Disentangling age and cohorts effects on home-ownership and housing wealth in Turkey

Abstract

This paper analyses the contribution of age and cohort effects on home-ownership and housing wealth in Turkey. I construct a pseudo-panel data set based on birth-year cohorts by using sixteen waves of the Turkish Statistical Institute Household Budget Surveys from 2003 to 2018. The empirical analysis reveals that young cohorts are less likely to own their homes, but they are more likely to have outstanding housing debt. Moreover, they are as willing to invest in second homes as older cohorts. I estimate a Heckman two-step selection model to distinguish the contribution of the improvement in the quality of new buildings to home values. As a result, I find that cohort effects on home values are significantly larger for young households even after controlling for age effects and the improvement in the quality of new buildings. Thus, the empirical findings show that young cohorts have a stronger housing demand than old cohorts.

Appendices

Appendix A

1.1 Home-ownership, second home investment and housing debt

See Figs. 7, 8, 9 and 10.

Fig. 7

Source: TURKSTAT Household Budget Surveys, Author’s calculations

Average cell size (Birth-year Cohorts).

Fig. 8

Source: TURKSTAT Household Budget Surveys, Author’s calculations

The probability of home-ownership.

Fig. 9

Source: TURKSTAT Household Budget Surveys, Author’s calculations

The probability of second home-investment.

Fig. 10

The probability of having outstanding housing debt

Appendix B

2.1 The contribution of quality improvements to home values

See Table 4.

Table 4 The contribution of the improvement of building quality on home values

I estimate a Heckman two-step selection model to distinguish the contribution of the improvement in the quality of new buildings to home values, while the selection criterion is home-ownership. I implement a two-stage estimation procedure as in Goodman (1988), since there is a significant degree of simultaneity between housing tenure choices and housing demand. There are 35,369 censored and 76,730 uncensored observations with a total of 112,099 household observations in the estimation. The dependent variable in the first stage of the Heckman two-step selection model is a dummy variable, D, which is 1 for households who live in their own homes and 0 otherwise (6). The probit model includes age and age-squared of the household head and dummy variables for education and income levels, employment status, employment sectors and health insurance coverage of the household head and family types, which are denoted by the Z matrix. Here, i represents household and t denotes time.

$$ D_{it} = \vartheta_{it} + \beta_{it} Z + \varepsilon_{it} $$

(6)

The dependent variable in the second stage of the estimation is home value in year-end 2003 TL prices (7). HW denotes home value as before and X represents features that might raise the value of the property such as the source of heating, construction time and the presence of an elevator in the building. As a result, the explanatory variables in the first and the second stages of the model are already different from each other. Moreover, I include time dummy variables for survey years in both stages of the Heckman two-step selection model.

The econometric results from the first stage of the Heckman two-step selection model show that the probability of home-ownership increases with age at a decreasing rate and women are less likely to be home-owners. I find that the probability of being a home-owner rises as household income level significantly increases, as expected. However, we observe that the probability of being a home-owner falls as education level increases and family size grows, except for extended traditional families that have a higher chance of owning their homes compared to a nuclear family without children (Table 4).²⁷

$$ HW_{it} = \mu_{it} + \delta_{it} X + \in_{it} $$

(7)

Empirical results clearly show that more amenities, space and features play an important role in home values. We observe that new buildings are more expensive, which is directly related to construction quality. Flats in smaller buildings are the most expensive form of accommodation. Moreover, houses that have central heating and natural gas as the source of heating are more valuable. The presence of an elevator in the apartment raises its value significantly. Houses that have larger living areas and higher number of rooms are more expensive (Table 4). Finally, time dummy variables indicate that home values increased significantly over time.

The regression coefficient of lambda from the Heckman two-step selection model is negative and statistically significant at the 1% confidence level. The predicted standard errors from this model are obtained and used as a proxy variable for home values, which are adjusted for quality growth in the next stage of empirical analysis. Intuitively, this approach is similar to the estimation of the permanent component of individual disposable income, but I utilize the residuals instead of the fitted values as if I were searching for its transitory component. The explanatory power of the second stage OLS regression is approximately 47%. As a result of that, the predicted standard errors constitute only a small fraction of home values.

Appendix C

3.1 Robustness checks

I group families into cohorts with respect to the birth year of their household heads. For this reason, the determination of household heads has great importance in the empirical analysis. According to the classification in the HBS, a family member who plays a greater role than the rest of the members in at least one issue is selected as the household head. The household head does not have to be the highest income earner in the family, but he/she is responsible for managing household income and consumption expenditures. As a robustness check, I select the highest income earner in the family as the household head and separate the sample set into cohorts based on the birth-year of the highest income earner. I perform random effects regression for unadjusted home values following the same approach from the previous sub-section (Table 5).

Table 5 Robustness checks

We observe that cohort effects acquired for both definitions of household head for unadjusted home values are very similar (Fig. 11). The gap between predicted cohort effects increases significantly as we move towards young cohorts. If I choose the highest income earner in the family as the household head, then I find that the cohort effect for households who were born in 1975 is 4.49 times larger than that of households who were born in 1945. This projection is slightly higher than the initial estimate using the definition of household heads from the HBS.

Fig. 11

Source: TURKSTAT Household Budget Surveys, Author’s calculations

The size of cohort effects (As a ratio of 1945 value).

Unfortunately, from the 2014 survey onwards, the HBS does not provide information about rural and urban households. However, location might be an important characteristic in the determination of home values. As another robustness check, I estimate the Heckman two-step selection model to find the contribution of quality growth on home values including a dummy variable for urban regions from 2003 to 2013. After that, I perform random effects regression using home values that are adjusted for higher build quality and urban regions from 2003 to 2013 following the same approach as in the previous sub-section (Table 5).

We observe that cohort effects on home values that are adjusted for quality growth and on home values that are adjusted for both quality growth and urban regions are relatively close (Fig. 12). The gap between predicted cohort effects rises as we move towards young cohorts, but its slope is considerably less steep, since I control for quality growth in home values as before. When I analyze home values adjusted for quality growth only from 2003 to 2013, I find that the cohort effect for households who were born in 1975 is 2.06 times larger than that of households who were born in 1945. However, if I analyze home values that are adjusted for both quality growth and urban regions, then I find that the cohort effect for households who were born in 1975 is 1.92 times larger than that of households who were born in 1945. This projection is actually lower than the estimate for home values when adjusted for quality growth from 2003 to 2013. I restrict the analysis period in this case to be consistent with available data on rural regions. This empirical observation suggests that living in big cities such as İstanbul, Ankara and İzmir could play a more important role on home values than the difference between rural and urban regions.

Fig. 12

Source: TURKSTAT Household Budget Surveys, Author’s calculations

The size of cohort effects (As a ratio of 1945 value).

Moreover, the empirical findings reveal that while cohort effects retained their importance, quality improvement became more widespread in the housing market in recent years. If I analyze unadjusted home values only from 2003 to 2013, then I find that the cohort effect for households born in 1975 is 4.42 times larger than that of households born in 1945. However, this ratio was previously estimated as 4.3 times for a longer sample period from 2003 to 2018. I discover that this downward trend is preserved, when I analyze home values adjusted for quality growth for a longer time period. In this case, the ratio of cohort effect for households born in 1975 to that of households born in 1945 drops from 2.06 to 1.55 when I extend the sample period from 2013 to 2018 as discussed previously (Figs. 6 and 12).

In this paper, I analyze home values, since I assume that housing wealth is a suitable proxy variable for housing demand, as in Mankiw & Weil (1989). However, housing wealth is a stock variable, while housing demand is a flow variable. Moreover, Deaton & Paxson (2000) focus on household consumption and saving, which are flow variables. I restrict the sample to homes purchased between 2003 and 2018 to remedy this situation, since the HBS is available for this period. If I analyze unadjusted home values, then I find that the cohort effect for households born in 1975 is 3.93 times greater than that of households born in 1945. If I perform the same empirical analysis using home values adjusted for quality growth, then I find that the cohort effect for households born in 1975 is only 1.4 times greater than that of households born in 1945. Thus, I find that the econometric results from the restricted sample are very close to the initial empirical findings from the whole sample.