Project Report - US Unemployment - A Panel Data Analysis
Descrição do Produto
U.S. Unemployment - A Panel Data Analysis 040079 UK Econometric Methods for Panel Data (MA) Michel Geller(1006560 ), David Zenz (9271378 ) June 2015 Abstract The purpose of this project paper is to report the implementation of the presented statistical models and econometric methods for panel data. For this project, data of the U.S. Census Bureau was used to build a panel for all U.S. counties respective all U.S. states. Hausman-tests were conducted and Fixed-Effects Models as well as Random-Effects Models were estimated in order to achieve meaningful results. Ambiguous findings were obtained regarding the different effects within the models.
1
Introduction
The structure of this report is as follows. Section 2 introduces the model, followed by the data description in section 4. Model selection and estimation methods are dealt with in section 5. Empirical results and predictions are presented and discussed in section 6; the final section 7 summarizes our findings. At first this project was aimed to be a simple panel regression analysis on unemployment. After formatting the data set and conducting some sample cross-section regression analyses, everything looked as expected and we decided to proceed to the panel case. Unfortunately from there on we encountered a lot of problems with the specification of the effects. During the whole project we did not find many ”good looking” fixed effects estimations while lots of random effects estimations reflected the findings from the previous cross-section analysis. This said the project is driven by curiosity. With the empowering tool of panel regression methods we wanted to estimate a simple unemployment regression model. Once we found the vast USA Counties database, the data-mining was quite tricky but promised a large 1
set of possible covariates. During the winter semester course in macroeconomics we encountered the right-to-work law. This is a state based policy forbidding the practice of closed shops, i.e. that firms require their employees to join a union. As it is a tool of republican politics we wanted to test for its influence.
2
The Model
The chief difference in the panel models lies in the assumptions about the errors u of the regression.
2.1
Fixed-Effects Model
The basic fixed-effects panel model, where no individual effects are allowed, can be described as yit = α + β T Xit + uit
(1)
uit = µi + νit where i = 1, . . . , N and t = 1, . . . , T , with the identifying condition
PN
i=1 µi
= 0, the
individual effects µi are assumed as unobserved and the errors νit fulfill the standard assumptions E(ν) = 0, V ar(ν) = σν2 and νit independent of regressors Xit and individual effect µit . Further the individual effects µi are not assumed to be independent from the errors νit .
2.2
Random-Effects Model
The basic random-effects panel model (also known as variance components mode), where individual effects are allowed, can be described as yit = α + β T Xit + uit
(2)
uit = µi + νit with µi ∼ i.i.d.(0, σµ2 ) νit ∼ i.i.d.(0, σν2 ) 2
(3)
for i = 1, . . . , N and t = 1, . . . , T , where µ and ν are mutually independent.
3
The Right-to-Work Law
The right-to-work law represents a legislative act to restrict the power of unions. In 1935, the second new Deal policies implemented several 4 types of union relations to firms. One of them is the closed shop, a firm that forces its employees to be member of a certain union. Furthermore it introduced some other therms on shops such as union shops, agency shops and open shops with decreasing importance for unionism. In 1947 the business people stroke back with the so-called Taft-Hartley Act. Against president Truman’s veto they implemented several anti-union policies, one of them being the right-to-work law. The right-to-work law allows a state, if embraced, to ban closed shops. The drawback is that unions are financed by contributions of their members. So these right-to-work laws actually reduce the financial and therefore the political power of unions. Today there are 25 states which embraced this law, until 2012 there were 22 so our study uses only 22. Let’s look a bit at the advantages and disadvantages of right-to-work: On partisans side there is the argument, that people should be allowed to act freely and decide whether to join or not to join a union. Additionally the proponents argue that there is a significantly lower unemployment rate and also claim that the right-to-work states have a higher economic growth. The unions and their supporters argue, that this growth comes partly from the fact that these states have a lower income. Moreover the reduced unemployment rate comes at the price of lower income (and lower wages). On the side of faltering unions, it is important to notice that unions, whether people are members or not, achieve a lot of positive agreements, and therefore create a significant free-rider effect. Unionized workers are more often (and better) health insured, have higher wages and benefit from way better pension plans.
4
The Data
The data used in this project is extracted from a publicly accessible large database called USA Counties, which is provided by the U.S. Census Bureau. The USA Counties database consists of more than 6,600 data items for the United States 3
and is assembled from a broad band of different sources, featuring for example data from the yearly U.S. census from 1980 to 2010. The database contains collected data from the U.S. Census Bureau, the Bureau of Economic Analysis, the Bureau of Labor Statistics, the Department of Education, the Federal Bureau of Investigation, the Social Security Administration and many other official sources. Information in USA Counties is derived from the following general topics: Accommodation and Food Services, Age, Agriculture, Ancestry, Banking, Building Permits, County Business Patterns, Civilian Labor Force, Crime, Earnings, Education, Elections, Employment, Government, Health, Hispanic or Latino, Households, Housing, Income, Manufactures, Non-Employer Statistics, Population, Poverty, Race and Hispanic Origin, Retail Trade, Social Programs, Survey of Business Owners, Taxes, Veterans, Vital Statistics, Water Use, and Wholesale Trade. It should be noted that the USA Counties database provides data for states as well.
4.1
The Sample
The chosen sample consists of 13 Variables: id identification for state unemp unemployment rate bac share of people with bachelor degree hisc share of people with high-school degree agemedian median age incmedian median income black share of black people urban share of people living in cities with more than 700 persons per m2 male share of males right-to-work dummy if right-to-work-law is implemented in state An overview and an excerpt of the first and last 3 entries of the panel is available in Table 5 on page 10.
4
4.2
Problems of Endogeneity
Over the project there arose many problems with endogeneity mainly with the Hausman test. Including or excluding some variables occasionally changes the Hausman test score completely. At first the study was planed to be executed entirely on county data. But over the study we encountered a lot of problems especially with the fixed effects estimator. Even (or especially) in situations where the Hausman test rejected, fixed effects delivered strange and very counter intuitive results. For example including bachelors degree in a fixed effects model, returns a significant positive correlation between bachelors degree and unemployment, which is highly unrealistic.
5
Model Selection
To decide between the correct specification of the model corresponding to the data, Hausman Tests (Table 1 on page 5) have been conducted. P-values for the models imply that the Random-Effects Model is consistent (though not necessarily efficient) for the first model (with a p-value of 0.05039, just above the 5% level), whereas the p-values for the 2nd and 3rd clearly reject the H0 that a Random-Effects Model is consistent and efficient (with p-values of 0.001519 and 0.002775). Lagrange-Multiplier Tests in the style of Breusch-Pagan, as shown in Table 2 on page 5, and Honda, available in Table 3 on page 7, generally tend to the Random-Effects Model, since the H0 of individual effects is rejected, but also have a lower power compared to the Hausman test. χ2 df p
(1) 14.0449 7 0.05039
(2) 23.28368 7 0.001519
(3) 23.5026 8 0.002775
Table 1: Hausman Tests
χ2 df p
(1) 488174591 1 < 2.2e−16
(2) 446846822 1 < 2.2e−16
(3) 444786859 1 < 2.2e−16
Table 2: Breusch-Pagan LM Tests
5
6 6.1
Empirical Results States
As we encountered counter intuitive results during lots of regression, we decided to add Ordinary Least Squares (OLS) regressions to each of our panel regressions. Furthermore we printed each random effects output next its respective fixed effects output. In our first regression (Table 6 on page 11), where the Hausman test does not reject, we focus on random effects results. A solid high school education lowers unemployment statistically significant, while median income pretends to increase unemployment, which is highly questionable. Moreover the labor force population reduces unemployment, which makes sense as people only join the labor force if they anticipate to find work. A quantile regression plot could find different findings at higher quantiles though. The right-to-work dummy behaves as expected, explicitly the right-to-work law reduces unemployment at an estimate of -1.524, due to the fact that low-wage jobs are generated, which would not be generated with closed shops (with compulsory union-membership and therefore more power of the unions). Being black increases unemployment, which is an expectable result due to the prevailing structural (job-)discrimination in the western world. When changing the bachelors degree variable to the high school degree variable (regression number 2, available in Table 7 on page 11), Hausman advises fixed effects. Here we have a further highly questionable result, the bachelors degree has a positive sign at 0.038 and is highly significant. Now the urban population turns into promoting unemployment at 0.562. The right to work dummy still resides in the same area as in the first regression. In the third regression, which can be found in Table 8 on page 12, we decided to add both bachelors degree and high school degree to see whether there are problems of multicollinearity, and if not to look at their impacts. But first it should be noted, that due to a rejecting Hausman test we need to rely on the fixed effects estimator. If there where multicollinearity we should find different estimators for high school respective bachelors degree. As one can see in the third regression table there is actually problems of multicollinearity. The high school estimator behaves differently than before. Still the remainder regressors behave similarly to before.
6
6.2
Counties
In contrast to the first panel regression of the states, the first county regression rejects the Hausman test, so we have to go with Fixed-Effects Model (as shown in Table 9 on page 13). Fixed effects indicates that high school increases unemployment, which is again very questionable. Fortunately the median income does not play a role in increasing unemployment. Median age statistically significant decreases unemployment, which also supports our findings. Furthermore urban population still increases unemployment, which is a reasonable and encouraging outcome. The statistically significant negative effect of an increase in the labor force population is also in line with the rest of our findings. Not surprisingly black people still suffer from discrimination. The second panel, as shown in Table 10 on page 13, again exchanged bachelors degree against high school degree. With a rejecting Hausman test, we are advised to use fixed effects. In this regression a bachelors degree again promotes unemployment at an estimate of 0.036 and being highly significant. The remainder of the results stays the same. In the final Table 11 on page 14 again the Hausman test rejects in favor of the FixedEffects Model. Now both of our educational estimates are positive and statistically significant implying that high-school education does not necessarily lead to employment. But if we compare fixed effects estimates to those of random effects and the ones of the yearly Ordinary Least Squares, we find that fixed effects is definitely d’accord with common findings, whereas the remaining outcomes, like the effect of a bachelor degree, are according the prior results. The Random-Effects Model in turn supports the overall findings of this project, where a bachelor degree, high-school education, labor force participation and the right-to-work law decrease unemployment and being black and/or female increases the risk of unemployment.
norm p
(1) 21634 < 2.2e−16
(2) 21138.75 < 2.2e−16
(3) 21153.89 < 2.2e−16
Table 3: Honda LM Tests
7
7
Conclusions
Finally we can say that our regression analysis was definitely not satisfying, as we found implausible results. On should be very cautious when stumbling upon regressions which imply statistically significant results saying that high-school education (or a high-school degree) positively influences the risk of unemployment. Not only it is completely counterintuitive, it is also completely antithetic to common economic theory. The same reasoning goes for bachelors degrees, where different results were found. Furthermore the results from fixed effects, which where at 5 of 6 regression settings recommended through the Hausman test, where quite unsatisfying. Comparing to the OLS output we added to the respective regressions, but random effects looked much more promising. In fixed effects we occasionally encountered implausible results such as median income driving the unemployment rate, just to immediately disappear in the next regression. Overall the random effects regressions seem a lot more stable to their initial findings. On the bright side we found the right to work laws to work in the expected way. Throughout the regressions, which were either simple cross-section Ordinary Least Squares or Random-Effects Models we found consistent negative effect of -1.5 on unemployment. It should be noted that random effects were only advised in the first regression, but still comparing to the yearly OLS estimates it seems that the effect is consistently estimated.
8
8
Appendix
Statistic id unemployment median age median income bachelor degree black population one-person households right-to-work urban population high-school degree male population labor force work in state
N
Mean
St. Dev.
Min
Max
204 204 204 204 204 204 204 204 204 204 204 204 204
28,960.780 6.247 33.855 34,556.280 21.892 0.107 0.251 0.392 0.088 77.906 0.490 0.495 0.696
15,715.400 1.437 3.520 14,297.480 6.071 0.118 0.037 0.489 0.284 8.838 0.009 0.033 0.095
1,000 3.500 24.200 12,096 10.400 0.002 0.170 0 0 53.100 0.463 0.387 0.427
56,000 11.000 42.700 69,475 47.100 0.703 0.468 1 1 91.100 0.530 0.562 0.930
Table 4: Summary Statistics of Panel
9
10
6.700 6.100 4.500
7.500 9.700 6.200
1, 000 2, 000 4, 000 .. .
54, 000 55, 000 56, 000
unemp
id
41.300 38.500 36.800
29.300 26.100 29.200
agemedian
37, 356 51, 569 51, 990
13, 669 25, 414 16, 448
incmedian
0.034 0.063 0.008
0.256 0.034 0.028
black
0.286 0.282 0.278
0.203 0.199 0.208
1phh
0 0 1
1 0 1
right-to-work
Table 5: Excerpt of Panel
18.8 21.9 23.2
12.2 15.7 19.0
bac
0 0 0
0 0 0
urban
81.600 89 91.100
56.500 82.500 72.400
hisc
0.493 0.496 0.510
0.481 0.530 0.492
male
0.441 0.539 0.515
0.428 0.509 0.446
labfor
0.612 0.663 0.856
0.683 0.742 0.805 .. .
workdom
States regression cova.unemp panel linear
OLS
RE
FE
OLS80
OLS90
OLS00
(1)
(2)
(3)
(4)
(5)
(6)
high-school
−0.056∗∗∗ (0.007)
−0.020∗ (0.011)
−0.145∗∗∗ (0.016)
−0.255∗∗∗ (0.013)
−0.296∗∗∗ (0.011)
−0.095∗∗∗ (0.010)
median age
0.00002∗∗∗ (0.00000)
0.00000 (0.00000)
−0.0002∗∗∗ (0.00002)
−0.0001∗∗∗ (0.00001)
−0.00004∗∗∗ (0.00001)
−0.0001∗∗∗ (0.00001)
5.269∗∗∗ (0.271)
3.965∗∗∗ (1.094)
2.212∗∗∗ (0.425)
2.705∗∗∗ (0.331)
1.267∗∗∗ (0.315)
8.122∗∗∗ (0.371)
black population
1.245 (0.784)
1.248 (1.110)
−14.945∗∗∗ (1.732)
−8.950∗∗∗ (1.291)
7.085∗∗∗ (1.265)
−11.356∗∗∗ (1.221)
male population
0.00001 (0.00002)
0.0002 (0.0001)
0.0001∗∗∗ (0.00003)
0.0002∗∗∗ (0.00003)
0.00004 (0.00002)
0.0001∗∗ (0.00003)
−2.093 (1.449)
−2.102 (1.873)
6.451∗∗ (3.224)
−7.180∗∗∗ (2.485)
−14.682∗∗∗ (2.028)
−9.239∗∗∗ (2.187)
labor force population
−19.592∗∗∗ (0.580)
−9.940∗∗∗ (0.762)
−19.006∗∗∗ (1.430)
−33.020∗∗∗ (1.153)
−26.137∗∗∗ (1.096)
−8.420∗∗∗ (1.008)
right-to-work dummy
−1.524∗∗∗ (0.076)
−2.786∗∗∗ (0.114)
−1.496∗∗∗ (0.088)
−1.043∗∗∗ (0.082)
−1.547∗∗∗ (0.101)
Constant
18.040∗∗∗ (0.827)
22.915∗∗∗ (1.830)
38.090∗∗∗ (1.470)
36.422∗∗∗ (1.284)
26.056∗∗∗ (1.340)
median income
urban population
∗
Note:
OLS10
p
Lihat lebih banyak...
Comentários