Log in / Register
Home arrow Computer Science arrow OECD guidelines on measuring subjective well-being.

Co-variates to include in the regression model

As noted earlier, any attempts to compare coefficients obtained through regression analyses need to consider the possible impact of correlations among independent variables. When using life satisfaction data to value non-market factors, Fujiwara and Campbell (2011) recommend that measures for all known determinants of well-being should be included in the model. Although they note the lack of consensus in this area, they list key determinants from the literature as: income, age, gender, marital status, educational status, employment status, health status, social relations, religious affiliation, housing, environmental conditions, local crime levels, number of children and other dependents (including caring duties), geographic region, personality traits (such as extroversion) and the non-market factor being valued.

The same authors also note that for policy purposes, there may be some indirect effects that need to be controlled in valuation regressions to fully estimate the impact that a marginal change in a non-market factor may have on subjective well-being. They take the example of pollution, noting that although pollution is expected to have negative effects on subjective well-being, individuals may be partially compensated for those effects through lower house prices and reduced commuting times. These offset the overall impact of pollution on subjective well-being and may cause the true value of a marginal reduction in pollution to be underestimated. Frey, Luechinger and Stutzer (2004) note the same difficulty in estimating the impact of living in terrorism-prone areas, where higher wages and lower rents potentially compensate individuals - and these authors conclude that all potential channels of compensation need to be controlled for.

The difficulty in attempting to control for all possible drivers and indirect effects is that this may crowd out the variables of interest. For example, including controls that also co-vary with income in the regression equation may shrink the coefficient for income, thus shrinking the increase in life satisfaction brought about by each additional per cent increase in income. Underestimating the impact of income can therefore risk over-valuing the impact of non-market factors. However, the same under-valuation risk is also present for the non-market factors. For example, if the effects of air pollution on life evaluations are mediated by respiratory health conditions, the coefficient for a measure of air pollution is likely to be substantially reduced if a measure of respiratory health conditions is included in the model. This would lead to a lower valuation of air pollution than if the health variable were excluded from the model. The choice to include or exclude other variables in the regression therefore depends on the assumed causal pathways - and these must be clearly described when conducting valuations. Mediational analyses therefore need to play an essential role in preparing models, to better understand how predictors interact with one another. In further developing this valuation technique, it is also important to establish a better overall consensus regarding which explanatory variables should be included in (and excluded from) a valuation regression, and under what circumstances. As noted previously, it may be helpful to report results as a range of values, derived from various different models with and without the presence of certain control variables.

Found a mistake? Please highlight the word and press Shift + Enter  
< Prev   CONTENTS   Next >
Business & Finance
Computer Science
Language & Literature
Political science