Household Financial Portfolios in an Emerging Economy⁠—The Case of Chile

ABSTRACT

This paper investigates household financial portfolios in Chile. We use latent class models to identify groups of households according to their financial behavior. The model reveals nine distinct behavioral groups. The two largest groups account for 40% of the population and represent mostly households lacking access to banking sector services. Overall, we find strong evidence of households mixing assets and debt, which contradicts the classical assumptions of the life-cycle theory. We demonstrate that a significant share of indebted households has credit in the informal sector even though they were able to save on regular basis and thus should seek credit in the formal market. Education debt seems to be equally present among different socio-economic groups.

KEY WORDS:

• Chile
• financial behaviour
• financial portfolio
• latent class models

JEL:

• G11
• G02
• O12

Notes

1. Empirical support for this finding has been presented by Flavin (1981).

2. For a comprehensive list of characteristics of financially capable persons refer to Atkinson et al. (2006).

3. It stems from two factors. First, banks are reluctant to accept household credit applications if those have lower incomes. Second, less wealthy households very often lack the ability to repay their current debts, which limits their transition to a different bank, even if potentially profitable thanks to lower interest.

4. It is important to mention that we do not directly use the probabilities in the third step, as this is not advisable (see Linzer and Lewis (2011)). Instead, we use the probabilities only to draw random samples and instead of creating only one assignment to latent class in step 2, we incorporate the uncertainty into our modeling framework by assigning randomly (but according to estimated class probabilities) class membership. In the third step of our procedure, we, therefore, use the drawn class membership as dependent variable and we repeat the procedure over 100 draws.

5. We use the standard notation of the multinomial logit model: $Prob(Y_{in}=i|X)=\frac{exp\left(X{\beta }_{in}\right)}{1+{\sum }_i\mathrm{e}\mathrm{x}\mathrm{p}\left(X{\beta }_{in}\right)}$, where the index i refers to different class memberships (dependent variable), X represents the vector of covariates (covariates are either coded as continuous or categorical – see Table 2), ${\beta }_{in}$ is the vector of estimates and the index n captures the n^th replication. Note that we obtain n * (i-1) vectors of coefficients, which are subsequently combined following the multiple imputation approach.

6. The question regarding bank accounts was conditioned on having a positive balance. Consequently, only 16.6% of Chilean households reported positive balance.

7. We started with a model of only two classes, but for the sake of readability, Figure 5 only includes models with between 6 and 10 classes. All models with less classes had substantially higher BIC values.