Authors: Phoebe Yi and Omkar Nayak
Humans are inherently social creatures, yet increasing rates of loneliness in the United States signal a growing societal concern. Loneliness, often linked to depression, anxiety, and heart disease, also influences political beliefs. According to Cacioppo, Cacioppo, and Boomsma’s evolutionary theory of loneliness, heightened loneliness triggers feelings of anxiety and vulnerability. This could make individuals more receptive to far-right ideologies that exploit fear of the “other” or resistance to societal change. This trend is particularly evident among young men, who have increasingly adopted far-right, gender-essentialist views, aligning with the broader “male loneliness epidemic” highlighted in the media.
To understand this phenomenon, this study examines the relationship between loneliness and political beliefs in adolescents, focusing on Social Network Size (structure) and the quality of social relationships (functional). Social Networks, combining both aspects, capture the extent and depth of an individual’s social connections. Data from the Monitoring the Future (MTF) survey, a long-running study by the University of Michigan, will serve as the basis for this analysis.
The central research question is: How does the quality and quantity of a person’s social interactions determine their Political Leaning? By leveraging MTF data and theoretical insights, this study aims to identify patterns linking loneliness indicators and social networks to political orientation. The results will inform a classification model predicting adolescents’ political leanings based on their social integration, offering insights into the relationship between social relationships and political ideology.
In order to find appropriate variables for out analysis, we needed to find the right symptoms to look for in our data. Naturally, this led us to research the symptoms of loneliness. After finding some inspiration from the litrature, we found the following variables from our data set by cross referenceing the total collecetions of symptoms we extracted to the available questions asked in the overall dataset. Finally, we found 8 relevant variables:
Because of the sheer scope and quantity of questions that the MTF survey contains, there are 7 datasets with a few overlapping questions for every year. For the purposes of this study, we needed to use variables located in two distinct datasets, but there was no guarantee that the same people who provided information for dataset 1 would also appear in dataset 6. Fortunately, every observation in the datasets had a “Responder ID.” Using the IDs of each observation, we were able to merge both datasets using an inner join (based on Responder ID) to get a merged set of people who responded to both the core questions and the extra questions relevant to this study.
df1 = pd.read_stata(form1_path)
df6 = pd.read_stata(form6_path)
df_merged = df1.merge(df6, on="RESPONDENT_ID", how="inner")
Once we merged our data, we still retained well over a thousand observations, confirming that it is a viable strategy for data collection! After we successfully combined all the data we needed into a single dataframe, we quickly noticed that Sex and Race were often redacted for security purposes. Since it made little sense to impute this information (which is generally not advised in social science research), we decided not to include these two columns in the training of our model. Additionally, by dropping the Sex and Race columns, we could focus on a specific responder’s Social Network rather than letting the model be influenced by any physiological or cultural differences between responders.
Due to this data being available to the public for academic use and the youth of the responders, many of the responses in critical variables were either unusable or redacted. We see this clearly in variables such as political leaning, where a significant portion of responses were “Unsure,” “Blank,” or “Redacted.” As a result, we dropped the observations that did not provide a direct response regarding their political leaning.
Cleaned data:
| RESPONDENT_ID | V1_x | SEX | POL_BELIEFS | MOTHR_PRES | LONELY | WISH_MORE_FRNDS | USLLY_FRNDS |
|---|---|---|---|---|---|---|---|
| 50001 | 2023 | 1.0 | 3.0 | 1.0 | 5.0 | 5.0 | 4.0 |
| 50002 | 2023 | 1.0 | 4.0 | 1.0 | 5.0 | 4.0 | 4.0 |
| 50005 | 2023 | 0.0 | 0.0 | 1.0 | 3.0 | 1.0 | 1.0 |
| 50006 | 2023 | 1.0 | 2.0 | 1.0 | 4.0 | 3.0 | 4.0 |
| 50007 | 2023 | 0.0 | 0.0 | 1.0 | 5.0 | 1.0 | 5.0 |
The dataset above is only an example of the overall dataset that is used the prediction model.
The figure above is a bar chart that shows the total number of people in each category from the survey. The bar chart reflects an intuitive result: most adolescents don’t know what their political leanings are. Another interesting observation is that the population of “Moderate” responders far eclipses the “Strong Conservative” and “Strong Liberal” populations, clearly depicting the importance of the “Undecided voter” during elections.
The survey included a vital variable that allowed people to report how lonely they felt and another variable that quantified their political leaning. With these two variables, we wanted to observe the correlation between them. To do this, we decided to use a heatmap to visually depict any possible tendencies in the data.
The heatmap suggests that the vast majority of students do not classify themselves as lonely, and they also don’t know what their political leaning is. Based on the heatmap, there are a few details to note. People who self-reported as “Strong Liberals” rarely classified themselves as “Very Lonely” (5) or “Rarely Lonely” (1). Rather, they seemed to be spread out in the middle of the “Loneliness scale.” By contrast, people in the “Strong Conservative” and “Conservative” categories tend to be more polarized, with a greater concentration in the “Not Lonely” and “Somewhat Lonely” ranges.
Considering these responders are in the 12th grade, it is reasonable to assume that the majority of their social network consists of their friends. Based on this assumption, we conducted a second bivariate analysis to explore how social network strength differed between the two sexes. We did this by isolating a variable that quantified the perceived strength of one’s social network and creating a box plot based on the number of people in each category.
Based on the figure, we find that females tended to have more established friend groups than males. This is evident from the lower fence of the Self-Reported Social Network Strength in the female category, which is the same as the median in the male category. Despite this, the upper fence for both males and females is 5. This indicates that, on average, both males and females have an established friend group by the 12th grade.
The survey included a vital variable that allowed people to report how lonely they felt and another variable that quantified the strength of a responder’s friend group. For most students in the 12th grade, it is reasonable to assume that their social network is largely grounded in their friend groups. Given these two variables, we wanted to observe the correlation between one’s perceived social network strength and their self-reported loneliness. To do this in a “Table” format, we used the pandas crosstab function to gain an empirical understanding of how social network strength and loneliness are related.
pivot_table = pd.crosstab(Alonescale, group_of_friends)
| Alonescale | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|
| 1 | 28 | 4 | 6 | 7 | 24 |
| 2 | 5 | 13 | 5 | 23 | 27 |
| 3 | 5 | 10 | 69 | 27 | 28 |
| 4 | 55 | 84 | 67 | 136 | 63 |
| 5 | 149 | 75 | 66 | 58 | 60 |
From the previous sections (especially Figure 2), it is clear that there is some correlation between how lonely people feel and their political leaning. Due to this trend, we want to explore any possible way to predict a student’s political leaning using these indicators. This naturally leads us to explore classification algorithms and how we may use multiclass classification to identify a student’s political disposition.
Formally, we are trying to use relevant variables that indicate the state of a responder’s social network, based on a survey, to train a multiclass classification model. On the micro scale, the algorithm would allow us to predict the political leaning of a single respondent, but on the macro scale, we can observe the sentiment of the entire class as a whole. In order to see how the overall sentiment of 12th graders changes over time, we must focus on the macro scale. Thus, our prediction problem is as follows: Can we predict the overall political leaning of the class of 12th graders based on each individual’s social network state?
Because the data is constructed in a way that prevents overlaps of conflicting categorical data points, we don’t need to modify our cleaned data for the purposes of the baseline model. In line with best model-building practices, we will use a 70-30 split, with 70% for training and the remaining 30% for testing. Since we are using a multiclass classification model, we will evaluate performance using accuracy and the F1 score for simplicity and clarity.
Finally, the training data for the prediction model is as follows:
Here is the data converted into an HTML table with black borders:
| SEX | POL_BELIEFS | ... | MOTHR_PRES | LONELY | WISH_MORE_FRNDS | USSLY_FRNDS |
|---|---|---|---|---|---|---|
| 2 | 4 | ... | 1 | 5 | 5 | 4 |
| 2 | 5 | ... | 1 | 5 | 4 | 4 |
| 1 | 8 | ... | 1 | 5 | 5 | 4 |
| 1 | 8 | ... | 1 | 4 | 1 | 4 |
| 1 | 1 | ... | 1 | 3 | 1 | 1 |
For the baseline Model to used a Random Forest Classifier of the 3 of the features, BR/SRinhouse , Lonely, and WishMoreFriends . Using the following three variables made sense as it focused on 3 relevant aspects of a 12th grader’s Social Network State: siblings, friends, and self perception. The RandomForest Classifier also had n_estimators be 100 to ensure that all the different combinations of trees could be used.
For the first simple model the accuracy was 0.28 or about 28% with a similar F1 score of 0.26 or about 26%. While this may seem low initially, we are trying to classify people into 5 categories. Thus random choice in our case is about 20% and we surpass that threshold. Additionally, this is evidence that the State of one’s Social Network is an indicator for their political leaning, but there needs to be more changes to the model before we can reliably make predictions using the model.
For the final model, we had to make some drastic changes. Despite our model giving us a better-than-random probability, the results were not reliable enough to warrant any meaningful utility. To address this, we explored how the RandomForest model works and identified an important feature we could leverage; Ensemble Learning. We realized that the RandomForest model is less sensitive to missing values, as it likely has trees that can derive some results even with missing categories. As a result, we changed our approach to data selection. We used all the columns that had valid inputs for the goal variable “PolBel” but included all other rows, even if their inputs were not “valid.” This significantly increased our usable dataset.
First, we wanted to focus on a responder’s social network rather than any other possible effects. For instance, whether a responder’s father or mother raised them would undoubtedly affect their political leanings. To neutralize this, we combined the “FatherPres” and “MotherPres” columns into a numeric category called “Parents_pres.” This numeric variable identified the number of parents a responder had. Additionally, we modified the “Lonely” column. Originally, we converted this categorical variable into a numeric one to process it in the decision tree, but this proved ineffective. To better highlight the influence of each loneliness category, we applied one-hot encoding.
Finally, we achieved an average accuracy of around 73%, which is a significant improvement from the 28% we were working with previously. Additionally, the F1 score increased due to the drop in the number of false negatives. These improved statistics gave us the confidence to move forward with the predictive modeling.
In essence, we have solved the micro-prediction problem discussed in Section 3. However, for the macro-decision problem, we had to conduct additional data processing. Specifically, we took the invalid responses to “PolBel” for every year and predicted the political leanings of those responders to get a more holistic view of their overall political preferences. To achieve this, we appended our predicted data for a given year to the “PolBel” series and calculated the average. This approach provided a cohesive view of the political leanings of responders for each year.
By applying this method to data from 2000 to 2023, we uncovered the following table:
| Year | Index | Mean | Accuracy | F1 |
|---|---|---|---|---|
| 2000 | 0 | 3.05 | 0.77 | 0.75 |
| 2001 | 1 | 3.03 | 0.78 | 0.77 |
| 2002 | 2 | 3.13 | 0.78 | 0.77 |
| 2003 | 3 | 3.06 | 0.71 | 0.69 |
| 2004 | 4 | 3.03 | 0.71 | 0.69 |
Finally, we answer the prediction problwem with the following graph: