My Weblog: umraniye elektrikci uskuadar elektrikci usta elektrikci sisli elektrikci

Page 2

Conceptual Priming vs Gender Bias in the “Surgeon Riddle”

Ece Hakim ‘21


The experiment (total N = 68) was conducted to investigate the impact of conceptual priming in the well-known “surgeon riddle.”  The surgeon riddle is as follows: “Father and son are driving a car. They get into a car accident, the dad dies, son gets rushed into hospital. The surgeon says, ‘I cannot operate, this is my son.’ Who is the surgeon?”

The correct answer is “the mother;” however, many participants fail to answer the question correctly. It was argued that the failure could be attributed to the priming effect rather than implicit gender stereotypes. In order to test this hypothesis, three conditions were created: a control group, a non-priming condition and a stereotype condition. For stereotype condition, the riddle was reformulated as “A mother and her daughter are driving a car…” to observe the outcomes when the correct answer, the father, is consistent with the stereotypes. For non-priming condition, the riddle was reformulated as “A father and his daughter are driving a car…” expecting the presentation of two sexes to override the priming effect observed in the original riddle. It was predicted that the change in the phrasing once the conceptual priming is overridden, the number of correct answers in the non-priming condition will be same as that of in stereotyping condition. Our results fell between these two extremes: Performance in the non-priming condition (father–daughter) improved (27%) but was far from the ceiling (89%).


Conceptual Priming vs Gender Bias in the “Surgeon Riddle”



A social stereotype is a mental association, which is not necessarily reflective of statistical reality, between a social group or social category and a trait. (Greenwald & Krieger, 2006) It is well-established by research that there are certain associations regarding the personal traits of men and women shared by the members of some social groups. (Ashmore & Del Boca, 1979; Cuddy et al., 2015; Eagly & Mladinic, 1989; Jackson & Cash, 1985) One of the most prevalent associations is the association between traditional domestic roles and collectivistic traits (i.e. nurturance, social sensitivity etc.) and women, as explained by social role theory. (Eagly, 1987)

Despite the increasing number of women joining in the workforce, social role theory seems to be pervasive. Particularly, the gender gap in areas such as math and science is attributed to the belief that women possess fewer of the traits that are associated with scientists such as analytical thinking and independence. (Carli, Alawa, Lee, Zhao, & Kim, 2016) Although this is a relatively recent finding in psychology, the belief itself has been deeply ingrained for a much longer time. A piece of evidence of this is the well-known surgeon riddle –which has become well-known because of a large number of people who get it wrong:


“A father and his son are driving a car. They get into a car accident. The dad dies, and the son gets rushed into emergency. The surgeon says, ‘I cannot operate, this is my son.’ Who is the surgeon?”

The mother of the son.


The New York Times columnist Stephanie Coontz recalls being stumped upon hearing the riddle in 1962. (Coontz, 2013) However, the origin of the riddle is unknown. In academic papers, it could be traced back to only 1985 –to Anthony J. Sanford’s book titled “Cognition and Cognitive Psychology.” In this book, Sanford uses the riddle to explain the phenomenon of presupposition. He attributes the failure to find the correct answer to the presupposition that surgeons are male. (Sanford, 1985) Likewise, later research –the ones that studied the riddle as well as the ones that referenced to it– explained the phenomenon with the implicit gender stereotype (Kollmayer, Pfaffel, Schober, & Brandt, 2018; Mineshima, 2008; Wapman, 2014) Although this explanation seems to make intuitive sense, which explains the limited research regarding this riddle, I argue that it is a quick conclusion to jump to. In this paper, I will explore an alternative explanation for the phenomenon of surgeon riddle, which is the priming effect.

A simple definition of priming effect would be the presentation of information designed to activate knowledge structures, such as trait concepts and stereotypes and hence make them more accessible. (Bargh, Chen, & Burrows, 1996; Gilovich, 2016) Research has shown that priming of a social category (i.e. elderly, women, African-American) impact social perception, and therefore a shift in judgement toward the primed category, assimilation effects of judgement, occurs. (Bargh et al., 1996; Herr, 1986, 1986; Herr, Sherman, & Fazio, 1983) I argue that the presentation of two male figures (father and son) and zero female figures in the riddle primes the social category of males. This leads to a shift in the judgement of participants toward males, and therefore makes it harder for them to think of a female surgeon and find the correct answer. So, if this priming effect were overridden, participants would not fail to find the correct answer.

In order to test this hypothesis, two different conditions were created. The first condition (non-priming condition) should have been presented with a question that would not create a priming effect. In order to achieve this, the question was reformulated in the following form:


“A father and his daughter are driving a car. They get into a car accident. The dad dies, and the daughter gets rushed into emergency. The surgeon says, ‘I cannot operate, this is my daughter.’ Who is the surgeon?”

The mother of the daughter.


I argue that the representation of both sexes, a male (father) and a female (daughter), would override the priming effect. In the second condition (stereotype condition), the question was reformulated in the following form:


“A mother and her son are driving a car. They get into a car accident. The mother dies, and the son gets rushed into emergency. The surgeon says, ‘I cannot operate, this is my son.’ Who is the surgeon?”

The father of the son.


This question also overrides the priming effect by presenting both sexes to the participant. However, in this condition, the expected answer from the participants was “the father.” So, if participants could find the answer to the question in this condition but not in the priming condition question, then the failure of the participants could be attributed to the implicit gender stereotypes. There was also be a control group which was given the original question. The success of the participants will be evaluated based on their accuracy (“Did they give the correct answer?) and on their response latency (“How many seconds did it take them to find the answer?”).




In total 68 (34 female, 34 male) participated in the study. The age of the participants ranges from 18 to 71, while the average is 29.9. The participants were randomly assigned to one of the three sample groups:

Control Group:

“A father and his son are driving a car. They get into a car accident. The dad dies, and the son gets rushed into emergency. The surgeon who comes into the emergency room and sees the son says, ‘I cannot operate, this is my son.’ Who is the surgeon?”

The mother of the surgeon.


Non-priming Condition:

“Father and daughter are driving a car. They got into a car accident, the father dies, daughter gets rushed into hospital. The surgeon says, ‘I cannot operate, this is my daughter.’ Who is the surgeon?

The mother of the surgeon.


Stereotype Condition:

“Mother and son are driving a car. They got into a car accident, the mother dies, son gets rushed into hospital. The surgeon says, ‘I cannot operate, this is my son.’ Who is the surgeon?

The father of the surgeon.


Each question was printed on a two-sided paper. On the front page, the participants were asked to provide information about their age, gender and education level. After they filled out the form, they were informed about the procedure. They were told to turn the page, read the question and write down their answer. The researcher started a timer to measure the response time. The timer was started immediately after the participants turned the page and stopped immediately before the participants started writing down their answers. Since the thought process starts while the participant is reading the question and ends right before s/he starts writing it, this specific method of time measurement captured the exact time spent finding the answer. If the participant still could not find the answer at the end of two minutes, s/he was stopped, and the answer was recorded as “N/A”. After the participants finished writing their answers, they were asked if they were familiar with the question. At the end of the experiment, the measured time was recorded by the researcher. The unit of data was seconds. The participant’s response was judged in terms of accuracy. The correct answers were assigned the value of 1, and the wrong answers were assigned the value of 0.



Among 68 participants, 10 of them were familiar with the question. So, their results were omitted from the calculations. In each condition, there were 19 participants who provided an answer. However, the number of participants who gave the correct answer varied in each condition group. In the control group, there were 4 correct answers. In stereotype condition, there were 17 correct answers. In non-priming condition, there were 9 correct answers. While calculating the average response times, only the response times for the correct answers were taken into account. The average response time for the control group was 33.8 seconds, while the average accuracy as 0.25. The average response time in non-priming condition was 38.5 seconds, while the average accuracy was 0.53. The average response time for sample 2 (Mother-Son) 30.6 seconds, while the average accuracy was 0.89 (Fig.1).




The differences in the number of correct responses across conditions were tested using chi-squared tests. These tests were conducted in R using the following syntax:


chisq.test(rbind(c(4, 16), c(9, 10))); chisq.test(rbind(c(4, 16), c(17, 2))); chisq.test(rbind(c(9, 10), c(17, 2)))


The results of these tests are as follows:

  • Control group (father–son) vs. non-priming condition (father–daughter condition):

χ2(1) = 2.17, p = .141

  • Control group (father–son) vs. stereotype condition (mother–daughter) condition:

χ2(1) = 16.23, p < .001

  • Non-priming condition (father–daughter condition) vs. stereotype condition (mother–daughter) condition:

χ2(1) = 5.97, p = .014


It was concluded that there is a clear difference between the control group and stereotype condition, whereas there is only a suggestive difference between the non-priming condition and stereotype condition.


Response Latency

The differences in average response times across conditions were tested by using t-test in R. The difference between non-priming condition and the control group was not statistically significant (p=0.86). Likewise, the difference between stereotype condition and the controlled group was not statistically significant (p= 0.33) Finally, the difference between non-priming condition and stereotype condition was not statistically significant (p=0.19).




The results of the study partially support the hypothesis. It was observed that when the priming effect was overridden through the presentation of two sexes in the question, male and female, the accuracy rate increased. Yet, the accuracy rate of stereotype condition was still significantly higher than that of the non-priming condition. Therefore, it was concluded that although priming effect has an impact on participants’ failure to find the correct answer of the original riddle (father-son), based on these results, the impact of implicit gender stereotypes cannot be neglected. The results also showed that the response time was an insignificant measure and did not contribute to the findings of the study. The participants who could eventually find the answer spent almost an equal amount of time across all conditions. The priming effect or the expected answer (whether it was “the mother” or “the father”) did not change the response time significantly.

In the discussion of the results, it is important to acknowledge the limitations of the study. To start with, the small sample size (N=68) is a significant limitation. In each condition, the sample size was around 20 participants. The fact that the surgeon riddle is a fairly well-known question since it was posted online by BBC caused a further reduction in the number of participants whose responses could be evaluated. Although there were 68 participants, 10 of those participants had to be omitted from the study due to their familiarity with the question. In the future, the use of a larger sample size could further improve the results of the study.

Another limitation is the willingness and persistence of the participants. Notably, the average response time of the college students who gave the wrong answer is 30.2 seconds, while the average response time of the participants knew the question is 19.8. This shows that the college students who gave the wrong answer spent very little time to find the answer. Maybe, if they were willing to spend more time, there could have been a higher number of correct answers. This limitation might disappear if the participants are offered a compensation for their participation in future studies.

Another limitation of the study is the problem of internal validity. It is not unusual for a participant to be asked a “riddle-like” question. However, it is regarded as unusual for a question that sounds like a riddle to have such an easy answer. So, the nature of the experiment leads the participants to think that the question is “tricky.” Therefore, it is possible that they automatically disregard the possibility of ‘the mother” as an answer. For future studies, this limitation can be overcome by giving the question in a survey, along with other questions that have very basic answers.




These conclusions are important in the sense that they challenge the widely accepted explanation that the people who are asked the surgeon riddle fail merely due to their implicit gender stereotypes. Acknowledging the fact that this explanation was so readily accepted by psychologists (see Hoagland, 1988; Kollmayer et al., 2018; Oakhill, Garnham, & Reynolds, 2005) communicates that as the discourse concerning stereotypes deepens, it becomes easier to build theories on assumptions. Therefore, the study reveals that it might be necessary to revisit certain fundamental theories with a critical eye.




Table 1. The Control Group:

“Father and son are driving a car. They get into a car accident, the dad dies, son gets rushed into hospital. The surgeon says, ‘I cannot operate, this is my son.’ Who is the surgeon?”

Table 2. Sample 1:

“Father and daughter are driving a car. They get into a car accident, the father dies, daughter gets rushed into hospital. The surgeon says, ‘I cannot operate, this is my daughter.’ Who is the surgeon?”

Table 3. Sample 2:

“Mother and son are driving a car. They get into a car accident, the mother dies, son gets rushed into hospital. The surgeon says, ‘I cannot operate, this is my son.’ Who is the surgeon?”


I would like to thank Fiery Cushman, Benedek Kurdi, and the fellow visitors of Harvard Square.



Ashmore, R. D., & Del Boca, F. K. (1979). Sex stereotypes and implicit personality theory: Toward a cognitive—Social psychological conceptualization. Sex Roles, 5(2), 219–248.

Bargh, J. A., Chen, M., & Burrows, L. (1996). Automaticity of social behavior: Direct effects of trait construct and stereotype activation on action. Journal of Personality and Social Psychology, 71(2), 230–244.

Carli, L. L., Alawa, L., Lee, Y., Zhao, B., & Kim, E. (2016). Stereotypes About Gender and Science: Women ≠ Scientists. Psychology of Women Quarterly, 40(2), 244–260.

Coontz, S. (2013, June 9). Progress At Work, But Mothers Still Pay a Price. New York Times (1923-Current File); New York, N.Y., p. SR5.

Cuddy, A. J. C., Wolf, E. B., Glick, P., Crotty, S., Chong, J., & Norton, M. I. (2015). Men as cultural ideals: Cultural values moderate gender stereotype content. Journal of Personality and Social Psychology, 109(4), 622–635.

Eagly, A. H., & Mladinic, A. (1989). Gender Stereotypes and Attitudes Toward Women and Men. Personality and Social Psychology Bulletin, 15(4), 543–558.

Gilovich, T. (2016). Social psychology (Fourth Edition.). New York: WWNorton & Company.

Greenwald, A. G., & Krieger, L. H. (2006). Implicit Bias: Scientific Foundations. California Law Review, 94(4), 945–967.

Herr, P. M. (1986). Consequences of priming: Judgment and behavior. Journal of Personality and Social Psychology, 51(6), 1106–1115.

Herr, P. M., Sherman, S. J., & Fazio, R. H. (1983). On the consequences of priming: Assimilation and contrast effects. Journal of Experimental Social Psychology, 19(4), 323–340.

Hoagland, S. L. (1988). Lesbian ethics: Beginning remarks. Women’s Studies International Forum, 11(6), 531–544.

Jackson, L. A., & Cash, T. F. (1985). Components of Gender Stereotypes: Their Implications for Inferences on Stereotypic and Nonstereotypic Dimensions. Personality and Social Psychology Bulletin, 11(3), 326–344.

Kollmayer, M., Pfaffel, A., Schober, B., & Brandt, L. (2018). Breaking Away From the Male Stereotype of a Specialist: Gendered Language Affects Performance in a Thinking Task. Frontiers in Psychology, 9.

Mineshima, M. (n.d.). Gender Representations in an EFL Textbook, 20.

Oakhill, J., Garnham, A., & Reynolds, D. (2005). Immediate Activation of Stereotypical Gender Information. Memory & Cognition, 33(6), 972–983.

Sanford, A. J. (1985). Cognition and cognitive psychology. New York: Basic Books.


Interview with Hannah Marcus, Professor in the Department of History of Science

Apurva Kanneganti ‘20, THURJ Writer


One of Harvard’s newest and brightest faculty additions, Professor Hannah Marcus is quickly making a name for herself in the History of Science department. Within a year of her arrival at Harvard, Marcus is already finishing up her first book, a translation of letters written by an enigmatic 16th century female apothecary and philosopher. A few of her numerous other projects including studying medical censorship in the Early Modern era and examining the literary proclivities of the great Galileo Galilei. THURJ writer Apurva Kanneganti had a chance to chat with Marcus about her story and her passion for research.


AK: Thank you for giving us this opportunity! To start off, could you give me some information about your educational background and career thus far?

HM: Absolutely! I did my undergraduate degree at the University of Pennsylvania, which was a big change from the small town in Maine where I grew up.  I also spent all of my junior year at the University of Bologna where I learned Italian and fell in love with the early history of science and doing research in archives. I took a year off after my undergraduate and worked at the State Archives in Maine on Civil War materials while I applied to graduate school.  Then I moved to California to do my PhD at Stanford, and finally spent a postdoctoral year based in Princeton before I started working here at Harvard in 2017.


AK: If you could name the single largest influence on your work until now, who or what would it be?

HM: My PhD advisor at Stanford, Paula Findlen, is not only one of the most brilliant people I’ve ever met, she’s also generous and a model of scholarship. I strive to emulate her in my academic career.

[Paula Findlen is the Professor of Early Modern Europe and History of Science at Stanford and recipient of the 2016 Premio Galileo prize. She and Marcus are currently working together on The Galileo Correspondence Project.]



AK: When you were an undergraduate, were there any classes that you think really had an impact on you? If so, what were they and why?

HM: Yes! Two classes hugely influenced my future career.  The first was Econ 001 where I learned about opportunity cost and decided that I should be starting to learn a new language rather than toiling with advanced French and biding my time for study abroad. As a result, I started taking Italian classes, which is why I ended up going to Italy and discovering the incredible joy of archival research. Every discipline has a necessary skill set. One of the unspoken skill sets of doing historical research is reading different languages.  I now work in six languages on a regular basis.

The other class that made a huge impact on me was actually not a class, but a book history seminar series that met on Monday evenings in the Rare Book Room at Penn. I was working in the library as a job, so I started going to the seminar with my colleagues after the library closed, and that’s where I learned about academic community and how academics engage in and share their research.


AK: Speaking of classes, what classes are you teaching right now, and do you have any ideas for classes you’d like to teach in the future?

HM: Ooh! I’m having such fun teaching History of Science 117—Inventing Science, right now. It’s a new course for me, and it’s a wonderful project to synthesize and weave together the many strands that make early science such a fascinating topic. (I’ll also be teaching it again next spring for those of you who missed it this year.) In the fall, I’m excited to be developing an undergraduate seminar on the early history of medicine in Europe.



AK: Turning to your work, could you talk briefly about your research focus?

HM: My research focuses on the interaction between science and religion in the period following the Protestant Reformation.  My first book looks at the Catholic censorship of scientific books in the 16th and 17th centuries, and also pays particular attention to books as physical, material objects. In addition to understanding how censors and early modern readers dealt with the content of the science, I’ve also spent a lot of time examining how censorship has left physical traces on books in the form of missing or blacked out pages, changed words, and passages that are pasted over with other paper to cover up illegal content. This is the kind of research methodology that was modeled in the book history seminar I sat in on as an undergraduate.



AK: What do you think are your most significant research accomplishments at this point in your career?

HM: I’m very proud of the book I’m finishing right now [Translation of Camilla Erculiani, Letters on Natural Philosophy], but probably my most sparkly finding is an article that will come out next year on Galileo writing in cipher, which has escaped notice for the past 350 years. This article is part of a larger project I’ve been working on that looks at Galileo’s correspondence and has related digital history components as well.



AK: What type of role do you think the Harvard environment has had in shaping your research?

HM: Harvard is all about research. Everything about my job here helps advance my research agenda. I have the incredible privilege of teaching classes that relate to my research interests, inspirational colleagues who support my work, and the world’s greatest libraries at my fingertips.  Harvard is a remarkable place!



AK: As for your specific department, what resources are in place to facilitate research and thought?

HM: Working groups are run by graduate students and are a venue for getting feedback on works in progress. It’s a place for forming intellectual community and engaging with our peers on a regular basis.



AK: Coming to a question I think many people have, how would you classify History of Science in the traditional scheme of classification (natural sciences, humanities, social sciences, etc), and why?

HM: The History of Science is a fundamentally interdisciplinary field.  It’s one of the aspects that’s most exciting about being in this department. As someone who works on early science (most of my research is on the 16th and 17th centuries), my own research is more humanistic than that of my colleagues working on more modern materials. Also, many historians of science have background in the hard science area that they study. For my department as a whole, I think aspects of all of these classifications apply to our work.



AK: Finally, is there anything you’d like to say to the students at the College wanting to pursue research or academia?

HM: I think for students in any field my advice would be the same: be clear about your research interests, and then find great mentors to help you pursue them. This is a big place with many opportunities, so look for allies among faculty, staff, and other students to support you in your goals.


Marcus’ first book, Translation of Camilla Erculiani, Letters on Natural Philosophy, is due for publication soon from the University of Toronto Press. Those interested in her area of research are also encouraged to check out the Early Sciences Working Group, which Marcus co-sponsors.

The interview was edited for clarity.

Just for laughs: Utilizing Machine Learning to Rate and Generate Humorous Analogies

Limor Gultchin ’17


Computational Humor: What’s At Stake

With recent advance in machine learning and data science, more and more fields that were once sealed off from computational approaches are opening up for the touch of artificial intelligence. Humor is one such field. Notoriously difficult to analyze, assess and rate even for the human agent, it provides a challenge for the adventurous computer scientists who dare to take it on. Literature that provides analysis of the roots and causes of humor has been circulating at least since ancient Greece, with the following major theories presented to date10:

1. Superiority theory. The superiority theory of humor states that the feeling of self containment associated with a sense of superiority in our status or well being over others is what makes us laugh. In other words, we laugh when we are better off than a fellow human, and our assertion of this fact. Notable philosophers proposed and bolstered this assumption, an act which did not help the status of comedians and humor: they were deemed as agents of evil mockery and pride. Among the famous supporters of this view were Plato, Descartes, Hobbes and the Bible.

2. Incongruity theory. Scholars endorsed a more modern and positive approach to humor was endorsed mostly since the renaissance. This reading of humor claims that what creates a comedic effect is a “benign violation of expectation”8 . Recent research have shown that many types of laugh outbursts could be explained by their surprising nature. Whenever an unexpected event happened, that turned out to be harmless – amusement was afoot. Two conditions had to be met to create humor:

(a) an incongruity, which first took an audience by surprise, perhaps even a tense, fearful one.

(b) a resolution of the contradiction of expectation had to happen, followed by relief, which had a vocal manifestation, of the form of “Ha Ha”, accompanied with an upward twitch of the lips.

Among the supporters of this understanding are James Beattie, Immanuel Kant, Arthur Schopenhauer and Søren Kierkegaard.

3. Stress release theory. Along the lines of the relief theme proposed by the “incongruists”, Sigmond Freud proposed a similar reading into humor, with a typical Freudian twist. To Freud, humor was a result of pent up stresses, many of whom related to societally taboo subjects, such as sexuality and violence, that are released when a joke teller refers to them directly or as an innuendo. By “opening for a conversation” a subject the listener has been putting active effort to suppress, a joke can signify a promise of relief, an appeasing statement, an “it’s OK” signal. It tells the listener others are contemplating and suppressing the same topics, and offers an opportunity to vent some of these accumulated suppressions. Lord Shaftesbury, Herbert Spencer and John Dewey also adopted a biological-psychological reading into humor as a stress reliever.

4. Mock-A ression play theory. From a biological-evolutionary point of view, laughter can be explained by its predecessors in primates. Observations by evolutionary biologists revealed that primates laugh too, usually during mock-aggression activity within a family of chimpanzees. Mock-aggression activity can be seen when members of the same family of monkeys are engaged in playing, that to an outside observer might look like a violent confrontation. When two or more such primates are “fighting” in this way, like children who are engaged in a mock-fight, they seem to be training for actual potential future conflicts, yet need to signify to each other that they are not being truly aggressive. The chosen signals are usually a quick, fast breath of air coming from the diaphragm and a movement of the mouth, in which the front teeth are exposed and the lips are drawn back and upwards. The sound this mock-aggressive signal is making is “ah ah”, which resembles human laughter, only in a backward direction: primates’ diaphragm is set up a little differently than that of humans (or rather the other way around?), to allow for appropriate locomotion. Thus, primates breathe in instead of out, when poking fun at their fellow primates.

Computational Humor

A computational approach to humor is a much more recent development, yet it has already produced its own modest history. Beginning in the 1990s, computer scientists attempted at understanding, and moreover producing humor automatically. Developments in machine learning made it possible to imagine an algorithm that could take in examples of humor and to try uncover their inner pattern. HAHAcronym14 was one of the first examples of the development of a humor-generation focused system, which aimed to produce humorous acronyms. In the mid 2000s, Twitter proved to be a useful platform for such investigations, due to its ease of access and its inclusion of ready-made initial classifications (e.g. hash-tags). Barbieri and Saggion utilized the popular social network to detect irony3 , while Yishay Raz focused on automatic classification of types of humor12, such as anecdotes, fantasy, insult, irony, etc. Some of these computational approaches referenced the voluminous traditional thinking of humor, presented above2. There was also much work done for the understanding and assessment of visual humor. In 2015, Shahaf et al. joined Bob Mankoff, cartoon editor of the New Yorker, to build a system that would be able to predict which one of the +5,000 weekly submissions to the newspaper’s cartoon caption competition is the funniest one13. These attempts and more have all been interesting and illuminating, but proved to have a limited amount of success. There is clearly much more to be done to achieve a more accurate understanding of what makes things funny and how we can “teach” humor to computers.

Increments to our knowledge of humor are gradual and modest, being a difficult task as it is, and yet there are more and more indications as to the potential of such investigations. In an attempt to add to the existing corpus of computational humor literature and experimentation, this thesis focuses on generating literal humor, through an examination of Google’s word2vec word embedding. In particular, we will examine humorous 4-word analogies generated based on the embedding’s representation of words.

Word Embeddings

For this literal approach to the composition of humorous 4-word analogies, we took advantage of the existing and relatively new technology of neural word embeddings, and in particular Google’s word2vec, which opened to public usage in 2013119. Using a neural net trained on instances of Google news articles containing 3 billion running words, Google’s machine learning engineers were able to create an embedding of 3 billion words mapped into vectors of 300 dimensions. The vectors are learned representations of the words in the training text corpus, which were crafted using continuous bag-of-words and skip-gram architectures. A neural net is trained to predict a word for the words that appear close by in the text, and the parameters learned by it are used as these word representations. In fact, a semantic field is created, such that words that tend to appear closer across the training texts appear closer to each other in this multidimensional space as well. The resulting vectors thus capture relations between words in the underlying training data, such as which words are similar to each other (and thus are ’neighbors’ in the semantic space) and allow us to complete analogies that capture the relations between pairs of words (such as Paris:France::Rome:Italy). These resulting vectors can therefore be used as features when training models in various natural language processing and machine learning applications. For example, through word2vec, Bolukbasi et al. uncovered gender biases in the underlying embedding6. In this study, analogies such as

Man : Computer programmer :: Woman : homemaker

were automatically created after taking the cross product and Euclidean distance measures of vector representation of single words, pairs and quadruples. We decided to utilize this approach to study the humorous nature of association of words. If word embeddings can uncover gender biases, why can they not uncover funniness of words and combinations of phrases?

Linear regression and SVM classification

This paper had three main goals: to generate humorous analogies, to predict ratings of humorous analogies and to perform a Turing test to assess our results. In order to achieve them, it focuses on binary classification of words and analogies into “funny” and “not funny” categories, and on linear regression to generate prediction of “funniness” scores, based on ratings given by Amazon Mechanical Turk users. Following is a brief explanation of these two methods, meant to dispel the magical nature of machine learning as a “buzz word.”

SVM classification

Once we had our data organized as numerical values, that represent features of different objects (provided by word2vec, in our case), we could train a classification model. Our approach required 4 different classifiers, which created somewhat of a “cascading” effect. SVM classifiers fit a linear classification separator between groups of data that have certain labels. In most tasks of binary classification, they are used to separate positive from negative examples. Our case was no different – we tried to separate positive funny examples from unfunny examples. The separator is fitted such that the least data points will be misclassified (having the opposite label than the desired one). SVMs, or support vector machines, are unique in their definition of a decision boundary which has a desired margin. In a classification problem treated with an SVM, we look at the points that are closest to the hyperplane as support vector (hence, the name). The certainty of a classification of a data point can be determined by its distance from the hyperplane (the farthest it is, the more certain the classification).5 Thus, the best possible hyperplane we could fit is the one that maximizes the distance of the closest points (or SVs) to the hyperplane. A basic hyperplane can be defined asThe goal is to fit the classifier with lowest loss rate, where labels of the classification itself will be described as 1 if h(x; w, w0 ) > 0 or -1 otherwise, and in our case as funny if h(x; w, w0 ) > 0 or unfunny if h(x; w, w0 ) <0. When determining the weight vector w we will try to maximize the distance of the closest points, the support vectors, from the hyperplane, on both sides, while still maintaining a correct classification. To find the optimal hyperplane we will need to minimize the following expression:

Linear Regression (Ridge)

The regression task was done with ridge regression, where a linear function was fitted to predict the scores (y axis) to match a data point, represented by the numerical value associated with a 4-tuple. The following is the definition of the regression1 :

And the loss function we aim to minimize is defined as:

Where X is our features matrix (aka the featurized data), w is the weight chosen for the production of the regression line, w0 is the included bias terms and λ is the ridge regularization parameter. h is therefore the function for a prediction generation. The generation and rating process discussed in this thesis uses these two ideas from machine learning theory5 . The implementation was made possible through the python library scikit-learn7 , which offers great support in putting machine learning theories into practice, and into actual predictions and classification models.


In this project we furthered the understanding of humor and the capabilities of producing it “on demand”. There are various benefits to the development of artificial humor capabilities:

1. Allowing the creation of smoother, more fun interfaces to use, which will surely play an even greater role in our lives in the coming years. Systems which include humorous components could be more congenial: making queries, tasks and warnings less repetitive, statements of ignorance more acceptable and error messages less patronizing4.

2. Facilitating a better understanding of humor itself, by asserting or disproving notions of what makes things funny.

3. Overcoming yet another interesting artificial intelligence challenge posed to computer science researchers.

To the best of our knowledge, there has been no attempt to use our newly gained understanding of word embeddings in the field of computational humor. Furthermore, there has been limited success in previous attempts of humor generation tasks. This is yet another attempt at providing a proof of concept, for future research. This paper shows that computers can not only construct a humorous structure, but also recognize humorous themes relatively well, implying that it might have implications on a fuller understanding of what makes things funny.

Joking Around, or, Learning to Generate Funny Analogies

Our main goal was to show that humorous analogies can be generated based on the word embedding word2vec. We started from any random combination of 4 words, and tried to later generate funny 4-tuples. In the process, we trained 3 different classifiers, which match the 3 phases of generation:

1. 4 random words → 4 funny words. To start building our data set we asked Amazon Mechanical Turk users to come up with to 5 humorous analogies on any topic, and created an initial pool of about 1000 human written analogies. We then asked other users to rate those analogies, in the following manner: each participant was asked to indicate whether a batch of 25 analogies was funny or not (such that they could provide a single up-vote for each joke they found funny). Each batch of analogies was rated by 10 different participants, for a total of about 1200 analogies rated (1000 of them human written; around 200 were analogies we found funny, presented as a check, to make sure raters won’t tire of repetitive analogies which might not be funny, and thus affect the quality of their rating). Since the score range of an analogy is between 0-10, and the average rating for analogy was around 2.5, we concluded an analogy had to gain 4 or more votes to be considered as funny. Next, we trained a classifier which treated as positive examples all the words that were used by AMT users in analogies they have written, and later were rated highly by their colleagues. The word embedding was used to obtain words representing the positive and negative examples, and to draw new words on which we fitted the classifier. Then, we could pull new words that were classified as funny, to create our collection of funny words.

We decided to treat each of the words from funny-rated analogies as ”funny” in itself for our initial classification, as we knew we needed a starting point for this demanding task. The negative examples were any word from the embedding, including verbs, generic names or propositions, which tend to be less likely to be part of a joke. We needed to create an initial collection of words that had a significant likelihood of appearing in a funny analogy. Thus, in a liberal yet effective manner, was treated all words already mentioned in funny analogies as having higher likelihood of appearing in jokes, and thus as generally ”funny words”.

2. 4 funny words → pairs of words. We made a new classifier to generate potentially funny pairs of words, from the pool of funny words. We trained another SVM model, and used the length of each word and the angle and distance between the vector representation of the words as features. After tuning the hyper-parameters of the model, we managed to classify quality pairs, using pairs from the Turkers analogies as positive examples, and random pairing of funny words as negative examples.

3. Generated pairs → classified matching of pairs. As a final stage in this cascading process, we trained a final SVM to tell the difference between random pairing of the 2-tuples and good pairings, which have an appropriate affinity between their first and second halves. The pairs we were using to create these full 4-word analogies were the generated pairs of the previous phase, and the features used were a combined 1200 dimension vector, made up of each words word2vec representation, as well as the distance and angle between the pairs. The Positive examples for training were, as expected, the full 4-tuple analogies rated funny among the AMT users, and the negative examples were random pairing of pairs generated in the last round.

Through this iterative yet evolving process, we managed to generate new analogies that could have now be put to the test of AMT users. The task that was presented to them was identical to the original rating task described above, with the sole difference that the analogies now presented were computer generated. We decided to use the following baselines (each consisting of 300 analogies tested) for our analysis, so that we could assess the progression of this method, one step at a time:

• 4-tuples of completely random 4 words from the embedding

• 4-tuples of funny words from the pool generated by classifier 1.

• 2-tuples of randomly matched generate pairs by classifier 2.

• AMT made analogies.

Figure 2.1: Genera on process schema
Figure 2.2: Average funniness scores for each baseline: random 4-tuples (rand-4), random 4-tuple of funny words (rand-4-fun), randomly matched classified pairs (rand-pairs) and fully generated pairs (GEN).


Let us compare these 5 results (including the complete, newly generated analogies, made using classifier 3):

Comparing the top ranked jokes can also give us a sense of what the trend looks like. Among the different baselines, the most highly rated score achieved was a 5, for the analogy

woman : jungle :: hair : fathers

which belonged to the random pairs. Here are a few examples, with their corresponding scores, per each baseline:

1. Random-4 – the highest score was a 3, given to 10 analogies out of 300. Among them were

store shelves : shipyards :: NZX : Bowl Championship renewed : knives :: policies : attempt implementations : socializing :: overspending : fix.

2. Random-4-fun – the highest score was a 4, given to two analogies:

meal : metamphetamine :: disillusioned : Hendrix bishop : appalled :: Australians : thermostat.

The first round of classification demonstrate a pretty good understanding of the kind of themes that drive humorous analogies, including references to pop culture, nationalities, religion, food and drugs. Other references that showed up in the pool were related to sex, family, animals (mostly dogs and cats) and sports.

3. Random-pairs – the random pairs baseline results showed a single analogy that was rated at 5, and a couple that were rated as 4. The 5-rated was

woman : jungle :: hair : fathers

and the next two, rated at 4, were

stunt : governor :: petty : sex ass : dems :: poodles : This.

Although most pairs seem promising, their combination is random, and therefore requires further classification.

Now let us examine the case of the fully classified analogies (following the 3-step process described above). The highest ratings given were 7s and 6s: the most highly rated analogy was

water : Kardashians :: toilet : Reality TV

and a few 6ers were

bear : diamonds :: Trump : empathy barking dog : McCain :: burglar : Vanilla Ice.

More results are summarized in Table 2.2.

Figure 2.3: Average funniness scores for fully generated pairs (GEN) versus Average funniness scores for Amazon mechanical Turk users written analogies (AMT).
Table 2.2: Breakdown of funniness ratings by type of baseline

Regression Towards a Laugh, or, How to Rate Funny Analogies

Predicting ratings AMT users would give to analogies (based on the same procedure that was described in the previous chapter) is another approach we could take to assess the types of humor understanding we can capture. Given an initial set of analogies and ratings provided by AMT users, we can define a regression line that would enable us to predict the scoring of a newly given analogy. The goal is to pick such a regression line y(x), such that the expected value of the loss will be minimized. For a more robust explanation of the training of such a model, please refer back to the section 1.4.2.

In our case, we started fitting the line with the data points representing the analogies given by AMT users, and using the ratings given by other AMT users as the data labels t. After tuning the weights in such a manner, we used the fitted regression line to predict the scores of new analogies generated by our system. The features for the regression were a 1200-dimension vector, the result of a concatenation of the word2vec 300-dimensions vector of each word. We chose ridge regression, which can lead to a sparse model that will be less likely to over-fit than a basic linear regression, and thus generalize better to new data points.

Our results showed positive correlation between our predictions and the scores which were eventually given by Turkers to our generated analogies (computer-generates ones). We performed a cross-validation, a technique which further helps us avoid over-fitting. To that aim, we separated the data to train and test sets (in our case, for a total of 5 groups) and could tune the hyper-parameters of the model without affecting the quality of our predictions, by modeling the noise in the training data so closely that it will negatively affect our accuracy for new data points. We chose a lambda value of 10 after performing a search for the value that will minimize the loss, and ran the regression. The results of this cross validation were giving a mean correlation score, between prediction and actual scores, of around 0.371 with a standard deviation of +/-0.038, for a 5-fold partition.

Though this positive correlation might be lower than we would aim for in typical regression, humor requires a different scale of assessment. Humor cannot be objectively agreed upon by different humans, let alone by a computational approach. Therefore, a positive correlation of 0.37 attests to an ability to predict scores of jokes in a satisfactory manner – one which performs much better than a random assignment of score predictions.

Comedy of Errors, or Applying the Turing Test to Our Jokes

Since the topic of this thesis is a machine learning question that is flirting with Artificial Intelligence, trying to make a machine “learn” a task that is deep within the realm of what humans define as intelligence. This chapter explores whether it is possible to fool humans into thinking some of the generated jokes were human made (and vice versa). To achieve this goal, we created another Amazon Mechanical Turk task which consisted of 69 analogies – half were computer generated and half were human made. We chose the 35 highest rated human-made analogies and the 34 highest rated computer-made ones, to make sure that all analogies were in fact relatively successful jokes, which would make the turing test more focused on the humor aspect of this project, rather than the logical quality of the analogies. As a result, we could also keep the AMT users who participated in this test slightly more amused. We asked 10 participants to mark whether each of them was, to the best of their assessment, the work of a computer or a human. We then took a majority vote of the 10 users – each vote for “human” was considered as 1, and each vote for “computer” was considered as -1 (if no choice was made by a user, we considered it as a 0 vote). Then, if an overall score was positive the analogy was considered as human-made by the majority vote, and vice versa was considered as computer-made if the obtained score was negative. Mostly, AMT users were pretty successful at predicting the identity of the generator. However, out of 69 analogies, AMT users failed to classify correctly 19 of them, for an error rate of 27.5%. Overall, 12 computer-generated analogies were thought of as human-made by a majority vote of 10 users, and 7 human-made were mistaken for computer-generated ones.

Our generated analogies were mostly recognizable as computermade. Yet, it provided evidence that this Turing test was far from trivial – out of 35 computer-made analogies, 12 were mistaken for a human-made analogy – a significant rate of 35%. The error rate of mistaking a human analogy to be a computer generated one was 20%, both quite surprising for a task that initially sounds rather intuitive.

Table 4.1: Table of Confusion: Turing test full results
Table 4.2: Examples of misclassified jokes in an AMT Turing test, separated by kind of mistake.


In this thesis, we have shown how new word embedding techniques can be used for the enhancement and advancement of the study of computational humor, and humor at large. We have provided a description of a process for the production of humorous analogies, which proved to have been more successful than three different possible baselines. We were able to generate analogies which were rated as funny by AMT users, who did not know the analogies were produced by a computer. Furthermore, we have shown that there is a positive correlation between our predictions and scores AMT users will provide to analogies. Finally, we performed a Turing test, in which users were asked to identify whether certain analogies were made by humans or by a computer program. In doing so, users were unable to identify 35% of the computer generated analogies.

After showing a “proof of concept” in the form described above, this thesis further opens the door for enlarging the capabilities of this humor-generating system. Our current system seems to be doing a good job at recognizing themes and generals imageries which are favorable to humor creation. We can recognize it in the significant spike in the “funniness” of our analogies, past our random-4-funny-classifier, and by the milder improvements presented by the rest of the classifiers which are involved in our cascading humor generation system.

However, a lot more can be done to put together the successful pieces we identified. We can assert that marriage, family, sex, religion, politics and food are topics favorable to humor, and that verbs are less likely to be used by our classifiers. Yet, we can do more to combine the words in an analogy that truly captures a good joke. For example, using solely qualitative assessments, it seems that the fourth word can have a great effect on the cohesiveness of a humorous analogy in the same way that a punch-line is qualitatively assessed by stand-up comedians to “make or break” a typical joke. Therefore, here are a few further steps we would like to take in the future:

1. Train a “punchline” classifier, to produce better analogies, and once again assert, in a quantitative manner, a well-known qualitative notion in the world of comedy.

2. Train a classifier to recognize the optimal ordering of an analogy. For example, the analogy we rated in the first chapter

twitter : Christmas :: Hitler : Santa

could arguably sound better if written in a different ordering, i.e.

Christmas: Santa:: Twitter: Hitler.

We could train a classifier based on orderings of recognized funny analogies, using features such as the angle between the vectors representing each word, pairs, and all possible pairing of the 4 words participating in the analogy. In doing so, we should be able to capture this subtlety rather well, and produce even “tighter” analogies.

3. After the suggested improvements to the generation process described above we can introduce a new Turing test, similar to the one described in chapter 4, such that we can achieve an even better error rate (or, as we like to think of it, “rate of confusion”). Eventually, a competition against professional comedians could give us another measure of our system’s performance.

Pursuing some of these future directions could produce funnier, tighter analogies, that will once again challenge our opinion on what computers can and cannot do.




[1] Alexander Rush, D. P. (2017). Linear regression and basis functions handouts. Harvard University, SEAS. cs181lectures/blob/master/03-lectures-lr.pdf.

[2] Bacciu, D., Gervasi, V., & Prencipe, G. (2016). Lol: An investigation into cybernetic humor, or: Can machines laugh? In LIPIcs-Leibniz International Proceedings in Informatics, volume 49: Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik.

[3] Barbieri, F. & Saggion, H. (2014). Modelling irony in twitter. In EACL (pp. 56–64).

[4] Binsted, K. et al. (1995). Using humour to make natural language interfaces more friendly. In Proceedings of the AI, ALife and Entertainment Workshop, Intern. Joint Conf. on Artificial Intelligence.

[5] Bishop, C. M. (2006). Pattern Recognition and Machine Learning (Information Science and Statistics). Secaucus, NJ, USA: Springer-Verlag New York, Inc.

[6] Bolukbasi, T., Chang, K.-W., Zou, J. Y., Saligrama, V., & Kalai, A. T. (2016). Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advanc in Neural Information Processing Systems (pp. 4349–4357).

[7] Buitinck, L., Louppe, G., Blondel, M., Pedregosa, F., Mueller, A., Grisel, O., Niculae, V., Prettenhofer, P., Gramfort, A., Grobler, J., Layton, R., VanderPlas, J., Joly, A., Holt, B., & Varoquaux, G. (2013). API design for machine learning software: experiences from the scikit-learn project. In ECML PKDD Workshop: Languag for Data Mining and Machine Learning (pp. 108–122).

[8] McGraw, A. P. & Warren, C. (2010). Benign violations: Making immoral behavior funny. Psychological Science, 21(8), 1141–1149.

[9] Miháltz, M. (2016). word2vec google news code implementation. github.

[10] Morreall, J. (2016). Philosophy of humor. In E. N. Zalta (Ed.), The Stanford Encyclopedia of Philosophy. Metaphysics Research Lab, Stanford University, winter 2016 edition.

[11] Radim Rehurek, Tomas Mikolov, G. (2013). word2vec.Google Code Archive.

[12] Raz, Y. (2012). Automatic humor classification on twitter. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologi : Student Research Workshop (pp. 66– 70).: Association for Computational Linguistics.

[13] Shahaf, D., Horvitz, E., & Mankoff, R. (2015). Inside jokes: Identifying humorous cartoon captions. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 1065–1074).: ACM.

[14] Stock, O. & Strapparava, C. (2003). Hahacronym: Humorous agents for humorous acronyms. Humor, 16(3), 297–314.


Ruben Reyes Jr. ’19


At his first State of the Union address, President Donald Trump beckoned to the parents of two girls murdered by members of the Salvadoran gang, MS-13 or la Mara Salvatrucha. The four adults were shown on national television in tears as the President spoke about their daughters and claimed that the young men who’d murdered them had been in the United States precisely because of the country’s immigration system. Against the backdrop of their tears, Trump called on Congress to increase a border wall and provide funding for Border Patrol and Immigration and Customs Enforcement. In 2017, in the same Long Island neighborhood where the two daughters Trump spoke of were murdered, ICE detained at least 32 teenagers on the assumption of affiliation with MS-13. Months later, after the teenagers had been taken to detention centers states away from their homes, 14 were released when a judge ruled that the federal government did not have adequate evidence to prove their affiliation. The link between immigrant policing and MS-13 in Long Island was not as straight-forward as Trump’s State of the Union portrayed it.

This paper attempts to track the way that MS-13 grew as a result of U.S. policies, including immigration reform passed in 1996 that sought to increase national security. The expanded definition of an “aggravated felony” was central to the criminalization of a broad array of Salvadoran immigrants who were detained and deported for offenses, unlike those of MS-13, that did not pose a serious national security threat. Through a close reading of speeches given by Presidents Obama and Trump, I will argue that, since at least 2008, MS-13’s actions have been conflated with those of the undocumented Salvadoran diaspora at large to justify an expansion of immigrant policing and deportation policies. Finally, a literary analysis of Salvadoran-American poet Yesika Salgado’s debut collection, “Corazón,” reveals that Salvadoran diasporic authors contra-dict these discourses. Salgado does this by centering the domesticity of the majority of Salvadoran immigrant communities, specifically through reframing the image of the machete in a positive light.


Salvadoran immigrants have been coming to the United States since as early as the nineteenth century, often working for transnational companies or providing desperately needed labor in the United States. Salvadoran immigration escalated during The Salvadoran Civil War which lasted from 1980 to 1992. The war and its aftermath led to the massive displacement of nearly three million Salvadorans (Rodriguez, 2009), the majority of whom resettled in the United States. According to the Pew Research Center, the 2.1 million Salvadorans living in the United States in 2015 made up the third largest Latino ethnic group in the country (Antonio et al, 2017).

In the 1980s, many Salvadorans fled the country to avoid political violence. The United States provided the repressive, authoritarian Salvadoran government with over 4 billion dollars throughout the 80s, even as the military government oversaw a multitude of human rights violations (Bourgois, 2001). A statistical survey from 1984 found that Immigration and Naturalization Service (INS) apprehensions soared in 1983 and 1984 when military sweep operations became “larger and more frequent” and that “fear of political violence appears to be the dominant motive for Salvadoran migration” (Stanley, 1984).

Though the link seemed clear, the United States government at the time did not acknowledge the immigrants as refugees. Salvadoran migrants qualified for refugee status under the United Nations 1951 Convention and the US Refugee Act of 1980, but the United States’ political and financial support for the Salvadoran government prevented the U.S. government from legally granting the distinction (Abrego, 2017). Fewer than 3 percent of Salvadorans who applied for political asylum in the 1980s were accepted. The Immigration Reform and Control Act of 1986 (IRCA) only offered a path to citizenship to immigrants who migrated to the United States before January 1st, 1982, which covered only a portion of Salvadoran immigrants given that cutoff was only two years into the war (Abrego, 2014).

This forced many Salvadorans to live in the United States as undocumented immigrants and made a large portion of the Salvadoran diaspora susceptible to restrictive U.S. immigration policy. The lack of legal status prevented undocumented Salvadoran immigrants from being able to “plan toward a stable future” (Abrego, 2017). An inability to work legally led most Salvadoran immigrants to live in “poor urban centers with high levels of racial and ethnic tensions” and to be “disproportionately employed in service, manufacturing, and construction industries characterized by instability and low wages.” (Dingeman-Cerda et al., 2015). Responding to “poverty, racism, and discrimination they experience as the stigmatized and marginalized of society” (Ward, 2013), many recent immigrants turned to gangs, their marginalization sparking, and sustaining, the rise of Latino gangs such as la Mara Salvatrucha.


Though formed in the United States, la Mara Salvatrucha, also known as MS-13 has become a transnational gang, spreading to thirty-one states and the District of Columbia in the United States and throughout Central America (Funes, 2008).  Between 1998 and 2005, the United States deported 46,000 people with criminal convictions to Central America, including gang members. Upon returning to their countries of origin, these deportees were marginalized in societies which were, very often, foreign to them. In El Salvador, deportees—regardless of whether or not they were involved with a gang—were profiled by police, intimidated by gangsters, and discriminated against by employers, often because of their tattoos or American accents. (Dingeman-Cerda et al., 2015). In El Salvador, this social marginalization was furthered by laws that associated men with tattoos with gangs and made membership to a gang alone grounds for imprisonment (Funes, 2008).

In response, some of the deportees to El Salvador were “compelled or forced to join gangs” (Dingeman-Cerda et al., 2015). This perpetrated gang violence further, spreading violence throughout the country and putting more Salvadorans at risk. To flee violence, many Salvadorans, particularly recently arrived deportees who are doubly targeted by law enforcement and gangsters, migrated back to the United States. Even after the Civil War, immigration from El Salvador to the United States continued as Salvadorans fled a nation increasingly susceptible to gang violence. Though nearly all Salvadoran immigration since the 1980s can be linked to the U.S.-funded civil war (Dingeman et al., 2009), the connection was seldom made. The American public viewed “illegal immigrants” and la Mara Salvatrucha as equally “criminal,” an opinion aided by immigration reform that expanded the definition of a “criminal.” This legislation laid the groundwork for politicians to construct discourses in the 2000s that criminalized the larger Salvadoran immigrant community.

According to scholars Ruben G. Rumbaut and M. Kathleen Dingeman, history shows that increased immigration “fuels nativist alarms, perceptions of threat, and pervasive stereotypes of the newcomers” (Dingeman et al, 2009). In response, immigration law has often been changed under the guise of  “national safety.” In 1996, immigration reform under Clinton’s presidency contributed to the further criminalization of undocumented Salvadoran immigrants. The passing of the Antiterrorism and Effective Death Penalty Act (AEDPA) and the Illegal Immigration Reform and Immigration Responsibility Act (IIRIRA) expanded the legal definition of an aggravated felony to include misdemeanors and minor infractions, such as undocumented reentry after deportation, shoplifting, and failure to appear in court (Funes, 2008). Petty crimes became grounds for deportation, contributing to the notion that Salvadoran immigrants were inherently criminal and opening up the possibility for all immigrants to be conflated with MS-13’s high profile crimes.

The 1996 legislation also included a provision that was used under both the Trump and Obama administration to further police undocumented immigrant communities. The provision, known as the 287(g) program, was the basis of the 2008 Secure Communities program which helped “deputize state and local authorities to enforce immigration law.” It increased the odds than an individual who had contact with a state or local authority would be confronted about their immigration status, and by the time Obama stopped the program in 2014, “the program had identified an astounding 2.4 million people as deportable,” a large chunk of the total 11 million undocumented immigrants in the United States. Additionally, nearly 50 percent of those deported under the program were deported for misdemeanor or minor charges, such as traffic violations, while another 18 percent were deported for non-criminal offenses (Abrego et al, 2017).

In 2017, Trump passed two executives orders meant to reinstate the main tenets of the Secure Communities program, and went further by eliminating the 2014 Priority Enforcement Program which had “shielded approximately 87 percent of the unauthorized population from removal” (Abrego et al., 2017). In 2017, the first year of Donald Trump’s presidency, ICE reported that 73.7 percent of total arrests were of people with criminal convictions, and 15.5 percent were of individuals with pending criminal charges (ICE, 2017). On the surface, these statistics imply that ICE was targeting criminal immigrants, like the MS-13 gang members who murdered four Latino teens in Long Island that year (Robbins et al., 2017), but a breakdown of the statistics complicates ICE’s actions.

In 2017, 68,346 immigrants were arrested for non-DUI traffic violations and 2,517 immigrants were arrested for immigration related reasons. Thousands of others were arrested for infractions that legally had criminal distinction, but a questionable effect on national security. This included 2,087 immigrants arrested for “Health/Safety,” 3,988 for “Liquor,” and 5,919 for ambiguous “Family Offenses.” The crimes that MS-13 has historically received the most media attention for made up a small subset of total arrests during the first year of the Trump presidency. Homicides accounted for less than one percent of total arrests and sexual assaults charges made up about one percent of arrests (ICE, 2017). The bulk of immigrants who were arrested by enforcement bodies were not individuals who posed a graphically violent threat to society.

Given that 46 percent of foreign-born Salvadoran immigrants are undocumented, and another 25 percent hold temporary or partial legal statuses without paths to citizenship (Dinegman-Cerda et al., 2015), Salvadoran immigrant communities have been disproportionately affected by the criminalization of undocumented immigrants. Though many deportable offenses were defined legally as “criminal,” statistics reveal that the majority of immigrants faced the risk of deportation for offenses that posed almost no threat to national security. In this context, MS-13 was incredibly important because their visibility was used to criminalize the broader Salvadoran diaspora, particularly the sector without permanent legal status.


Though separated by two countries, El Salvador and its citizens have a strong connection with the nation’s diaspora.  Benedict Anderson’s seminal work on nationalism argues that national identity is based on  “an imagined political community.” For El Salvador, this imaginary allowed a sense of “a deep, horizontal comradeship” (Anderson, 1983) between Salvadorans in El Salvador and those in the United States, regardless of inequalities between the two groups. The trope of el hermano lejano, fueled by El Salvador’s politicians and press, squeezed the entirety of the “aggregate of various immigrant waves, generations, and identities” under a single label (Rodriguez, 2009). In practice, this built a shared identity between all Salvadoran immigrants, regardless of citizenship status and class. This constructed, transnational identity empowered American politicians to form dominant political discourses that criminalized all Salvadoran immigrants by focusing on the violent crimes of MS-13.

I use the term “dominant political discourse” in this article to refer to the narrow way la Mara Salvatrucha is talked about by politicians and the manner in which MS-13’s actions are conflated with those of the broader Salvadoran diaspora. These narratives offer a single, incomplete narrative to explain MS-13’s connection to immigration, yet ultimately influence the way many people in United States viewed Salvadoran immigrants. In writing about Dominican literary traditions, Lorgia Garcia-Peña’s argues that hegemonic historical narratives, which I refer to as “dominant political discourse” in this article, are sustained by historical documents that support a “national ideology.” Borrowing from that framework, I focus on political speeches as the historical documents that sustain an ideology surrounding Salvadoran immigrant criminality.

Salvadorans gangs have been used by media and politicians “to stereotype the Salvadoran migrant community as criminal” (Dingeman et al., 2009). These stereotypes drive politician’s actions, particularly in the way they frame immigration, often using MS-13 to justify immigration policy decisions. Most insidiously, the criminalization of Salvadoran immigrants hides the degree to which policies affect the entire diaspora, not just members of MS-13. As explored earlier in this paper, the increasingly broad legal definition of  “criminal” has inflated the gravity of offenses that immigrants have been deported for. Obama-era speeches occulted the fact that “criminal” offenses often included infractions that were of low-risk to public safety.

Though Obama’s deportation policies from 2008 to 2012 were based on policing a wide net of offenses, he consistently claimed that his administration had prioritized “criminals.” On July 15, 2012, President Barack Obama gave a speech that invoked his administration’s deportation of “criminals” to justify Deferred Action for Childhood Arrivals (DACA), an executive order meant to protect youth who were brought to the U.S. as children from deportation. In the White House Rose Garden, Obama claimed, “We focus and use discretion about whom to prosecute, focusing on criminals who endanger our communities rather than students who are earning their education. And today deportation of criminals is up 80 percent” (Obama, 2012). In wielding the statistic chronicling an increase in deportations of “criminals,” Obama clouded how expansive the legal definition of a “criminal” was. He attempted to juxtapose the image of hard-working students and young professionals with those deported for criminal charges or convictions, hiding the fact that immigrants deported under his administration included those removed from the country for and traffic violations and other less heinous “crimes.”

In a televised speech to the nation concerning immigration on November 20th, 2014, Obama repeated the statistic, saying that “That’s why over the past six years deportations of criminals are up 80 percent, that’s why we’re going to keep focusing enforcement resources on actual threats to our security” (Obama, 2014). Again, Obama framed his deportation policy as only targeting immigrants who were a “threat,” a dubious claim at best given the spectrum of crimes that counted as criminal under U.S. law at the time.

Another quote from that same speech ties immigrant communities to MS-13 very directly. In describing deportation priorities, Obama claimed that his administration was removing “Felons, not families. Criminals, not children. Gang members, not a mom who’s working hard to provide for her kids” (Obama, 2014). La Mara Salvatrucha is invoked as the “criminals” and “felons” that the Obama Administration is targeting, furthering the inaccurate idea that the only immigrants being deported were those committing the kinds of crimes the Salvadoran gangs were. Obama also said that his policies would help non-criminal immigrants. He said that, “If you meet the criteria, you can come out of the shadows and get right with the law. If you’re a criminal, you’ll be deported” (Obama, 2014). Again, he implies that these “criminals” are the likes of MS-13 gang members, the Salvadorans most explicitly “constructed as criminal” (Cienfuegos, 2016).

Presenting MS-13 as the gold standard of immigrant criminality in the context of immigration reform pushed the idea that gang violence was inherently tied to immigration and that deportations were only used to ensure national security. The dominant political discourse that Obama constructed in his speeches relied on a link between MS-13 and deportations. All the while, it hid the degree to which non-gang members were being deported. Research from 2009 found that 76 percent of Salvadoran immigrants were removed from the United States because of immigration violations, not criminal convictions (Dingeman et al., 2009).

The way the Obama Administration used gang violence to justify the policing of immigrant communities set the stage for the outwardly cynical, nativist, and racist rhetoric Donald Trump furthered. On July 28th, 2017, President Trump gave a speech to law enforcement in Long Island outlining his administration’s plan to address violence perpetrated by MS-13 in the borough. Throughout the speech, Trump graphically narrated the high-profile crimes committed by MS-13 members, and continued to dehumanize them by repeatedly referring to them as “animals.”

“I was reading—one of these animals was caught—in explaining, they like to knife them and cut them, and let them die slowly because that way it’s more painful, and they enjoy watching that much more. These are animals” (Trump, 2017).

Furthermore, though the speech was about MS-13 specifically, Trump conflated the issue of MS-13 with broader immigration policy. He claimed that the Obama administration, “enacted an open-door policy to illegal migrants from Central America. ‘Welcome in. Come in, please, please.’ As a result, MS-13 surged into the country and scoured” (Trump, 2017). In this quote, Trump contributed to a discourse that ignored the origin of MS-13 and the ways U.S. deportation policies fueled their rise. Additionally, he conflated all forms of immigration with gang violence, saying that “a failure to enforce our immigration laws had predictable results: drugs, gangs and violence.” At another point in the speech, Trump frames his proposed border wall as a “vital tool” to curb gang violence, again marking immigration as the source of a variety of social issues.

President Trump’s discourses painted all undocumented immigrants as the source of violent criminal activity, but the claims were rooted in more speculation than fact. Research from the early 2000s showed that Latin American immigrants, even those that live in poverty, were less likely to commit violent crimes than U.S. born individuals. Among various ethnic groups, incarceration rates for U.S. born individuals were nearly five times as high as their immigrant counterparts, and the lowest incarceration rates among Latin American immigrants were for Salvadorans and Guatemalans (Dingeman et al., 2009). Though Trump’s speeches more explicitly conflated all Salvadoran immigrants with MS-13, Obama’s speeches did so as well by equating “criminals” with gang members. Both politicians constructed dominant political discourses that hid the way their respective immigration policies criminalized the Salvadoran immigrant community as a whole, leading to the removal of many individuals who were in no way affiliated with MS-13.


It was in this context that, in 2017, Salvadoran-American poet Yesika Salgado released her collection, “Corazón.” Salgado is a Los Angeles based poet whose collection is primarily about romance and heartbreak, but often makes references to her experiences as a Salvadoran-American diasporic author. Though her ethnicity is not the main concern of her debut collection, references to her Salvadoran-American identity are mentioned throughout her poems. These references are significant because they are antithetical to the the kind of criminalization of the Salvadoran diaspora that politicians like Trump and Obama engage in. Scholar Lorgia García-Peña argues that, “literature works, at times, to sustain hegemony, while at others, it serves to contest it” (Garcia-Peña, 2016) leading her to coin the term “contra-diction.” In contra-dictions, “dictions” are the “performance of language and meaning” that run counter to hegemonic narratives to offer a subaltern perspective of historical “truth.”

Salgado’s poetry is written in contra-diction precisely because it challenges the dominant political discourses that conflated MS-13 with the broader Salvadoran diaspora. As Lindsey Cienfuegos does in her work on cultural representations produced by MS-13,  an analysis of Salgado’s poems will use “fiction to discuss the controversial topic of gangs because I concur that the socio-historical process of gang formation and our knowledge of that process cannot be separated from fiction because historical narrative is already ‘one fiction among others’” (Cienfuegos, 2016). Specifically, a literary analysis reveals that Salgado challenges the way la Mara Salvatrucha was used to criminalize the broader Salvadoran diaspora by reframing the image of the machete and presenting it as a positive, and central, piece of Salvadoran-American life and identity.

The machete appears in multiple poems from Yesika Salgado’s collection, “Corazón,” which is notable because of the machete’s historical significance. Machetes are commonly used in the rural, agricultural departments of El Salvador for various types of farm work, but they have come to attention in the United States for more sinister reasons. Through the 2000s, “Salvadoran gangs have since been the subject of media sensationalization and politicization” (Dingeman et al., 2009), a trend marked by a heavy focus on instances of la Mara Salvatrucha using machetes to commit murder.

In 2003, a pregnant, seventeen-year-old Honduran immigrant named Brenda Paz was stabbed multiple times with a machete by members of MS-13, leading the Washington Post to call it a “‘savage’ and ‘brutal’ affair.” Paz’s story “put the Mara Salvatrucha at the forefront of American television and radio” (Cienfuegos, 2016). In addition, a 2005 editorial published in Human Events, a conservative political newspaper, claimed that MS-13 “the El Salvador based syndicate has a love affair with machetes” (Malkin, 2005).

The focus on the machete as a murder weapon lasted, prompting even President Trump to focus on the brutality of MS-13’s machete killings. In his speech to Long Island law enforcement officials, Trump argued that la Mara Salvatrucha, “shouldn’t be here. They stomp on their victims. They beat them with clubs. They slash them with machetes, and they stab them with knives” (Trump, 2017). This latest narration of MS-13’s violent crimes followed a trend of “savage violence” dominating discourses surrounding la Mara Salvatrucha, with an emphasis on the primitive nature of the machete (Cienfuegos, 2016). These dominant political discourses fed into the general fear that MS-13 gang members would infiltrate middle-class life with violence (Cienfuegos 2016), an economic-based anxiety that applied to the Salvadoran diaspora more broadly.

In response, Salgado’s poem “Los Corvos” centers the machete in a positive light, completely separated from la Mara Salvatrucha. The poem starts by narrating the way the narrator’s mother and grandmother garden with a machete and ends by comparing the narrator’s lover to a weed that can be similarly cut out. Instead of viewing the machete as a threat to commonplace, domestic life the way President Trump’s speech constructed it, Salgado includes the machete as a centerpiece of her home life. The poem opens with the lines, “we keep a machete in our home/Mami uses it to cut the weeds/on Saturday afternoons” (Salgado, 2017).

Here, the machete is stripped of the brutality and primitiveness that dominated political discourses by describing the blade’s use in the context of the homey pastime of gardening. Additionally, the machete is centered as part of the household. The use of “home,” instead of “house” or another synonym, is important because “home” carries a much more positive connotation, one closely associated with comfort and intimacy. Diction further conveys a sense of comfort found within the machete when Salgado writes, later in the poem, that “the blade is our friend.” The word friend similarly conveys a sense of comfort and intimacy.

The second stanza of the poem focuses on the way the narrator’s grandmother used a machete in El Salvador, subverting the danger of the machete by reframing it as a non-threatening tool instead of a savage weapon. The narrator’s grandmother uses the machete to clear “tree branches” and “overgrown bamboo” and Salgado writes that, “Mamita, mami’s mami,/used to grab her own machete/back in El Salvador…/her tiny frame in the distance/her right arm extended/the blade catching the sunlight” (Salgado, 2017).

First, referring to her grandmother as “Mamita” is critical to portraying the machete as non-threatening because it softens the depiction of the machete’s wielder. In the Spanish language, the suffixes “-ita” or “-ito” are commonly added to names as a term of endearment. The suffixes are diminutive, implying that Mamita is small in stature, further conveyed by the descriptor of her “tiny frame.” Salgado rids the machete of its perceived danger by framing her grandmother, through both description and diction, as non-threatening.

Lastly, Salgado presents the machete as an integral part of the Salvadoran immigrant identity, finding a sense of pride and strength in its image rather than a source of criminality. Salgado frames the machete as part of the tradition of Salvadoran immigrant families, saying that “I come from women/who fend for themselves.” The narrator then uses the image a source of power in her own life, saying that “you are a weed/I know to slice you out of me.” The metaphor of a toxic man in her life as a weed continues to reframe the machete for use in the domestic sphere, not for murder, while simultaneously presenting the machete as a staple of the Salvadoran immigrant identity.

Salgado’s poem “A Salvadoran Heart,” is divided into four different sections that talk about her relationship with her Salvadoran-American ethnicity. The very first section offers the narrator’s genesis story. She writes, “I come from women of corn and cotton fields/of machete/and fire” (Salgado, 2017). The machete is presented as a foundational part of her identity, alongside “corn” and “cotton fields.” Corn is important as both the staple of Mesoamerican diets and religious belief system held by the indigenous people of what is now El Salvador (Rodriguez, 2006). For many years, cotton was a main Central American export (Rodriguez, 2009). Both cotton and corn are foundational to the history of the Salvadoran nation as the source of economic and cultural sustenance. Including them in this paragraph implies that cotton and corn are, metaphorically, a central part of the narrator’s Salvadoran-American identity.

Later in the poem, Salgado metaphorically uses two fruits, jocotes and mangos, in the same way. Jocotes and mangos are both popular fruits cultivated in El Salvador. When the narrator says that “my tongue came to me through a jocote seed” and that “I am the daughter of a river and mango tree” (Salgado, 2017), she frames these objects positively, as the sources of her voice and life respectively. By including machete alongside the images of corn, cotton, jocotes, and mangos, she implies, positively, that the machete is a critical aspect of her Salvadoran-American identity.

By making the machete personal, she decriminalizes the tool and presents in a positive light. Salgado’s use of the machete in both these poems serves as a harsh contra-diction to the images of male, machete-waving, barbaric MS-13 gang members perpetrated by politicians seeking to justify broad immigration policies. She highlights the law-abiding, domestic lives that the majority of Salvadoran immigrants live, throwing into question the dominant political discourses that paint Salvadoran immigrants as criminals. Salgado’s use of the machete is not necessarily reactionary or consciously sparked by the specific speeches given by Trump and Obama. It is likely that she writes about the machete because her childhood home prominently featured one with a practical purpose—gardening. But given that the machete is so loaded for the Salvadoran diaspora given its sensationalized relation to MS-13, the reframing of the machete in her poetry is important to illustrate the way the Salvadoran diaspora understands itself outside of dominant discourses of criminality. 


Due to U.S. immigration law and deportation policies, Salvadoran immigrants have been extremely susceptible to policies that sought to police and remove unauthorized individuals from the country. Immigration laws meant to improve national security contributed to the broad criminalization of Salvadoran immigrants, with the expanding definition of a “criminal” fostering the growth of MS-13. The gang received more visibility and scrutiny from both the press and politicians. Laws are not created in a vacuum and the discourses used to justify them are often as important and significant as the laws themselves. Eventually, MS-13 would be used by both the Obama and Trump administration to justify their respective immigration policies, conflating the violent crimes of MS-13 and the minor offenses of the Salvadoran diaspora in the process.

In response to this broad criminalization of the Salvadoran diaspora, Salvadoran-American authors seek contra-dict dominant political discourses in their poetry. Yesika Salgado’s reframing of the machete in Corazón offers a rich case study, but it is only a sliver of an expanding literary tradition. Further work needs to be done to understand, for example, the discourses that motivated a Salvadoran-American poet to write lovingly of the tattoos on a MS-13 gang member (Zamora, 2017). Political discourses, magnified by popular media and pundits, often drown out the individual experiences of the Salvadoran immigrants who are being implicated in national debates. Given that hegemonic discourses hide multitudes, simplifying complex sociopolitical situations and obscuring the voices of the oppressed, literature is a valuable and necessary site to construct a fuller history of the Salvadoran-American experience.





Abrego, Leisy. (2017). “On silences: Salvadoran refugees then and now.” Latino Studies 15, 75. 76.

Abrego, Leisy. (2014). Sacrificing Families: Navigating Laws, Labor, and Love Across Borders. 15.

Abrego, Leisy, Mat Coleman, Daniel Martínez, Cecilia Mejívar, and Jeremy Slack. (2017). Making Immigrants into Criminals: Legal Processes of Criminalization in the Post-IIRIRA Era. Journal on Migration and Human Security 5, 703-704. 698.

Anderson, Benedict. (1983). Imagined Communities: Reflections on the Origin and Spread of Nationalism. 6-7.

Bourgois, Philippe. (2001). “The power of violence in war and peace: Post-Cold War lessons from El Salvador.” Ethnography 2, 9.

Cienfuegos, Lindsey. (2016). Cyclical Productions of Gang Aestheticization; Dominant and Counter-Rhetorics Surrouding the Mara Salvatrucha. M.A. dissertation, Georgetown University, 12. 23. 35. 37. 30. 40.

Dingeman-Cerda, Katie and Rubén Rumbaut. (2015). Unwelcome Returns: The Alienation of the New American Diaspora in Salvadoran Society. The New Deportation Delirium, 232. 235. 231.

Dingeman, M. Kathleen and Rubén Rumbaut. (2009). The Immigration-Crime Nexus and Post-Deportation Experiences: En/Countering Stereotypes in Southern California and El Salvador. University of La Verne Law Review 31, 387. 365. 338. 388. 376-377.

Flores, Antonio, Gustavo Lopez, and Jynnah Radford. (2017). “Facts on U.S. Latinos, 2015.” Pew Research Center.

Funes, Freddy. (2008). Removal of Central American Gang Members: How Immigration Laws Fail to Reflect Global Reality. University of Miami Law Review 63, 304. 312. 301.

García-Peña, Lorgia. (2016). Borders of Dominicanidad: Race, Nation, and Archives of Contradiction. 13.

Malkin, Michelle. (2005). Alien Gangstas Menace D.C. Suburbs. Human Events,

Obama, Barack. (2012). “Transcript of Obama’s Speech on Immigration Policy.” New York Times.

Obama, Barack.  (2014). Transcript: Obama’s immigration speech. The Washington Post,

Robbins, Liz and Nadia Rodriguez. (2017). The Gang Murders in the Suburbs. The New York Times.

Rodriguez, Ana Patricia. (2009). Dividing the Isthmus: Central American Transnational Histories, Literatures, and Cultures. 168-171. 176. 5.

Salgado, Yesika. (2017). Corázon, 4. 63.

Stanley, William. (1984). Economic Migrants or Refugees from Violence? A Time-Series Analysis of Salvadoran Migration to the United States. Latin American Research Review 22, 134. 144.

Trump, Donald. (2017). Remarks by President Trump to Law Enforcement Officials on MS-13. White House transcript of remarks.

U.S. Immigration and Customs Enforcement. (2017). Fiscal Year 2017 ICE Enforcement and Removal Operations Report. 4.

Ward, T.W. (2013). Gangsters Without Borders: An Ethnography of a Salvadoran Street Gang. 3, 13.

Zamora, Javier. (2017). Unaccompanied. 9-10.

Zamora, Javier. Unaccompanied. Port Townsend: Copper Canyon Press, 2017.

Why Bother? – The Effect of Perceived Life Control on Life Outcomes

John P. Nolan ’19

This paper provides evidence for how cultural values surrounding the perceived degree of freedom in life influence real outcomes in the market. I proxy the perception of control in life as being the perceived returns to effort and other inputs. I first demonstrate significant between-country differences in the cultural belief surrounding one’s perception of control and freedom in life. Controlling for a great deal of covariates, I then use global panel data from two waves of the World Values Survey to demonstrate a strong relationship between the perception by an individual of her control over her life, and her outcomes in the labor market. These results in the global panel are robust to controlling for a variety of individual-level characteristics as well as time-fixed effects.

In a panel of second-generation immigrants from the United States, I give evidence for cultural transmission of perceptions of personal life control affecting educational outcomes in the same manner as the global panel. I then demonstrate that the vertical transmission of this cultural value has persisting influence on outcomes for offspring in the markets for both labor and private health insurance. Results from the United States panel are robust to controlling for individual-level characteristics, as well as effects fixed for both time and geography, by county.


I. Introduction

There is a large and growing literature in economics studying the real returns in the market to the treatment of schooling. However, Jensen (2010) contends that it is not real returns that influence schooling decisions, but rather perceived returns. In his work, he demonstrates using a survey of eighth grade boys in the Dominican Republic that the perceived returns to attending secondary school are quite low, while its actual returns in the market are quite high (Jensen, QJE 2010). Over 80% of youth in the Dominican Republic complete primary school, while under 30% complete secondary school. This gap persists despite the fact in the Dominican Republic that workers completing secondary school earn on average 40% more than workers completing only primary school.

Further, his paper shows the significance of this information gap between actual and perceived returns to schooling. Randomly selected Dominican Republic schools where children were given information about the actual returns to attending secondary school observed their students receive on average between an additional 0.20 and 0.35 years of schooling over the next four years, compared to schools not receiving a shrinkage of the information gap between perceived and actual returns (Jensen, QJE 2010).

But how does this information gap persist? Agents may pass to their descendants a cultural package of expectations about their perceived returns to many elements of life. This could occur through explicitly expressed views, or by example from the agent’s demonstrated preferences in life choices surrounding education, work, and lifestyle, and these demonstrated preferences are then passed to offspring. I provide evidence for the role of culturally distinct values surrounding the perceived returns to effort and input, proxied by one’s perception of life control, in determining an individual’s real outcomes in educational attainment as well as the markets for labor and private health insurance. My thinking for this proxy is that the perceived returns to an action exist in the degree to which an individual believes that factors beyond the action in question influence his or her subsequent life outcomes. The weighting of these other factors that are not the action in question are represented in the perception of life control. My conjecture for this specific empirical pathway between perceived life control and perceived returns to effort is rudimentary and requires more research, but I do believe the intuitive connection holds generally.

This paper demonstrates that mean levels of the individual perception of personal life control are different between nations at a highly significant level. Subsequently, in a global panel I demonstrate the role that personal perception of life control plays in determining outcomes for educational attainment, controlling for individual-level characteristics as well as geography and an interaction of geography and gender. Finally, in a panel of second generation immigrants in the United States, I give evidence for the persistence of these effects of perceived life control in determining real outcomes for offspring in educational attainment as well as the markets for labor and private health insurance.

In these relationships, when approaching the highest levels of perceived personal life control, the inclusion of quadratic terms in my regressions gives strong evidence for an inverted U-curve relationship between perceived life control and the outcomes in question. This means that levels of perceived personal life control that are both relatively high and relatively low can exhibit lower perceived returns. Lower levels of perceived returns observed at relatively higher levels of perceived personal life control perhaps occur from naïveté in agents, where excessively high perceptions of personal control in determining one’s life outcomes make an individual naive or arrogant about the role of relevant factors that are not the agent herself.

In the global panel, individuals who perceive less control over their lives and fewer returns to the input of effort are less likely to complete secondary school and college. These results are robust to numerous individual-level controls and fixed effects for geography. In the United States panel, second generation immigrants whose paternal cultures perceive fewer returns to effort, as proxied by the mean level in the father’s home country, are actually more likely to complete high school, college, and graduate school. Further, second generation immigrants with this cultural heritage exhibiting lower perception of life control are actually more likely to be in the labor force, but those who are have lower total income. Additionally, when controlling for income, lower mean perception of life control in the home country of the father predicts that an individual is more likely to have private health insurance. This may result from risk aversion in the face of greater uncertainty. Though the real returns in the market are nevertheless an incredibly significant issue of economic study, policymakers should consider the implications of perceived personal life control and perceived returns to better understand how individuals make economic decisions, and these insights might more efficiently shape economic policy.


  1. Data

I combine data from waves three and four of the World Value Survey, exploiting the inclusion of the same question in both administrations of the survey. I also make use of the distinctions between them in the specific nations surveyed, allowing for a larger and more balanced global panel. The specific language of the question of interest asserts that “[s]ome people feel they have completely free choice and control over their lives, while other people feel that what they do has no real effect on what happens to them.” The respondent is then asked to respond with an indexed value from one to ten, with a reported value of ten being associated with perceiving “a great deal of freedom,” and a reported value of one being associated with perceiving “none at all.” Respondents are instructed to use the index “to indicate how much freedom of choice and control you feel you have over the way your life turns out.”

I proxy an individual’s response to this index of perceived personal life control to represent perceived returns to one’s effort and other inputs in life. Between two individuals, a relatively lower value of perceived life control indicates that a person believes his or her actions and decisions matter relatively less in determining life outcomes. Mean levels of individual perception of freedom between nations are different to a high degree of significance. Tables 1, 2, and 3 demonstrate that highly significant differences in national mean levels of perceived freedom exist even between bordering countries.

Data on individual educational outcomes and controls are from the WVS. To study cultural transmission and its effects, I use data from years 1995-2015 of the March Supplement of the Current Population Survey, exploiting its inclusion of immigrant status and paternal nativity. The CPS has rich data on individual income and labor force status, as well as health indicators and many other controls.

III. Empirical Strategy

             i. WVS

For the global panel data, I perform three sets of regressions: one respectively for each of waves 3 and 4 of the World Value Survey, and another set performed on the combination of the two waves. In all models for the global panel, I regress two respective educational outcomes of attending secondary school and attending college on an individual’s response to the index of perceived life control, as well as the squared value of that index response. I include individual-level controls for age, age-squared, and being female. For all three global panel sets, in the secondary regression on attending college I include an indicator for attending secondary school to help control for unobserved, confounding factors that might influence one’s probability of receiving any higher education whatsoever.

I include country-fixed effects. Standard errors are clustered by country. As the experience of growing up and living in respective countries is likely not homogenous between genders, I include controls for not only country-fixed effects, but also an interaction between country-fixed effects and a dummy variable set to one if the individual is female. This allows differential country-fixed effects between males and females. The intuition behind this seems solid, in that being a woman in Pakistan likely has different effects on life outcomes than being a Pakistani man. Alternatively, this may be understood as geographic conditioning of gender fixed effects, rather than gender-based conditioning of geographic fixed effects.

Religiosity and spirituality may be related to both educational attainment and one’s perception of control over his or her life. To explore this possibility, in the regressions performed on exclusively wave 3 of the WVS, I exploit the inclusion of a question about religious upbringing, and recode it to be an indicator for an individual having been raised in a religious household. In these regression for WVS3, this indicator is included as an individual control to inspect for potential omitted variable bias. The WVS does not include individual-level data for labor force participation, income, or private health insurance. As a result, in the global panel I am only able to study the relationship between perceptions of life control and educational attainment.

            ii. CPS

            Evidence in the global panel is strong but cannot rule out an issue of reverse causality. Could it be attending school that makes one feel more commanding of his or her own existence, rather than the other way around? Does higher perceived life control incentivize an individual to attend school, or does attending school empower one with a greater perception of life control? Further, the two could feed into each other, whereby those with initially high levels of perceived life control are more likely to attend school, and schooling further influences one’s perceived life control. The resulting simultaneity bias would confound the estimated causal effects of perceived life control on real outcomes.

To rule out the potentials for reverse causality and for simultaneity bias, I work in the spirit of Alesina et al. (2013), employing an epidemiological method to isolate the channel of influence concerning this cultural value. The method of this paper, too, considers second-generation immigrants in the United States. The March supplement of the Current Population Survey offers this capability by including a question for parental birthplace in addition to that of the respondent. The motivation to use these individuals is that they were born in the United States, under a fairly standard external environment of markets, institutions, laws, and policies (Alesina et al. 2013). The main distinction between them exists in the different cultural backgrounds brought from the home nations of an immigrant parent. For this secondary empirical strategy, this paper assigns to a second generation immigrant the mean value of perception of freedom of the nation of his or her father, as measured from the combination of waves 3 and 4 of the World Value Survey.

To avoid unobserved but potentially confounding elements of market factors that could exist between second-generation immigrants and those with ancestry in the United States, I only compare between second-generation immigrants. A caveat of this method is that to be included in this United States panel, the father of a second-generation immigrant must have come from a country for which data is provided on either wave 3 or wave 4 of the World Value Survey.

To explore evidence for the persistence of relationships observed in the global panel between perceived personal life control and educational outcomes, I regress the educational attainment of United States second-generation immigrants on the mean level of perceived life control in the respective home countries of their fathers, taken from the WVS. As in the WVS panel, I include not only the mean value assigned to an individual, but also its value squared. Further, in the CPS panels I include the natural log of the mean value of perceived freedom in the father’s home country in order to further explore the inverted U-curve relationship observed in the WVS global panel.

The first set of outcomes under study are dummy variables for graduating high school, completing a bachelor’s degree, and completing graduate school—whether a master’s degree, professional degree, or doctorate. I include controls for age, age-squared, gender, race, having ever been married, and an interaction of having ever been married with gender. I interact my dummy variable for gender and the indicator for having ever been married because it seems unlikely that homogenous effects exist between genders on labor market outcomes from ever having been married in the past or present. I include time-fixed effects in all regressions, and explore robustness of effects to additional fixed effects for geography by county in the United States. In the regressions on completing college, I include the dummy variable for having completed high school, and in the regressions on completing some graduate school, I include the dummies for both high school and college attainment.

Subsequently, I regress ancestral mean perception of personal freedom on outcomes of labor force participation, total income for those in the labor force, and having private health insurance. I include the original panel of individual-level controls, as well as dummy variables for graduating high school, college, and graduate school. I raise total income to the natural log as is generally standard. To study the relationship between education and cultural heritage in determining real outcomes in the market, I include interactions of an individual’s assigned value of cultural heritage with the dummy variables for educational attainment. I include time-fixed effects in all regressions, and explore robustness of effects to additional fixed effects for geography by county in the United States. I include controls for income when modeling outcomes for private health insurance. The fourth regression on having private health insurance considers only those second generation immigrants in the labor force, as a robustness check, and this eliminates over 30,000 individuals from the sample in that regression.

The cultural value in question may not itself affect individual outcomes in the market, but potentially could be picking up discrimination in the markets of the United States. In this scenario, it is not that the cultural value itself surrounding perceived freedom and returns influences individual outcomes; rather, that this person’s culture being different from that of the general mass of the United States effects discrimination for them in the market. I need to rule out the possibility that the relationship between the ancestral cultural value in question and the outcomes for education and labor and health insurance exists from this discrimination, rather than from influence on perceived returns. I include as a control the squared distance of an individual’s ancestral mean value of perceived freedom from the mean value of perceived freedom in the United States, provided by the WVS.

IV. Results

            i. Results from WVS Global Panel

In the WVS global panel for wave 3 from Table 4, an indicated increase by an individual in the indexed value of perceived freedom by one out of ten predicts that an individual is about 4% more likely to complete secondary school, indicated by the positive coefficient on the individual response. The negative coefficient on the squared term for the cultural value suggests an inverted U-curve relationship between the perception of life control and secondary school attainment. These two coefficients for the perception of freedom are each significant at the 0.001% level. The inverted U-curve relationship suggested between perception of personal life control and secondary school attainment in Table 4 indicates the greatest probability of secondary school attainment for an individual indexing his or her perception of personal life control at a value of 8.94 out of 10. For nearly identical individuals in the model who indicated respective values of 7 and 8 on the index of perceived personal freedom, we should observe the individual indexing a value of 8 to be about 1% more likely to complete secondary school than his counterpart.

These results are robust to a series of individual-level controls as well as fixed effects for geography and interactions between geography and gender. The controls for age, age-squared, and being female are all significant at the 0.001% level, while controls for having been raised religious remain significant at the 0.05% level. Many country fixed effects are significant at the 0.001% level, while many others are still notably significant though at lesser levels. The same pattern of significance occurs for the secondary country fixed effects which are interacted with gender, suggesting that allowing fixed effects for geography to be dynamic to gender picks up notable variation between individual perceived life control and secondary school attainment, and therefore may reduce omitted variable bias. In regression 2, the same increase from 7 to 8 in indexed response to perceived life control also yields nearly a 1% increase in the probability of an individual attending college. Controlling for having attended secondary school in regression 3 changes the predicted effect of the same increase in perceived life control on probability of college attainment to be positive but roughly a third of the 1% effect indicated without the control.

In the WVS global panel for wave 4 from Table 5, the same incrementing of the index of perceived life control from 7 to 8 predicts a 0.5% increase in the probability of completing secondary school. The results in WVS4 are consistent with trends observed in WVS3 between perceived personal life control and educational attainment, but the magnitude of the effect is roughly halved. Also, controlling for secondary school attainment in the second regression on college attainment makes coefficients on perceived personal freedom insignificant. These results are robust to the individual-level controls and both sets of fixed effects for geography.

In the WVS4 panel, the controls for age-squared and female are all significant at the 0.001% level, but controls for raw age in years are now insignificant. As in WVS3, in WVS4 many country fixed effects are significant at the 0.001% level, as well as many secondary geography-fixed effects dynamic to gender. In regression 2 of Table 5, the increase from 7 to 8 in indexed response to perceived life control also yields over a 0.6% increase in the probability of an individual attending college. This pattern is consistent with the results in WVS3. Controlling for having attended secondary school in WVS4, regression 3 changes the predicted effect of the same increase in perceived life control on probability of college attainment to be negligible, and these coefficients on the variable under study are no longer significant. Data are not provided on religious upbringing in WVS4.

From Table 6, the relationships observed in the full global panel combining waves WVS 3 and WVS 4 are consistent with the trends seen respectively in each of the two panel waves alone. However, the combined global dataset has a more balanced panel of countries, and roughly double the sample size of either panel wave alone. In the full global panel, an increase by one out of ten in the indexed value of perceived life control indicated predicts just under a 1% greater probability in attaining completion of secondary school. This is consistent with trends observed in both of the separate panels. Coefficients on personal perception of freedom and its indexed value squared are all significant at the 0.001% level throughout, except in regression 3 where the coefficient on perceived freedom-squared is significant at the 0.01% level. These results are robust to a series of individual-level controls as well as fixed effects for geography and interactions between geography and gender.

The control for being female is consistently significant at the 0.001% level, while controls for age and age-squared vary in significance between the 0.001% and 0.01% levels. As in both WVS3 and WVS4, in the full panel from Table 6 many country fixed effects and their gender-conditioned complement fixed effects are significant at the 0.001% level. The full panel from the World Value Survey observes that the same increase in individual response to perceived life control from 7 to 8 yields just over a 0.5% increase in the probability of an individual attending college. Including a control for having attended secondary school changes the predicted effect of the same increase in perceived life control to increase the probability of completing college by 0.03%, negligible compared to the 0.5% indicated without the control.

            ii. Results from the CPS Panel for the United States: Education

From the Current Population Survey, Table 7 demonstrates the relationship between paternal cultural heritage surrounding one’s perception of personal freedom in life and the attainment of secondary school by offspring in the United States. The values of perceived freedom assigned to an individual are the respective mean values of perceived personal freedom observed in the home country of that individual’s father in waves 3 and 4 of the World Value Survey. Consistent with the global panel relating an individual’s indicated perception of personal freedom and his or her educational attainment, in the panel on second-generation United States immigrants I observe an inverted U-curve relationship between ancestral beliefs surrounding life control and secondary school attainment in progeny in the United States.

Consistent with regressions on secondary school from the WVS panel, in the CPS the coefficient on ancestral perceived freedom is positive, while the coefficient on its squared value is negative. Perhaps surprising is that while the raw mean assigned cultural value observes a positive coefficient, the natural log of that raw mean cultural value observes a negative coefficient. All coefficients on the assigned ancestral perception of personal freedom are significant at the 0.001% level. The results for high school completion are robust to controls for age, race, and gender, as well as fixed effects for year and US county. All controls are highly significant.

Though the coefficients for high school attainment in the CPS observe the same pattern as the coefficients in the WVS panel, their distinct magnitudes produce different net effects from a cultural value surrounding perceived life control. In regression 3, incrementing mean ancestral perceptions of freedom in life from 7 to 8 out of 10 as before now yields a 9% decrease in the probability of the individual completing high school. This likely results from a greater magnitude  in the coefficient on the ancestral culture-squared term, and is consistent with a negative coefficient on the natural log of the ancestral cultural value.

Table 8 is the model regressed on the outcome of receiving a bachelor’s degree. The trends with respect to the educational outcome are consistent with those concerning high school attainment in Table 7. There is an inverted U-curve relationship between ancestral beliefs surrounding life control and college attainment in progeny in the United States. The coefficient on ancestral perceived freedom is positive, while the coefficient on its squared value is negative. The natural log of the raw mean cultural value from the father’s home country observes a negative coefficient with respect to college attainment. Controls for completing high school, age, age-squared, and being native american are all significant at the 0.001% level. The control for gender is insignificant in the regression on college attainment, and what are initially quite high significance levels for controls on being black and being white are not robust to the inclusion of county fixed effects, as seen in distinctions between regression 1 and 2 versus regressions 3 and 4. In regression 3, incrementing the ancestral cultural value from 7 to 8 as before reduces an individual’s probability of completing college by 13%, a negative net effect of greater magnitude than the effect of the same incrementing on one’s probability of completing high school.

Table 9 gives the model’s fit for the outcome of completing graduate school. In these regressions, the coefficient on the raw value of ancestral mean perceived person freedom is not nearly as significant as in previous regressions. It does become significant in regression 3 with the inclusion of county-fixed effects, but observes a surprising relationship whereupon its coefficient is negative, but the coefficient on its squared value is positive. This relationship is inconsistent with the inverted U-curve trend observed in all previous models. Instead, this relationship is a standard U-curve. In regression 3, incrementing ancestral mean perception of personal freedom from 7 to 8 predicts that an individual is just over 1% less likely to complete graduate school. These results are robust to time and county-fixed effects. Controls for completing high school, completing college, age, age-squared, and gender are all significant at the 0.001% level.


            iii. Results from the CPS Panel for the United States: Market Outcomes

Table 10 shows the relationship between culturally transmitted perceptions of personal freedom and labor force participation in second generation immigrants in the United States. Consistent with trends observed in the WVS and CPS panels on educational attainment, a positive coefficient on the raw value of ancestral perceptions of life control in conjunction with a negative coefficient on that value squared indicates an inverse U-curve relationship between the probability of one being in the labor force and the mean level of perceived personal freedom in the home country of that individual. However, the numerous interactions between educational attainment outcomes and the assigned value for ancestral perceived life control may convolute what had previously been a fairly simple trend. The large majority of these interactions are highly significant at the 0.001% level, indicating that conditioning of ancestral cultural influence and individual educational attainment has significant explanatory power for variation in labor force outcomes in the United States. Yet, the significance of these relationships between cultural perceptions of freedom and participation in the labor force is not robust to county-fixed effects. Controls for age, age-squared, gender, and being white are highly significant throughout. Upon their inclusion, controls for have ever been married and its interaction with gender are each significant at the 0.001% level.

In regression 1 from Table 10, a high school dropout would be 4% less likely to be in the labor force if their ancestral package of cultural values surrounding perceived personal freedom in life was incremented from 7 to 8 out of 10. Alternatively, the same incrementation of the assigned cultural value surrounding perception of freedom would predict that a college graduate is 2% more likely to be in the labor force when modeled using regression 1.

In regression 3 the inclusion of the measurement of cultural distance from the United States is shown to be significant at the 0.001% level, but this is not robust to the inclusion of county-fixed effects. Regression 4 demonstrates that nearly all of the effects of culture, education, and their interactions observed to be highly significant in previous regressions are not robust to the inclusion of county-fixed effects. However, notable still is the incredibly high level of significance (0.001%) for the coefficient on the interaction of the ancestral mean cultural value for perceived personal freedom with the completion of high school in predicting individual labor force participation.

Table 11 shows the relationship for those second-generation immigrants in the labor force between total income and ancestral perceptions of personal freedom in life. Regression 1 demonstrates that though my quadratic treatment of the cultural value in question is insignificant in predicting individual income, when raised to the natural log in regression 2 the data element has significant explanatory power for income raised to the natural log. Further, this relationship is robust to the inclusion of my set of individual controls, as well as robust to fixed effects for not only time but also geography by county.

As in Table 10, the interactions in Table 11 between educational attainment and the assigned ancestral cultural value under study are highly significant, and here the significance of coefficients on interactions between the cultural value and respectively completing high school and college are robust to both time and county-fixed effects. The inclusion of my measure of cultural distance in the value under study from its level in the United States does not have explanatory power in the model for income among those in the labor force. Controls for age, age-squared, gender, and both elements of the gender-dynamic control for marital status are all significant at the 0.001% level. Racial controls are largely insignificant in this model for income, but may be absorbed by assigning an individual the cultural value of his or her paternal ancestry, as this would undoubtably be correlated with ethnicity.

In the fully-controlled regression 4 in Table 11, a 1% increase in ancestral perception of personal freedom, taken from the mean value of it observed in the home country of the individual’s immigrant father, results in a 1.26% increase in total income for a high school dropout. Alternatively, incrementing the ancestral cultural value from 7 to 8 out of 10 for a black, female college graduate who is thirty years old and has been married at least once predicts a decline in total income of over $2700 in regression 4. The interactions between educational attainment and culture allow for this fairly dynamic effect in market outcomes from cultural values surrounding freedom and life control.

Table 12 shows the relationship for second-generation immigrants between being covered by a private health insurance plan and ancestral perceptions of personal control over one’s life and life outcomes. Regression 1 demonstrates that though my quadratic treatment of the cultural value in question is significant in predicting individual outcomes in the market for private health insurance, when raised to the natural log in regression 2 the data element has more significant explanatory power for the probability of being covered by private health insurance. The measure I constructed of cultural difference from the general population of the United States is also highly significant (0.01%) in predicting outcomes for private health insurance. Controls for the natural log of total income, finishing high school, finishing college, age, age-squared, gender, race, and marital status are all highly significant at the 0.001% level.

Coefficients on the interactions of the raw assigned cultural value with educational attainment respectively in high school and college are each highly significant at the 0.001% level. Except for the coefficient on college, the significances of all these controls are robust in regression 4 in Table 12 to both county-fixed effects and the choice to study private health insurance outcomes for only those individuals from the sample who are in the labor force. From regression 4, a high school dropout whose assigned ancestral value of perceived freedom incremented by one from 7 to 8 out of 10 is 14% less likely to have private health insurance. Alternatively, for a college graduate, this incrementation only reduces the probability that the individual has private health insurance by about 0.5% in regression 4.



            i. WVS

At a high level of significance, the results from the WVS indicate that individuals who exhibit a higher level of perceived control over their life outcomes also exhibit a greater probability of having attended secondary school and college. These results are robust to many individual-level controls that could affect the empirical relationship between perceived freedom and educational attainment, and they are robust to simple geographic as well as gender-conditioned geographic controls. Between individuals of the same age, sex, nationality, and religious background, those persons with higher levels of perceived personal freedom and control over their life outcomes are significantly more likely to be educated. As discussed previously, the causal channel here could exist in either direction between perceived control over one’s life and educational attainment. Also, an essential unobservable here is parental income for the individual, as family financial means might directly affect both perceived personal freedom and one’s educational attainment.

The quadratic relationship between education attainment and perceived life control indicates that there may be a degree of indicated perceived control over life outcomes that is so great as to be naive, or arrogant. Under this thinking, an individual is perhaps relatively oblivious to the structural elements influencing his or her outcomes in life. This evidence suggests that my conjecture for perceived life control proxying perceived returns to effort is fairly strong, but clearly too simplistic. The highest levels of perceived life control might actually correspond to relatively lower perceived returns in life to inputs and effort, from an agent being naive about the role of factors external to oneself in determining life outcomes.

            ii. CPS

The secondary CPS analysis isolates the causal channel by providing evidence that the cultural transmission of values surrounding perceived personal freedom exhibits lasting influence on real outcomes for offspring in educational attainment and the markets for labor and private health insurance. The same quadratic relationship observed in the global panel is found among second-generation immigrants in the United States between perceived freedom and its value squared, and completing high school, college, and graduate school. However, in the CPS panel the coefficient on the squared term of the assigned cultural value is of greater magnitude than in the WVS panel. As a result, the standard incrementation of 7 to 8 out of 10 on the index of perceived personal freedom actually predicts a lower probability of educational attainment. Nevertheless, the quadratic relationship observed in the global panel persists. Paternal transmission to offspring of the cultural values from the father’s home country explains significant variation in the educational attainment of that offspring. This causal relationship exists controlling for a consistent institutional environment, time, geography, and a number of individual-level characteristics.

I extend the pathway of cultural heritage affecting educational attainment to explore its effects on real outcomes in the markets for labor and private health insurance. By including interactions of the assigned cultural heritage value for perceived personal life control with indicators for several levels of individual-level educational attainment, I give evidence that the effects of cultural heritage in the market are dynamic to one’s level of education. Generally, the trend appears that greater levels of education dilute in magnitude the effect of cultural heritage surrounding perceived control over life outcomes.

At lower levels of education—specifically those who did not complete high school, second-generation immigrants in the United States whose parental ancestors observe greater average perceptions of life control (7 vs. 8 out of 10) are slightly less likely to be in the labor force; however, if they are, they are expected to have substantially higher income. Controlling for income, they also are much less likely to have private health insurance, perhaps from risk aversion and information gaps. This thinking is that those who are culturally conditioned to perceive relatively less control over life outcomes consequently perceive greater uncertainty in their lives, and are more likely to pursue private health insurance as a method of risk aversion.


VI. Brief Closing

This paper provides evidence for the role played by culturally transmitted values surrounding personal life control in determining an individual’s outcomes surrounding education, the labor market, and the market for private health insurance. The United States panel may observe distinct effects on educational attainment from the perception of personal freedom from those effects in the World Value Survey because of the simultaneity bias between perceived personal life control and educational attainment. In the WVS panel, individuals indicate both their own educational attainments and their perceived personal life control, whereas in the CPS I assign an individual a mean value of this perceived freedom from respondents in his or her father’s home country.

As a result, further work should explore the highly statistically significant but distinct empirical relationships in predicting educational attainment between the global panel and the panel for the United States from perceived control over one’s life outcomes. Individuals selecting their perceived level of control over life outcomes represents a different data element entirely than assigning an individual the national mean level observed from all respondents in their father’s home country; further, such distinct elements are subject to different confounding factors in the study of their effects on real outcomes. Also, future work should include the necessary unobservable in this data set which is parental income, as its addition would ameliorate bias between perceived control over life outcomes and treatments to schooling.

Additionally, the use of perceived personal freedom and life control to proxy perceived returns to one’s actions in the market represents a specification that is perhaps intuitive a priori, but which remains incredibly crude and limited. Any real connection between perceived personal life control and the perceived returns of an action requires further, more thorough research to determine.



Alesina, Alberto, Paola Giuliano, and Nathan Nunn. “On the Origins of Gender Roles: Women and the Plough.” The Quarterly Journal of Economics 128, no. 2 (2013): 469-530. doi: 10.1093/qje/qjt005.

Jensen, Robert. “The (Perceived) Returns to Education and the Demand for Schooling.” The Quarterly Journal of Economics 125, no. 2 (2010): 515-48. doi:10.1162/qjec.2010.125.2.515.