The Causes of Poverty Author(s): Lester C. Thurow Source: The Quarterly Journal of Economics, Vol. 81, No. 1 (Feb.

, 1967), pp. 39-57 Published by: The MIT Press Stable URL: http://www.jstor.org/stable/1879672 Accessed: 10/09/2009 20:10
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/action/showPublisher?publisherCode=mitpress. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that promotes the discovery and use of these resources. For more information about JSTOR, please contact support@jstor.org.

The MIT Press is collaborating with JSTOR to digitize, preserve and extend access to The Quarterly Journal of Economics.

http://www.jstor.org

THE CAUSES OF POVERTY

LESTERC. THUROW

The poverty model, 40.- Farmers, 43.- Families with no one in the labor force, 44.- Education, 46.- Alaska and Hawaii, 47.- Full-time employment, 47.- Industrial structure, 48.- Nonwhites, 51.- Principal component analysis, 52. Implications, 56.

From 1947 to 1963 the percentage of families with incomes of less than $3000 fell from 31.4 per cent to 18.5 per cent (in 1963 dollars).1 Acceleration of this decline requires an analysis of the different factors which contribute to the incidence of poverty. Since many of the possible explanations are interrelated, both their qualitative and quantitative importance are difficult to determine from descriptive classifications of those living in poverty. One method of isolating the relative importance of the different factors is to specify the relationships more precisely in an econometric model and to analyze the results. To increase the number of observations and to provide for greater variation in the incidence of poverty, this paper presents a model to explain the incidence of poverty in each of the fifty states and the District of Columbia. In 1960, 21.4 per cent of the nation's families were living in poverty, but this average masked a dispersion in the incidence of poverty ranging from a high of 51.6 per cent in Mississippi to a low of 9.8 per cent in Connecticut. The parameters from the poverty model provide estimates of potential gains from altering explanatory variables, but they do not indicate whether these factors can be easily altered. To investigate this problem each of the explanatory variables is examined in detail. Where possible, the different explanatory factors are examined in regressions to see how and why each variable changes over time. Since the explanatory variables used in the poverty model are not completely independent of each other and since many other variables are closely related to those actually used in the model,
* The author would like to thank Lester D. Taylor and John R. Meyer for many excellent suggestions. 1. This paper accepts but does not attempt to defend the use of the $3000 poverty line. Data availability limits the usefulness of more sophisticated definitions on a state-by-state basis. While only families are considered in this paper, the same techniques could be used for unrelated individuals. All of the data in the paper came from the following sources: U.S. Department of Commerce, Bureau of the Census, Current Population Reports, Series P-60 and Series P-50; and 1960 Census of Population, Characteristics of the Population, Vol. 1, Part 1, U.S. Summary.

40

QUARTERLY JOURNAL OF ECONOMICS

principal component analysis is used to isolate the independent dimensions which are present among the various possible explanatory factors. This technique allows a further reduction in the number of factors necessary to explain the incidence of poverty, identifies the basic dimensions which are present, and permits a ranking of their relative importance.
THE POVERTY MODEL

The independent variables used in the model are: (1) the percentage of families living on farms, (2) the percentage of families headed by a Negro,2 (3) the percentage of families with no one in the labor force, (4) the education of the family head, (5) the percentage of the population working full time, (6) the industrial structure of the state, and (7) a dummy variable necessary to correct for the particular circumstances of Alaska and Hawaii. The model for 1960 is specified by the following equation: (1) P = a+bF+cN+dL+ eE-fW-fI-hD+u where P = percentage of families in poverty (income less than $3000), F = percentage of families living on farms, N = percentage of families headed by a nonwhite, L = percentage of families with no one in the labor force, E = percentage of family heads with less than eight years of school completed, 'W P = percentage of population fourteen years old and above who worked 50-52 weeks per year, I = an index of the industrial structure of the state3 D = dummy variable for Alaska and Hawaii, u = error term. In this equation the incidence of poverty is explained without having explicit recourse to state median incomes or wage levels. Differences in median incomes explain over 90 per cent of the vari2. All data in this paper actually refer to the category which is officially called nonwhite. Since this group is 92 per cent Negro, the text will use Negro and nonwhite interchangeably. 3. The index is defined as follows: n I =:E Xi WI where Xi = percentage of the state's labor force in industry i, WI = the ratio of the U.S. median income in industry i to the general U.S. median income. This index measures the prevalence of high-wage industries in the state.
i=1

THE CAUSES OF POVERTY

41

ance in the incidence of poverty by state, but median incomes and poverty are partially measuring the same phenomenon. To explain poverty by median incomes is to introduce a near tautology. The two variables undoubtedly move together and could be used in forecasting, but this is due to the fact that the same causes are influencing the behavior of both.4 While equation (1) is an attempt to explain poverty, the same variables could equally well be used to explain median incomes (though the parameters would differ). The poverty model is designed to determine some of the factors influencing both poverty and median incomes. Although the poverty model partially escapes from the problems of relating poverty and median incomes, the escape is not complete by any means. The industrial structure (farming is also part of the complete industrial structure) and the number of full-time workers both have a very direct connection with median incomes since they are two of the component elements that go into determining a state's median income. What the model does avoid is having to use the observed differences in state wage levels within a particular industry. This is significant, since wage scales in most industries have a range of 2 to 1 from the highest wage state to the lowest wage state, but this procedure does not eliminate the problem. The model uses explanatory variables more basic than median incomes, but it has not answered the question as to why some regions have a poor industrial structure and few full-time jobs. Ideally, these factors should be represented in the model.5 Given our current theoretical and empirical knowledge of poverty, equation (1) cannot be considered a structural model since causation presumably does not move solely from independent to dependent variables. Low education levels and a poor industrial structure lead to a high incidence of poverty, but poverty also leads to low education levels and to a poor industrial structure. Eliminating this circle is difficult, both conceptually and from a policy standpoint. Using equation (1) in a weighted regression gives the following results: 6
4. Gallaway has shown that changes in poverty are strongly related to changes in median family incomes over time. In his simplest formulation a $100 increase in real median family incomes reduces poverty by about 0.5 percentage points. L. E. Gallaway, "The Foundations of the 'War on Poverty'," American Economic Review, LV (Mar. 1965), 127. 5. Probably the factors that are missing in the equation are those factors which determine location decisions after adjustments have been made for the quality of the human labor force (education, etc.). If locational factors were known, they could probably be substituted for the industrial index of a region. 6. A regression using equation (1) needs to be weighted by the popula-

42 (A)

QUARTERLY JOURNAL OF ECONOMICS

P = 96.5125 + .2978 F + .1133 N + .5410 L (.1677) (.0544) (23.1516) (.0978) + .4345 E(.0480)
ax *

.5368 W(.1117)
**

.7600 I(.1978)
*;*

10.3777 D (4.8210)
**

2.3 .98 The poverty regression indicates that most of the variance in the incidence of poverty from state to state can be explained in terms of these seven variables (see Table I). All but one of the explanatory variables is significant at the 1 per cent level and it is significant at the 5 per cent level. All of the coefficients have the signs that would be expected theoretically. While regression (A) gives impressive cross-sectional results, these results need to be tempered with the knowledge that the various possible explanatory variables are so closely related to each
** =2

significant at 5 per cent level = significant at 1 per cent level

d.f. = 43

TABLE I
RANGE OF DATA USED IN POVERTY REGRESSION

(Percentages)
United States High Low

Percentage of families living on farms Percentage of families headed by a nonwhite Percentage of families with no one in the labor force Percentage of family heads with 0-7 years of school completed Percentage of population 14 and above who worked 50-52 weeks per year Index of industrial structure Incidence of poverty
Source: U.S. Census, 1960

7.4 9.4 10.7 21.9

31.3 64.2 24.7 45.0

0.0 0.1 7.5 8.0

34.8 98.9 21.4

47.4 103.1 51.6

28.2 83.9 9.8

tion of each state since the dependent variable and most of the independent variables are in percentage terms. Since a large state provides more of the total number of families living in poverty, it needs to have a larger weight in the regression. The large intercept term is a scaling factor in the weighted regression. It does not indicate anything about the incidence of poverty.

THE CAUSES OF POVERTY

43

other that a wide variety of variables could be used to do an equally good job of explaining poverty. Thus the variables in the poverty model need to be interpreted with caution since they may be serving as proxies for a set of related variables.
FARMERS

Part of the significance and size of the coefficient of farmers in the poverty model (see regression A) occurs because the $3000 poverty line does not allow for income in kind, but adjusting for this problem does not eliminate the importance of farm poverty. If the average farmer had $500 worth of income in kind, the incidence of poverty among farm families would have fallen only from 47.6 per cent to 41.2 per cent in 1963. From 1947 to 1963 the percentage of families headed by farmers fell from 12.8 per cent to 6.5 per cent.7 This decline can be explained by the relative income opportunities in farming and by the availability of alternative nonagricultural employment opportunities.8 When farmers' relative income goes up, the process of leaving the farm is delayed, and when more nonagricultural jobs are available the process is accelerated. From 1947 to 1963 the following results occur: 9
(B) F = 228.4661 + .0420 Ir
-

25.6083 In J

(18.1589)
**

(.0229)

(2.0565)
**

+ 4.7276 D1 + 2.6835 D2 (.4894) (.2979)

** **

12 .31 d.w. = 2.44 R2 = .99 Se d.f= where F = percentage of families headed by farmers Ir = median farm family income / median nonfarm family income J = civilian nonagricultural employment D1 and D2 = dummies to account for changes in the definition of farmers in 1949 and 1959.
7. These percentages have been corrected for changes in the definition of a "farmer." Before correction the number of farmers drops from 17.5 per cent to 6.5 per cent. 8. For a more complete discussion of farm mobility see: Brian Banbury Perkings, "Labor Mobility between Farm and Nonfarm sector," unpublished Ph.D. thesis, 1964, Michigan State Universitv. 9. The large intercept in this regression is a scaling factor necessary to adjust the regression for the use of the natural log of civilian nonagricultural employment.

44

QUARTERLY JOURNAL OF ECONOMICS

In this equation the elasticity of the percentage of farm families with respect to relative farm income is 0.2 (evaluated at the means), while the elasticity with respect to nonagricultural employment is 2.2.1 The impact of alternative employment opportunities clearly dominates the income effect, but both of these variables have contributed to the postwar decline in the percentage of farm families as well as to increases in recessionary years such as 1958. While the number of farm families declined by 200,000 per year in the postwar period, the number actually rose by 200,000 during 1958. The positive coefficient for farmers in the poverty model indicates that reducing the number of farmers and thus the amount of farm poverty does not result in an equivalent increase in urban poverty. Reducing rural poverty by moving people into urban jobs reduces the general level of poverty as well as rural poverty.
FAMILIES WITH NO ONE IN THE LABOR FORCE

Since 70 per cent of the 4.8 million families with no one in the labor force in 1960 were accounted for by families headed by females or by persons over sixty-four years of age, these two variables were used as substitutes for the percentage of families with no one in the labor force in the poverty model (equation 1), but they proved to be poor substitutes, that is of lesser statistical significance. Other factors such as job opportunities, played an important role in bringing individuals into the labor force. The importance of job opportunities can be seen both over time and across states. After accounting for demographic factors, a one percentage point rise in the number of full-time workers lowers the percentage of families with no one in the labor force by 0.4 percentage points across states, and a one percentage point decline in unemployment reduces the number of families with no earner by 0.2 percentage points over time.2 (See regressions C and D.)
1. This elasticity is calculated by dividing the regression coefficients by the mean of the percentage of farm families. When regression (B) is computed with a trend term, the income elasticity drops to zero and the employment elasticity drops to 0.8, but this still leaves the problem of determining what is causing the negative trend of 0.3 percentage points per year. 2. Two different economic variables and two slightly different dependent variables have to be used since the percentage of full-time workers and the percentage of families with no one in the labor force are not available annually for the full postwar period. The percentage of families with no one in the labor force differs from the percentage of families with no earner since the labor force figure refers to a survey week while the earner figure refers to the whole year.

THE CAUSES OF POVERTY

45

(C)

Across states in 1960: L = 15.6297 + .3277 F + .4766 A (.1767) (5.0543) (.1783)

** **

.3859 W + 16.4915 D (4.0877) (.0952)

** **

Se-2.7 d.f. = 46 percentage of families with no one in the labor force percentage of families with female head percentage of families with head over sixty-four percentage of population fourteen years old and over who work 50-52 weeks per year D = dummy for Alaska and Hawaii Over time with national data: 3 1947-1963 11.6880 + .5716 F + .9819 A + .2204 U Le = (D) (.0968) (.3713) (3.4875) (.1397) R2= .86 where L = F = A = W=
** ** * *

d.w.= 1.51 = .80 Se = .37 Le = percentage of families with no earner where U = unemployment rate While the coefficient of the percentage of families outside of the labor force (0.54) is very large in the poverty model (Regression A), both time series and cross-sectional results indicate that a major proportion of the families headed by females or by persons over age sixty-four are unable to benefit directly from general economic progress, though they are able to benefit from rising pensions and welfare payments made possible by economic growth. The same point is emphasized by the rising importance of this group in the total incidence of poverty. From 1950 to 1963, the proportion of poverty-stricken families accounted for by those with no one in the labor force rose from 29.3 per cent to 48.9 per cent. The number of families permanently outside of the labor force could be reduced by special training programs and by tailoring jobs to their specific capabilities, but many still could not or should not be in the labor force. For example, from society's point of view it is not
?2

3. Previous studies have indicated that the size of the labor force is cyclically sensitive, but these results also indicate that some of the cyclical growth of the labor force comes from families that are outside of the labor force during depressed periods. For more on sensitivity of labor force see: Alfred Tella, "The Relation of Labor Force to Employment," Industrial and Labor Relations Review, Vol. 17 (April 1964).

46

QUARTERLY JOURNAL OF ECONOMICS

at all obvious that mothers with young children should have jobs or that older workers should be forced to work in their old age. All of these factors indicate that if the United States is serious about eliminating poverty, and not just reducing it, a program of income redistribution will have to become a central part of the attack on poverty among those now outside of the labor force.
EDUCATION

If the coefficient of education in the poverty model (Regression A) is correct, improvements in education are one of the most effective ways of eliminating poverty. According to the most optimistic projections of the Census Bureau,4 the percentage of family heads with less than an eighth-grade education will fall from 21.9 per cent in 1960 to 15.9 per cent in 1970 and to 11.1 per cent in 1980. This suggests that education alone will reduce the incidence of poverty to 18.8 per cent in 1970 and to 16.8 per cent in 1980, if the economy is able to generate enough capital to equip a more highly educated labor force. Accelerating the decline of poverty by improving education levels will be difficult, however, since low educational attainments are concentrated among older workers. Reducing high-school dropouts might be an excellent investment in the war on poverty, but this program will not have a big impact on the number of poorly educated family heads in the short run. Any dramatic additional reduction in the number of poorly educated workers will depend on the establishment of efficient adult education programs. While the largest reductions in poverty would occur by improving the lower end of the educational spectrum, the percentage of
TABLE II
AND POVERTY EDUCATION (Percentages)
Education of Head of Family Incidence of Poverty in 1963

Elementary: High School: College:

less than 8 years 8 years 1 to 3 years 4, years 1 to 3 years 4 years or more

43.9 25.1 17.9 1Q.1 9.3 4.7

Source: Bureau of Census, Low Income Families, Series P-60, No. 45, 1964.

heads is approximated from estimated changes in male education levels.

4. Current Population Reports, Series P-20, No. 91. Education of family

THE CAUSES OF POVERTY

47

family heads with less than an eighth-grade education should be partially regarded as a proxy for all levels of education.5 Since the incidence of poverty progressively falls as education rises (see Table II), improvements in the entire structure of education should be important in eliminating poverty.
ALASKA AND HAWAII

A dummy variable is necessary for Alaska and Hawaii since neither state fits into the pattern of the rest of the United States. Both states have a very high proportion of "nonwhites," but they are not Negroes and do not suffer from the same kinds of discrimination. Differences in price levels, cultural backgrounds, and a whole host of other characteristics separate these two states. The coefficient of the dummy variable (Regression A) indicates that if Alaska and Hawaii suffered from the same problems as the rest of the United States, their average incidence of poverty would be 10.3 percentage points higher than it actually was in 1960.
FULL-TIME EMPLOYMENT

State unemployment rates were substituted for the percentage of the population aged fourteen and above who worked 50-52 weeks per year in the poverty model (equation 1), but proved to be statistically insignificant with regression coefficients close to zero. Many of those living in poverty are not unemployed, but are working on part-time jobs or are not in the labor force. In 1963 families with an unemployed head accounted for 4.5 per cent of those in poverty, but families with a head employed part time accounted for 23.1 per cent and families without a head in the labor force accounted for 48.9 per cent. Since the percentage of full-time workers takes into account the number of unemployed workers, the number of parttime workers, and the labor force participation rates of the region, it is a much broader measure of the employment characteristics of a labor market than the unemployment rate and consequently a better explanatory factor. In the poverty regression, the coefficient for full-time workers is
5. The median number of school years completed was substituted for the percentage of family heads with less than an eighth-grade education, but proved to be a much less effective variable. Other more complex variables (such as the percentage of family heads with less than an eighth-grade education minus those with a college education) were just as effective but were not included since they did not improve on the results of the simpler index in the poverty regression.

48

QUARTERLY JOURNAL OF ECONOMICS

large (0.54), but rising productivity leading to higher wages also means that full-time work is becoming increasingly effective as a means of escaping poverty. The incidence of poverty among families with a head employed full time dropped from 12.2 per cent in 1956 to 6.9 per cent in 1963 (1963 dollars).6 While the proportion of full-time workers is a better explanatory variable of the incidence of poverty than state unemployment rates, the same economic factors influence both. (The simple correlation between these two variables is 0.81.) In the five states with the lowest unemployment rates in 1960, the average percentage of the population who were full-time workers was 9.5 percentage points higher than in the five states with the highest unemployment rates.7 Over time the same relationship appears, but changes in the age-sex composition of the labor force also have to be considered. Increases in female participation rates increase the proportion of full-time workers in the population, even if they decrease the proportion of full-time workers in the labor force, while increases in the teenage population decrease the proportion of full-time workers. In a time series regression each of these variables has the expected effect: 1950-1963 .5129 Te Wh = 33.7617 - .5996 U + .2951 Pf(g) (.2122) (.1525) (3.5836) (.1159)
** ** *

Se = .39 d.w. = 1.62 .82 d.f. = 10 where Wh = percentage of population aged fourteen and above who worked 50-52 weeks per year and who worked 34 or more hours per week. U = unemployment rate Pf= participation rate for females Te = proportion of the population aged fourteen and above who are teenagers.
R2=

STRUCTURE INDUSTRIAL The industrial structure of a region is a powerful force in the incidence of poverty. The regression coefficient from the poverty model is very large (0.76), but this variable is difficult to change over time. A similar index was constructed from the period from
6. Time series data refer to those who work 50-52 weeks per year and who work more than 34 hours per week. 7. The average unemployment rates for these two groups was 2.9 per cent and 8.3 per cent.

THE CAUSES OF POVERTY

49

1948 to 1964 using 1960 weights. It rose from 94 to 99 with a pronounced cyclical pattern.8 A faster rate of growth would probably have led to a larger change, but the industrial structure is not a variable that can be rapidly altered for the entire economy. Over time the growth of output per worker (resulting from gains in productivity within each industry as well as from shifts to more productive industries) and the level of utilization of the available labor force are the important sources of higher family incomes for the entire economy. In addition to these two factors, an explanation of changes in family incomes has to consider cyclical variations in the share of output going to persons. If median family incomes are made a function of gains in output per employee, the level of utilization of the labor force, and the share of total output going to personal income (see equation 2), lnSt + ut (2) lnMt = a + b 1nPt + clnEt + d where Mt = median family income Pt - GNP/empIoyee Et = percentage of labor force employed St personal income/GNP median family income with respect to output per of elasticity the 1.15 for both nonwhites and whites (see approximately is employee of family income with respect to the the elasticity but Table III), of labor utilization differ sharply. level the and share personal income
TABLE III
MEDIAN a FAMILY b INCOME, c

1947-1964
d R2 d.w. d.f. Se

Median White Family Income

- 0.882

1.140 (.031)

1.209 (.520)

2.353 (.462)

.998 .990

.78 .70

14 14

.02 .04

Median Nonwhite - 1.967 Family Income

1.603 1.172 2.285 (.070) (1.157) (1.023)

The elasticity of the personal income share is 1.6 for nonwhites and is not statistically different from zero, but it is 2.4 for whites and highly significant. Since the share going to personal income rises in recessions and falls in expansions, the much larger positive
8. An occupational index was constructed in the same manner as the industrial index, but the two were so closely related that both variables could not be used. The industrial index seemed to be slightly better and when the two were used together, the occupational index was not significant. The average rate of increase was 0.1 percentage points in recession years, 0.4 percentage points in nonrecession years, and 0.6 percentage points during the Korean War.

50

QUARTERLY JOURNAL OF ECONOMICS

elasticity for whites means that their incomes rise relative to thosc of Negroes during recessions and fall during expansions. During a recession the incomes of both whites and Negroes can fall due to a decline in the utilization of the labor force (see below), but a larger part of this decline is offset for whites than for Negroes. Unemployment insurance, the consistency of dividends, lower tax payments, and other factors contributing to the shift toward personal income in recessions are of much more cyclical importance to whites and lead to a relative decline in nonwhite income during recessions. The cyclical income shift is exacerbated by different coefficients for the utilization of the labor force. The results show that a fall in the level of employment has almost twice as much effect on Negro as on white incomes (the elasticities are 2.24 versus 1.21). Negro incomes fall almost twice as fast as white since Negro unemployment rises twice as fast as white. In expansions the relations are reversed and Negroes make large relative income gains. When these two cyclical factors are combined, nonwhite incomes can be falling in recessions while white incomes are still rising. This phenomenon occurred in both 1958 and 1961. Between 1947 and 1952 median Negro family incomes rose from 51 to 57 per cent of median white family incomes, but by 1958 all of the gains were lost and Negro incomes were back to 51 per cent. With high unemployment and fewer job opportunities, Negroes found jobs increasingly difficult to find and increases in their incomes fell behind those for whites. During every recession in the postwar period, the ratio of Negro to white incomes dropped sharply. With improving employment opportunities in 1964, Negro incomes rose from 53 per cent to A6per cent of white incomes, but they were still below 1952 levels. As the economy moves toward full employment, Negro incomes rise more than proportionately but the equalization does not continue once full employment is reached. At full employment there is no tendency for general economic progress to narrow the gap in relative terms and it widens in absolute terms.9
9. If the economy were able to stay at full employment for several years, the parameters of equation (2) might change and lead to higher relative incomes for Negroes, but as the equation stands it conflicts with Locke Anderson's evidence that the gap between white and nonwhite family incomes would continue to close after full employment is reached. In his equation family income is simply made a function of per capita personal income. Since the coefficient for nonwhites is much larger than the coefficient for whites, Anderson concludes that general economic growth would reduce the difference between white and nonwhite median family incomes. Anderson's conclusions result from his failing to distinguish between trend and cycle. When allowance is made for cyclical changes, the coefficients of secular growth do not differ for whites and nonwhites. If Anderson's coefficients were correct, the postwar growth in per capita income should have led to much higher relative

THE CAUSES OF POVERTY

NONWHITES

51

The coefficient for the percentage of families headed by a nonwhite is relatively small in the poverty model (see regression A). If this variable were the only factor in the equation which was responsible for Negro poverty, it should account for 4.3 percentage points of the total incidence of poverty; yet it actually accounts for only 1.0 percentage points. Since many of the other regression coefficients are also important determinants of Negro poverty, the coefficient for Negroes primarily represents the sheer handicap of being a Negro. Negroes are poor in part because they suffer from a lack of equal opportunities, but they are also poor since they have low education levels, are more apt to be outside of the labor force, and are crowded into part-time jobs and into the lowest occupations. Consequently, the coefficients of the other variables in the regressions include many of the long-run effects of discrimination that would have to be eliminated before Negroes achieve true equality. The hypothesis that the different variables in the poverty regression have a homogenous effect across race as well as across states could be tested by calculating the same poverty regression with data for Negroes and then for whites. If the regression coefficients were the same, the hypothesis would be substantiated. Lacking this test because the data are not available, an alternative is to insert national data for Negroes and whites into the poverty regression (regression A) and see how the actual and predicted incidences of poverty differ. The actual and predicted values could be similar due to offsetting errors among the different explanatory factors, but agreement would create a presumption that the effects were similar. Most of the explanatory variables have very different values for Negroes and whites (see Table IV), but the equation does a very good job of projecting the incidence of both Negro and white poverty. The projected and actual rates are identical for whites and differ by only 0.1 percentage points for Negroes (48.0 per cent versus 47.9 per cent), despite the fact that the incidence of Negro poverty is more than twice as high as that for whites. Both qualitatively and quantitatively the same variables seem to explain Negro and white poverty. Most of the difference between the incidence of white and Negro poverty is explained by the handicaps of being a Negro and low education levels (see Table IV).
incomes for nonwhites. It clearly has not done so. Locke Anderson, "Trickling Down," this Journal, LXXVIII (Nov. 1964), 552.

52

QUARTERLY JOURNAL OF ECONOMICS

TABLE IV WHITES AND NONWHITES FOR REGRESSION POVERTY (Percentages)
Contributions to Difference between White and Nonwhite Poverty

Data Whites

Nonwhites Nonwhite minus White

Percentage of families living on farms Percentage of families headed by nonwhite Percentage of families with no one in the labor force Percentage of family heads with 0-7 years of school completed Percentage of population 14 and above who worked 50-52 weeks per year Index of industrial structure Incidence of povertyProjected Actual
Source: U.S. Census, 1960. 1

7.5 0.0 10.5 19.2 35.5 98.8 18.6 18.6

6.6 100.0 12.9 48.5 28.9 97.8 48.0 47.9

-0.3 11.3 1.3 12.7 3.6 0.8 29.4

1 The index of industrial structure is calculated by weighting the index of each state by the number of white and nonwhite families living in that state.

PRINCIPAL

COMPONENT

ANALYSIS

The seven explanatory variables in the poverty regression explain the variation in the incidence of poverty by state, but they do not indicate that there are seven independent factors. Several of these variables are closely related to each other and are probably slightly different manifestations of more general explanatory variables. As might be expected, the percentage of families with no one in the labor force is closely related to the percentage of the population who worked full time, and both of these factors are closely related to the industrial structure of the state; the percentage of Negro family heads and heads with less than an eighth-grade education also move together. Ideally, the explanatory variables should all be completely independent of each other (orthogonal). By using principal component analysis it is possible to evaluate more clearly the independent dimensions which might be present among the various explanatory factors. A new variable is constructed (the first

THE CAUSES OF POVERTY

53

principal component) which is a linear combination of the explanatory variables. It explains the maximum amount of variation of the explanatory variables, and is completely independent of every other principal component. The same procedure is followed in calculating the second principal component after the influence of the first principal component has been removed. Among the seven variables in the poverty regression, the first four principal components account for 98 per cent of the total variation of all seven variables (See Table V where the percentage of the trace gives the percentage of the total variation explained by each component), and only the first three are really important. The first component accounts for 55 per cent of the total variation, the second for 28 per cent, and the third for 13 per cent. Basically, there would appear to be only three independent dimensions among the seven variables which appear in the poverty model (equation 1). The seven variables in this model could be reduced to three independent variables without sacrificing much information. An obvious further step is to see whether it is possible to identify the first four principal components. This objective can only be accomplished by correlating the original explanatory variables with their principal components, thereby finding out which variables move which principal components. These correlations are given in Table VI.
TABLE V
PERCENTAGE OF TRACE OF SEVEN PRINCIPAL COMPONENTS OF THE INDEPENDENT VARIABLES IN THE POVERTY EQUATION Principal Component Per Cent of Trace

1 2 3 4 5 6 7

54.68 28.29 13.39 2.26 .89 .50 .00

The first principal component can be identified with a general measure of the quality of the potential labor force, as shown by the negative correlations with the percentage of families headed by a Negro, the percentage of families with no one in the labor force, and the percentage of families with low education levels. All three of

54

QUARTERLY JOURNAL OF ECONOMICS TABLE VI

VARIABLES BETWEEN THE INDEPENDENT SIMPLECORRELATION ANDTHEIR FIRSTFOURPRINCIPAL COMPONENTS
Principal component Variables 1 2 3 4

Percentage of families living on farms Percentage of families headed by a nonwhite Percentage of families with no one in the labor force Percentage of family heads with 0-7 years of school completed Percentage of population 14 and above who worked 50-52 weeks per year Index of industrial structure Dummy for Alaska and Hawaii

.09 -.95 -.54 -.65 -.04 .04 -.54

.71 -.27 -.05 .70 -.70 -.60 -.51

.69 .14 .08 -.30 .38 -.73 .28

-.03 .01 .71 -.06 -.51 -.11 .22

these groups suffer from handicaps in competing for jobs and in taking advantage of economic progress. Both the second and third principal components are positively related to the percentage of families headed by farmers and negatively related to the industrial structure of a region. Since farming is part of the industrial structure of a region (the quality of the industrial structure goes up as the percentage of farmers goes down), both of these principal components can be identified with different aspects of the industrial structure of a region. The second principal component is also positively correlated with the percentage of families with low education levels and negatively correlated with the percentage of the population who work full time. These correlations are consistent with the previous identification of the second principal component. Education is related to the industrial structure in the long run, since occupational demands depend on which industries are located in the region, but education is also related because high wage industries do not locate in regions with low education levels. Over time, good employment opportunities, which provide a large number of full-time jobs, also increase labor force participation rates by attracting workers into the labor force. Thus, the correlation with education levels and full-time workers fits into the identification of the second principal component as a measure of the industrial structure. This identifica-

THE CAUSES OF POVERTY

55

tion is confirmed by ranking the states according to the coefficients of the second principal component. States with a poor industrial structure are at the bottom of the ranking and those with a good structure are at the top.' While the third principal component also represents the industrial structure, a ranking of states by the coefficients of this principal component reveals that it is primarily distinguishing between farm and nonfarm states. Both southern and northern nonfarm states have high rankings while the low rankings go to the farm states of the midwest and plains.2 The fourth principal component is much less important but it is related to the percentage of families with no one in the labor force, and it is negatively related to the percentage of the population who work full time. These correlations would seem to identify this component as a general measure of the labor force participation rates in an area. Principal component analysis can also be used to see whether any of the other variables considered for inclusion in the poverty equation add an independent dimension to the study of the causes of poverty. Principal components were computed with six additional variables,3 but the first four principal components still explained 98 per cent of the total variation of all thirteen possible variables and the correlations found between these thirteen variables and the first four principal components confirmed the original identification of the principal components. In sum the incidence of poverty seemingly can be explained by four general dimensions: (1) the quality of the human labor force as viewed through discriminatory eyes, (2) the quality of the industrial structure, (3) the amount of farming, and (4) a measure of labor force participation rates.
1. Ignoring the special cases of Hawaii, Alaska, and the District of Columbia, the top ten states are Utah, California, Washington, Delaware, Massachusetts, Connecticut, New Jersey, New York, Nevada, and Colorado, and the bottom ten states are South Dakota, Alabama, West Virginia, South Carolina, North Dakota, Tennessee, Arkansas, Mississippi, North Carolina, and Kentucky (proceeding from top to bottom). The indexes of the industrial structure are high for the first group and low for the second group. 2. The top ten states are West Virginia, Rhode Island, Pennsylvania, Massachusetts, Louisiana, New Jersey, Connecticut, New Hampshire, New York, and Maine, and the bottom ten states are Wyoming, Montana, Kansas, Oklahoma, Mississippi, Idaho, Iowa, Nebraska, North Dakota, and South Dakota. Farm population is low in the first group of states and high in the second. 3. The other six variables considered were (1) the percentage of families headed by a female, (2) the percentage of families with heads over age sixtyfour, (3) median school years completed, (4) percentage of families where head has less than eighth-grade education minus those with college education, (5) state unemployment rates, and (6) an occupational index.

56

QUARTERLY JOURNAL OF ECONOMICS

IMPLICATIONS

The variables used in the poverty regression are essentially the standard economic factors for explaining why particular workers have low marginal productivities. The variables can be interpreted as proxies for the amount of capital per worker (farmers and the industrial structure), the amount of human investment per worker (education and outside of the labor force), and the adequacy of government policies in translating potential productivity into actual productivity by fully utilizing the available human and capital resources (full-time work and outside of the labor force). The real productivity from these factors is distorted by subjective evaluations of productivity which interfere with market adjustments (discrimination against Negroes). Both the regression and principal component analysis reveal that the incidence of poverty can be explained with the use of very few general explanatory variables. These are basically the same variables that have been selected on the basis of a priori knowledge and descriptive classifications of families living in poverty. The numerical values of the parameters from the poverty equation are also comforting. They indicate that investment in human resources, equal rights for Negroes, and the push towards full employment can potentially make significant reductions in the number of families living in poverty. Potential reductions do not become actual reductions, however, unless deliberate policy actions can effect the explanatory variables and through them the incidence of poverty. Some of the factors are relatively easily affected. The unemployment rate was 5.6 per cent in 1960, but if a more expansionary fiscal policy had resulted in a 4.6 per cent rate of unemployment and if the parameters from the poverty model are accurate, the incidence of poverty would have been approximately 0.7 percentage points lower through the effects of a tighter labor market on the number of farmers, the number of families outside the labor force, the number of full-time workers, and the industrial structure. A 3.6 per cent unemployment rate would have reduced poverty by 1.4 percentage points. Other variables such as education and discrimination have large potential effects, but the cost of changing these variables has not yet been determined. Preschool classes which result in fewer school drop-outs will not increase education levels among family heads for fifteen years. Adult education programs will have to convince all individuals that they need to learn and can successfully do so.

THE CAUSES OF POVERTY

57

Both of these things may take time. This paper indicates that many of the expected results and benefits do occur, but it does not attempt to determine whether government programs can effect these variables or at what cost. Eliminating poverty by accelerating the rate of growth of family income will depend on the ability of the economy to accelerate the growth of productivity and to utilize fully the available labor and capital resources. This is true for both Negroes and whites. After full employment is reached, an accelerated growth of productivity will not eliminate the gap between white and Negro incomes (though it will make training programs and the elimination of discrimination easier and more effective), but this fact would seem to be no reason for sacrificing the large (and equal) gains that would accrue to both whites and Negroes from a faster rate of growth.
HARVARD UNIVERSITY