Math 636 - Mathematical Modeling Fall Semester, 2006 Discrete Modeling - U. S. Population © 2001, All Rights Reserved, SDSU & Joseph M. Mahaffy San Diego State University -- This page last updated 07-Sep-06

Discrete Modeling - U. S. Population

Outline of Chapter

This section compares and contrasts three discrete models for simulating the population of the U. S. We develop the simplest Malthusian growth model, then extend this model in two directions to show a nonautonomous Malthusian growth model and the classic logistic growth model. These models are all fit to census data using a nonlinear least squares best fit. Procedures using Excel and MatLab are presented for finding the best fits and projections are given for each of the models.

Population of the United States

The United States takes a census of its population every 10 years. The census has important ramifications for many aspects of our society, such as budgeting federal payments and representation in Congress. The method of taking the census and how it is analyzed has been a very hot issue pitting the Republicans against the Democrats in 1999 with the issue landing in the Supreme Court. The Republicans wanted a strict interpretation of the Constitution, knowing that a direct head count always undercounts minorities and the poor, who vote predominantly Democratic. The Democrats claimed that the Constitution framers wanted an accurate count of the populace, so that modern statistical methods should be employed. This would naturally give them an advantage in the voting. (The Supreme Court came down in the middle pleasing neither party and saying that the Constitution requires a head count, which will be used for allocating Congressional seats and districting, while the more accurate statistical count may be used for apportioning the money for Federal funding.) The arguing over the numbers will go on for several years as each group tries to use the numbers to their best advantage to gain federal money and political power. Despite the political controversy over the numbers, accurately predicting these demographic data are important for planning our communities in the future. At the base of all calculations for the future population predictions is some type of mathematical model. Current models are quite sophisticated, but first we must appreciate the basic models behind them. Below we present the census data for the history of the U. S.

U. S. Census Data

1790
3,929,214
1870
39,818,449
1950
151,325,798
1800
5,308,483
1880
50,155,783
1960
179,323,175
1810
7,239,881
1890
62,947,714
1970
203,302,031
1820
9,638,453
1900
75,994,575
1980
226,545,805
1830
12,866,020
1910
91,972,266
1990
248,709,873
1840
17,069,453
1920
105,710,620
2000
1850
23,191,876
1930
122,775,046

1860
31,433,321
1940
131,669,275

The growth rate between each decade can be determined by dividing the census at one date by the census a decade earlier and subtracting one. The growth rate for the decade of 1790-1800 35.1%. The javascript below performs this computation to give the growth rate for any decade in the history of the U. S.

 Select the range of Census Data 1790 - 1800 1800 - 1810 1810 - 1820 1820 - 1830 1830 - 1840 1840 - 1850 1850 - 1860 1860 - 1870 1870 - 1880 1880 - 1890 1890 - 1900 1900 - 1910 1910 - 1920 1920 - 1930 1930 - 1940 1940 - 1950 1950 - 1960 1960 - 1970 1970 - 1980 1980 - 1990 1990 - 2000 Growth rate of the range selected

From the javascript above, we find that the growth rates for decades following 1790, 1800, and 1810 are 35.1%, 36.4%, and 33.1%, respectively, which averages 34.9% per decade. This growth rate remains almost constant until 1860, so this information should allow us to estimate the census data up until 1860 using a model with a constant growth.

Discrete Malthusian Growth

The simplest mathematical model is the Malthusian growth model, which states that

Pn+1 = Pn + rPn = (1 + r)Pn,

where r is the average growth rate. Thus, the population in the next decade is equal to the current population plus the current population times the average growth rate, r, of the population. This equation is the Discrete Malthusian Growth model (named after the work of Thomas Malthus (1766-1834)). If we begin the model at the population in 1790 and use the average growth rate, r = 0.349, from above, then the table below shows the predictions by this model through 1870 along with the percent error from this model.

 Year Census Model Pn+1=1.349Pn % Error 1790 3,929,214 3,929,214 1800 5,308,483 5,300,510 -0.15 1810 7,239,881 7,150,388 -1.24 1820 9,638,453 9,645,873 0.08 1830 12,866,020 13,012,282 1.14 1840 17,069,453 17,553,569 2.84 1850 23,191,876 23,679,765 2.10 1860 31,433,321 31,944,002 1.62 1870 39,818,449 43,092,459 8.22

Notice that the error remains small until 1870 because of the fairly constant rate of growth. Most of the predicted populations are a little high, especially in 1870, suggesting that throughout the 19th century the growth rate declined. The declining growth rate is a general trend that we observe for the U. S. and is something that needs to be accounted for in any improved model. Between 1860 and 1870, the Civil War occurred, which is one cause of the dramatic decline in the rate of growth of the population in the U. S. Thus, the 1870 prediction is one that you would expect to be poor. In fact, the shift from the primarily agricultural society to the industrial revolution is more significant in causing the decline in the rate of growth. The crowding effects are what the logistic growth model considers.

Clearly, the Malthusian growth model is limited to a range of dates where the growth rate remains relatively constant. If you attempt to continue using this model until 1920 or 1970, then the model produces the census values of 192,365,343 and 859,382,645, respectively. (These estimates are 82% and 323% too high.) Thus, this model becomes increasingly bad if we assume the constant growth rate of 34.9%. A calculation of the growth rate in 1920 gives around 15%, which further drops to only 13% in 1970. (The lowest growth rate can be seen to have occurred during the Great Depression (between 1930 and 1940) with only a 7.2% growth rate.) Thus, this simple model can only predict populations for a limited time into the future, but certainly provides good estimates for some community planning that is required.

Solution of Discrete Malthusian Growth Model

There are not many discrete models that have an explicit solution. However, it is easy to solve the discrete Malthusian growth model. From the model above, we see that

P1 = (1 + r)P0,

P2 = (1 + r)P1 = (1 + r)2P0, ...

Pn = (1 + r)Pn-1 = ... = (1 + r)nP0.

Thus, the general solution of this model is given by

Pn = (1 + r)nP0.

This shows why Malthusian growth is also known as exponential growth. The solution to the model that is given by the equation above is an exponential function with a base of 1+ r and power n representing the number of iterations after the initial population is given.

Improving the Malthusian Growth Model

The section above presents a discrete Malthusian Growth model based on the U. S. population from census data. From the solution to the model above, we see that the Malthusian growth model has only two parameters that can be adjusted, the initial condition, P0, and the growth rate, r. In this section, we examine two extensions of the Malthusian growth model that increase the number of modeling parameters to three. One model extends the Malthusian Growth model to include time varying reproduction rates, and the other model is the classical logistic growth model. We will compute the least squares best fit of all models in this section to see how they do at matching the census data and discuss the model predictions for the future. A hyperlink is provided to the Excel worksheet with the U. S. population models, and below we present MatLab computations for finding the least squares best fit to the logistic model.

The general discrete dynamical population model is given by

Pn+1 = f (Pn) ,

where f is a function depending only on the population P at time tn. This difference equation is said to be autonomous as it does not have a temporal or time dependence. A more general difference equation is given by

Pn+1 = f (tn, Pn) ,

which is a nonautonomous difference equation.

As noted above, the general solution of the Malthusian growth model is given by

Pn = (1 + r)nP0.

A nonlinear least squares best fit to the census data gives the best fitting parameters to this model as the initial condition, P0 = 15.07, and the growth rate, r = 0.1524. Thus, the best fitting model is given by

Pn = 15.07(1.1524)n.

This model poorly fits the early and late census data, but it is extremely simple with only two parameters. The sum of square errors is 2297.

We know that human populations vary with technology and medical care much more than other animal populations. This suggests that a time varying growth rate would provide a better model. It is clear from the census data that the growth rate over the last two centuries has been declining. A linear growth rate in time adds only one parameter, but proves to create a substantially better fit to the census data. The nonautonomous Malthusian growth model with linear time varying growth is given by

Pn+1 = (1 + b+ atn )Pn,

where a and b are parameters to be fitted to the data and tn = 10n is the time after 1790 in years. We now have three parameters: the initial condition, P0, and the growth parameters, a, and b. A nonlinear least squares best fit to the census data gives the best fitting parameters to this model as the initial condition, P0 = 5.877 , and the growth parameters, a = -0.001062 and b= 0.3092. Thus, the best fitting model is given by

Pn+1 = (1.3092 - 0.01062 n )Pn,

with P0 = 5.877. This model fits the census data quite well, and the sum of square errors is only 300.1.

The earliest and most famous of the autonomous models that extend beyond Malthusian growth is the logistic growth model. This model has been shown to fit many animals populations with limited resources quite well. So we can ask how well the human population follows this model. The logistic growth model also has three parameter. The logistic growth model is given by

Pn+1 = Pn + rPn(1 - Pn/M),

where r and M are parameters to be fitted to the data with the initial condition, P0. A nonlinear least squares best fit to the census data gives the best fitting parameters to this model as the initial condition, P0 = 7.992 , and the growth parameter, r = 0.2342 and carrying capacity, M= 408.4 . Thus, the best fitting model is given by

Pn+1 = (1 + 0.2342(1 - Pn/408.4 )Pn,

with P0 = 7.992. This model fits the census data quite well, and the sum of square errors is only 491.9. The sum of square errors shows that the logistic growth model is only a slightly poorer fit to the data than the nonautonomous Malthusian growth model.

The graph below gives a comparative study of these three models. The Malthusian Growth model performs poorly as noted above because the growth rate varies too much over the time interval being considered. The Nonautonomous Discrete Malthusian Growth and Logistic Growth models match the data quite well.

The analyses above and the graph above were computed using Excel's nonlinear Solver routine, which is a complex algorithm for finding minima (or maxima) of functions by varying parameters. The complete worksheet is available for downloading. Similar calculations can be easily done with MatLab using its fminsearch routine.

The simulations of the models given above into the future project that the population in 2050 and 2100 are 602.3 and 1,224 million according to the Malthusian growth model, 378.4 and 401.2 million according to the nonautonomous Malthusian growth model, and 358.7 and 393.6 million according to the logistic growth model. The nonautonomous growth model predicts that the maximum population in the U. S. will occur between 2080 and 2090 with the maximum being about 405 million. The logistic growth model predicts that the population will increase less and less rapidly, leveling off around 408 million. It is rather remarkable how similar the predictions of the nonautonomous Malthusian and logistic growth models are.

Nonlinear Least Squares Fit with MatLab

It is not difficult to create a matlab program to find the least squares best fit to the U. S. census data. Begin by loading the U. S. Population data. In MatLab, we write

We create function for simulating the logistic growth model.

function pop=dislog(p)
pop(1) = p(1);
for n = 1:20
pop(n+1) = pop(n)+p(2)*pop(n)*(1-pop(n)/p(3));
end

This function depends on the three parameters given above, and they are stored in the vector p with P0 = p(1), r = p(2) and M = p(3). A function is created to compute the sum of square errors

function err=lstsquares(p)
pop = dislog(p);
e1 = usdata(:,2)'-pop;
err = e1*e1';

Next we use fminsearch to compute the least sum of square errors and store the information in the variable p1 as follows:

p1 = fminsearch(@lstsquares, [3.9, 0.25, 400])

If we type the following, then we obtain a nice plot of the data and model:

pop = dislog(p1);
t = usdata(:,1);
plot(usdata(:,1),usdata(:,2),'o',t,pop,'-')

References:

[1] Statistical Abstracts of the United States (1993) 113 th ed., U. S. Department of Commerce, Bureau of the Census, Washington, DC.