about 52% higher than that in Mississippi (Christie, 2009).

Conclusion

A multivariate regression was performed in order to evaluate the relationship between the

following independent variables obesity, cancer, smoking, income, and dependent variable life

expectancy. The regression equation Life Expectancy = 91.1658 - Smoking Adults (20.7315) - Cancer

Death Rate (0.0311) + Average Income (0.00001391) - Adult Obesity/Overweight Rate (5.0349) - North or

South (-1.1314) shows that there is a correlation between our independent variables and our dependent

variable. With this relationship we can prove that there is a strong dependency on our variables with

effect to our dependent variable. With every $1 increase to average income, life expectancy increases

by a factor of 0.00001391. With every 1% increase in obesity rate there is a 5.0349% decrease in life

expectancy. With every 1% increase in cancer death rate there is a 0.0311 decrease in life expectancy.

With every 1% increase in smoking adults there is a decrease in life expectancy by 20.7315%. When

analyzing the R2 value of .856 we conclude that 85.6% of the variability in life expectancy is explained by

20
 Multivariate Assessment for Life Expectancy

the independent variables. The R2 value along with the series of hypothesis tests to determine the

significance of the overall model, slope, and intercept all prove that the independent variables are in

fact significant when compared to the dependent variable (life expectancy). While 85.6% of the

variability is explained by the regression analysis, there is 14.4% that is not explained. This 14.4% could

be attributed to many different variables, for example, genetics, extracurricular activities, and life style.

This tells us that S.C. Insurance Agency can conclude that average income, obesity rate, cancer

deaths, and smoking are in fact significant variables when related to life expectancy, and will help them

better the company as a whole. It will help show which states S.C. Insurance Agency will profit most

from and which states will lose the most money. For example, in the states that have a lower life

expectancy and have a lower significance to all of the four variables will in return give S.C. insurance the

most profit. On the other side, if the variables for life expectancy show high and have a strong

correlation to all four variables, S.C. insurance will lose money.

21
 Multivariate Assessment for Life Expectancy

Bibliography

Behavioral Risk Factor Surveillance System. (2013, 08 15). Retrieved 05 31, 2014, from Centers of Disease and
Control Prevention: http://www.cdc.gov/brfss/data_documentation/index.htm

Number of Cancer Deaths per 100,000 Population. (2013, May 8). Retrieved June 6, 2014, from The Kaiser Family
Foundation: http://kff.org/other/state-indicator/cancer-death-rate-per-100000/

State Health Facts. (2014). Retrieved May 31, 2014, from The Henry J. Kaiser Family Foundation:
http://kff.org/other/state-indicator/smoking-adults/#

Cannon, M. W. (2010, April 13). UCLA Study Proves Mormons live longer. Retrieved June 3, 2014, from Deseret
News: http://www.deseretnews.com/article/705377709/UCLA-study-proves-Mormons-live-
longer.html?pg=all

Christie, L. (2009, September 22). Where to find the fattest paychecks. Retrieved June 6, 2014, from CNN Money:
http://money.cnn.com/2009/09/21/news/economy/highest_income_census/?postversion=2009092203

Doane, D., & Seward, L. (2013). Applied Statistics in Business and Economics. New York: McGraw-Hill/Irwin.

McCarthy, J. (2014, March 13). In U.S., Smoking Rate Highest in Kentucky, Lowest in Utah. Retrieved June 3, 2014,
from Gallup Well-being: http://www.gallup.com/poll/167771/smoking-rate-lowest-utah-highest-
kentucky.aspx?utm_source=rss&utm_medium=rss&utm_campaign=in-u-s-smoking-rate-lowest-in-utah-
highest-in-kentucky-smoking-rate-in-alaska-has-dropped-the-most-since-2008

Noss, A. (2012, September 1). Household Income for States: 2010 and 2011. Retrieved June 6, 2014, from U.S.
Census: http://www.census.gov/prod/2012pubs/acsbr11-02.pdf

22
 Multivariate Assessment for Life Expectancy

Appendix A

23
 Multivariate Assessment for Life Expectancy

Appendix B

24