Abstract

Several studies have described the correlation between welfare dependency and factors such as welfare conditionality, gender, and high school or college graduation rates. Using Annual Social and Economics Supplement Data (ASEC) from 2009 through 2019, downloaded from sources such as IPUMS CPS, this paper crafts an OLS regression model to find the relationship that years of completed education have on welfare dependency status. This paper concludes that there is a negative correlation between higher education levels and lower participation in the welfare system, with the completion of one additional year of schooling suggesting a decrease in the probability of needing welfare by 0.1%. While this correlation is small, it is still statistically significant in the linear probability model due to a large sample size (n = 145,431). After adding other explanatory variables, such as measures for race, biological sex, and employment status to control for endogeneity, further regressions confirm that there is still a statistically significant negative relationship between education and welfare dependency. These results suggest that policymakers should focus on educational subsidies over welfare subsidies to increase social mobility. 

I. Introduction

Education is often referred to as an essential mechanism in promoting social mobility (Haveman, 2006). However, the rising costs of education in America have forced many individuals to require more income to pay off student loans. As a result, families who are enrolled in welfare programs are spending a larger portion of their income on student debt, correlated with an increased reliance on such welfare programs and a positive feedback loop that makes it more difficult to climb out of welfare dependency (Johnson, 2019). In addition, most welfare programs have substantial requirements that, rather than helping recipients to get out of poverty, restrain recipients from escaping the welfare system (Rupp et al, 2020). This, and other societal pressures, have forced lots of students to put a pause on their education and work at low-skilled jobs with minimal pay, keeping them reliant on welfare programs (Johnson, 2019). This vicious cycle will only cause more people to remain trapped within welfare programs, preventing them from escaping poverty and improving their livelihoods. 

Previous studies have shown that education levels are correlated with the probability of a welfare recipient returning to welfare in the future (London, 2008). Other studies have also shown how changes in the welfare system have improved welfare recipients' education qualifications and subsequently their employment opportunities (Hernaes, 2017). London’s (2008) study focused on how attaining a higher educational degree allows welfare recipients to improve their employment opportunities, reduce their welfare dependency, and reduce their overall family poverty levels by 63%. Meanwhile, Hernaes et.al (2017) found that more conditionality in welfare programs helped Norwegian teenagers from welfare-recipient families reduce their reliance on welfare programs; and lower the country’s high school dropout rate by 21%. In addition, Pacheo & Maloney (2003) found that intergenerational welfare participation differs between genders due to family characteristics such as household size and parents’ welfare dependency. As a result, young females tend to have lower educational attainment and are nearly two times as probable of relying on welfare in the future when compared to their male counterparts (Pacheo et al, 2003). 

Based on the insights offered by the studies above, this paper aims to contribute to this field by investigating the hypothesis that years of schooling completed reducing the probability of receiving welfare in the future. Factors of endogeneity will also be analyzed through the implementation of explanatory variables such as race (Courtney, 1996), sex (Bakas, 2014), number of children (Arulampalam, 2000), marital status (Hoffman, 1997), hours worked (Bick et al, 2018) and employment status (Arranz, 2004) into the regression model. These variables were chosen due to past publications finding possible links between this psychographics and demographics to welfare benefits. Preliminary hypotheses predict that there will be a negative relationship between the education level attained and the probability that an individual will receive future welfare. Using simple and multi-linear OLS regression analysis and the IPUMS-CPS annual data from 2009-2019, it has been observed that individuals with more years of schooling completed are less likely to be on welfare in the future. Data was chosen from this period because the American economy was beginning to recover from the 2008 Financial Crisis during this time. This allows us to observe correlations between education levels attained by individuals and whether they ended up in welfare programs more clearly. 

This paper will be presented as follows: Section II will cover previous research on how welfare conditionality, gender, and education levels affect welfare dependency. Section III will present information on how this data was obtained and explain the data-cleaning process along with the types of variables used throughout this paper. This section ends with explanations of how the data is verified through the four OLS assumptions. Section IV will cover econometric methodology which includes alternate functional forms explored and additional X-Variables tested through multiple regression along with methodologies we’ve used to control the endogeneity of independent and explanatory variables to ensure the fairness of the regression model. Section V highlights the results, the sample regression line, and statistically significant information regarding the regression analysis. Section VI contains the paper's conclusions where the results are evaluated and put into context within the field. The paper concludes with section VII, the appendix, where all tables, figures, diagrams, and supporting calculations are represented for reference. 

II. Literature Review

The literature works that are presented here serve as important foundations in the field and provide extensive insight into the relationship between welfare dependency and education levels, along with how other variables might affect this relationship. The study conducted by Hernaes et al. (2017) found that the strict welfare conditionality, linking welfare to certain characteristics or traits in Norwegian welfare programs, has reduced welfare dependency while increasing the high school graduation rate among Norwegian welfare recipients. In the process, they used a logarithmic regression model (LRM) and regressed a dependent dummy variable that identifies welfare recipients who are 21 years old onto an independent variable that consists of family characteristics such as parent’s education background and cumulative income, to control for endogeneity. The study resembles this approach because the dependent variable that they’ve used is also a dummy variable that indicates welfare recipients. In addition, the study used other explanatory variables, in particular the recipient’s parental background, to control for the endogeneity of those variables on the probability of returning to welfare. However, Hernaes et al. (2017) emphasizes how family background affects teens’ probability of returning to welfare in the future through explanatory variables that focus on family characteristics. Whereas this study focuses more on how other individual characteristics such as education level, labor condition, and family status of the welfare recipient have affected the welfare recipient’s probability of returning to welfare programs in the future. 

Notably, a previous study indicated that there is a correlation between welfare recipients who have obtained a higher education degree with reducing their reliance on welfare programs, but only if they receive additional financial aid to support their college expenses. London (2006) uses data such as college attendance, college graduation rate, and personal characteristics, such as extraversion and race demographics, to predict the welfare recipient’s three outcomes: employment, return to aid, and poverty status. By controlling influencing factors that change over time – such as the rate of college enrollment – and making sure all omitted variables, – such as familial culture and personal motivation – are factored into the result, the study employs instrumental variable econometric models to calculate predictions. The study found that “college attendance, more than graduation, is an important predictor of future employment. At the same time, college graduation better predicts the probability of returning to aid or being poor within five years of leaving welfare” (London, 2006, p. 491). Specifically, the study quoted “college graduation rather than enrollment without graduation has an effect on recidivism, and only in the five-year interval” (London, 2006, p. 489). Their findings support this hypothesis that the education level a welfare recipient attains is crucial to the probability of returning to welfare in the future. Despite the similarities in the use of variables to investigate the issue, predicting the probability of return to welfare using college graduation and attendance is only a part of this study’s objectives. The study also conducts an investigation into how college graduation and attendance affect employment opportunities and family poverty levels. 

Another earlier study showed that genders might have different levels of welfare participation and education attainment. Pacheco & Maloney (2003) learned that females “have an estimated intergenerational correlation coefficient that is more than double that for males.” (Pacheco & Maloney, 2003, p. 371). The study uses simple regression models and inputs such as the number of years in formal education completed by age 21, family background characteristics (parent’s education qualifications and the number of children in the household), and the proportion of years where parents obtain welfare benefits to produce their findings. In addition, Pacheco & Maloney (2003) found that female welfare recipients whose families have a history of welfare dependency tend to remain in welfare programs. The study uses the same regression model to offer insight into how familial and cultural forces affect male and female probabilities of returning to welfare in the future. Nevertheless, Pacheco & Maloney (2003) offered insight into how gender might have altered the relationship between education levels and probability in return to welfare. 

III. Descriptive Statistics

All of the raw data was downloaded directly from the CPS portion of the IPUMS website, which is a reputable federal source for time series and cross-sectional data. Annual Social and Economic Supplement Data (ASEC) from 2009 to 2019 was downloaded. These years were selected to obtain the most up-to-date data while also analyzing enough observations to create the best regression analysis possible. Twenty-one variables were analyzed within these years, the most important of which were EDUC and INCWELFR, the two variables that were altered and then used for the regression analysis. These variables were raw and included nearly 150,000 observations over the 11 years. The data was meticulously cleaned before running any regressions to test the hypothesis. 

The first variable cleaned was EDUC. The raw EDUC variable could hold any coded value from 1 to 125. These coded values did not reflect the true years of schooling any individual had, so a new variable was created: EDUC_REV, to accurately reflect the true years of schooling each individual has completed. The values for this new variable were generated using the observations for the EDUC variable alongside the specific numeric code utilized by CPS. For example, an individual who has obtained a high school diploma through 12 years of completed education would receive a value of EDUC=73 within the CPS data set. The data was cleaned so this specific value would now be EDUC_REV=12. This cleaning procedure was used for all possible levels of education within the data set. Individuals who were too young to receive any education at all were also removed from the data set (they were identified through EDUC=1 in the original data set). 

The focus then shifted toward the INCWELFR variable from CPS. This variable measures the dollar value of the income an individual receives from any source of government welfare benefits. In this study, the focus is on the effect that education has on the reception of welfare at all, not the amount of welfare that was received. This means the analysis is valid if an individual receives any form of welfare payments, and not focusing on the actual dollar value of said payments. So, for this reason, another new variable was created: WELFARE. This variable is a dummy variable that gets its values from the information in the INCWLFR variable. If the individual receives no form of welfare they will be assigned INCWLFR=0 in the data set. This same individual would be assigned a value of zero for the newly created variable (WELFARE=0 when INCWLFR=0). However, if an individual receives welfare in any form, regardless of the amount, they will be assigned a value of one for the new variable (WELFARE=1 when INCWLFR>0). Any individual who was not eligible to receive welfare in any form was denoted by INCWELFR=999999. These observations, many of which were individuals under 18, were removed from the data set to generate a less skewed, and more accurate, sample. 

Additional variables were also analyzed for the multiple regression analysis. These variables tested the effects of not only education, but also employment status, income, hours worked, marital status, gender, and number of children on the reception of welfare. These variables were used to try and control for endogeneity within the model and are further described in Table 1 of the appendix. 

Before the new variables could be put through a proper regression analysis, the four assumptions of an Ordinary Least Squares Regression Line had to be tested. If all of these assumptions hold true then the estimators of b1 and b2 would be BLUE (Best Linear Unbiased Estimators) and all of the calculations done through STATA would be completely accurate. 

The first OLS assumption is that the expected error within a sample will be zero. This is noted as E[WELFARE_RES/EDUC_REV]=0 and this does hold true in this sample. The 95% confidence interval for WELFARE_RES does include zero so it is likely that the expected value of the error is zero and therefore the first OLS assumption is met. 

The second OLS assumption is that the data is homoscedastic. This is noted as Var(WELFARE_RES/EDUC_REV)=Sigma^2. However, since the dependent variable is a dummy variable, this regression takes the form of a linear probability model (LPM). By definition, every linear probability model has heteroscedastic data. Therefore, the second OLS assumption is not met. 

The third OLS assumption is that the data is free of clustering. This is noted as 

Cov(WELFARE_RES_i,WELFARE_RES_j)=0, meaning that the value of WELFARE for one value does not directly influence the value of any other observation within the data set. This influence usually occurs when two observations are within the same geographical unit. While there is no way to test if any observations are within the same geographical unit (such as the same household) due to confidentiality, the sample size is large enough and pulls from each region almost equally, so it would be extremely unlikely for any two observations to come from the same household. Therefore, for the sake of the regression, the third OLS assumption will be met. 

The fourth and final OLS assumption is that Y is normally distributed. This was tested by creating a histogram for WELFARE and seeing if it roughly resembled a bell curve. When this was done, it was obvious that the data was not normal. This is apparent through a multitude of factors but is most clearly shown by the high skewness, a value of over 16. Therefore, it was concluded that welfare was not normally distributed. However, since the sample size consists of 145,431 observations, the central limit theorem (CLT) is met. So, while the fourth OLS assumption failed to be met for this particular regression, it will not have a significant impact on the regression since the sampling distribution for WELFARE will still be normally distributed. In conclusion, the regression met two of the four OLS assumptions. Therefore, while the regression analysis will not be BLUE, it will still be significant since it is free of serious sampling errors.

 IV. Econometric Methodology 

While this paper mainly focuses on the linear probability model and the effect that education has on welfare dependency, other functional forms that could better fit the regression analysis were also considered to develop a more thorough analysis. This was done through the experimentation of the functional forms that the independent variable took. While the previous section discussed the linear form of EDUC_REV, exponential and logarithmic forms of this variable were also considered. The independent variable was only altered since the dependent variable is a dummy variable. Altering the value of the variable will not generate any different results since its domain is limited to {0,1}. Other explanatory variables, and the results they produced, are summarized in Table 5 of the appendix. 

While all of the functional forms tested would have produced statistically significant interpretations that support the hypothesis, although their interpretations would have been different, the original regression was still the most accurate for this particular data set. Other functional forms included EDUC_REV in quadratic, cubic, and log forms. These functional forms are used to emphasize the effects of EDUC_REV in order to match the data points. The original is the most accurate because it has the highest R-Squared value, a measure of how well the data points fit the linear regression line. These R-Squared values can be found in Table 5 but the linear model has the highest value of .0014. Since the linear regression between WELFARE and EDUC_REV has the most accurate regression line relative to the data set, this regression model was the basis from which all conclusions were drawn. 

Interaction terms were also analyzed by creating the term EDUC_UNEMP which was EDUC_REV multiplied by UNEMPLOYED. By using this interaction term, the possible effect of EDUCATION on WELFARE varying with UNEMPLOYED can be studied. The regression showed that when UNEMPLOYED is 0, the likelihood of WELFARE is constant plus b2. When UNEMPLOYED is 1 then the likelihood of being on WELFARE increases. This means that individuals who are unemployed are more likely to be receiving benefits from welfare. The motive that drives this is individuals who are unemployed do not receive any form of compensation or income outside of their welfare payments. 

 Slope and Intercept dummy variables are additional variables added to this study. In this situation, the intercept dummy variable is UNEMPLOYED. The presence of UNEMPLOYED is represented with a 1 and causes an increase in the intercept, which translates to an increase in the probability of welfare. When describing this relationship on a graph there are two parallel lines and the difference between them is caused by the slope dummy variable. Both lines have the same slope and the probability gap of being on WELFARE remains the same at all levels. This is not the main difference between someone who is unemployed and someone who is employed. This supports the claims made through interaction term analysis in the previous paragraph.

However, while the simple regression analysis supports the hypothesis, there could be other confounding variables that underlay such correlation seen between WELFARE and EDUC_REV. If these possible confounding variables are correlated with both WELFARE (controlling for EDUC_REV) and EDUC_REV, then it could make EDUC_REV an endogenous variable, indicating that EDUC_REV does not necessarily cause the decrease in the probability of an individual on welfare. 

To test this claim, a multiple regression analysis was run, including both EDUC_REV and a variety of other possibly confounding variables, for their possible effects on WELFARE. The results showed that the three variables with the largest effect on WELFARE were BLACK, MALE, and UNEMPLOYED. These are variables created within the data set describing an individual's race, gender, and employment status, respectively. All of these are strong contenders for possible confounding variables and the true reason the regression effect on welfare was observed, and therefore put EDUC_REV at risk of being an endogenous variable (Courtney, 1996; Bakas, 2014; Arranz, 2004). A full list of the additional X-Variables tested along with the multiple regression output can be found in Table 10. 

That being said, this is not enough evidence to conclude that education levels are definitely an endogenous variable when describing the probability of receiving welfare. These possible confounding variables could be further analyzed if a more in-depth regression analysis was performed in future studies. 

V. Results

After the data had been completely cleaned and verified for OLS assumptions, the regression of EDUC_REV on WELFARE was run. This regression showed the noncausal effect that years of completed education have on the probability of receiving welfare. If the hypothesis holds true, the Least-Squares Regression Line should have a negative slope, denoting that the more years of education an individual completes, the less likely it is that the individual receives welfare. 

The output for the regression analysis, as well as the full, scatter plot showing the Least Squares Regression Line for EDUC_REV against WELFARE, can be seen in Table 8 and Table 9 of the appendix. However, these figures can be summarized by the equation for the sample regression:

WELFARE_hat = b1 + b2 EDUC_REV t-statistic = -14.27

WELFARE_hat = .015 - .001 EDUC_REV n = 145,431

(SE) (8.08e-4)    (5.68e-5) p-value = 0 *** 

The most important value within the sample regression line for the hypothesis is -.001, or the slope of the regression line denoted as b2. Since b2 is a negative value, there is a negative correlation between the number of years of completed schooling (EDUC_REV) and the reception of welfare (WELFARE). While this value seems too small to have any real effect, it is still statistically significant. This is because the 99% confidence interval for b2 does not include zero because the standard deviation is extremely close to zero based on the large sample size. A hypothesis test at the critical level of .01 was also run to see if the value generated for b2 could be equal to zero. This test gave a critical value for b2 of -14.27 and a probability of B2 being equal to zero of zero. These results lead to the conclusion that it is statistically significant that as EDUC_REV increases, WELFARE decreases within the regression. In conclusion, while increases in education could have a small effect on the probability of relying on welfare, it is still a statistically significant effect. However, this does not prove that increases in education will decrease the probability of relying on welfare since ceteris-paribus does not hold true for this collected data set and a causal relationship is not established. 

This regression analysis supports the hypothesis that as an individual's education increases, the probability that said individual will rely on welfare as a source of income decreases (since b2 is a statistically significant negative number). By applying these findings, it was determined that as an individual’s years of completed schooling (EDUC_REV) increases by 1 year, the probability that the individual will receive welfare (WELFARE as a dummy variable) decreases by .001 or 0.1%. This is because the slope of the linear regression model, with a dependent dummy variable, is -.001 and the functional form analyzed is a linear probability model. While this relation is not inherently strong, and years of completed schooling do not have a large impact on the probability of receiving welfare, it is still statistically significant. 

Within the regression, b1 is also statistically significant. The value of b1 in this sample regression line is .015, or an applied .15%. By applying this value to the context of the study, it was found that the probability of an individual receiving welfare given that they have completed zero years of schooling is .15%. This number is positive so it is technically feasible and within the domain of the study. However, it is extremely unlikely that an individual has received zero years of schooling and is also eligible to receive welfare (Stephens, 2014). For this reason, the value of b1 was not a focus within these results.

VI. Conclusion

As stated above, this study shows a minor, yet the statistically significant, effect of EDUC_REV on WELFARE. These results indicate that there is evidence to support a possible relationship between higher levels of completed education and lower chances of an individual receiving welfare in the future. The thought process behind this regression is that individuals with higher education are more likely to land better jobs and therefore make more money, thus decreasing their need for welfare. 

While focusing on the simple regression model for the majority of the paper, important results when controlling for endogeneity through a multiple regression model were also found. This multiple regression analysis was performed while controlling for multicollinearity. Since none of these variables share a strong correlation (r > .8) with each other, it is okay to run a regression model with all of these X-Variables. The full correlation results can be seen in Table 11 of the appendix. AIC, BIC/SC, and R_Squared were also analyzed and are summarized in Table 12. Since the multiple regression model has more X-Variables, it has a larger potential to explain any variation in Y and is likely to be a better fit for the data. 

Even with the introduction of these additional X-Variables, the initial variable tested in the multiple regression analysis, EDUC_REV, was still statistically significant, as seen in Table 10. Thus, even with controls for endogeneity, there is still a statistically significant negative correlation between the highest level of completed education and the probability of receiving welfare, only strengthening this paper’s claims. 

In relation to previous studies in part II, this study aligns with London’s (2006) conclusion that welfare recipients who have received a higher education degree have a lower probability of receiving welfare in the future, with the assumption that both genders fit into the conclusion. However, to what extent education attainment is beneficial to both genders and race remains questionable since the data lacked suitable information to investigate how omitted variables might have affected the relationship between the education level attained and the probability of receiving welfare. This paper has also failed to reproduce the findings that Pacheco & Maloney (2003) found. This paper did not control the age and time of welfare received by the recipient, whereas Pacheco & Maloney (2003) did. In addition, Pacheco & Maloney (2003) factors in the background of the welfare recipient’s parents, such as their income received from welfare, educational background, and race. 

This study, on the other hand, did not factor family characteristics into the regression model. This paper also failed to reproduce the results that Hernaes et. al (2017) produced because the nature of the data is different from Hernaes et. al (2017). First, Hernaes et. al’s (2017) dataset had the location of each welfare recipient’s municipality. The location variable allows Hernaes et. al (2017) to determine whether the welfare recipient was in a municipality that has stricter welfare policies or not. Second, Hernaes et. al (2017) was able to capture each municipality’s level of conditionality through survey responses collected in a report by a research institute. These are some of the features that the data, unfortunately, do not possess. 

This paper supports the theory that there is a correlation between the highest level of education completed and the probability of receiving welfare. Thus, more educated individuals are less likely to be dependent on welfare. In a broader context, policymakers could use this information to find more effective means for increasing social mobility, rather than investing heavily in welfare payments. Since there is possibly an inverse relationship between education and welfare, the federal government could create a new program to subsidize education rather than simply making payments to disadvantaged citizens. This would provide an economic incentive for individuals who were previously on welfare to attend school, making the entire nation more educated and more productively efficient as a result (Brown et al, 1991). However, while this paper could be used from a policy perspective, there are some drawbacks. The relationship between education and the probability of welfare is not proven to be causal after this analysis. This is because the ceteris-paribus condition does not hold true throughout the data and regression. In addition, this dataset has a limited scope regarding population characteristics. The dataset indicates the highest education level attained by the individual but does not indicate when they achieved that education. For example, some individuals might have dropped out of high school during their youth and returned to complete their high school degree after a long period of time. If that information is also provided in the dataset, that would open new frontiers on how education-level attainment influences the probability of receiving welfare. Before any change is enacted, especially on a governmental level, first proving a causal relationship would be recommended. This paper merely lays the framework for possible studies regarding welfare analysis in the future. 

This paper did support the hypothesis that as education levels rise, the probability that an individual becomes dependent on welfare decreases. Through the regression analysis, it was determined that there is a small, yet statistically significant, difference that education has on the probability of receiving welfare in the future. This trend could be utilized by policymakers to stimulate education as a means of reducing welfare dependency, creating a population that is not only less dependent on welfare payments, but more educated, and more productive as a result. 

Note: see "Full Editions," Volume IV Issue I for appendix.

VIII. References 

Arranz, Jose Ma, and Muro, Juan. "Recurrent Unemployment, Welfare Benefits and Heterogeneity." International Review of Applied Economics. 18, no. 4 (2004): 423-41.

Arulampalam, W. "Unemployment Persistence." Oxford Economic Papers. 52, no. 1 (2000): 24-50.

Bakas, Dimitrios, and Papapetrou, Evangelia. "Unemployment by Gender: Evidence from EU Countries." International Advances in Economic Research. 20, no. 1 (2014): 103-11.

Bick, Alexander, Fuchs-Schündeln, Nicola, and Lagakos, David. "How Do Hours Worked Vary with Income? Cross-Country Evidence and Implications." The American Economic Review. 108, no. 1 (2018): 170-99.

Brown, Phillip, and Lauder, Hugh. "Education, Economy and Social Change." International Studies in Sociology of Education. 1, no. 1-2 (1991): 3-23.

Cliff, Aiden, Rupp, Matthew, Lieng, Owen. “A Study on the Relationship Between Education and Probability to Receive Welfare Assistance.” Boston University (2020): 204

Courtney, ME. "Race and Child Welfare Services: Past Research and Future Directions." Child Welfare. 75 (1996): 99.

Gooden, S. (2000). Race and Welfare. Journal Of Poverty, 4(3), 21-24. https://doi.org/10.1300/J134v04n03_02 

Haveman, Robert, and Timothy Smeeding. "The Role of Higher Education in Social Mobility." The Future of Children 16, no. 2 (2006): 125-50. Accessed April 28, 2021. http://www.jstor.org/stable/3844794.

Hernæs, Ø., Markussen, S., & Røed, K. (2017). Can welfare conditionality combat high school dropout. Labour Economics, 48, 144-156. https://doi.org/10.1016/j.labeco.2017. 08.003 

Hoffman, Saul. "Marital Instability and the Economic Status of Women." Demography 14, no. 1 (1977): 67-76.

Johnson, D. (2019). What Will It Take to Solve the Student Loan Crisis. Harvard Business Review. Retrieved 29 April 2020, from https://hbr.org/2019/09/what-will-it-take-to-solve-the-student-loan-crisis. 

Kim, Hwanjoon. "Anti‐Poverty Effectiveness of Taxes and Income Transfers in Welfare States." International Social Security Review. 53, no. 4 (2000): 105-29.

London, R. (2005). Welfare Recipients' College Attendance and Consequences for Time-Limited Aid. Social Science Quarterly, 86, 1104-1122. https://doi.org/10.1111/j.0038-4941.2005.00338. 

London, R. (2006). The Role of Postsecondary Education in Welfare Recipients' Paths to Self-Sufficiency. The Journal Of Higher Education, 77(3), 472-496. Retrieved 28 April 2020, from https://www.jstor.org/stable/3838698 

Pacheco, G., & Maloney, T. (2003). Are the Determinants of Intergenerational Welfare Dependency Gender-specific. Australian Journal Of Labour Economics, 6(3), 371-382. Retrieved 28 April 2020, from https://www.researchgate.net/ publication/46557521_Are_the_Determinants_of_Intergeneration al_Welfare_Dependency_Gender-specific 

Stephens, Melvin, and Yang, Dou-Yan. "Compulsory Education and the Benefits of Schooling." The American Economic Review. 104, no. 6 (2014): 1777-792.