Skills: Data Cleaning, Visualization, Pivot Tables, Regression
The objective of this project is to understand the association between an individual's ownership of specific types of financial products (e.g. savings account, health insurance) and their 'financial wellbeing', a measure of the extent to which one's financial situation provides security and freedom. In addition to exploring whether their is a correlation between ownership of certain types of financial produts and financial wellbeing, I will also investigate whether their might be a causal effect on financial wellbeing (e.g. owning a savings account 'causes' an individual's financial wellbeing to improve) by controlling for confounding variables.
As a measure financial wellbeing, I use The Consumer Financial Protection Bureau's Financial Well-Being Scale. The Financial Well-Being Scale scores individuals from 0-100 based on "the extent to which someone’s financial situation and the financial capability that they have developed provide them with security and freedom of choice". For more information on the scale, see the CFPB website.
The financial products under consideration are:
- Savings account
- Health insurance
- Retirement savings account
- Pension
- Investment
- Education savings account
- Student loan
- Financial products do not explain much of the variation in financial wellbeing among respondents. By themselves, most products explain less than 10% of the variation in financial wellbeing scores. However, the effect of products on financial wellbeing scores is significant.
- Retirement and investment accounts had the greatest (positive) effect on financial wellbeing
- Student loans, which has a negative effect of financial wellbeing, followed close behind
- The impact of all products reduces once income, employment status, education and health are accounted for. Yet, the effect of the financial products on the financial wellbeing score remains significant.
Python Version: 3.7
Packages: pandas, numpy, statsmodels, matplotlib, seaborn
I use data from the CFPB's 2016 financial wellbeing survey, downloaded from the CFPB website. The survey dataset includes respondents’ scores on the CFPB Financial Wellbeing Scale, the financial products that they use as well as other measures of individual and household characteristics that research suggests may influence adults’ financial well-being (e.g. income and employment, education).
The National Financial Well-Being Survey was conducted in English and Spanish via web mode between October 27, 2016 and December 5, 2016. Overall, 6,394 surveys were completed: 5,395 from the general population sample and 999 from an oversample of adults aged 62 and older.
After retrieving the CFPB survey data, I clean the data to improve the accuracy of my findings. Changes made to the raw dataset include:
-
Removing irrelevant columns: The survey collects 217 characteristics of each respondent that go far beyond their financial wellbeing scores and the financial products that they own. To speed up and smooth out analysis, I remove columns with information that is not essential to the current question. The columns that are kept are as follows:
- Financial Wellbeing Score: The outcome of interest
- Financial Products: Whether the respondent owns one of the listed financial product types
- Controls: As I am interested in not just observing the correlation between product ownership and financial wellbeing, but also investigating causal effects, I included several characteristics that might be correlated with both financial wellbeing and product ownership. Controlling for these will help unearth causal relations. The characteristics I include are: income, employment status, education, financial skill and health status.
-
Removing Junk Values: Several columns included 'junk values' that represented a respondent's data being unavailable. After checking that only around 200 out of 6394 responses included junk values I decide to remove rows inclduing junk values, since little data (relative to the size of the compelete dataset) is lost as a result.
-
Assigning categories to numeric labels: Based on CFPB's provided Codebook, I replace the numeric labels used to code education, employment and income status with their associated categorical values. This makes creating dummy variables for each category easier down the line.
-
Re-labeled columns for easy comprehension
To gain an initial understanding of the data, I create several visualizations. For the numerical data, I create histograms to understand their distribution. For the categorical data, I create bar charts for the categorical data to understand the balance of classes.
Here are two examples of the visualizations included:
In addition, I transform the categorical variables into dummy variables. Using these, I create pivot tables that, for each product type, show the mean and medians associated with owning and not owning the product type. I add a third row that calculates the difference between the median and the mean for each product type, to see which products are associated with the greatest differences in financial wellbeing.
A short code like this can be used to create informative pivot tables:
for i in df_products.columns:
table = pd.pivot_table(df, values=['financial_wellbeing'], index=[i],
aggfunc=[np.mean,np.median])
table.loc['difference'] = np.abs(table.loc[0].sub(table.loc[1], fill_value=0))
print(table)
print()
| Retirement Account | Mean | Median |
|---|---|---|
| Yes | 60.1 | 50 |
| No | 50.4 | 60 |
| Difference | 9.7 | 10 |
| Investment Account | Mean | Median |
|---|---|---|
| Yes | 63.7 | 63 |
| No | 52.4 | 52 |
| Difference | 11.2 | 11 |
Retirement accounts are associated with the greatest difference in means, while investment accounts see the greatest difference in medians.
I begin my regression analysis by running univariate regressions of the financial wellbeing score on each product type. The outcome of thsi regression indicates whether owning certain types of financial products is associated with differences in financial wellbeing.
Using statsmodels, one can quickly run a univariate regression by using the following type of code:
for i in df_products.columns:
X = df_products[i]
y = df['financial_wellbeing']
X = sm.add_constant(X)
est = sm.OLS(y, X).fit()
print(est.summary())
I add visualizations of each regression and create a table with the R-squared value of each model to easily compare fit across models.
An example of a boxplot used to visualize one of the univariate regression models:
The results of the univariate regression are as follows:
| Product | Coefficient | R Squared | p-value |
|---|---|---|---|
| Savings Account | 8.6 | 0.046 | 0.00 |
| Life Insurance | 4.8 | 0.030 | 0.00 |
| Health Insurance | 6.9 | 0.048 | 0.00 |
| Retirement Account | 9.8 | 0.118 | 0.00 |
| Pension | 9.0 | 0.092 | 0.00 |
| Investment Account | 11.1 | 0.136 | 0.00 |
| Education Savings Account | 5.22 | 0.008 | 0.00 |
| Student Loan | -6.1 | 0.023 | 0.00 |
There are several things to note about these results:
- The R Squared values for most products are low, meaning that the associated univariate regression model explains little of the variation in financial wellbeing. The model that includes investment accounts explains the most variation, around 13.6%.
- While the explanatory power of the models is low, the p-value indicates that for all financial products the effect on financial wellbeing is statistically significant
- The products that seem to have the greatest impact on financial wellbeing are: investment accounts, retirement accounts and pensions
- Student loans are the only product type that has a negative effect on financial wellbeing. This is unsurprising as it is the product type that represents a liability.
- The R squared of the model including education savings accounts is very low. A contributing factor might be that the data is unbalanced, with very few respondents owning education savings accounts.
Before adding the control variables, I examine the effect that the control variables individually have on financial wellbeing. Running univariate regressions of the financial wellbeing score on each control variable. The results of these regressions include:
| Product | Coefficient | R Squared | p-value |
|---|---|---|---|
| Income | 0.0(0008) | 0.131 | 0.00 |
| Health | 4.5 | 0.244 | 0.00 |
| Financial Skill | 0.6 | 0.244 | 0.00 |
| Product | R Squared |
|---|---|
| Education | 0.071 |
| Employment | 0.092 |
For each dummy variable representing a type of employment or education level, p < 0.5.
These results indicate that the control variables individually explain some, but not a lot, of variation in financial wellbeing scores. Health and financial skill perform the best in this regard. In addition, for all control variables, the effects on financial wellbeing are significant.
While significance or strong explanatory power are not required for control variables to be included (we are interested in their effects when our variables of interest are taken into account, after all), these results indicate that our control variables might help improve the explanatory power of the model regressing financial wellbeing on product types.
Finally, I run a multivariate regression of financial wellbeing on each product type plus the explanatory variables:
| Product | Coefficient | Adjusted R Squared | p-value |
|---|---|---|---|
| Savings Account | 3.4 | 0.427 | 0.00 |
| Life Insurance | 1.3 | 0.422 | 0.00 |
| Health Insurance | 2.0 | 0.423 | 0.00 |
| Retirement Account | 4.4 | 0.438 | 0.00 |
| Pension | 3.2 | 0.49 | 0.00 |
| Investment Account | 4.3 | 0.435 | 0.00 |
| Education Savings Account | -0.2 | 0.419 | 0.70 |
| Student Loan | -4.1 | 0.430 | 0.00 |
There are several interesting changes that occur when control variables are added:
- The coefficients for all product types decreases
- The explanatory power (R squared) of the models improves (more than 40% of the variance can know be explained by most models)
- In our new models, student loans have a greater effect on financial wellbeing than pensions, savings accounts and health insurance
- The effect of ownership of each product type remains significant, except for education savings accounts (again, this could be due to unbalanced data).
This straightforward regression analysis does by no means complete the examination of the effect of financial products on financial wellbeing. There are several ways in which the modelling might be improved to allow for more accurate results and a correct characterization of causal influence. Some of these include:
- Adding additional control variables: Financial wellbeing is a multi-dimensional concept that is likely influenced by many factors. Therefore, to be confident of the causal effect of different product types on financial wellbeing, more control variables would likely have to be added to account for confounding factors that our small set of control variables does not measure.
- Adding Interaction Terms: I did not check for interaction effects between products and control variables. If it is the case that the impact of product ownership on financial wellbeing varies for different levels of the control variables, introducing interaction terms would improve the moel.
- Accounting for Multi-colinearity: I did not check for multi-colinearity (with, for example, a VIF test). There is chance that multi-colinearity is distorting are results as it might be the case that multi-colinearity occurs between some of the product variables and the control variables (e.g. income might be correlated with whether one has an investment account). While there might also be multi-colinearity between the control variables, this is less of a concern, as it does not affect the values we are interested in.