Is Technology Becoming Essential for Employment?- 8 mins
Last November, two of my peers, Noah Sebek and Yiran Xu, and I wrote a paper called Predicting Employment through Internet, Social Media, and Gaming Activity: Comparing CART and Random Forests about the performance of CART and Random Forests. In the context of predicting employment status of US residents through survey data we hypothesized that with the growing use of technology in employment applications, attitudes about technology would strongly predict employment status. As the project unfolded, however, we placed a strong emphasis on the machine learning aspect of the problem instead of letting the data guide us–in retrospect, this makes me uncomfortable. As a result, I decided to write a new version of this paper. While the conclusions of the new analysis are similar, the analysis more rigorous in my opinion. This is my first post in my effort to work on independent data analysis projects more often, so I hope you enjoy!
Is Technology Becoming Essential for Employment? Technology vs. Demographics in Employment
Analysis of survey responses to questions regarding technology and employment suggest three notable ideas: disabilities continue to hinder employment despite regulations for accomodations, given US policies on maternal leave mothers are more likely to leave the workforce after their first child-birth, and people who use the internet for job searches are more likely to have higher wages. Demographic information continues to affect employment status more despite the increasing usage of technology for employment.
The 2008 Financial Crisis was the worst financial crisis since The Great Depression, decreasing employment by 8.8 million in just 14 months1. The US has since dropped its overall unemployment rate to 4.5% in March, 20172. Knowing what factors decrease unemployment rate helps the government enact social policies more effectively. Given the increased use of the internet for job postings and applications, the resurgence of the employment rate can be partially attributed to the development of technological resources and use by employers. A natural conclusion is that those who are more comfortable with technology can benefit more greatly from these resources than those who are not. We use the Pew Research Center’s Gaming, Jobs, and Broadband survey data to examine whether this conclusion holds true or not.
Cleaning the Data
In the survey, 2001 people answered 140 questions about their demographics, personal information, and attitudes about gaming, internet, and technology. The data-cleaning process mainly involved merging conditional questions and removing variables correlated or associated with employment status. We reduced the variable for employment status to
Unemployed by removing students and those have have already retired. Since the frequency of underemployment/part-time employment was low, we merged those reponses with the
Employed factor. We recognize, however, that underemployment/part-time employment can reveal more details that we can not otherwise uncover. In the analysis of this dataset, we use the term
Unemployed loosely to imply lack of employment instead of the technical definition of unemployment because the survey did not distinguish lack of a job from unemployment. The final dataset consisted of 1395 observations and 45 features.
Building Random Forests
We trained a Random Forest model with a 10-fold repeated cross validation, repeated 3 times, on a randomly sample of 80% of the data and 43 non-id features. The model produced 84.96% accuracy on the training data, and more importantly 83.73% accuracy on the testing data (area under the ROC curve: 0.8069). The model’s variable importance showed that the features with the largest weights were
Internet Usage for Job Search,
Parental Status, and
The results are inconsistent with our hypothesis that attitudes and activity of technology are strong predictors of employment. We analyze more closely how some of the variables of highest importance are related to employment.
Accomodations for persons with disability remain insufficient in the US
Approximately a quarter (24.2%) of people in the weighted data identified as being unemployed. However, 71% of people with disabilities were unemployed. In contrast, only 17% of people without any disabilities were unemployed. Only 13% of people in the survey had disabilities, but they comprised of almost 40% of the unemployed group.
Title I of the Americans with Disabilities Act Amendments Act (ADAAA), the updated version of the Americans with Disabilities Act, requires employers to provide reasonable accomodations to qualified applicants or employees, with the exception of companies with fewer than 15 employees3. The data suggests that either accomodations are insufficient, employers are not complying with this law, or people with disabilities are unqualified for most available jobs. We find it highly unlikely that people with disabilities are unemployable. Furthermore, based on the income distribution (this feature was not included in the Random Forest), people with disabilities consistently represent a larger proportion of the lower income class and smaller proportion of the higher income class.
According to the US Census Bureau, nearly 1 in 5 people in the US have a disability. Surely, better accomodations and/or stronger regulations to provide accomodations for people with disability are necessary. The fact that disability status is the strongest indicator of employment status in our model (or any resonable model, for that matter) is highly problematic.
Gender roles pertaining to parenting may drive women out of the workforce
The current maternity leave policy in the US, the Family and Medical Leave Act, (FMLA) provides most US employees with upto 12 weeks of unpaid maternal leave4. In contrast, in 2015, Slovak Republic offer up to 160 week of paid maternity or paternity leave5. Among the 35 Organisation for Economic Co-operation and Development (OECD) countries, the US is the only country with no paid maternity leave. Exploring employment status by sex and age, we find that the only deviation of pattern for all factors is the relative increase in unemployment for women between ages 26 to 30 while unemployment for men in the same age group decreased.
In 2014, the average age of mothers at first birth was 26.3 years6, where women’s unemployment rate increases in our data. This could suggest that women’s drop in employment may be attributed to the lack of support for maternity leave. Exploring how unemployment differs between men and women who are or are not parents, the men and women who are not parents share similar unemployment distributions.
In contrast, mothers have an unemployment distribution more concentrated from ages 25 to 50, a common age group to care for children. While we lack sufficient information to make stronger claims about child care needs and employment, we suspect that providing paid maternity and/or paternity leave would increase mothers’ employment rate.
Technology used for employment purposes affect higher income jobs more than lower income
In additional to technological advances contributing to the US’s record-low unemployment rates since the 2008 Financial Crisis, those who use technology and the internet for assistance for job searches tend to have higher salaries. While the income distribution for those who do not use technology for job searches is skewed to the right, the distribution for those who do is skewed to the left.
We suspect that the ability to use the internet for job searches is closely associated with having more technical skills. Since jobs with more technical skills tend to have higher wages, this result is not surprising.
Demographics are still better predictors of employment than technology
While technology seems to have some indication of employment status and income level, demographic information still predicts employment status more robustly. Given the abundance of information in the survey about different topics, further research can reveal relationships between features not discussed here.
Code for Cleaning and Analysis
Please feel free to follow the code for this analysis! https://github.com/leejunta/Employment
1 Bureau of Labor Statistics (Apr 2017). Labor Force Statistics from the Current Population Survey. Retrieved from: https://data.bls.gov/timeseries/LNS14000000
2 Bureau of Labor Statistics (Apr 2011). Employment loss and the 2007–09 recession: an overview. Retrieved from: https://www.bls.gov/mlr/2011/04/art1full.pdf
3 ADA National Network. What is the Americans with Disabilities Act (ADA)?. Retrieved from: https://adata.org/learn-about-ada
4 United Census Bureau (Jul 2012). Nearly 1 in 5 People Have a Disability in the U.S., Census Bureau Reports. Retrieved from: https://www.census.gov/newsroom/releases/archives/miscellaneous/cb12-134.html
5 OECD (Apr 2016). PF2.5. Trends in parental leave policies since 1970. Retrieved from: http://www.oecd.org/els/family/PF2_5_Trends_in_leave_entitlements_around_childbirth.pdf
6 Centers for Disease Control and Prevention (Jan 2016). Mean Age of Mothers is on the Rise: United States, 2000–2014. Retrieved from: https://www.cdc.gov/nchs/products/databriefs/db232.htm