Exploring Income Patterns- A Comprehensive Analysis of the UCI Adult Census Dataset
Abstract
This project explores income patterns using the UCI Adult Census Dataset from 1994. This analysis aims to uncover intricate relationships between various demographic factors as well as looks at the impact that hours worked have on income bracket and marital status.
1. Introduction
The UCI Adult dataset, also known as the “Census Income” dataset, is a collection of data aimed at predicting whether an individual’s income exceeds $50,000 per year based on census data. This dataset was donated to the UCI Machine Learning Repository on April 30, 1996, and has been a significant resource for research in social science, particularly in the field of machine learning.
1.1. Data Characteristics
• Multivariate Nature: The dataset is multivariate, which includes multiple variables or features for analysis.
• Subject Area: It falls under the domain of social science.
• Associated Tasks: Primarily used for classification tasks.
• Feature Types: Includes both categorical and integer data types.
• Number of Instances: Contains 48,842 instances or records.
• Number of Features: Comprises 14 features
1.2. Datapoints Overview
The dataset includes various demographic and employmentrelated features. Some of the key features are:
• Age: A continuous variable representing the individual’s age.
• Workclass: Categorical variable indicating the employment status (e.g., private, federal government,etc.)
Education: Includes both a categorical variable indicating the level of education (e.g., Bachelors, HS-grad) and a continuous variable representing the number of years of education.
Marital Status: Categorical variable indicating marital status.
Occupation: Categorical variable describing the individual’s occupation.
Relationship: Categorical variable indicating family relationship status.
Race: Categorical variable indicating race.
Sex: Binary variable indicating gender.
Capital Gain and Loss: Continuous variables representing income from investments.
Hours per Week: Continuous variable indicating the number of hours worked per week.
Native Country: Categorical variable indicating the country of origin
2. Project Question Goals
What is income bracket distribution across different demographics like age, gender, education level, and occupation?
This question seeks to understand how income levels are distributed among different demographic groups based on age, gender, education level, and occupation. The motivation here is to uncover patterns or disparities in income distribution, which could be critical for economists. Analyzing these factors can help understand economic inequality dynamics and various demographic factors’ role in income determination.
Does the number of work hours per week impact income significantly?
This question explores the relationship between labor input and financial compensation by exploring whether the number of work hours per week significantly impacts income. This is important for understanding market dynamics, especially the valuation of labor in terms of hours worked. It can provide insights into whether working longer hours correlates with higher income and might reveal cases where this isn’t true.
Is there a correlation between hours worked, marital status, and income bracket?
This question combines three variables to explore their interrelationships. The motivation here is to investigate how personal life and professional life interplay. This can offer insights into a work-life balance and its economic implications.
Overall, answering these questions using the dataset can provide a deeper understanding of socio-economic patterns and their implications, which is essential for developing targeted policies, business strategies, and further academic research.
3. Tools Used
Tableau
An intuitive and versatile tool for data visualization, was the primary tool used to conduct this visual analysis. The initial stage of the project involved a comprehensive data preprocessing and refinement phase dedicated to ensuring the dataset’s integrity and accuracy.
Python
We used python for tasks such as managing missing data and standardizing data formats, establishing a solid foundation for seamless visualization within Tableau’s environment. This groundwork enabled the subsequent phase of detailed exploration and visualization creation.
4. Data Visualization & Analysis
4.1. What is income bracket distribution across different demographics like age, gender, education level, and occupation?
Figure 1 - Break down of respondents by income bracket
As displayed in Figure 1, more than 75% of the dataset is individuals with income below 50K (<=50K) & less than 25% of the dataset is individuals with income above 50K (>50K). This might be due to the average average income in 1994 stood at $26,939 according to the Bureau of Labor Statistics.
Note: From here onwards the Income below 50K will be denoted as <=50K & income above 50K will be denoted as >50K.
Figure 2 - Distribution by age
The distribution plot in Figure 2 clearly illustrates a significant concentration of individuals aged 18 to 25 with incomes <= $50K. This trend reflects the early stage of their careers. However, as individuals progress in age from 25 to 50, there’s an exponential rise in the count of those earning >$50K. This shift likely corresponds to increased experience and career advancement during this period.
Figure 3- Distribution by sex
The bar graph in Figure 3 reveals a notable discrepancy: there were approximately six times as many males earning >$50K compared to females in the same income bracket. This disparity might be due to the gender biases & the glass ceiling prevalent during the 1990s.
Figure 4 - Income by highest level of education
Figure 4 shows a correlation between education level and income level as education level increases, and the count of individuals with income >50K also increases. For example, individuals with the highest level of education as Professional school or Doctorate have more than 90% of individuals in that group who have income >50K. This correlation might be due to the skill sets, expertise, and specialized knowledge that accompany higher levels of education. Individuals with advanced education often qualify for professions that command higher salaries, thus leading to an increase in the number of individuals earning above the $50K.
Figure 5 - Income by Race
Figure 5 reveals a notable discrepancy: there were approximately 23 times as many whites earning >$50K compared to blacks in the same income bracket. This disparity might be due to the racial biases & the glass ceiling prevalent during the 1990s.
4.2. Does the number of weekly work hours impact income significantly?
Figure 6 - Income by Occupation
We first sought to explore the income distribution when it came to different occupations. Figure 6 illustrates skill based or managerial occupations have higher numbers of individuals with income >50K as observed in occupations like Craft Repair, Sales & Prof - Specialty. This might be due to the higher compensation associated with roles that require specialized skills, managerial responsibilities, or expertise within these occupational categories. Occupations demanding specific technical prowess or positions involving managerial duties often command higher salaries, resulting in a greater number of individuals within these categories earning above $50K.
Figure 7 - Avg weekly hours by income
Figure 7 demonstrates a direct relationship between the number of hours worked per week and income levels: as the weekly hours increase, so does the income. Moreover, ignoring the outliers, individuals who worked more than or equal to 40 hours had income >50K. This is due to the structure of many job roles and labor practices. Full-time positions, typically requiring 40 hours per week or more, often come with higher wages or salaries.
4.3. Is there a correlation between hours worked, marital status, and income bracket?
Figure 8 - Avg weekly hours by marital status and income
Figure 8 & 9 suggests a correlation between marital status and average hours worked. Those married, especially in a civilian (CIV) capacity or with an armed forces (AF) spouse, tend to work longer hours on average than other marital statuses. This might be due to family responsibilities or shared financial goals. What’s particularly interesting is the behavior of individuals who’ve never married. Despite their younger age, they exhibit a lower average hour worked compared to other marital statuses. Understanding the rationale behind this could shed light on diverse priorities, career ambitions, or a distinct approach to work-life balance among this demographic.
Figure 9
Figure 8 reaffirms the findings of motivating question 2, that higher the number of hours worked weekly, higher the income level. Across the marital statuses there is a clear gap in average hours worked weekly between the income groups. What’s particularly interesting is the consistent trend observed across all groups: individuals surpassing the average hours worked within their respective groups tend to have an income >50K.
This striking consistency underscores a significant pattern: those who dedicate more hours to work than their peers within the same marital status consistently tend to earn above 50K annually. This suggests a strong correlation between higher average work hours and higher income across the board.
5. Results
Based on this analysis, we extract the following key points:
As the age increases the income also increases until the age 50
Males tend to earn more than their female counterparts.
Whites earn more than any other race
Higher the education level, higher the income
Individuals with specialized skills, managerial, or expertise within a specific field have higher income.
There is a strong correlation between average hours worked weekly and income.
Individuals who have never married, despite their young age & start of their career tend to work for less number of hours weekly.
6. Discussion
One additional note that should be made about our findings is that they show a potential skew in the UCI Adult data we are working with. Our dataset has far more data points for younger white males than other demographic groups. While this may not fully reflect US demographics, it is also a potential source for skewing within the data and, therefore, our findings.
7. Conclusion
The UCI Adult dataset, a foundational resource in machine learning, predicts income levels exceeding $50,000 annually based on demographic and employment data. A recent analysis by Ding, Hardt, Miller, and Schmidt at NeurIPS 2021 highlighted limitations in the dataset, particularly the potential skewing effect of the binary income threshold on fairness analyses for specific demographic groups. Their study underscores the significant impact of income thresholds on fairness criteria violations, urging a reconsideration for a more comprehensive portrayal of income distributions. It is this background context through which we tried to answer multiple questions about our data. As discussed in the results section, we found several factors that seemed to impact income bracket distribution, such as age, gender, education level, occupation, number of hours worked, and marital status. While no definitive conclusions can be made, especially given the age and context surrounding the dataset, we still find that these visualizations provide meaningful and exciting discoveries.
Looking ahead, future studies could delve deeper into the nuanced interplay between demographic factors and income bracket distributions. Exploring these factors in conjunction with other societal dynamics may unravel deeper insights into income disparities and their relationship with socio-economic variables. Additionally, considering the evolving societal landscape, investigating temporal shifts in these relationships across different eras or within diverse cultural contexts could provide a more comprehensive understanding. Furthermore, employing advanced analytical methodologies and machine learning models may offer enhanced predictive capabilities while addressing fairness concerns associated with income threshold categorizations.