The effects of the distribution of mortgage credit on the wage share: Varieties of residential capitalism compared

Abstract

The financialisation of advanced economies has developed in part through the widespread engagement of workers with the financial sector, specifically through the increasing adoption of household debt. Taking on debt establishes “investor identities” for borrowers, which disciplines borrowers to demand employment and marginalises worker bargaining power. Although increased levels of household debt have been linked with a declining wage share of national income, there has been little examination of the effects of the specific channel of mortgage credit, the largest source of household debt, on the wage share. A series of regression models were implemented to test the relationship between the total mortgage stock and the wage share of GDP in four different countries, across three different Varieties of Residential Capitalism. While the panel data results demonstrate a negative relationship between mortgage credit and the wage share across the typologies, the relationship is concentrated in the liberal markets of the UK and the USA, and does not hold independently in Sweden and Denmark. There are four important differences that may explain the divergent results: (1) mortgages in liberal markets are provided at a higher cost than in non-liberal markets; (2) the use of mortgage bonds in non-liberal markets facilitates renegotiations in periods of borrower uncertainty; (3) the level of collective bargaining in Social Democratic states may strengthen worker wage negotiations; and (4) workers in liberal markets are reliant on mortgage finance to access the home as a financial asset in systems of asset-based welfare. The combination of which may explain how the increased disciplinary effects of mortgage finance have resulted in the reproduction of the conditions for capital accumulation in liberal capitalist states.

Appendices

Appendix A: Expanded Model Specification

Although the models developed in this paper have sought to adhere to Achen’s (2002) ROT to refine the accuracy of the model specification, the literature has identified key variables that may influence the formation of the wage share that were not included in the regression models. Therefore, a wider specification model has been developed to test the validity of the results of the refined models used in this analysis. Technological change and globalisation have been identified as key determinants of the wage share of national income (IMF, 2007). Technological change has been modelled using the capital-labour ratio, which is measured using the logarithm of net capital stock divided by number of persons employed in domestic industries, and the data were obtained from the AMECO database (Stockhammer, 2013, p. 21). Exports as a share of GDP, and foreign direct investment (FDI) inflows as a share of GDP have been used as proxies for globalisation, and these data were obtained from the World Bank (2015) (Stockhammer, 2013, p. 20). Labour productivity has also been identified as a significant causal factor in the formation of the wage share of national income (e.g., OECD 2015a, p. 7), which can be measured in terms of GDP per hour worked, and the data were obtained from the OECD (2015b). The post-Keynesian analysis of financialisation is measured in terms of financial globalisation, which is the logarithm of external assets plus external liabilities divided by GDP, and as per Stockhammer (2013) the data were obtained from Lane and Milesi-Ferretti (2007). The deregulation of the financial sector is another way to account for the financialisation of the economy, which has been modelled using the summary financial reform index developed by Abiad et al (2008). The financial reform dataset only accounted for years up to 2006, therefore, specification models with and without this variable were included to maximise the number of available observations and degrees of freedom. Table A1 shows the results of the specification models, and in both cases the sign and significance of the OMB variable is consistent with the results from the main models used in this analysis, which is all that is necessary as the variable has been converted to a natural log. Therefore, one may conclude that the efficient model specification in this analysis, according to Achen’s ROT, may be considered valid.

Table A1 Specification model results

Appendix B: Data Sources

An adequate pre-collated dataset was not readily available to examine the relationship between the wage share of GDP and mortgage credit growth in the USA, the United Kingdom, Sweden, and Denmark between 1979 and 2012; therefore, an original dataset compiling data from each country had to be constructed. Data to construct the wage share variable, employee compensation and nominal GDP at factor cost for each case were obtained from the European Commission’s annual macro-economic database (AMECO, 2015). The data for the OMB variable were obtained from a variety of sources: a complete dataset of British OMB data for the selected time period was not available from a single source, and the data were compiled from two sources; data from 1979 to 1991 were obtained from the Council of Mortgage Lenders’ (CML) compendium of housing finance statistics (CML, 1997, p. 88), and data from 1992 to 2012 were collected from the British Building Societies Association annual yearbook (Rex, 2013, p. 102). All OMB data for the Swedish case were obtained from the Swedish statistics office (Statistics Sweden 2015); while the Danish OMB data from 1979 to 2011 were obtained from the Danish Central Bank (Abildgren, 2012), and the missing 2012 figure was obtained from the Association of Danish Mortgage Banks (Realkreditraadet, 2015). The OMB data for the USA were collected from the Federal Reserve Bank of St Louis (2015) statistical database. The data for the unemployment rate for each country were also collected from the AMECO (2015) database. The RILE data were collected from the Comparative Manifesto Project (CMP) database (version 2015a), which is a subjective qualitative content analysis of party election manifestos from domestic democratic elections, and contains data for each case covered by this analysis (Volkens et al, 2015). The data were transformed to the weighted RILE index using the Stata do-file provided on the CMP website.

Appendix C: Data Validation and Diagnostic Tests

A Levin–Lin–Chu test was run on each variable in the pooled dataset to test for the presence of unit root, and in each case the null hypothesis that all panels contain a unit root could be rejected, which suggests that the panels are stationary for each variable. A cointegration test was also run for the key variables in the pooled data models, the wage share of GDP and OMB, and the tests reject the null hypothesis of no cointegration, which suggests that there is a stable relationship between the key variables in this analysis. Post-regression diagnostic tests were performed on the four OLS models pertaining to the individual VORC (models 3–6), where each model was tested for the normality of the residuals, heteroskedasticity, serial correlation, and multicollinearity. In each case, the residuals were distributed normally, and each variable in each case had a VIF score of less then 5, which demonstrates that there is an absence of multicollinearity in the models (Belsley et al, 1980). A Breusch–Godfrey test was run for each OLS model to test for serial correlation, and the results show that in each case the null of no serial correlation must be rejected, therefore, first-order autocorrelation (AR1) is present in each model. A test for heteroskedasticity was run on the OLS models and in each case, except for the British case (model 5), we fail to reject the null hypothesis of homoscedasticity being true; therefore, it can be stated with at least 95 percent certainty that the errors are homoskedastic in the Swedish, Danish, and US models, and are drawn from a distribution with a constant variance. To account for the serial correlation in each model and heteroskedasticity of the British case, each individual model was re-run to include robust standard errors, which are included in the results table. As stated in the empirical design, the results from the pooled data models (1, 2, 7 and 8) were obtained using the AR1 autocorrelation structure and the assumption of panel-level heteroskedastic errors to address problems of serial correlation and heteroskedasticity respectively. Each of the OLS models was also run using the bootstrap resampling method, which replaces the unknown population distributions with the known empirical distributions originated by the regression to establish precise statistical estimates that allow for robust hypothesis testing using relatively few assumptions (Chernick, 2008). The bootstrap models were re-run with 2,000 repetitions, and each variable retained its statistical significance as reported in the results section (Table 3), with identical parameter estimates to the non-bootstrap models. Diagnostic tests were also run for each bootstrap model with results that were similar to the corresponding non-bootstrap models above. These bootstrap models, and the subsequent diagnostic tests performed on them, suggest that the results from the individual VORC models (models 3–6) may be considered valid and robust.