Modeling population growth in the United States

Luke Paulsen
OpenIntro
openintro.org
CC BY-SA∗

1 Introduction

Population growth can define a community. Communities that grow rapidly may see increased
investment, while contracting communities may find local assets, such as homes and businesses,
falling in value relative to those of nearby communities. In this investigation, we wish to determine
which demographic factors relate most closely to population growth in U.S. counties from 2000
to 2010.

2 Data Exploration

Data for each county is available from the US Census website, including age, gender, race, and edu-
cation, along with other relevant demographics such as homeownership, employment, and income.
Five counties for these data are summarized in Table 1, and the data were originally collected from
the US Census website.1 This investigation will only consider a subset of variables and be limited
to counties where those variables are complete. The resulting data set represents 3,083 counties
on 23 different variables. A complete list of the variables under consideration along with variable
descriptions is available at
www.openintro.org/stat/data/cc.php

growth pop2000 age under 5 age under 18 female black hs grad bachelors
1 24.96 43671 6.6 26.8 51.3 17.7 85.3 21.7
2 29.80 140415 6.1 23.0 51.1 9.4 87.6 26.8
3 -5.44 29038 6.2 21.9 46.9 46.9 71.9 13.5
4 10.03 20826 6.0 22.7 46.3 22.0 74.5 10.0
.. .. .. .. .. .. .. .. ..
. . . . . . . . .
3143 8.49 6644 5.7 21.8 47.4 0.3 91.1 17.9

Table 1: Five rows from the countyComplete data set with 8 of the 23 variables.
∗
This document is released under a Creative Commons Attribution-ShareAlike 3.0 license.
1
These data were collected from the US Census website. The data are available in the openintro R package and
also as a tab-delimited text file at openintro.org/stat.

1
A variable called growth that represents the population growth rate for each county from 2000 to
2010 is included in Table 1, and this variable represents the response variable for the analysis. This
variable is summarized in Table 2 and Figure 3. Growth rates for the ten-year period average 5.4%,
with the middle half ranging from -2.2% to 10.4%.

Mean Median St. Dev. IQR Min Max

5.42% 3.29% 13.18% 12.70% -46.6% 110.40%

Table 2: Statistical summaries of population growth in US counties from 2000

to 2010.

300

250
>30
200
Frequency

150

100 0
50

0
−50 0 50 100
<−30

Growth (percent)

Figure 3: Population growth across the United States from 2000 to 2010.

Several other variables in the data set are worth exploring. The pop2000 variable measures the
population of the county during the 2000 census and is shown in Figure 4. The variable is very
skewed, so we will use the natural logarithm of population in the model. Taking the natural
logarithm of population allows us to measure population differences in terms of multiplication
rather than addition. For example, a difference of 1,000 people would be important in a county
with population 10,000 but less so in a county with population 1,000,000. Using the natural
logarithm for population means differences are compared geometrically, e.g. comparing counties
with populations of 1,000 and 10,000 will be analogous to a comparing two counties with 10,000
and 100,000 people in the model.

600 200 1,000 10,000 100,000 1,000,000

500
150
Frequency

Frequency

400

300 100
34 counties with a population
200 greater than 1,000,000
are not shown 50
100

0 0
0 500,000 1,000,000 4 6 8 10 12 14 16

Population, Year 2000 Natural Log of Population, Year 2000

Figure 4: Distribution of populations. Left: original populations. Right: log-

transformed populations.

2
 100
40 70

Percent With Bachelors Degree

60
80

Percent White (not Hispanic)

Percent under Age 18

30
50
60
40
20
40 30

10 20
20
10
0
0
0 2 4 6 8 10 12 0 20 40 60 80 40 50 60 70 80 90 100

Percent under Age 5 Percent Black Percent With High School Degree

Figure 5: Three figures that highlight the collinearity of several predictors.

There are also several groups of variables that divide the population with respect to a particular
statistic: age, race, or education level. We expect these variables to be related to one another,
and this relationship must be considered when interpreting the results. Figure 5 highlights the
relationships among some of these variables.
The first plot in Figure 5 suggests that the variables age under 5 and age under 18 are strongly
correlated. The diagonal line in the second plot represents the fact that the percentages of each
racial group in the population cannot sum to more than 100%. The percentage of the population
that self-identifies as some race other than black or non-Hispanic white is represented by the
distance of a point from the downward-trending diagonal. The relationship in the third plot is
somewhat weaker, but it shows that the percentage of the population with a bachelor’s degree is
always smaller than the percentage that completed high school, as would be expected.

3 Analysis

Two variables can be linearly related. For example, the left panel of Figure 5 shows a positive trend
relating age under 5 and age under 18. This trend looks linear, and it can be modeled, even if
imperfectly, by using a straight line. Such a line would have error for individual observations, but
it would capture the overall structure of the relationship.

3.1 Modeling population growth

When working with many variables, the principles of the linear model can be generalized, where
here we simultaneously fit many variables against a response rather than one variable at a time.
We begin by writing a formula that models the growth rate as a linear combination of all the other
variables that we are considering:
d = β0 + β1 × log(pop2000) + β2 × female
growth
..
.
+ β21 × poverty + β22 × sales per capita

3
Statistical software may be used to identify the best fitting model, where point estimates of β0 , β1 ,
..., β22 would be estimated in the model.
To improve the model, we perform model selection, eliminating variables using backwards selection,
until all remaining variables are found to be statistically significant. The model following backwards
selection is summarized by Table 6.

Estimate Std. Error t value Pr(>|t|)

(Intercept) 5.3168 6.3879 0.83 0.4053
log(pop2000) 1.6556 0.2007 8.25 0.0000
age under 5 3.3118 0.4179 7.92 0.0000
age under 18 -0.6160 0.1605 -3.84 0.0001
age over 65 -0.2666 0.0816 -3.26 0.0011
female -0.3026 0.1080 -2.80 0.0051
hispanic 0.1320 0.0327 4.03 0.0001
white not hispanic 0.0803 0.0145 5.54 0.0000
no move in one plus year -0.5614 0.0520 -10.79 0.0000
foreign born 0.2509 0.0632 3.97 0.0001
foreign spoken at home -0.1744 0.0471 -3.70 0.0002
bachelors 0.5281 0.0431 12.24 0.0000
mean work travel 0.7156 0.0434 16.48 0.0000
housing multi unit -0.4930 0.0347 -14.22 0.0000
median val owner occupied 0.0000 0.0000 4.20 0.0000
persons per household 9.9882 1.3261 7.53 0.0000
per capita income -0.0003 0.0001 -3.73 0.0002
poverty -0.3266 0.0506 -6.46 0.0000
sales per capita 0.0002 0.0000 5.74 0.0000

Table 6: Model summary for the regression model predicting population growth
after model selection. See page 2 for a link that provides variable descriptions.

The variables black, hs grad, and density were eliminated during model selection. However,
variables that we would expect to be closely correlated with these variables – hispanic and
white not hispanic with black, and bachelors with hs grad – still appear in the model. As we
saw in the Data Exploration section, some variables are highly correlated, i.e. they are collinear.
When predictors are collinear, having one in a multiple regression model may be about as good as
having both, and this may explain why black and hs grad were eliminated during model selec-
tion.
In the age variables there is a surprise of a different type. The variables age under 5 and age under 18
are highly collinear, but both are still included in the model, and the model suggests they have
opposing effects on population growth. It may be tempting to make a standard interpretation of
the coefficients, however, that could be misleading. These two variables are collinear (see Figure 5),
and this complicates interpretation. For example, dropping age under 5 results in the coefficient
of age under 18 changing from -0.62 to 0.40. The practical interpretation of these variables has
been complicated by other variables in the model.

4
3.2 Diagnostics

In order to assess the multiple regression model, we check conditions on the model’s residuals. The
general requirements are that the residuals are roughly normal, have approximately the same vari-
ance, and are independent. We leave it to the reader to check whether any nonlinear relationships
exist between the predictors and growth variable.
Figure 7 is a normal probability plot of the model’s residuals. There is clear curvature, and the tails
at the corner of the graph indicate that some of the observations have unusually distant residuals
from zero. While this would be a substantial concern for a model with only a small number of data
points, over 3,000 counties are being used here, so the influence of these outlying residuals should
be very limited.

50
Sample Quantiles

−50

−3 −2 −1 0 1 2 3

Figure 7: Normal probability plot for the residuals following model selection.
There is clear curvature, but the outliers are probably reasonable for the size
of this data set.

Figure 8 shows that the residuals plotted against their fitted values. The variance is approximately
consistent, with perhaps a small increase in variability with larger predicted values. One county,
Kalawao County in Hawaii, had a predicted value far from the cloud at (-59.4%, 20.6%). This
small and isolated county was previously a quarantine for leprosy patients; no new residents are
allowed to move to this county. Due to the unusual nature of this county, this observation should
be excluded in future analyses.
Figure 9 is useful for checking spacial independence of the residuals. In a model that fully explained
the observations, we would expect the residual values to be randomly distributed geographically;
instead there are definite geographic patterns and clusters of similar residuals. For example, the
model fails to account for variables such as climate, which may help explain why adjacent counties
tend to have similar residual values. This figure indicates there are additional features remaining
within the data that were not captured by the multiple regression model presented in here, violating
the independence condition for the residuals.

5
 50
Kalawao County, Hawaii
Residuals

●
0

−50

−60 −40 −20 0 20 40

Predicted Values

Figure 8: Residuals versus fitted values from the regression model.

>18

<−18

Figure 9: Residuals plotted by their location. Empty spaces represent counties

that had missing data and were not included in the analysis.

6
3.3 Practical interpretation of model coefficients

We will proceed in estimating the impact of many variables on population growth, but we want to
highlight that these findings may be somewhat unreliable due to the violation of the independence
condition for the residuals. Each variables’s coefficient was multiplied by the variable’s IQR to get
a scaled impact for the variable, shown in Table 10. The proper way to interpret each value is,
“The growth rate for a county at the 75th versus the 25th percentile in this variable, other things
being equal, would be estimated as higher over ten years.”

log(pop2000) age under 5 age under 18

2.8% 4.3% -2.3%
age over 65 female hispanic
-1.3% -0.4% 0.9%
white not hispanic no move in one plus year foreign born
2.2% -3.2% 1.0%
foreign spoken at home bachelors mean work travel
-1.3% 5.0% 5.1%
housing multi unit median val owner occupied persons per household
-4.8% 1.1% 2.6%
per capita income poverty sales per capita
-1.9% -2.6% 1.4%

Table 10: The values in this table represent the estimated difference in growth
rate for a county at the 75th versus the 25th percentile in each variable, other
things being equal.

4 Conclusion

In this investigation, we attempted to model a U.S. county’s population growth based on readily
available demographic data, a potentially useful tool for economic and other applications. We
found strong statistical evidence that many of the demographic variables measured by the 2010 U.S.
Census (including age, racial, and demographic distribution, economic conditions, and household
makeup) were important in modeling a county’s population growth between 2000 and 2010. Taken
together in a multiple regression model, the measured variables appear to explain nearly half of the
variation in growth rate among counties.
Of the variables measured, the percentage of the population with a bachelor’s degree may be
especially important in terms of population growth. We suspect the modestly large estimated
coefficient of mean commute time is not a driver of population growth but a result of a migration
to suburbs, which often require larger commute times. It is also important to consider that many
of the variables examined are related to one another, which complicates the interpretability of
many model coefficients. This makes it especially difficult to conjecture causal conclusions from
the current model.
Further analysis of how the model’s variables are related to one another, possibly including trans-
formations of some variables in the model, may be helpful in eliminating this source of error and
in providing more definite results. In addition, this model includes no information on geographic

7
location and does not distinguish between urban, suburban, and rural areas. These types of infor-
mation appear to be important in determining county growth rates and should likely be included
as variables in future investigations.