Differential income taxation and household asset allocation

Abstract

This article empirically investigates the effects of differential income taxation on households’ portfolio choice and asset allocation, applying a two-stage budgeting model of asset demand to German survey data. The model is structured into the discrete and the continuous asset choice. Cross-sectional variation in marginal tax rates, appropriately instrumented, as well as over-time variation from a major tax reform are used to identify the tax effects. Households with higher tax rates are found to have relatively greater demand for tax-privileged assets, such as nonowner-occupied housing, mortgage repayments, building society deposits, stocks, insurances and consumer credits, than households with lower tax rates. Demand at higher tax rates is lower for owner-occupied housing, bank deposits and bonds.

Keywords:

• household asset allocation
• portfolio choice
• two-stage budgeting
• capital income taxation

JEL Classification:

• C35; G11
• H31

This article greatly benefited from valuable discussions with Viktor Steiner, Martin Browning, participants of the ‘Ageing, Savings and Retirement’ conference at DIW Berlin, and audience of Economic Policy Seminars at DIW Berlin and Free University of Berlin and helpful comments from two anonymous referees. Financial support from the Fritz Thyssen Stiftung, through project ‘Taxation and Asset Allocation of Private Households – Empirical Analyses and Simulations of Policy Reforms for Germany’, is gratefully acknowledged. The usual disclaimer applies.

Notes

¹ See the survey by Poterba (2002) on the relevance of taxation effects on household portfolio choice and asset allocation. On the theory of taxation and portfolio allocation, see the survey by Bernheim (2002), Rapp and Schwetzler (2008) point out in a theoretical approach that the effects of capital income taxation on portfolio choice depend on whether redistribution of taxes is accounted for in the household utility or not.

² I use only elements that were implemented before 2003. See Homburg (2000), for details on this reform. Wagenhals (2001) estimates incentive and redistribution effects of this reform.

³ See, for example, the latest reflections on the Mirrlees Review on the design of capital income taxation (Keuschnigg, 2011); or, estimates for economic effects of adding a flat-rate capital income tax to a consumption-based flat tax in the United States (Diamond and Zodrow, 2007). For effects at the extensive margin, see, for example Fehr et al. (2008), who analyse savings incentives of tax-favoured retirement accounts.

⁴ There are a couple of studies, in this field, applying similar analyses for various countries. Dicks-Mireaux and King (1983) as well as Alan et al. (2010) find significant, but rather modest, effects for Canada. Hochguertel et al. (1997) find positive tax effects on risky assets using data from the Netherlands. Agell and Edin (1990) find significant effects for stocks, housing assets and mortgages, using data for Sweden. For Germany, only Lang (1998) conducts a comparable analysis. He uses the same data set that is applied in the study at hand, but from older cross sections. For a recent study on effects of dividend taxes on the portfolio choice of institutional investors in the United States, see Desai and Dharmapala (2011).

⁵ For a basic portfolio choice model, see Brainard and Tobin (1968). There are several macroeconomic extensions of the Brainard–Tobin model, where asset demand is modelled in a two-stage budgeting model (e.g. Conrad, 1980; Taylor and Clements, 1983). For a similar approach with micro-data, see Lang (1998).

⁶ These separability assumptions may be very restrictive and only approximately valid. A test for separability suggested by Browning and Meghir (1991) provides evidence here that separability of the asset allocation decision from the consumption-savings decision should be rejected, while separability from the labour supply decision cannot be rejected.

⁷ Housing assets may include loans raised to buy a house. Reinvestments of business profits are excluded from the analysis and considered exogenous.

⁸ Separability is again tested for by the test from Browning and Meghir (1991). For this test, the demand equations are augmented with variables for all asset shares from the respective other cluster, and the shares are tested for significance. Many of the shares are found insignificant at the lower stage of the 2SBM, providing evidence for separability, while only some shares are found significant.

⁹ The AIDS goes back to Deaton and Muellbauer (1980a) and the QUAIDS to Banks et al. (1997). For applications of the financial AIDS portfolio model, see e.g. Blake (2004) or the recent work by Ricciarelli (2011).

¹⁰ Note that the present characterization of the German income tax system refers to the time before taxation of capital income taxation has been subject to reform in 2009. At that time, Germany implemented a dual income tax for income from capital and labour, in a trend with many other European countries (Genser and Reutter, 2007; Weiss, 2009), intensifying the differential treatment of capital and labour income.

¹¹ In 1998, gross dividends at the shareholder level were subject to PIT, and there was a KEST prepayment of 25%, while the corporate tax payment was considered as a tax credit (Anrechnungsverfahren). However in 2003, net dividends were subject to PIT, but with only 50% of the net dividend taxable, and the KEST was reduced to 20% (Halbeinkünfteverfahren). Net dividends are net of corporate taxes and net of withholding tax on capital income at the shareholder level.

¹² The literature suggests various arguments for incomplete portfolios, such as differential tax treatment altering relative prices of assets, e.g. Feldstein (1976), fixed or unique transaction costs, monitoring costs, inter alia Perraudin and Sørensen (2000), and borrowing constraints, e.g. Auerbach and King (1983).

¹³ Such a two-step approach is common in the portfolio choice literature, see inter alia Poterba and Samwick (2002); King and Leape (1998). Alternatively, a multivariate tobit model could be applied, see Amemiya et al. (1993), which however relies on identical tax effects at the discrete and at the continuous choice. Fixed transaction costs would certainly contradict such an assumption.

¹⁴ In the MNL, a cluster-specific, but alternative-invariant set of regressors, , implies alternative-specific coefficients given the cluster, .

¹⁵ Note that estimating this nested structure of MNLs in two steps resembles a two-step limited information maximum likelihood estimation (LIML) of a nested logit model (NLM). In Ochmann (2010), I find evidence for the appropriateness of the nested structure.

¹⁶ I have tested interaction effects of the tax rate with several controls. They are not significantly different from zero for most of the controls. Only for the budget, the time dummy, and the self-employed dummy, significant interaction effects were found. Still, none of the interactions turns out significant in each of the demand equations. Thus, results for the significant interaction effects are left as a robustness check.

¹⁷ The validity of this exclusion restriction cannot be tested. There is evidence in the literature that education should fulfil the necessary conditions. King and Leape (1998), for example, find that education significantly affects the probability of accumulating an asset, resulting from information costs, while there is no evidence for a significant effect of education on conditional asset demand.

¹⁸ Variation from the second source is needed here in the specific regression design. This is because asset demand is modelled as a function of several socio-demographic characteristics that are highly correlated with the MTR, as for example the age of the household head and household composition. As a support for the significance of the remaining variation, I conduct a regression of the MTR on the ratio of capital income to gross household income, controlling for all other variables that are included in the asset demand equations. The coefficient for this ratio is found to be highly significant (t-statistic of 19.7), providing evidence that there is significant remaining variation in the MTR across the structure of income.

¹⁹ The unweighted (conditional) mean MTR corresponds to: 22.4% (weighted 18.6%) in the unconditional sample, 23.3% (19.9%) at the upper stage, 28.1% (26.1%) in the housing cluster, and 23.4% (19.9%) in the financial cluster. Detailed estimation results for the discrete and the continuous asset choices can be found in Ochmann (2010).

²⁰ Heteroscedasticity-robust SEs are generated from maximum likelihood estimation. They shall be regarded as a lower limit, as they have not been adjusted for the fact that the discrete and the continuous choice have been estimated in two steps.

²¹ This result appears plausible, as at the upper stage, only some 10% of all households select themselves into zero demand for compound assets, and only slightly more into zero demand for financial assets, while almost of all households do not demand any housing assets.

²² It should though be noted, that the own-price effects do not have a structural interpretation in this simplified demand system, as the price effect on the budget, , is neglected. They should rather be interpreted as empirical tax elasticities. If the effect on the budget was considered, the corresponding compensated own-price elasticities would follow from the Slutsky equation: , where the unconditional income effect in the 2SBM is calculated from: , see Edgerton (1997).

²³ Note that all elasticities refer to the asset shares. The corresponding budget elasticities on the levels follow from: . They allow the interpretation that owner-occupied housing, nonowner-occupied housing, bank deposits, stocks, and bonds are luxury goods () whereas all other assets are necessities (). Moreover, in a complete demand system with cross-price effects, uncompensated own-price elasticities for the levels would follow from , implying the price effect on the budget. If assumed the price effect on the budget is zero, the compensated own-price elasticities for the levels, following from , see Edgerton (1997), would be nonpositive for all assets and thus theoretically consistent with a negative semi-definite Slutsky matrix.

²⁴ Still, comparability of the results, with the tax effects found in this literature, is limited for two reasons. First, in the literature, the condition for conditional shares is on positive demand for the asset, not the asset cluster as it is the case here. Second, differences in the tax treatment of assets across countries can affect the underlying tax incentives in asset demand significantly.

²⁵ All unconditional tax effects have been evaluated for an average household, with an average conditional MTR. In another specification, the effects have been found to vary slightly over the distribution of the MTR for bank deposits, insurances, mortgages, and consumer credits.