Debt and Health: The Impact of Over-indebtedness on Mental Well-being in Sweden

(1)

Debt and Health: The Impact of Over-indebtedness on Mental Well-being in Sweden

By MARIA RÖNNGREN

Abstract

Household borrowing is a key element for consumption-smoothing over the life cycle.

However, over-indebtedness may induce negative health impacts through uncertainty, worries, and shame for example. This paper examines how over-indebtedness affects the mental well- being in Sweden between 2010-2018. The data is collected from several Swedish authorities at the municipal and county level. In the attempt to estimate the causal relationship between debt and health, a Bartik-like instrumental variable approach is used as an empirical strategy.

The main finding from the results is that an increase in the degree of over-indebtedness improves mental health conditions but worsen excessive alcohol consumption. Nonetheless, most of the estimates are imprecise and should not be interpreted as causal.

Keywords: debt, mental health, health behavior, instrumental variable, Bartik instrument

Master’s Thesis in Economics Uppsala University

Spring 2020

(2)

1. Introduction

Household indebtedness has been rising for a long time period and is now on its highest levels ever in many countries. Since the mid 1990’s debts have increased more than disposable income causing the debt ratio to climb above 100 in some countries, such as Sweden (Chmelar, 2013). While household borrowing is a key element in the economy allowing for consumption- smoothing over the life cycle, over-indebtedness implies higher sensitivity to changes in interest rates, which in turn jeopardize future disposable income and negatively affect people’s well-being (Keese & Schmitz, 2014). Being over-indebted in a situation where individuals are unable to manage financial commitments may lead to worries and uncertainty, shame, low relative social status, social isolation, and other negative factors (Ahlström & Edström, 2015;

Ridley et al. 2020). High repayments of debt can hence provoke stress and negatively affect the mental health as well as promote unhealthy behaviors, such as alcohol consumption (Clayton, Liñares-Zegarra & Wilson, 2015), and physical health (Keese & Schmitz, 2014; Grafova, 2007).

On the other hand, being indebted can also be a consequence of having bad health. Both sets of mechanisms can appear. Overall, the relationship between debt and health is raised as an important social issue which could have policy implications both regarding health services and financial services (Lindén, 2016; Meltzer et al., 2013).

Besides the potential simultaneous causality with debt both as a cause and a consequence of health statuses, some additional endogeneity challenges need to be addressed to identify the causal relationship between debt and health. One of them is the issue of omitted variable bias.

In general, it is difficult to control for everything that is correlated with both debt and health to isolate the relationship of interest. Measurement also poses a challenge since it is not obvious how to measure over-indebtedness or what markers of health outcomes to focus on, that may potentially capture different aspects of individuals’ true health.

In this paper, I examine the effect of over-indebtedness on mental health in Sweden between 2010-2018. To do this I use data structured as a panel over Swedish municipalities and counties.

The data include measures of debt based on people in the register of the Swedish Enforcement Authority (hereafter the SEA), which should provide a good proxy for over-indebtedness. As there is a long process for ending up with debts in that register, indicating severe financial problems (Swedish Data Protection Authority, 2020), this measure thus presumably provides an accurate and valid measure of over-indebtedness in its most serious form. The health outcomes are categorized into four types of measures; self-reported, medical prescriptions,

(4)

throughout the entire disease chain from having own symptoms to being forced into sick leave. To estimate the causal impact of debt on health I use an instrumental variable approach.

The instruments are Bartik-like using the initial distribution of debt combined with the exogenous Euro market interest rate to construct a “synthetic” path to the original debt trends over time.

The main results show that higher average amounts of debt improve mental health conditions but worsen excessive alcohol consumption. Some of the estimates are imprecise but even economically significant effects cannot be interpreted as causal considering the sensitivity analysis. However, the analysis of heterogeneous effects establishes interesting associations between debt and higher ages together with high-income regions.

My results contribute to several different literatures. First, there is a vast literature on the relationship between socio-economic status and health where especially the income-health gradient in income is well investigated. For example, both Smith (1999) and Cutler, Lleras- Money & Vogl (2011) discuss the varying income effects by age and that early childhood is the most critical period for the impact on the health stock. Overall, higher income appears to improve health outcomes. Up until recently the health effects of different low socio-economic factors were unexplored (Richardson, Elliot & Roberts, 2013). The relationship between poverty and mental health issues is investigated where causal evidence from poverty causing

depression and anxiety is found from randomized-controlled trials (e.g. Ridley et al. 2020).

Among such effects of low income and closely related employment aspects, which economists mainly have been interested in, the literature of indebtedness and health is a quite new

research area with an increasing interest (Brown, Taylor & Wheatley Price, 2005). The existing literature on this topic is spread out over several disciplines of research and the results

consistently find that high debts are associated with worse health conditions. There are studies in medical and health related areas that distinguish between different maturities of debt (Clayton, Liñares-Zegarra & Wilson, 2015), different types of debt (Meltzer et al., 2013; Jenkins et al., 2008) and whether the debt is secured or unsecured (Richardson, Elliot & Roberts, 2013;

Hojman, Miranda & Ruiz-Tagle, 2016) for example.

Similarly, within the economics literature, Brown, Taylor & Wheatley Price (2005) find that unsecured debt, such as arrears from credit cards, has a greater impact on psychological well-being than mortgages, where no effect was found for households in the UK. Using subjective measures of both debt and mental health according to the GHQ12 score (General Health Questionnaire), they estimate the effect with a commonly used ordered probit model.

(5)

Interestingly, the results are confirmed additionally through the opposed investigation of how increased savings lead to better health. Although the use of subjective measures of debt from household surveys is common in many studies, they have also limitations. Both Gathergood (2012) and Bridges & Disney (2010) discuss the threats to make inference based on self- reported financial variables in the UK since the well-being could be influenced by perception and lead to unreliable debt measures. Also, Dackehag et al. (2019) argue for a distinct

difference in using objective or subjective measures of health. In the Swedish context, they only found effects of debt on self-reported measures while objective measures from

administrative registers rather show links from bad health to later payment difficulties. This direction of the relationship is also discussed in Zimmerman & Katon (2005) after no effects of debt on depression were found. Hence, the causal direction of the correlation has been

problematic to claim in different designs. In the attempt to estimate the causal effect of over- indebtedness on health, a handful of studies use an instrumental variable approach (e.g.

Gathergood, 2012; Zimmerman & Katon, 2005). Other studies use less complicated strategies such as excluding individuals (Keese & Schmitz, 2014) or using lagged debt variables (e.g.

Dackehag et al., 2019).

This thesis presents novel evidence from the Swedish context, which is of particular interest given its low levels of income inequality (OECD, 2020). Further, I consider a more coherent analysis using comprehensive data that covers a variety of health outcomes and a reliable proxy of over-indebtedness. Lastly, to account for endogeneity concerns I exploit the method of Bartik-like instruments. This combination contributes with a new broad overview of the question of interest, which is possible when using data on the aggregate level in Sweden while most of the previous studies discussed above use individual survey data in comparison.

The rest of the thesis is structured as follows: in section 2 the data is presented and described, in the third section I explain the empirical methodology in the context of the thesis, in section 4 the results are presented together with analyses, and finally, section 5 concludes and

contains a more broad discussion about the results.

2. Data & Descriptive statistics

In this section, I describe the panel data that is used to answer the research question. The first two parts consist of detailed descriptions of the different variables. The third part summarizes

(6)

2.1 Measuring over-indebtedness

The main explanatory variables of over-indebtedness are collected from the Swedish

Enforcement Authority (2020b) at the municipality level. The process from having an invoice to pay to have debt in their register consists of several steps. To begin with, it is common to get a reminder from the creditor before the errand goes to a debt collection demand from a

collection agency. The next step is for the creditor to contact the SEA to apply for a payment injunction. In turn, the debtor is contacted by the authority to get fully informed about the injunction and time to act. If there are still no payments transmitted the authority issues a decision, a verdict. At this point, the debtor gets a record of non-payments (Swedish Data Protection Authority, 2020). To proceed with the issue, the creditor can request execution of a verdict; a request of help getting paid, and that the debt collection is now in the hands of the agency. After receiving the “Debt to Pay”-letter without taken any actions, the SEA investigates the assets of the debtor and finally moves forward with seizure orders or any other type of debt collection. Between all these steps the debtor has the chance to act, either by paying or

objection. Hence, having debt at the SEA indicates really severe financial problems and it thus seems plausible that the individual can be considered to be over-indebted (Swedish

Enforcement Authority, 2020a).

The different measures of debt that are used in my analysis are the share of debtors and the average amount of debt per person in each municipality, the share of first debtors, and the share of people being debtors for at least 20 years, so called infinite debtors. Besides the key debt variable ‘average debt amount’ measuring the degree of over-indebtedness, the data allows me to study whether there is a significant difference in being a debtor for the first time and being a debtor for a long time. All debt variables are reported by gender and the most crucial variables ‘average debt amount’ and ‘share of debtors’ are divided by age as well. This makes it possible to split the sample into subsamples and make further analyses about potential heterogeneous effects. In addition to the ideal experiment of exploiting random allocation of debts, a desirable scenario would be to have further information about the sources of debt, what the underlying nature and occasion is of the debt, for deeper analyses and to be able to rule out some of the potential simultaneous causality (Meltzer et al., 2013).

The data covers the period 2010-2018 and is restricted by the lack of data for earlier years.

Observations named “missing” or “other” are excluded from the data. These represent individuals not registered in any of Sweden’s 290 municipalities, such as foreign residents or estates, and constitutes about 25% of the total number of over-indebted each year.

(7)

Additionally, in contrast to the previous studies that use subjective measures of debt from surveys, I exploit the advantage of using an objective proxy for over-indebtedness. Since this measure of debt is not influenced by perceptions or certain conditions of the individuals, such as health statuses, it can be considered as reliable (Gathergood, 2012).

In the analysis, I include several control variables to increase the precision of the estimates;

age, gender, population size, the share of individuals with two or more years of upper

education, unemployment rate, divorce rate, the share of foreign born residents, and average disposable income. All of them are conducted by Statistics Sweden (2020).

2.2 Health outcomes

The mental health outcomes can be categorized into four subgroups of measures; self-

reported, prescriptions of drugs, diagnoses, and sickness payments from the insurance system.

The self-reported measures are collected from the Public Health Agency (2020) and their survey “The National Public Health Survey” (Nationella folkhälsoenkäten). The variables show the share of people that experienced ‘stress’, ‘reduced mental well-being’, ‘anxiety’, ‘sleeping problems’, and ‘alcohol risk consumption’. They are also in the form of moving averages in 4- year intervals, which yield five observed time periods between 2010-2018.¹ Prescriptions and diagnoses are both collected from the National Board of Health and Welfare (2020).

Prescriptions of medical drugs measure the proportion of consuming people and are

‘antidepressants’ (ATC code N06A), ‘anxiolytics’ (ATC code N05B), ‘hypnotics’ (sleeping pills, ATC code N05C), ‘alcohol abuse’ (ATC code N07BB) and ‘antibiotics’ (ATC code J01). The diagnoses are measured in a similar way where ‘alcohol related diagnoses’, ‘mood disorders’

(ICD10: F30-39), ‘stress related diseases’ (ICD10: F40-48), and further also ‘tumors’ (ICD10:

C00-D48) are included. These will capture the effect on hospitalizations due to depression and general mental illness. Hence, tumors together with antibiotics will act like a placebo in the sensitivity analysis where we do not expect to find an effect from debt. It is, however, possible that these health outcomes capture other underlying health aspects, so called “co-morbidities”.

Both the self-reported, drugs and diagnoses are measured on the county level divided by gender while drugs and diagnoses are divided by age as well.

Finally, as Dackehag et al. (2019) argue, to control for more general health issues, received benefits from the Social Insurance (2019) are included. These variables are measured by gender in the form of the share of sickness cases, both specifically for people with severe stress (ICD10:

(8)

F43), more broadly psychological illness (ICD10 F00-99), and sickness in general. A sickness case is defined as a period of continuously received compensation for work absence longer than 14 days, including both sickness, rehabilitation, and work injuries. Insurance payment for F43 and in general are measured on the municipal level, while psychological F00-F99 is on the county level. The former has some missing values for some of the municipalities and years. The reason for being missing cannot be noticed, hence they are considered to be missing at

random and not causing bias in the estimates.

In summary, the health outcomes are both subjective and objective measures on different levels of illness to capture the width of health, and to be able to distinguish between the two types of measures as Dackehag et al. (2019) show importance of.

2.3 Descriptive statistics

To get a general overview of the data, some descriptive statistics are presented in this section.

The trends of the two main explanatory variables of debt are shown in figure 2.1 below. The total number of people in over-indebtedness and their total amount of debt in MSEK are displayed with different y-axes.

In the first 4 years of the studied period, up until 2013, these two variables are following the same patterns. After that point in time, interestingly the number of indebted people starts to decrease while the debt amount continues to increase. According to the Swedish Enforcement Authority (2020c), the decrease in debtors can partly be explained by the increase in decisions of debt reliefs and a period with a booming economy and low interest rates. In such periods it is easier for people to get out of indebtedness and stabilize tough situations by starting a new job for instance. At the same time, the ones who already are having severe financial difficulties have growing debts, and as a result, the average amount per debtor has increased since 2013. Additionally, the FIGURE 2.1 Number of debtors and their total amount of debt

per year in Sweden.

(9)

number of people in over-indebtedness is quite volatile and is expected to increase again once the economy turns into a recession with high unemployment for example. With that in mind, I choose to mainly focus on the average debt amount per person as the endogenous variable.

To get a deeper insight into the regional differences that are used in the analysis, with account for the population as well, the distribution of debt changes over the studied period 2010-2018 within all municipalities is presented in a histogram in figure 2.2 below.

In figure 2.2 the distribution of changes in debt amount is shown. It is centered around slightly above 0 in log changes where most of the municipalities have increased the average amount of debt per resident. The peak is shown for near 100 municipalities to have an increase of 10-20 % between 2010-2018. Only about 25% of the municipalities perform a decline in average debt amounts. On both sides of the distribution, there are a few outliers. Lomma in Skåne län is the municipality with the largest decrease in average debt amount with -1.69 log changes,

corresponding to around 80% decrease during the studied time period. On the other hand, Tjörn in Västra Götalands län has the most increase by 177 %, almost twice the amount of debt

FIGURE 2.2 Distribution of change in debt amount within municipalities between 2010-2018 in Sweden. Bin width = 0.1 log.

(10)

per person. However, the variation between years within the time interval 2010-2018 is not captured in the figure but used in the analysis.

TABLE 2.1 Summary statistics.

Full sample Debt < Median Debt  Median

Variable Mean SD Mean SD Mean SD

Stress (%) 12.48 3.48 15.25 2.42 9.70 1.70

Reduced mental well-being (%) 12.42 3.21 14.88 2.35 9.96 1.71

Anxiety (%) 31.63 6.63 37.67 3.14 25.60 2.25

Sleeping problems (%) 33.16 5.94 38.58 2.43 27.73 2.39

Alcohol consumption (%) 15.31 4.17 11.71 1.94 18.91 2.23

Antidepressants (%) 9.24 0.92 9.20 0.82 9.29 1.01

Anxiolytics (%) 5.39 0.72 5.38 0.83 5.41 0.59

Hypnotics (%) 8.19 0.66 8.35 0.51 8.01 0.75

Alcohol abuse (%) 0.24 0.04 0.25 0.05 0.23 0.04

Antibiotics (%) 18.82 2.32 18.96 2.44 18.68 2.18

Mood disorders (%) 0.88 0.19 0.85 0.18 0.92 0.20

Stress related (%) 0.89 0.24 0.86 0.21 0.94 0.26

Alcohol related (%) 0.22 0.08 0.19 0.05 0.25 0.10

Tumors (%) 3.43 0.51 3.36 0.46 3.51 0.56

Sickness: severe stress (%) 1.03 0.54 1.03 0.56 1.03 0.51

Sickness: psychological illness (%) 2.55 1.48 3.57 1.35 1.43 0.45

Sickness: general (%) 19.67 4.96 19.29 5.02 20.05 4.88

Debtors (%) 3.70 1.12 3.11 0.77 4.30 1.10

Avg debt amount/person (KSEK) 6.58 3.14 4.67 0.82 8.48 3.43

First debtors (%) 0.72 0.16 0.66 0.13 0.78 0.17

Infinite debtors (%) 1.15 0.42 0.92 0.27 1.37 0.43

Population (log) 9.84 0.96 9.94 0.94 9.74 0.97

Education (%) 20.48 6.47 21.71 6.91 19.27 5.75

Marital status (%) 0.42 0.11 0.40 0.10 0.45 0.12

Employment status (%) 3.90 0.72 3.87 0.68 3.94 0.76

Ethnicity (%) 12.98 6.05 11.48 4.54 14.46 6.93

Disposable income/person (MSEK) 0.18 0.03 0.18 0.03 0.18 0.03

Table 2.1 Notes. Percentages of the population. Mean is based on the region for each variable; municipality or county.

FIGURE 2.3 Number of prescriptions of mental illness drugs in Sweden.

FIGURE 2.4 Number of mental illness diagnoses in Sweden.

(11)

The use of medication for mental health issues has had an increasing trend through all years since 2010. The diagnoses within the same area follow a similar pattern. This is shown above in figure 2.3 and 2.4. Clearly, the total mental well-being is worsened throughout the investigated period according to the objective health registers in Sweden. Overall, the underlying

relationship of interest between debt and mental illness is positive. In table 2.1 of the summary statistics, the mean and standard deviation of all variables used are presented. In column 1 the mean is based on all years between 2010 and 2018 and on the specific regional level each variable is measured. In line with expectations, it can be seen that the share of people

experiencing mental issues decreases from self-reported to medical prescriptions, and again to diagnoses as the level of health gets more severe. This cannot entirely be seen for the fourth category of sickness payments since these variables measure a broader sense of psychological and general health. In columns 2 and 3 the sample is divided by the median in average debt amount per person. Among the self-reported measures, it is noticed that worse health conditions are perceived among the counties with low debt amounts, except for excessive alcohol consumption. The opposite is seen for antidepressants and anxiety among the drugs and for all diagnoses, where the counties with higher debts are having worse mental health.

The pattern among sickness cases is more ambiguous with no change between debt regions in severe stress, a larger share of psychological cases in low debt regions, and a higher share of general cases in high debt regions. As different health statuses differ between low and high debt regions it is shown that health is driven by other factors as well, for which regression analysis is needed.

3. Empirical method

In the following section, the empirical strategy is explained in more detail. Further, a discussion about the identifying assumptions, potential threats and limitations are also included.

3.1 General specification

The aim of the empirical strategy is to estimate the causal effect of over-indebtedness on mental health outcomes. The general model of interest has the following form:

𝑀𝑒𝑛𝑡𝑎𝑙 ℎ𝑒𝑎𝑙𝑡ℎ_𝑖𝑡 = 𝛼 + 𝛽𝐷𝑒𝑏𝑡_𝑖𝑡 + 𝛾𝑋_𝑖𝑡+ 𝑓_𝑖+ 𝜆_𝑡+ 𝛿_𝑗𝑡𝑅𝑒𝑔 × 𝑌𝑒𝑎𝑟 + 𝜀_𝑖𝑡 (1)

(12)

𝑀𝑒𝑛𝑡𝑎𝑙 ℎ𝑒𝑎𝑙𝑡ℎ is the outcome variable for municipality or county 𝑖 and year 𝑡, and 𝐷𝑒𝑏𝑡 is the main explanatory variable leading to 𝛽 is the coefficient of interest. 𝑋 is a set of covariates, which are specified in section 2.1. 𝑓 denotes municipality/county fixed effects which capture all region specific characteristics that are constant over time and have an impact on health, such as local persistent levels of risk aversion. 𝜆 denotes time fixed effects and capture effects with an impact on health that changes over time but not across regions, such as technological changes. The interaction term between counties and time is the region-by-time fixed effect and controls additionally for specific changes in counties in different time periods in case some time effects are affecting the counties differently. This effect is only included in those

specifications estimated at the lower municipal level (where 𝑖 = municipality and 𝑗 = county), which is for two of the 15 different outcomes; sickness cases for severe stress and general cases.

When exploiting this panel structure with included fixed effects it uses the within region and time variation to rule out the potential omitted variable bias due to unobserved heterogeneity.

This relies on the identifying assumption of strict exogeneity of debt, meaning no correlation between the error term 𝜀 and debt remains. The remaining concerns regarding omitted variable bias would only be from those varying both over time and between municipalities or counties. Besides the observed characteristics included as controls, there are likely several unobserved time varying variables determining both indebtedness and mental health.

Dackehag et al. (2015) discuss for example that individual specific factors, such as financial knowledge and expectations or confidence and attitudes to debt, can explain why people run into debts. Genetics is further argued by Zimmerman & Katon (2005) to have an important role in the risk for mental disorders. This may also be related to family background, social norms, and peers in the surroundings (Dackehag et al., 2015) and hence influence both the financial situation and health status.

For these reasons, equation (1) is likely not sufficient to estimate a causal relationship between debt and health. To be able to interpret the results causally, the problems with simultaneous causality and unobserved confounders need to be addressed in some way.

3.2 Instrumental variable approach: Bartik-like

One potential way to deal with the above endogeneity problems is to instrument the measure of debt by using an instrumental variable (IV) design. Here, over-indebtedness is instrumented by using shift-share instruments, also called Bartik-instruments. This type of instrument can be considered to create a “synthetic” distribution of average debt amounts and the number of debtors. Following Boustan et al. (2013), the initial distribution of amounts of debt and

(13)

indebted people in 2010 is used to predict the distribution in subsequent years when

combining the initial distribution with national exogenous shocks in interest rates. An increase in the interest rate should intuitively lead indebted people to fall into over-indebtedness and the share of debtors in the register of the SEA increases. For those already indebted should also their debt amount increase in case of a rise in the interest rate correspondingly. By holding the local area of debt distribution fixed at 2010 in each municipality or county as the share-

variable, and then let the prediction vary based on national patterns; the shift-variable, the instrument is constructed by interacting the two variables. First, the two different share- variables (𝜃_𝑖¹, 𝜃_𝑖²) that are used are calculated as the following:

𝜃_𝑖¹ =

𝐷𝑒𝑏𝑡 𝑎𝑚𝑜𝑢𝑛𝑡𝑖,2010 𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑖,2010 𝐷𝑒𝑏𝑡 𝑎𝑚𝑜𝑢𝑛𝑡𝑆𝑊𝐸,2010

𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑆𝑊𝐸,2010

𝜃_𝑖² =

𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑑𝑒𝑏𝑡𝑜𝑟𝑠𝑖,2010 𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑖,2010 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑑𝑒𝑏𝑡𝑜𝑟𝑠𝑆𝑊𝐸,2010

𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛𝑆𝑊𝐸,2010

(2)

𝑖 = municipality or county

In the first share-variable, the initial distribution of debt amount per resident is used where each share corresponds to one municipality or county. Similarly, for the second share-variable, the distribution of the share of debtors is used. Below are the designs of the final instruments (𝑍_𝑖𝑡¹, 𝑍_𝑖𝑡²) expressed:

𝑍_𝑖𝑡¹ = 𝜃_𝑖¹× 𝐸𝑈 𝐼𝑛𝑡𝑒𝑟𝑒𝑠𝑡 𝑅𝑎𝑡𝑒_{𝑆𝑊𝐸,𝑡} 𝑍_𝑖𝑡² = 𝜃_𝑖²× 𝐸𝑈 𝐼𝑛𝑡𝑒𝑟𝑒𝑠𝑡 𝑅𝑎𝑡𝑒_{𝑆𝑊𝐸,𝑡} (3)

The first instrument in equation (3) uses the initial distribution of debt amounts together with the shift of the Euro market interest rate. The rate is measured as a yearly average of the 6 months maturity rate (Riksbanken, 2020). This is assumed to predict ‘average debt amount’, thus be the main instrument used throughout the analysis. In the same way, the second instrument uses the distribution of debtors in 2010 to interact with the Euro interest rate. This is a better predictor when the number of debtors is estimated and will only be used in the part of the analysis where the effects of first and infinite debtors are distinguished.

There are two assumptions to fulfill for instrument validity; relevance and exogeneity. The exogeneity assumption consists of two parts where the instrument is assumed to be both randomly assigned and excludable. Only the first assumption about relevance is testable through the first stage regression in the IV approach.

(14)

𝐷𝑒𝑏𝑡_𝑖𝑡 = 𝜋₀+ 𝜋₁𝑍_𝑖𝑡^𝑘 + 𝜌𝑋_𝑖𝑡+ 𝜆_𝑡+ 𝑢_𝑖𝑡 (4)

To fulfill relevance, the instrument needs to have a significant impact on the endogenous variable, hence 𝐶𝑜𝑣(𝐷𝑒𝑏𝑡𝑖𝑡, 𝑍_𝑖𝑡^𝑘) ≠ 0 in equation (4). For this reason, the instrument based on amounts are likely to be a better fit for ‘average debt amount’, while the one based on numbers

of debtors may rather predict ‘share of first and infinite debtors’. 𝑋 is the same set of covariates used in the general specification in the previous subsection.

Year dummies are also included. The results from the first stages for the main

endogenous debt variable are presented in table 3.1 to the left. The instrument is standardized to have a mean equal to 0 and a standard deviation of 1 and show positive significant coefficients in the specifications separated for both counties and

municipalities. An increase by 1 standard deviation “Bartik-shock” results on average in approximately 1600 SEK increase in debt amount per person at the municipal level and 1500 SEK at the county level. Also, the F-statistics are high enough to exceed the rule of thumb F >

10, indicating that the instrument is not weak at any regional level. Hence, the relevance assumption is fulfilled.

The other part of instrument validity, the exogeneity condition, is specified as

𝐶𝑜𝑣(𝑍_𝑖𝑡^𝑘, 𝑢_𝑖𝑡|𝑋_𝑖𝑡, 𝜆_𝑡) = 0. Conditional on the covariates and time effects, the instruments should be uncorrelated with the error term. Hence, when there are no omitted variables related to both the instrument and health, the instrument is as good as randomly assigned. Additionally, the instrument should have no direct effect on health to fulfill the excludability; the only channel to health should be through debt. By construction, the shifts in the shift-share methodology can be considered as shocks, from where exogeneity is assumed. The Euro market interest rate is induced by international macroeconomic factors, nothing the Swedish households are able to manipulate. Hence, the EU interest rate is considered to be as good as randomly assigned. It drives the Swedish market interest rate because of macroeconomic connections such as trade for instance, and is in turn affecting household debt but is not directly related to health. In addition, the interest rate is assumed to only affect health through

TABLE 3.1 First stage estimates.

Average debt amount - municipal 𝑍¹

- county 𝑍¹

1.612***

(0.184)

1.509***

(0.363)

F-statistic 76.83 17.32

No. observations 36540 2646

Table 3.1 Notes. Significance level: *** = 1%, ** = 5%, * = 10%.

Robust standard errors in parentheses. Control variables: age, gender, population, education, employment status, marital status, ethnicity, and disposable income. Year dummies are included. All regressions are average-weighted by population.

(15)

the impact on debt. No other factor than debt is supposed to be affected by the Euro interest rate and in turn, has an impact on mental well-being. Using these instruments and assuming the validity assumptions are true, then the IV-approach accounts for all omitted factors at the regional level stated in the general specification. Thus, there is no need to control for regional fixed effects. I do however include time fixed effects to increase precision.

Assuming instrument validity, the general causal estimate of interest consists of two parts; the reduced form estimate, where I regress the health outcomes on the instruments, and the first stage estimate presented above. These are expressed in equation (5) below.

𝛽_𝐼𝑉 = ^{𝑑𝐻𝑒𝑎𝑙𝑡ℎ/𝑑𝑍}^𝑘

𝑑𝐷𝑒𝑏𝑡/𝑑𝑍^𝑘 (5)

3.3 Limitations and potential threats

As mentioned in the introduction of the thesis, some potential issues have been considered and addressed in terms of identification and causality for this research question. Though, there might still be limitations in the methodology and potential threats to identify the causal effect of interest. Posing the highest threat to the instrumental variables approach, is the use of invalid instruments. Such instruments can generate more bias in the IV-estimates than in OLS where the endogeneity problem remains. Even though the instrument is relevant enough (see table 3.1) the exogeneity condition might be violated. If so, the endogeneity problems of both potential omitted time- and region-varying confounding variables and simultaneous causality would not be solved, and therefore, the IV estimates will be biased.

Generally, it is difficult to find valid instruments in practice. Since the exogeneity condition is not statistically testable, it only relies on theoretical intuition and knowledge within the context. Even though the interaction between the initial distribution of debt amounts together with the Euro interest rate should be randomly assigned to Swedish households, the most problematic part is to ensure that the Bartik instrument exclusively affects health via the measure of debt. Yet, all possible channels between the interest rate and health are hard to consider. Another limitation associated with the use of instrument, given that it is assumed to be valid, is when the estimates only reflect the average treatment effect of those who increase their debt or run into debt because of the instrument, not the overall average effect between different amounts or between indebted and unindebted. Such “LATE”-interpretation (local

(16)

The instrument is then assumed to affect the individuals in the same direction, meaning that all who are affected by an increase in the interest rate will increase their debt amounts. It is difficult to identify these people who get affected by the instrument, thus for whom the effect represents. Even though it would be possible to observe the subgroup of treated individuals, another related weakness is the difficulty to distinguish between always-takers and compliers;

the ones who always increase their debt amounts independently of the instrument and the ones who increase the amounts due to the interest rate, respectively. This is also associated with the fact that the estimates of 𝛽𝐼𝑉 may differ depending on the choice of instruments due to different specific groups affected. If another instrument would have been chosen it is possible that another subgroup of individuals would be affected, and the effects of debt on health showing different signs or magnitudes. Hence, IV estimates have strong internal validity for the specific groups but may have little external validity for the whole population.

4. Results & Analyses

Section 4 presents the results from the models specified in the previous part and my analysis of the effect of over-indebtedness on health. In the following subsections, I show results also for heterogeneous effects and robustness tests.

4.1 Main results

In the first step of the analysis an OLS specification (1), without area fixed effects and no account has taken to the bias problems, is estimated for each health outcome variable shown in table 4.1. This is a summarized table from the complete tables A1-A4 in the appendix. Robust standard errors are used in all analyses. In table A5 in the appendix a similar summary is presented but with clustered standard errors to control for potential correlation between observations within regions. However, these results do not differ remarkably and thus only the robust standard errors are used hereafter. The effects of the average amount of debt per person on the self-reported measures in panel A are mostly negative. For reduced mental well-being and risk consumption of alcohol, the effects are statistically significant at the 5% level and 10%

level, respectively. Only on sleeping problems debt show a positive impact where an increase in average debt amount of 1000 SEK leads to a 0.35%-points increase on average in the share of people using hypnotics, holding all controls constant. Further, in the second step of the

analysis, the IV regressions are estimated. These are presented in model (2). Then, all

coefficients turn insignificant besides the effect on stress. At the 5% significance level, higher

(17)

TABLE 4.1 Summarized OLS and IV estimates of the relationship between average debt amount and psychological outcomes.

Panel A. Self- reported

Stress Reduced mental well-being Anxiety Sleeping problems Alcohol consumption

(1) OLS (2) IV (1) OLS (2) IV (1) OLS (2) IV (1) OLS (2) IV (1) OLS (2) IV

Average debt amount (KSEK)

-0.078 (0.096)

1.158**

(0.510)

-0.213**

(0.085)

0.140 (0.313)

-0.144 (0.131)

0.550 (0.474)

0.354***

(0.116)

0.257 (0.420)

-0.247*

(0.143)

0.209 (0.547) R-sq.

No. observations

0.884 210

0.779 210

0.878 210

0.867 210

0.946 210

0.936 210

0.938 210

0.937 210

0.863 210

0.851 210

Panel B. Drugs Antidepressants Anxiolytics Hypnotics Alcohol abuse

(1) OLS (2) IV (1) OLS (2) IV (1) OLS (2) IV (1) OLS (2) IV

-0.297***

(0.020)

-0.436***

(0.153)

-0.173***

(0.013)

-0.291**

(0.120)

-0.294***

(0.023)

-0.637***

(0.207)

0.014***

(0.001)

0.021***

(0.007) R-sq.

No. observations

0.930 2646

0.927 2646

0.912 2646

0.908 2646

0.938 2646

0.925 2646

0.866 2646

0.860 2646 Panel C.

Diagnoses

Mood disorders Stress related Alcohol related

(1) OLS (2) IV (1) OLS (2) IV (1) OLS (2) IV

-0.024***

(0.004)

0.008 (0.029)

-0.029***

(0.005)

-0.058 (0.039)

0.023***

(0.003)

0.070***

(0.015) R-sq.

No. observations

0.667 2646

0.649 2646

0.634 2646

0.625 2646

0.791 2646

0.595 2646 Panel D. Sickness Sickness: severe stress Sickness: psychological Sickness: general

(1) OLS (2) IV (1) OLS (2) IV (1) OLS (2) IV

0.021***

(0.005)

0.135*

(0.071)

0.011 (0.030)

-0.422 (0.401)

0.078***

(0.021)

0.099 (0.112) R-sq.

No. observations

0.742 1764

0.620 1764

0.861 378

0.773 378

0.786 5220

0.786 5220 Controls

Regional dummies Year dummies

Yes No Yes

Yes No Yes Table 4.1 Notes. Significance level: *** = 1%, ** = 5%, * = 10%. Robust standard errors in parentheses. 𝑍¹ is used as instrument. Control variables included: age, gender, population, education, employment status, marital status, ethnicity, and disposable income. All regressions are average-weighted by population.

(18)

amounts of debt by 1000 SEK per person increase the share of people with high stress levels by over 1 %-point on average. In a calculation example, assume 150 000 people are living in a region. Between 2010-2018 the mean of the moving average shows that around 12.5 % were feeling stressed (see table 2.1 over summary statistics). If a 1000 SEK increase in average debts among the residents leads to about 1.2%-points increase in the share of stressed people, that means an additional 1800 people on average in that region feel problems with stress. However, the other estimates are still positive, which may indicate an overall positive relationship between amounts of debt and worse perceived mental health.

Panel B shows the impact of average debt amounts on the use of drugs for mental health issues. The effects on antidepressants, anxiolytics, and hypnotics are significantly negative at the 10% level in the first OLS estimation. When regressing with the Bartik-like instrument the coefficients remain negative and the effect gets larger in magnitude. Though, the effect on medical prescriptions for alcohol abuse is significant and positive but small in both

specifications. The IV estimate shows that if debt amounts per person increase by 1000 SEK the use of pharmaceuticals for alcoholics increases by 0.02 %-points on average, given everything else constant. In relation to the mean of alcohol abuse between 2010-2018 per region this is an increase of 9% and thus economically significant. The health outcomes of diagnoses are presented in panel C. Consistently with the drugs, all are negatively related to amounts of debt while alcohol related diagnoses are shown to have a positive relationship with higher debts.

Comparing the OLS estimates and the IV estimates, only the coefficient on alcohol related diagnoses remains significant. Similarly as before, the magnitude is larger in the IV

estimations. This is also the case in the final panel D over insurance payments from sickness cases. Both the specific cases due to severe stress and the general sickness show positive effects in the OLS regressions, while there is only on severe stress that the average amount of debt has a precise impact in IV estimation.

To summarize, the estimates in the first specification in table 4.1 establish there is a relationship between the amount of debt and mental health. In the attempt to causally interpret the results by using instruments, an effect of debt is only found for a few of the outcomes. Most of them also show the reverse sign compared to the hypothesis. However, comparing the two specifications the IV estimates are often larger in magnitude than the OLS.

One potential explanation could be measurement error in the endogenous variable. In this context, utilizing rich register data from the SEA, that does not seem probable. Another more likely reason could be that the instrument estimates the “LATE”; local average treatment effect, discussed previously in section 3.3, i.e. the average effect of debt for those who run into debt

(19)

because of the instrument. The effect is estimated for the people who increase their debt amount in the register of the SEA because of their reaction to the national trends in the interest rates. If this group of individuals, the compliers, are more responsive in terms of health changes in over-indebtedness compared to the general population, then this may explain why the IV estimates are larger than the OLS. In table 4.2, the reduced form regressions of mental health on the different Bartik-instruments are given. If the final IV estimates do not hold, it is yet interesting to study the reduced forms as they do not rely on the exclusion restriction.

Among the self-reported outcomes, the instrument has a positive effect on stress, while the other perceived health statuses are not affected significantly from the exogenous variation in the instrument. The direct effect on many of the drug outcomes in panel B is negative. An increase by one standard deviation of “Bartik-shock” in the instrument 𝑍¹ lead to a decrease in the share of people using antidepressants by 0.66%-points for example. The corresponding value on the share of anxiety drug consumption is a decrease by 0.44%-points. These declining results are consequences of the indirect effect the instrument has on the average amount of debt, which in turn affects the use of mental related drugs. Since the reduced form estimates are one part of the causal estimation, these negative effects may explain the negative effects

Table 4.2 Notes. Significance level: *** = 1%, ** = 5%, * = 10%. Robust standard errors in parentheses. Control variables included: age, gender, population, education, employment status, marital status, ethnicity, and disposable income. Year dummies are included. All regressions are average-weighted by population.

TABLE 4.2 Reduced form estimates.

Stress Reduced mental well-being

Anxiety Sleeping problems

Alcohol consumption Instrument: 𝑍¹ 0.648***

(0.151)

0.078 (0.169)

0.308 (0.232)

0.144 (0.260)

0.117 (0.284)

No. observations 210 210 210 210 210

Panel B. Drugs Antidepressants Anxiety Hypnotics Alcohol abuse Instrument: 𝑍¹ -0.657***

(0.238)

-0.438**

(0.178)

-0.961***

(0.297)

0.031***

(0.012)

No. observations 2646 2646 2646 2646

Panel C.

Diagnoses

Mood disorders Stress related Alcohol related Instrument: 𝑍¹ 0.012

(0.043)

-0.087 (0.061)

0.106***

(0.024)

No. observations 2646 2646 2646

Panel D.

Sickness

Sickness: severe stress

Sickness:

psychological

Sickness:

general Instrument: 𝑍¹ 0.109*

(0.056)

-0.285*

(0.164)

0.173 (0.203)

No. observations 1764 378 5220

(20)

found in the IV estimation in table 4.1. One potential explanation of why negative effects are found as a consequence of that the interest rate has positive effects on debt (see the first stage estimates in table 3.1), is the possibility that the excludability assumption is violated. The interest rate may influence mental health through other channels than debt exclusively. Other national macroeconomic factors, such as unemployment rates, could potentially be affected by interest rates and in turn have an impact on health.

However, besides these negative effects, there are some positive coefficients found for two of the sickness payment outcomes as well as for two of the diagnoses. Noticeable is also the positive relationship between the instrument and alcohol variables for all degrees of

categories. Just as the IV results in table 4.1, the reduced form effects show precise estimates specifically for the alcohol abuse drugs and alcohol related diagnoses.

Overall, most of the significant main results are negative. Hence, an increase in the average amount of debt per person lowers the level of bad mental well-being which is a result against the hypothesis. But the results also show how increased debt amounts lead to excessive alcohol consumption. It is rather the health behavior that is affected and changes in the lifestyle act like a consequence of increased amount of debt.

4.2 Heterogeneous effects

The advantage of having data divided by gender and age allows me to study potential heterogeneous effects of debt. Also, in addition to the main variable of debt that is used throughout the thesis, I analyze data on the share of first debtors and infinite debtors to investigate whether the period of indebtedness has a different impact on mental health

outcomes. Below, in table 4.3, these results are presented. First, I run the first stage regressions of these two endogenous debt variables on the instrument Z² for both regional levels to test for relevance. Here, it is plausible to assume Z² to predict the share of debtors rather than Z¹ using debt amounts. In the bottom right corner of the table, the estimates of the instrument show strong enough and positively significant impact on both the share of first and infinite debtors. A 1 standard deviation increase of “Bartik-shock” in the instrument results in approximately 0.1 percentage points increase in the share of first debtors and 0.15%-points increase in the share of infinite debtors, both at the county level.

The estimates in panel A on self-reported outcomes are very imprecise for the effect of first debtors. However, the effect of infinite debtors is economically significant and positive on

(21)

stress and anxiety. A 1%-point increase in the share of debtors among those who have been indebted for at least 20 years leads to an increase of near 12%-points in the share of people feeling stressed, which corresponds to an increase of 100% on average. Panel B shows the effects on medical drugs. For both first debtors and infinite debtors, there is a positive impact on the use of anxiolytics, significantly different from zero at the 10% level. Although, being a first debtor has a larger impact than being a debtor for a long time. A 1%-point increase in the share of first debtors leads to 4.3%-points increase on average in the share of people

consuming anxiolytic drugs. Further, there is also a positive impact of infinite debtors on alcohol abuse medicine. In panel C the estimates on stress related diagnoses are also shown to be positive for both types of debtors. Similarly, the magnitude of the effect is larger for the first

TABLE 4.3 IV estimates of being first and infinite debtor on psychological outcomes.

Stress Reduced mental well-being

Anxiety Sleeping

problems

Alcohol consumption First debtor (Z²) 1845.0

(34085)

172.9 (3362)

993.4 (18433)

363.2 (6814)

171.9 (3564) Infinite debtor (Z²) 11.67***

(2.862)

1.093 (2.606)

6.281*

(3.683)

2.297 (3.990)

1.087 (4.587)

No. observations 210 210 210 210 210

Panel B. Drugs Anti-depressants Anxiolytics Hypnotics Alcohol abuse First debtor (Z²) 1.415

(2.285)

4.313*

(2.540)

-0.637 (1.811)

0.196 (0.135) Infinite debtor (Z²) 0.924

(1.431)

2.817*

(1.511)

-0.416 (1.191)

0.128*

(0.075)

No. observations 378 378 378 378

Panel C.

Diagnoses

Mood disorders Stress related Alcohol related 1^st stage:

county Z² First debtor (Z²) -0.680

(0.534)

1.505*

(0.855)

0.032 (0.215)

0.097***

(0.023) [16.99]

Infinite debtor (Z²) -0.444 (0.357)

0.983*

(0.541)

0.021 (0.138)

0.148***

(0.040) [13.48]

No. observations 378 378 378 378

Panel D. Sickness Sickness: severe stress

Sickness:

psychological

Sickness: general 1^st stage: municipal Z²

First debtor (Z²) 1.740 (1.396)

-1.922 (2.377)

-4.975*

(2.947)

0.086***

(0.009) [98.37]

Infinite debtor (Z²) 1.892 (1.669)

-1.255 (1.685)

-2.129*

(1.266)

0.120***

(0.016) [153.10]

No. observations 1764 378 5220 5220

Table 4.3 Notes. Significance level: *** = 1%, ** = 5%, * = 10%. Robust standard errors in parentheses. F-statistic in brackets. Control variables included: gender, population, education, marital status, employment status, ethnicity, and disposable income. Year dummies are included. All regressions are average-weighted by population.

(22)

debtors. For the sickness outcomes in panel D, the effects of being first and infinite debtors are similar in magnitude and positive on payments received due to severe stress and negative on psychological sickness cases. Only the effects on general sickness cases are statistically significant, where being a first debtor has more than twice as large negative impact than for the infinite debtors. A 1%-point increase in first debtors lead to almost 5%-points decrease in general sickness cases on average.

Henceforth, the main results presented in the previous subsection are deeper analyzed by splitting the sample by gender, age, and income regions. In table 4.4 the effects of the average amount of debt on a smaller set of outcomes, drugs and diagnoses, are presented, separated for women and men using the instrument. The main results in table 4.1 present negative effects on antidepressants, anxiolytics, and hypnotics, while positive on alcohol abuse. The effects on antidepressants and anxiolytics are here seen to be driven by men. On the other hand, the

TABLE 4.4 IV estimates of debt amount on drugs and diagnoses separated for women and men.

TABLE 4.5 IV estimates of debt amount on drugs and diagnoses separated for young and old.

Women Men

Panel A. Drugs Anti- depressants

Anxiolytics Hypnotics Alcohol abuse

Anti- depressants

Anxiolytics Hypnotics Alcohol abuse Average debt

amount (Z¹)

-0.894 (0.559)

-0.562 (0.492)

-1.588**

(0.633)

0.049***

(0.015)

-0.168**

(0.068)

-0.132**

(0.063)

-0.237*

(0.129)

0.013 (0.008) Panel B.

Diagnoses Mood

disorders Stress

related Alcohol

related Mood

disorders Stress

related Alcohol related Average debt

amount (Z¹)

-0.025 (0.120)

-0.316*

(0.174)

0.097**

(0.050)

0.012 (0.017)

-0.022 (0.019)

0.073***

(0.015)

No. observations 1323 1323 1323 1323 1323 1323 1323 1323

Table 4.4 Notes. Significance level: *** = 1%, ** = 5%, * = 10%. Robust standard errors in parentheses. Control variables included: age, population, education, marital status, employment status, ethnicity, and disposable income. Year dummies are included. All regressions are average-weighted by population.

Age: 26-34 Age: 55-64

Panel A. Drugs Anti- depressants

Anxiolytics Hypnotics Alcohol abuse

Anti- depressants

Anxiolytics Hypnotics Alcohol abuse Average debt

amount (Z¹)

0.299 (0.772)

-0.365 (0.625)

-0.604 (0.815)

-0.057 (0.089)

-0.132 (0.177)

0.078 (0.099)

-0.224*

(0.131)

0.014 (0.013) Panel B.

Diagnoses

Mood disorders

Stress related

Alcohol related

Mood disorders

Stress related

Alcohol related Average debt

amount (Z¹)

0.192 (0.396)

-0.039 (0.372)

-0.129 (0.195)

-0.040 (0.030)

-0.110**

(0.055)

0.036*

(0.020)

No. observations 378 378 378 378 378 378 378 378

Table 4.5 Notes. Significance level: *** = 1%, ** = 5%, * = 10%. Robust standard errors in parentheses. Control variables included: gender, population, education, marital status, employment status, ethnicity, and disposable income. Year dummies are included. All regressions are average-weighted by population.

Debt and Health: The Impact of Over-indebtedness on Mental Well-being in Sweden