"Has debt", but whose debt? Homeowners, renters, and interaction
Contents
When a survey asks "Do you have any debt?", homeowners answer "yes" more often than renters. Most homeowners' debt is a mortgage, and the amounts are large. The debt that renters report, on the other hand, is more likely to be unsecured: consumer loans, credit card balances, and the like. A single variable, "has debt", therefore holds two groups whose debts are different in nature.
This article treats the situation as a question of interaction (effect modification) and uses a simple simulation to show how the conclusion changes with the way the analysis is set up.

Types of debt and mental health
Most studies of debt and health focus on unsecured debt. A systematic review and meta-analysis of 65 studies found that unsecured debt was associated with mental disorders (odds ratio 3.24), depression (2.77), and suicide (7.9), among other outcomes1. The authors noted, however, that causality is difficult to establish.
A study using the British Household Panel Survey (BHPS) found that household heads with non-mortgage debt, and larger amounts of it, were less likely to report good psychological well-being2. The study's separation of mortgage debt from other debt bears directly on the question of this article.
Mortgages matter too when they become hard to pay. Using the same BHPS data, household heads who had difficulty paying their mortgage or fell into arrears showed worse mental health, over and above the effect of general financial hardship3. In a US survey of 8,400 young adults, a high ratio of debt to assets was associated with stress, depression, worse self-rated health, and higher diastolic blood pressure4. In a survey centred on older adults in Miami, having debt was associated with depression, anxiety, and anger, and the association was explained by worry about not being able to pay the debt off5.
Taken together, these findings suggest that the type of debt and the prospect of repaying it shape its effect on health. Home ownership may be working as a marker of the type of debt.
Stating the hypothesis
HypothesisThe strength of the association between having debt and psychological distress differs between owners and renters
In epidemiology, a change in the effect of one factor across levels of another is called effect modification, and the effect of changing two factors together is called interaction6. The question here, whether the effect of debt differs between owners and renters, is a question of effect modification. In a regression model, both are represented by a product term (debt × owner).
Simulation setup
I created fictitious data on 10,000 people. Sixty percent are homeowners. Among owners, 50% have a mortgage and 8% have unsecured debt. Among renters, 25% have unsecured debt. The survey is assumed to capture only a "has debt" variable, which is 1 if a person has either kind, and the total amount of debt.
The outcome is a continuous score standing in for a psychological distress scale (K6). The data are built so that unsecured debt raises distress by 3.0 points and a mortgage has no effect on distress. Renting itself adds 1.0 point. In other words, the correct answer is known.
With this setup, 53% of owners and 26% of renters have debt. Among people with debt, the median amount is ¥17.42 million for owners and ¥0.76 million for renters. Of owners with debt, 15% have unsecured debt; the rest have only a mortgage.
Across the four combinations, only renters with debt stand out (8.0 points). Owners score in the 4s whether or not they have debt.
How the model changes the conclusion
I fitted regression models with different specifications to the same data.
- In the model with "has debt" alone, the coefficient was 0.80. Owners are more likely to have debt and have lower distress to begin with, so the harm of debt looks diluted.
- Adjusting for home ownership gave 1.28. This is still an average of the effects in the two groups.
- Adding the product term (debt × owner) gave an effect of 3.17 among renters and 0.33 among owners. The coefficient of the product term was −2.84.
The difference between the two groups becomes visible only after the product term is added. The 3.17 among renters is close to the true effect of unsecured debt, 3.0. The 0.33 among owners reflects the fact that only 15% of owners' debt includes unsecured debt.
The 1.28 from the adjusted model fits neither group. Reporting it as "the effect of debt" would understate the harm for renters and overstate it for owners.
Checking with simple slope analysis
Simple slope analysis takes a model with a product term, extracts the slope of the predictor at each value of the moderator, and tests it. The 3.17 and 0.33 above are the simple slopes of "has debt" for renters and owners.
If the model is written as "distress = b0 + b1 × debt + b2 × owner + b3 × debt × owner", the slope among renters (owner = 0) is b1 and the slope among owners (owner = 1) is b1 + b3. The standard error of b1 + b3 comes from the variance–covariance matrix of the coefficients: √(Var(b1) + Var(b3) + 2Cov(b1, b3)). The test of b3 tells you whether the two slopes differ; whether the slope among owners differs from zero is answered by the test of this simple slope.
| Group | Slope of "has debt" | Standard error | 95% CI | p value |
|---|---|---|---|---|
| Renters | 3.17 | 0.11 | 2.95 to 3.39 | <0.001 |
| Owners | 0.33 | 0.08 | 0.17 to 0.49 | <0.001 |
The slope among owners is also different from zero, but it is about one tenth of the slope among renters. This is because some owners' debt also includes unsecured debt.
Swapping the roles of predictor and moderator shows the same interaction from the other side. Among people without debt, owners score 0.81 points lower than renters; among people with debt, they score 3.65 points lower. The difference between these two gaps (−2.84) equals the coefficient of the product term. Which direction to present depends on the research question.
When the moderator is continuous, such as income or age, slopes are computed at representative values such as the mean and the mean ± 1 standard deviation, or the Johnson–Neyman technique is used to find the range of the moderator over which the slope is significant7.
Debt amount reverses the association
What happens if debt amount is used instead of whether a person has debt? Regressing distress on debt amount for everyone gave a coefficient of −0.03 per ¥1 million. The more debt, the less distress: a result that runs against intuition.
Looking by category shows why. Renters' debts range from a few hundred thousand to a few million yen, and their distress is around 8 points. Most owners' debt is a mortgage of more than ¥10 million, and their distress is in the 4s. People with larger amounts are increasingly mortgage holders, so the overall association slopes downward.
A reversal between the association within groups and the association in the pooled data is known as Simpson's paradox8. The data alone cannot tell which view is correct. Here, because we know the mechanism by which the type of debt produces distress, we can judge that the stratified view is the appropriate one.

What to do in an analysis
- Include the product term and show the effect in each group. Beyond the coefficient of the product term, use simple slope analysis to tabulate the effect and confidence interval for owners and renters separately9.
- Measure the type of debt. Asking separately about mortgages, car loans, student loans, and consumer loans removes the need to rely on home ownership as a marker.
- Look at ratios and repayment status. The ratio of debt to assets or income4 and arrears3 capture the link with distress more directly.
- Pay attention to the scale. Whether interaction is present depends on whether it is assessed as a difference in means (additive scale) or as an odds ratio (multiplicative scale). When logistic regression is used, presenting both scales is recommended6,9.
The R code for the simulation is below.
set.seed(2026)
n <- 10000
own <- rbinom(n, 1, 0.6) # 1 = owner
mortgage <- ifelse(own == 1, rbinom(n, 1, 0.50), 0) # mortgage
unsecured <- rbinom(n, 1, ifelse(own == 1, 0.08, 0.25)) # unsecured debt
debt <- as.integer(mortgage == 1 | unsecured == 1) # all the survey captures
k6 <- 4 + 1.0 * (1 - own) + 3.0 * unsecured + rnorm(n, 0, 3)
summary(lm(k6 ~ debt)) # crude association
summary(lm(k6 ~ debt + own)) # adjusted for ownership
summary(lm(k6 ~ debt * own)) # with the product term
# Simple slope analysis
m <- lm(k6 ~ debt * own)
b <- coef(m); V <- vcov(m)
b["debt"] # slope among renters
b["debt"] + b["debt:own"] # slope among owners
sqrt(V["debt", "debt"] + V["debt:own", "debt:own"] + 2 * V["debt", "debt:own"]) # its standard error
summary(lm(k6 ~ debt * I(1 - own))) # switching the reference makes the debt coefficient the slope among owners
The numbers vary slightly with the random draws. Because the data are fictitious, the size of the effects has no meaning in itself. The point is that when the content of a variable differs between groups, the conclusion depends on how the analysis is set up.
References
- Richardson T, Elliott P, Roberts R. The relationship between personal unsecured debt and mental and physical health: a systematic review and meta-analysis. Clin Psychol Rev. 2013;33(8):1148-1162. https://doi.org/10.1016/j.cpr.2013.08.009
- Brown S, Taylor K, Wheatley Price S. Debt and distress: evaluating the psychological cost of credit. J Econ Psychol. 2005;26(5):642-663. https://doi.org/10.1016/j.joep.2005.01.002
- Taylor MP, Pevalin DJ, Todd J. The psychological costs of unsustainable housing commitments. Psychol Med. 2007;37(7):1027-1036. https://doi.org/10.1017/S0033291706009767
- Sweet E, Nandi A, Adam EK, McDade TW. The high price of debt: household financial debt and its impact on mental and physical health. Soc Sci Med. 2013;91:94-100. https://doi.org/10.1016/j.socscimed.2013.05.009
- Drentea P, Reynolds JR. Neither a borrower nor a lender be: the relative importance of debt and SES for mental health among older adults. J Aging Health. 2012;24(4):673-695. https://doi.org/10.1177/0898264311431304
- VanderWeele TJ, Knol MJ. A tutorial on interaction. Epidemiol Methods. 2014;3(1):33-72. https://doi.org/10.1515/em-2013-0005
- Preacher KJ, Curran PJ, Bauer DJ. Computational tools for probing interactions in multiple linear regression, multilevel modeling, and latent curve analysis. J Educ Behav Stat. 2006;31(4):437-448. https://doi.org/10.3102/10769986031004437
- Hernán MA, Clayton D, Keiding N. The Simpson's paradox unraveled. Int J Epidemiol. 2011;40(3):780-785. https://doi.org/10.1093/ije/dyr041
- Knol MJ, VanderWeele TJ. Recommendations for presenting analyses of effect modification and interaction. Int J Epidemiol. 2012;41(2):514-520. https://doi.org/10.1093/ije/dyr218