Incomemediating variables should not be included in the equa-
tion. It would be nice if we could say how far the other
factors are “mediating” an effect of income or are simply
correlated with it (i.e., confounders). If we assume they are
all confounders we can infer that if we double someone’s
income (cet. par.) we could raise their life- satisfaction by
about 0.14 points.^11 Equally if we raised their income by
10% we would raise their life- satisfaction by 0.02 points—
not a huge amount for a substantial cost.^12
These are fairly standard results of cross- sectional anal-
ysis. But cross- sectional analysis is always at the mercy of
omitted personal variables. Those omitted variables that do
not change over time can be controlled for by including a
personal fixed effect, provided we can obtain two or more
observations on the same individual. The BCS provides
such observations at ages 34 and 42, and Table 2.1 reports in
the next column the results of a fixed effects analysis using
these data. As expected, the estimated effects are reduced,
and the impact of log income falls from 0.20 to 0.13.
However, much better panel data can be found in the
household surveys carried out repeatedly on the same house-
holds in Britain, Germany, and Australia— the British
Household Panel Survey (BHPS),^13 the German Socio-
Economic Panel (SOEP), and the Household, Income and
Labour Dynamics in Australia (HILDA) survey.^14 These are
not birth cohorts, so we cannot include childhood charac-
teristics as we can in the BCS. Nor do these panels include
data on criminality. But the other variables are defined to be
as close as possible to those in the BCS. In addition we in-
clude US data from the Behavioral Risk Factor Surveillance
System (BRFSS). This is not a panel survey but a large an-
nual survey of different samples of people each year. In this