Unemployment and Psychological Distress: An Analysis Using NHIS Data

Co-authors: Aziza Altyyeva & Ameer Abedy

Studying the relationship between unemployment and mental health is important because unemployment and poor mental health are both costly for individuals and society (Arena et al., 2023). People who are unemployed face more stress, and stress can affect families, communities, and healthcare costs (Goodman, 2015). Mental distress can cause things such as violent crime and substance abuse for individuals. This would in turn, require communities and governments to spend money on treating those social problems. Furthermore, parents who are mentally distressed and/or unemployed can cause their children to experience similar problems in the future, thus requiring more spending from governments to treat those issues and create a cycle of economic inefficiencies (Mattejat et al., 2008). This cycle of inefficiency can be particularly destructive in developing nations which are unable to cope with severe waves of decline in mental health and their effects. Our research hopes to examine the relationship between these two variables to serve as an attempt to solve this problem.

Unemployment affects more than just people’s finances, it influences how they feel, think, and function in daily life. Losing a job can lead to depression, anxiety, and a loss of purpose or self-worth. These emotional struggles harm families and communities by increasing healthcare costs, lowering productivity, and creating social strain. When people experience psychological distress, they may withdraw from social life, struggle to care for their families, or find it harder to return to work. Understanding this link is important because supporting mental health is also part of supporting a stable economy and a healthier society (Arena et al., 2023; Goodman, 2015).

Recent years have brought major changes to the U.S. job market. The COVID-19 pandemic, rising costs of living, and rapid shifts in work patterns have made job insecurity more common. Although earlier research has shown that unemployment harms mental health, little is known about how strong that link is today after the disruptions of the early 2020s. Studying this relationship in the post-covid context can provide useful evidence for policymakers and communities who aim to reduce both unemployment and its emotional costs. By examining 2024 data from the National Health Interview Survey (NHIS), this study aims to uncover how unemployment and psychological distress are connected in the current U.S. context and highlight why this issue remains a central concern for both economic and public health policy.

Literature Review The existing scholarship on unemployment and mental health consistently shows that job loss is associated with psychological distress. The severity and persistence of this relationship can vary depending on economic context, institutional support, and social factors. Yang et. al (2024) conducted a comprehensive global study, examining data from 201 countries between 1970 and 2020. Using a fixed-effects panel regression they found that unemployment significantly increases the prevalence of severe mental disorders, particularly depression, anxiety, and bipolar disorder. The study reported that a one-percent rise in unemployment correlates with a 0.01–0.014% increase in overall mental disorders. Furthermore, their results also show that while unemployment has a universal negative effect, its intensity is affected by factors such as institutional strength, globalization, and the quality of governance. Countries with robust democratic institutions and social safety nets experience weaker negative effects. This emphasizes the substantial increase in the severity of this issue for developing countries, as they lack social protection structures that mitigate the effects of mental distress.

This is further elaborated on by Houssemand and Meyers (2011) who examined unemployment and mental health in Luxembourg, a country characterized by generous welfare policies and strong economic growth. In their longitudinal study of 384 newly unemployed individuals, they measured self-esteem, psychological distress, perceived stress, and depression symptoms over a period of 12 months. Contrary to prior research, they found no significant deterioration in mental health over time, nor did mental health measures predict reemployment. The authors concluded that in a strong economy, the typical negative mental health effects of unemployment may substantially be reduced as individuals experience reduced financial strain, shorter unemployment durations, and better job prospects. Economically stable countries with better social safety nets are able to cope better with the adverse social and economic effects of unemployment than more volatile economies that lack substantive social support systems. This further highlights the motif of feedback loops.

Paul and Moser’s (2009) meta-analysis reviewed over 300 studies and found a moderate, but consistent negative effect of unemployment on mental health. Their results confirm that the unemployed report significantly higher psychological distress than the employed, though the magnitude varies across countries and time periods. Importantly, they identified social context as a key variable: strong welfare states tend to reduce, but not eliminate, unemployment’s psychological burden. Earlier meta-analyses by McKee-Ryan et al. (2005) support this view, emphasizing that well-being during unemployment is shaped by financial strain, coping resources, and the availability of social support.

Another important aspect to consider is that unemployment and preexisting economic conditions have had a compounding effect. Weich and Lewis (1998) investigated the combined impact of poverty, unemployment, and social disadvantage on common mental disorders in the United Kingdom. Using population-level data from the British Medical Journal, they showed that people experiencing both unemployment and poverty were at a much higher risk of depression and anxiety than those facing only one of these conditions. Their results underscore that unemployment interacts with economic deprivation not just on the national level, but also on personal one. This amplifies stress and limits access to mental health resources. Together, these studies show that socioeconomic context plays a key role in shaping mental health outcomes during periods of joblessness.

In conclusion, the existing literature provides strong evidence that unemployment affects mental health negatively, although the severity of the effect varies across different economic and social contexts. Across multiple empirical and meta-analytic studies, unemployment is consistently linked to higher rates of depression, anxiety, and psychological distress, driven by financial strain, social isolation, and loss of purpose. However, research also shows that these effects are not inevitable: strong welfare systems, low inequality, and favorable labor markets can substantially reduce the psychological costs of joblessness. Existing literature emphasizes that the effects of unemployment are context-specific and can be mitigated by a variety of factors. Our research considers some of these variables to determine how they affect a person’s well-being while going through a period of unemployment.

Methodology Our research question asks whether unemployment causes psychological distress. Consistent with the literature on the subject, we expect that an individual’s mental health can be influenced by several economic and social factors. This idea is summarized in the following economic model:

Psychological Distress = f [Unemployment, Male, Race, Income, Education, Marital Status, Age, Children]

The econometric interpretation is summarized in the following measurable form of the economic model:

depfreq_ord = β0 + β1 unemployed + β2 employed + β3 male + β4 black + β5 native + β6asian + β7mult_race + β8 povlev + β9 hseduc + β10mediumeduc + β11higheduc + β12partnered + β13nevermarried + β14age + β15agesq + β16hhkids1 + β17hhkids2 + β18hhkids3plus + β19unempXpovlev_c** + u**

Dependent variable To capture continuous distress, depfreq_ord , which measures how often respondents felt depressed, is used. Although depression and stress are not equivalent, they are closely connected, and in the literature, sustained stress often appears through more frequent depressive symptoms (Praag, 2009). Because the NHIS does not include a direct measure of ongoing stress, depfreq_ord serves as a practical proxy for distress in our analysis. The variable was collected for sample adults that completed the NHIS Adult Functioning and Disability Supplement. The original questionnaire associated smaller numbers with more frequent depression, however, the scale was reversed in our model to reflect higher values as indicative of more frequent depressive episodes. This makes the regression results easier to interpret as larger values consistently represent higher levels of distress.

Independent Variables Our main independent variable, unemployed, reports the primary reason a respondent was not working during the past week. The variable was created by isolating the responses captured in whynowk2, which measures the reason why someone was not working in the past week, for people who were unemployed in the economic sense. This allows us to distinguish people who were truly unemployed — laid off, or looking for work — from those who were not working for other reasons — school, disability, retirement, illness, or home responsibilities. The latter is considered as the reference group “not in the labor force (nilf)” in our analysis.

Respondents were coded as unemployed = 1 if they were jobless and actively seeking employment or laid off. All other non-working individuals were coded as unemployed = 0, since they were not part of the economically unemployed population. This approach gives a clearer and more accurate measure of unemployment.

In addition to the unemployment measure, we include employed taken from empstat, coded as 1 if the respondent was working during the past two weeks and 0 if they were not working. This is important to consider as the literature shows that employment can improve mental health as it can lead to self-reliance and self-confidence (Drake & Wallach, 2020). It is important to note that “not working” in empstat does not mean the respondent is unemployed; it simply indicates they were not engaged in paid work at the time of the interview. This group includes unemployed individuals as well as people who are retired, disabled, in school, or out of the labor force for other reasons. Including the employed variable allows for a differential capture between the mental-health of those currently working and those not working, regardless of unemployment status.

Gender is another main indicator of mental well-being. It may influence stress levels in complex ways due to societal expectations or biological factors. Studies have confirmed that whether a person is male or female has a discernable impact on how they respond to daily stressors, with males being more likely to display things such as violent behavior, drug abuse, susceptibility to infectious diseases, etc. (Verma, 2011). Gender is captured in this analysis by using a dummy variable called male, which is coded as 1 if the respondent is male and 0 if female. We expect this variable to have a positive coefficient, indicating a positive correlation with distress.

Income captures economic class, as financial struggles make stress worse for people in lower-income groups. Kessler and Cleary (1980) show that individuals with lower socioeconomic status experience significantly higher psychological distress, largely because financial strain increases exposure to stressors and reduces coping resources. Income is measured using povlev, which is a ratio of family income to the federal poverty line for that specific family size. A higher ratio means higher income relative to the poverty threshold. We expect povlev to have a negative sign since having more financial resources in the household can decrease the financial burden associated with unemployment.

Racial differences capture broader structural and social disparities, such as discrimination, unequal access to resources, or differing cultural coping mechanisms, that directly shape mental health outcomes. A study based on the National Comorbidity Survey, which collects data through a series of face-to-face interviews with respondents at their homes, found that racial minorities experience greater mental-health distress during periods of joblessness, largely due to structural disadvantages such as lower household wealth, weaker employment protections, and exposure to discrimination as well as cultural expectations (Diette et al., 2018). Furthermore, even when we disregard employment as a determinant, we see that racial minorities still experience mental disorders differently from Caucasians. According to a study published in the National Library of Medicine, racial minorities are “less likely to suffer from acute episodes of Major Depressive Disorder (MDD) than Caucasians, they are more likely to suffer from prolonged, chronic, and severely debilitating depression with heavy consequences on their level of daily functioning” (Rhan Bailey et al., 2019). A serious issue to consider when dealing with race as a factor is that people who are part of minority groups are “less likely to report psychological symptoms or remain compliant with initiated treatment” (Rhan Bailey et al., 2019).

In this analysis, race is measured by a series of dummy variables titled black, native, asian, and mult_race (multiple race). These variables are respectively coded as 1 for the specific racial minority they represent, with White being the control group. We expect all these variables to have positive coefficients.

Education is yet another main variable we consider. The literature shows that “overeducation” (having more education than your job requires) is positively associated with depressive symptoms, and that this effect is stronger in contexts with higher national unemployment rates (Dudal & Bracke, 2019). Including education in our model will help us account for differences in expectations, stress, and frustration that are not captured by income alone. Education is measured by a series of dummy variables that capture different education levels. Hseduc is coded 1 for respondents who only have a high school degree or a GED. Mediumeduc is coded 1 for respondents who have an associate’s degree or some college education. Lastly, higheduc is coded 1 for respondents who have a bachelor’s degree or higher, with the control group being respondents who have no degree at all. We expect that the coefficients of the highest education levels will be negative.

Marriage is considered as an important marginal variable in this model as it may either alleviate stress through emotional and financial support or increase it through relationship conflicts and feelings of dependency. Studies show that there are benefits to being married when it comes to mental health and happiness (Stutzer & Frey, 2005), granted that there is also evidence that people who are already happy and better adjusted are more likely to get married. Marital status is captured by including two dummy variables: partnered, coded 1 for respondents who are currently living with their partner regardless of marriage, and nevermarried, coded 1 for respondents who were never married in the first place. The control group is respondents who were previously partnered (divorced, widowed, or separated). This would allow us to evaluate stress levels of partnered adults and never-married adults relative to those who may have experienced relationship loss or disruption.

Age is another marginal factor to keep in mind. A longitudinal study found that as individuals age, the amount of stress they experience can change, with older individuals generally experiencing less stress (Almeida et al., 2023). In this study, age is a continuous variable, measured in years. A quadratic, agesq, is also considered to capture possible life-cycle effects, such as stress levels that rise during middle age and decline later in life. We expect age to have a negative coefficient.

Lastly, number of kids in the household is worth considering, as having more children could add financial and emotional pressures, but it might also offer emotional fulfillment, reducing distress. However, the literature seems to suggest that having a greater number of children has a negative impact on mental health, specifically for mothers (Pearson, 2018). The number of kids in our model is accounted by having three dummy variables: hhkids1, hhkids2, and hhkids3plus. These variables are coded as 1 for respondents who have 1 child in the household, 2 children in the household, or 3 or more children in the household respectively, with 0 children being the reference group. We expect a higher number of kids to have a positive relationship with mental distress.

Data and Summary Statistics

The data for this analysis come from the National Health Interview Survey (NHIS), which provides nationally representative cross-sectional 2024 data for U.S. adults. Each observation in the dataset represents one individual respondent. The dependent variable and all independent variables are taken from the NHIS public use files provided by IPUMS Health Surveys. After cleaning, the dataset includes about 29,318 adults aged 18 and older.

A review of the summary statistics shows that most variables display sufficient variation for regression analysis. However, several dummy variables appear infrequently in the data. For example, only about 1.5% of respondents are unemployed, 1.1% identify as Native American, and 1.5% identify as multiracial. The indicator for households with three or more children is also relatively rare, at about 4.9%. Although these categories are small, they still capture distinct subgroups and remain appropriate to include in the model. Table 1 shows the summary statistics for the variables.

Table 1: Descriptive Statistics (rounded to two decimals)

To assess potential multicollinearity, we examined pairwise correlations among all independent variables included in the econometric model. None of the coefficients approached the commonly used threshold of 0.75 in absolute value. Some moderate correlations do appear among related categories, but they are not surprising. For example, high school education (hseduc) is negatively correlated with higher education (−0.47), and medium education shows a similar negative correlation with higher education (−0.51). These patterns occur because the education categories are mutually exclusive: individuals can only belong to one group. Meaning, being in the lower-education category necessarily means not being in the higher one. A modest negative correlation between age and employed (−0.53) is also observed, which is reasonable given that older individuals are more likely to be retired and therefore not working. Additional moderate correlations include povlev and higheduc (0.44), indicating higher-income households tend to have more education, and age and never-married (−0.40), consistent with younger adults being more likely to have never married. Overall, these correlations remain well below levels that would raise concerns, indicating that multicollinearity is unlikely to be a problem in the models we estimate and that the regression coefficients should not be distorted by strong linear relationships among the predictors.

Regression 1: Main effects We ran four different regressions, each building on the one before it. Regression 1 estimates the relationship between depression (depfreq_ord) and the main explanatory variables, unemployed, employed, hseduc, mediumeduc, higheduc, male, black, native, asian, mult_race, and povlev. All education categorical variables were positively associated with depression. The signs of several coefficients were the opposite of what we expected. In particular, the estimated coefficients show that male (−0.183), black (−0.153), and asian (−0.199) individuals report lower depression frequency compared to their respective reference groups, female and white. The lower correlation coefficients for African American and Asian populations could be because racial minorities are less likely to report depression, as highlighted in Rhan Bailey et al, 2019. The R-squared for this model is 0.033, indicating that the regression explains approximately 3.3% of the variation in depression frequency.

The model is statistically significant at the 1% level, as reflected by an F-statistic of 90.65 (p < 0.001), indicating that the set of predictors jointly contribute to explaining variation in depression frequency. Several of the coefficients are individually statistically significant. At the 1% level, these are unemployed, male, black, asian, mult_race, povlev. At the 5% level, no additional variables become significant. At the 10%, native is significant, and higheduc is marginally significant with a p-value of 0.102. All remaining variables, including employed, hseduc, and mediumeduc, are not statistically significant. The regressions comparison table is shown in Table 2.

Regression 2: Extended with Marginal Controls For Regression 2, our goal is to build on the main-effects model by adding the remaining variables that may reasonably influence depression frequency. The model is expanded by adding partnered, nevermarried, age, and hhkids1, hhkids2, hhkids3plus. Regression 2 represents a more complete specification that controls for these marginal factors. Although several estimators in Regression 1 were statistically insignificant, we chose not to drop any of them because their estimates could change once marginal factors are included. If they are still insignificant in Regression 2, they will be dropped in the next regression.

In Regression 2, the direction of almost all coefficients remains the same. The only sign changes were the coefficients for hseduc and employed, which switched to negative. In the case of employed, this new sign is in line with what the literature predicted: employment improves mental health. The sign on the coefficient for the newly added variable age is negative, which is also consistent with the literature: as individuals grow older, they experience less depression symptoms and frequency. The signs for partnered and nevermarried came out to be negative. This is also consistent with the literature for partnered as we reviewed that being married and living together have benefits in terms of mental health. We do not have direct literature on never-married adults, but the result makes sense once we consider the base group. People who were never married may not face the same partnership-related stressors as those who are widowed, separated, or divorced. Finally, for the categorical household-children variables, we initially expected that having more children would be associated with higher depression. However, all child categories turned out to have negative coefficients, indicating that respondents with children reported slightly lower levels of depressive frequency compared to those without children.

Regression 2 shows an increase in R² (+0.033) and adjusted R² (+0.033). The F-statistic is still significant at 1% level, and most variables retain their individual significance. The variable employed became significant at the 1% level unlike the first regression where it was not significant. This is likely due to the addition of age, which controlled for the number of people above 65 which are eligible for retirement and represent about a third of the respondents in the dataset. The variable native, which was previously significant at the 10% level (p ≈ 0.068), is no longer significant. The variable mult_race also became insignificant. The variables that are insignificant in both regressions are hseduc, and mediumeduc. With the addition of the new marginal variables, higheduc, which was previously significant, also became insignificant. Whereas the newly added marginal variables themselves are all significant at the 1%.

In both regressions 1 and 2, hseduc displayed a p-value above 50%, whereas mediumeduc and higheducare were also insignificant in the second regression, although with p-values below 50%. We ran joint significance tests to evaluate whether the education variables should remain in the model. When hseduc, mediumeduc, and higheduc were tested together, the F-statistic was significant at the 5% level (p = 0.0394), indicating that the education variables have joint explanatory power.

The variables mult_race and native are also individually insignificant in the second regression, eliciting a joint significant test on the two, using the F-Wald test. The p-value came out to be 0.297, indicating the two do not have joint explanatory power. The same test was then ran for all the race categories. The results show that they are highly jointly significant at the 1% level, influencing the decision for their retention in the model.

Regression 3: Controls with Quadratic Adjustments For the third regression, a quadratic term, agesq, was added to allow for the possibility that the relationship between age and depressive frequency is non-linear. All variables from Regression 2 retained their signs in Regression 3 except age, and the newly added agesq showed a negative coefficient. Together with the positive sign on age, this indicates an inverted-U relationship, where depressive feelings rise slightly at younger ages and decline later in adulthood.

Regression 3 also shows a modest improvement in the model’s fit, with small increases in both R² and adjusted R². Most variables preserve their significance levels, although nevermarried and unemployed became significant at the 5% level instead of 1%. Overall, including the quadratic term provides a better description of age patterns in depressive symptoms without changing the main relationships identified in Regression 2.

Regression 4: Specification with Interaction Term An interaction term between unemployment and poverty level, unempXpovlev_c, was introduced using the mean-centered version of povlev. Centering povlev allows the coefficient on unemployed to be interpreted at the average poverty level rather than at povlev = 0, which is not a realistic point in this dataset. After adding the interaction, the coefficient on unemployed becomes statistically significant at the 1% level, showing that an unemployed person at the average poverty level reports higher depressive-symptom frequency than someone who is not in the labor force. The interaction term itself is positive but not statistically significant, suggesting that a person’s relation to poverty does not meaningfully change how unemployment relates to depressive frequency.

The signs and significance of the other variables remain stable and consistent with Regression 3. Since Model 4 provides a clearer interpretation of unemployment and strengthens its statistical significance while keeping the overall structure of the model unchanged, it is the most informative model for understanding the unemployment effect. At the same time, the R² and adjusted R² remain nearly identical to Regression 3 (about 0.069), indicating that the interaction improves interpretation but does not add explanatory power.

All regressions are compared below in Table 2:

Table 2: Regressions Comparison

Heteroscedasticity Check White’s test was used to test for heteroskedasticity. This test does not require specifying a particular pattern for the error variance and is appropriate when the structure of heteroskedasticity is unknown. The White test strongly rejects the null hypothesis of homoskedasticity (χ²(175) = 1679.17, p < 0.000), indicating the presence of heteroskedasticity in the model. The best regression is re-estimated using robust standard errors to correct this issue.

The robust version of Model 4 was examined to confirm that the main results do not depend on the assumption of equal error variance. The coefficient for unemployed became significant at the 5% level in the robust version, which gives more confidence that its effect is genuine and not driven by uneven variation in the data. The rest of the variables kept the same general significance pattern as before, and no major conclusions changed. This check shows that the main results of this analysis hold even when a more reliable method is used for estimating the uncertainty around the coefficients.

Conclusions The labor market variables (unemployed and employed) show a strong and consistent pattern across the four models. Individuals who are employed report substantially lower depressive frequency than those who are not in the labor force, and this relationship is large, negative, and highly stable in every model. Unemployment, in contrast, shows a smaller but persistent association with higher depressive frequency. Its coefficient is positive in all models, and while the strength of its significance changes slightly across models, the relationship remains statistically significant throughout. In Model 4, unemployment is significant at the 1% level, and it remains significant at the 5% level when robust standard errors are applied. This indicates that the link between unemployment and depressive symptoms is fairly reliable, though not as strong or consistent as the pattern observed for employment. Overall, employment is strongly associated with lower depressive frequency, whereas unemployment shows a more modest but still meaningful association with higher depressive frequency.

Poverty level (povlev) is one of the most consistent predictors in the model. Individuals with lower income levels report feeling depressed more often, and this relationship remains highly significant across all versions of the analysis. The strength and stability of this finding point to a strong link between financial strain and mental health.

Demographic characteristics also show important patterns. Men report fewer depressive episodes than women, and Black and Asian respondents report lower depressive frequency relative to their White counterparts. These relationships remain statistically strong after using robust standard errors, indicating that these demographic patterns are stable features of the data and not driven by model specification choices.

Relationship status, age, and household structure reveal some of the most interesting patterns in the data. Both partnered and never-married individuals report fewer depressive symptoms relative to the base group of previously partnered people. This suggests that people who are currently in a relationship, as well as those who have never been in one, tend to have better mental well-being than those who have experienced a relationship disruption. The age pattern adds another layer: depressive symptoms rise modestly in early adulthood, peak around age 23, then decline as people get older, creating a curved pattern that only appears once age is modeled nonlinearly. One of the more surprising findings comes from household children. Across all models, individuals with children report fewer depressive episodes than those without them, in contrast to what we initially expected. This recurring pattern suggests that family structure, or the presence of children specifically, may offer forms of support or connection that relate to better mental well-being.

Limitations and Future Work Mental health outcomes depend on a complex variety of factors. As the American Psychological Association notes, measures of depression, stress, and well-being are “multidimensional constructs shaped by many unobserved characteristics” (APA, 2019). As a result, it is very common in empirical studies explaining mental health outcomes to get low R² indicators. These factors include social, biological, and environmental influences. Many of them are very hard or even impossible to measure. For example, internal thought patterns and cognitive processes, as noted by Nolen-Hoeksema (2000), cannot be directly observed in survey data. A low R² does not necessarily mean the results are unimportant or the model is insufficient, it is rather a reflection of the complexity of the explained variable. As long as explanatory variables maintain statistical significance, the model provides a worthwhile assessment of whether a specific predictor or explanatory variable has a significant effect on the explained (Ozili, 2022).

Another limitation comes from how the dependent variable is measured. The variable, depfreq_ord, is based on self-reported answers about depression. Self-reports can include reporting errors, differences in how people understand questions, or mood at the time of the survey. Ryff and Keyes (1995) show that people with similar backgrounds can still respond differently when asked about their well-being. Because depfreq_ord only looks at the frequency of depression, it misses out on other important aspects that shape mental health. Future work should use a more complete measure that covers several parts of mental health. Studies can also employ mixed methods that combine survey answers with behavioral or physiological characteristics to assess and categorize individuals, as suggested by Arena et al. (2023).

Lastly, many of the key variables in the model are measured using dummy categories rather than continuous or more detailed measures. For example, employment status is reduced to “employed,” “unemployed,” or “not in the labor force,” which does not capture important differences such as job stability, part-time versus full-time work, job quality, or how long someone has been unemployed. The same issue applies to education, marital status, race, and household composition, these dummy variables summarize broad groups but lose the variation that exists within them. Future work could use datasets with more detailed information, or construct continuous variables where possible, to better reflect the complexity inside these categories.

References and Data

The studies cited in this paper along with the data used and any analysis conducted is available upon request.