Estimating Households’ consumption for surveys of income and wealth |
|
| Chair | Dr Junyi Zhu (Bundesbank ) |
Adjustable rate mortgages (ARMs) made a dramatic entry into the U.S. mortgage market subsequent to the passage of the Garn - St. Germain Depository Institutions Reform Act of 1982. By the mid 1990’s over 20 percent of first mortgages were adjustable and were used across a wide range of urban market types. Going forward to 2007 the percent of families with an ARM as a first mortgage had declined to 11% and the use of ARM’s was often then described as a tool for affordability – since the rates are commonly lower than for a fixed rate mortgage for a given repayment risk class of borrower. By 2013 the percent of ARMs as a first mortgage had fallen still further to just under 9%, and the use of an ARM for a second mortgage had fallen as well. The longer run attractiveness of ARMs from the supply side was reduced by extensive securitization of fixed rate mortgages, allowing lenders to avoid correlated risk in local housing markets. Moreover, with the overall decline in mortgage rates in the context of zero lower bound conditions and low expected inflation, the spread between fixed and adjustable rates narrowed, making the ‘affordability’ component to ARM far less significant.
The field of household finance examines financial circumstances of households and behaviours by which they seek to shape those circumstances. This is a very active area of research but also one in which critical issues remain unresolved. For example, there is considerable uncertainty over identifying which types of people are poor. Currently, monitoring and policy interventions are focussed on income-based measures of poverty. However survey data from the UK, the US and Canada suggest that households with very low income spend more than households with moderately low income (Bee, Meyer and Sullivan 2015; Brewer, Etheridge and O’Dea in press; Brzozowski and Crossley 2011; Meyer and Sullivan 2003). We do not know whether this is because of (i) reporting error, with low-income households under-reporting income and/or over-reporting spending, (ii) low income households smoothing their consumption in periods of temporary low income by using savings, or borrowing in a rational and sustainable way, or (iii) very low income households engaging in unsustainable borrowing.
The main obstacle to resolving this and other puzzles is lack of data. There is a household budget identity: for a given household, in a given period, flows in (income, dissaving and borrowing) must equal flows out (spending, new saving and retirement of debt). However surveys typically specialise in the collection of income, or expenditure, or assets and debts, (or possibly two of these three.) As a result, current data provide only very indirect evidence: it is not possible to distinguish reliably between competing hypotheses since covariances between income, expenditure and assets and debts are not observed. Moreover, there is some experimental evidence that reconciliations and balance checks can improve the quality of household finance data collected in surveys (Brzozowski and Crossley 2011; Fricker, Kopp and To 2015; Hurd and Rohwedder 2010; Samphantharak and Townsend 2010). If it were possible to collect all the components of the household budget identity in a single survey, this would not only provide the missing covariances, but should also deliver better data quality.
In 2016 we fielded two versions of an experimental module in Understanding Society’s Innovation Panel, a representative sample of 1,500 households in Great Britain. We collected information from singles and couples on their income and spending. In addition, one version of the module asked about changes in stocks of assets and debts, while the others asked flows of saving and borrowing. In both cases, respondents were confronted, in a summary screen, with the household budget identity, and where their responses failed to balance, were given the opportunity to update their results. We record both initial responses and subsequent revisions.
We report on the degree to which responses satisfied the household budget identity and the revisions that households made when confronted with imbalance. We also compare the two versions of the module. Finally, we examine how the process of budget reconciliation alters the relationship between reported income and spending, especially at low income levels, and how this affects the apparent incidence of poverty.
In empirical research is often that we are interested in the relationship between two variables, but that those two are not contained in a single data set. For example, a key question in macroeconomics is the effect of income or wealth (or changes in income or wealth) on consumption. Traditionally, consumption has been measured in dedicated household budget surveys which contain limited information on income or wealth. Income or wealth, and particularly changes in income and wealth are measured in panel surveys with limited information on consumption. A common strategy to overcome such problems is to impute the dependent variable into the data set containing the independent variable. For example, Skinner (1987) proposed a method for using the U.S Consumer Expenditure Survey (CE) to impute a consumption measure into the Panel Survey of Income Dynamics (PSID). In this paper, we consider the consequences of estimating a regression with an imputed dependent variable, and how those consequences depend on the imputation procedure adopted.
We show that the Skinner procedure leads to biased and inconsistent estimate of the regression coefficient of interest. We show that the bias is equal the R-squared on the first stage regression of the variable to be imputed on predictors (which are common to both surveys). This leads us to suggest a “rescaled-Skinner” procedure. We then show that with a single predictor, the rescaled-Skinner procedure is identical to a procedure developed by Blundell et al. (2004, 2008) in which the first stage involves regressing the predictor on the variable to be imputed. However, the rescaled-Skinner procedure generalizes naturally to multiple predictors.
Lusardi (1996) combines CE consumption data with PSID income data with the 2- sample IV approach proposed by Angrist and Krueger (1992) and (Arellano and Meghir, 1992). We clarify the relationship between that approach and the imputation procedures we study. We also show how the precision of these procedures can be improved using an adjustment for finite sample differences between the datasets that is analogous to the advantage of 2-sample-2-stage least squares over 2-sample-IV described in Inoue and Solon (2010). We illustrate these with a Monte Carlo study
and an empirical example using data from CE and PSID.