<<

. 4
( 11)



>>

strongly suggest that women are most disadvantaged in countries such as
those in southern Europe and eastern, Asia, where female labor force par-
ticipation rates are low, strati¬cation on the labor market is high, and the
distribution of domestic work is very unequal. If access to paid work and
the ability to form autonomous households are the fundamental interests
of women, as Orloff and others argue, one would expect gender con¬‚icts to
be most intense in these countries. Yet, these are the countries in which the
policy preferences of men and women appear the most similar, and where
there does not appear to be a strong gender gap in electoral politics (Iversen
and Rosenbluth 2003). One plausible explanation is that the family as an
institution is heavily protected, legally and normatively, in these countries.
The likelihood of a ¬rst marriage ending in divorce in Italy is less than one
in ten “ even lower than the rate of divorce in the United States in the
1950s. Following Becker, if divorce is a highly unlikely prospect, men and
women are much less likely to adopt con¬‚icting policy preferences.
Another recent controversy surrounds the role of the public-private
sector division in Scandinavia. According to some, this division, which
concerns issues of public sector size, relative pay, and public sector job
protection, has emerged as a salient cleavage in electoral politics. The high
gender segregation in the public sector also helps explain a widening gen-
der gap. Paul Pierson points out (Pierson 2000) that because men in the
private sector tend to be married to women in the public sector, there is no
89
Political Foundations of Social Policy


compelling reason that spouses should quibble over issues of relative pay.
At the end of the day, the income of both spouses simply adds to family
income. But this logic only applies when husband and wife have few rea-
sons to concern themselves with outside options. Because pay in the public
sector is ¬nanced by taxing the private sector, policies affecting relative pay
are a perfect example of an area where gender con¬‚ict is likely to be intense.


3.2. Testing the Model

3.2.1. Statistical Model
In Section 3.1.3, it was demonstrated in Model IV that the relationship
between the “preferred” level of R and the two exogenous variables y (ex-
pected income) and s (skill speci¬city) is given by the implicit equations:
VR (R, s , g) = 0
y =±·s ·g +β ·g (3.14)
And from (3.14) was derived Result II that
‚R sg
<0 0 < RRA <
if
‚y s g ’ w/2
and Result III
‚R
>0 0 < RRA
if
‚s
It is shown in Appendix 3.B that
‚R ‚R
R=K+ ·y+ ·s
‚y ‚s
is the ¬rst-order Taylor expansion of (3.14). Thus, the regressions take the
form
R =k+b ·y +c ·s (3.15)
By implication, if the estimate of b is signi¬cantly different from zero
and negative, we can infer that 0 < RRA < s g/(s g ’ w/2). If c is signif-
icantly different from zero and positive, 0 < RRA, so that skill speci¬city
increases the demand for social protection. This is the main argument and
hypothesis.5

5 The model also implies that the coef¬cients of y and s are independent of cyclical variations
in the unemployment rate. This implication can be tested through multilevel modeling as
discussed in the next section and Appendix 3E.

90
Explaining Individual Social Policy Preferences


More generally, Model IV encompasses Models I, II, and III. Hence,
we can test for these models as well. Model I (Meltzer-Richard without tax
disincentives) implies that b = c = 0. Model II (Meltzer-Richard with tax
disincentives) implies that b < 0 and c = 0. And Model III (the insurance
model with RRA > s g/(s g ’ w/2)) implies that b > 0 and c = 0.6

3.2.2. The Data
The empirical analysis is based on individual-level data from eleven ad-
vanced democracies obtained from two sets of national mass surveys con-
ducted under the auspices of the International Social Survey Program, one
in 1996 and the other in 1997 (ISSP 1999; 2000).7 These surveys offer by
far the best individual-level data on skills and preferences for social protec-
tion. These individual-level variables are supplemented with economy-wide
unemployment data. The following two sections describe the operational-
ization of the dependent and independent variables.

Dependent Variables The 1996 survey contains four spending questions
that are closely related to the three protection variables introduced in
Chapter 1. Three of the four are used in a cluster of questions that asks
whether the respondent would like to see more or less government spend-
ing on (a) unemployment bene¬ts, (b) healthcare, and (c) pensions (see
Appendix 3.C for details). Re¬‚ecting an assumption in the model, each re-
spondent was warned that more spending may require higher taxes. The
fourth variable is based on a question that asks whether the respondent
favors government spending on declining industries for the purpose of
protecting jobs (see Appendix 3.C for details). Although the respondent
was not explicitly told about the potential costs of government subsidies,
such subsidies are widely acknowledged to be problematic for economic
ef¬ciency.
The four variables are closely related to the conceptual framework pre-
sented in Chapter 1. The ¬rst item is obviously a measure of unemployment

6 Any general insurance model is, of course, consistent with the present model. The hypothesis
about the insurance model being tested is the particular version (Model III) together with
the assumption of RRA > s g/(s g ’ w/2). This solves the inequality-spending puzzle by
implying a positive relationship between income and support for redistributive spending.
7 The eleven countries are Australia, Britain, Canada, France, Germany, Ireland, Italy,
Netherlands, Norway, Sweden, and the United States. Japan, used in both ISSP surveys,
could not be included because of missing data on a key occupational variable (explained
later).

91
Political Foundations of Social Policy


protection, while healthcare insurance and pensions are the two main
sources of publicly provided wage protection. Together, healthcare and
public pensions make up the bulk of total transfers, and from Chapter 1 we
know that transfers reduce pretax and transfer income inequality. Hence,
these types of spending smooth out income for workers moving between
different employment states and, therefore, constitute a form of wage pro-
tection. Finally, the question about job protection is a reasonable proxy for
the concept of employment protection, and we would expect speci¬c skills
workers to be more concerned than general skills workers with keeping
their present jobs.
The survey also asked people whether they favored more or less spend-
ing on “culture and the arts” and “the environment.” These policy areas are
clearly unrelated to social protection, but they are nevertheless relevant to
the argument because general education is often argued to increase support
for spending on “postmaterialist” activities, whereas the theory says that it
reduces support for spending in the social policy area (cf. Kitschelt 1994).8
Because one might object that the ¬ndings for skills re¬‚ect general ideo-
logical opposition to government spending among those with long formal
educations, it is useful to be able to show that the relationship between skills
and support for spending varies by policy area.
For presentational purposes I used con¬rmatory factor analysis (CFA) to
create two spending indexes: one for social spending and one for postmate-
rialist spending (both constructed to have a standard deviation of one). The
adjusted goodness of ¬t index of the CFA model applied to all eleven coun-
tries is 0.94 and varies little by country (the range is 0.90“0.98).9 The social
spending index is closely related to the concept of income protection (R)
in the theoretical model, and it makes for a parsimonious presentation of
results. I subsequently show the ¬ndings for each of the four component


8 Duch and Taylor (1993) make a similar argument concerning postmaterialist attitudes
(though they do not directly discuss spending).
9 LISREL Version 8.5 was used to conduct the con¬rmative factor analysis, using the resulting
factor loadings to construct the two indexes from the individual spending variables. The
model was estimated from the covariance matrix for the six spending variables, assuming
that the variables are indicators of two latent spending variables: social and postmaterialist
spending. The factor loadings for each latent variable are as follows: (i) social spending: .52
(subsidies to protect jobs), .48 (health insurance), .58 (pensions), and .55 (unemployment
insurance); (ii) postmaterialist spending: .58 (environment) and .51 (culture and the arts).
Alternatively the indexes can be constructed from the results of ¬tting con¬rmatory factor
models to the covariance matrixes for individual countries, but the regression results are
only marginally affected.

92
Explaining Individual Social Policy Preferences


items in the overall protection index, which allows for a discussion of each
of the three protection areas discussed in Chapter 1.

Independent Variables Two different approaches to the measurement of
skill speci¬city are employed, each re¬‚ecting different aspects of the theo-
retical model. The ¬rst is to classify workers™ skills, or the skills required to
perform certain jobs, according to their degree of specialization or speci-
¬city. This approach is an attempt to gauge s directly. The second starts
from the model assumption that the dif¬culty of ¬nding a job where one™s
skills are needed is proportional to their speci¬city. This approach is an
attempt to gauge s indirectly through rq: the probability of reemployment
into State I.
The ¬rst approach is based on the ILO™s detailed classi¬cation of peo-
ple™s occupations: the International Standard Classi¬cation of Occupations
(ISCO-88). ISCO-88 classi¬es workers in “occupations” based on two
criteria: the level of skills required for an occupation and the degree of spe-
cialization of those skills. ISCO-88 distinguishes between four broad skill
levels, which are a function of “the range and complexity of the tasks in-
volved” and are explicitly dependent on informal as well as formal training
(ILO 1999, p. 6). Skill level, thus, corresponds to (s + g) in the model. All
other distinctions between occupations are based on the specialization of
skills required to carry out particular jobs, re¬‚ecting “the type of knowl-
edge applied, tools and equipment used, materials worked on, or with, and
the nature of the goods and services produced” (ILO 1999, p. 6). Guided
by this logic, the subdivision of skills proceeds through four levels of ag-
gregation until a high degree of skill homogeneity is reached within each
group.10 At the most disaggregated level, called the unit level, there are 390
occupational categories with highly speci¬c job descriptions.11
Because the occupation of every respondent in the ISSP surveys was
classi¬ed according to ISCO-88 at either the most detailed or second most
detailed level (for exceptions, see Appendix 3.C), one can exploit the skill-
based hierarchical structure of ISCO-88 to capture the specialization of

10 There is no claim that homogeneity is equivalent in every unit group. Yet, skills that
are clearly distinct from one another are unlikely to be in the same group at the most
disaggregated level, and major groups with a highly diverse skill structure therefore will
tend to have more minor and unit groups.
11 Unit group 3144, for example, represents “air traf¬c controllers,” a member of the minor
group “ship and aircraft controllers and technicians,” which is itself one of ¬ve categories
in the major group called “technicians and associate professionals.”

93
Political Foundations of Social Policy


workers™ skills. This is accomplished by comparing the share of unit groups
in any higher-level class to the share of the workforce in that class. The
logic is that the number of unit groups in any higher-level class will be a
function of the size of the labor market segment captured by that class and
of the degree of skill specialization of occupations found in that particular
labor market segment. For example, 8 percent of the workforce across all
countries are classi¬ed as “plant and machine operators and assemblers”
(major group 8), whereas this group accounts for 70 out of the 390 unit
groups, or 18 percent of all unit groups. If occupations at the unit group
level are, on average, equally homogeneous in terms of skills, the dispropor-
tionate share of unit groups in major group 8 will re¬‚ect a greater degree of
specialization of skills found within that major group. By dividing the share
of unit groups (.18) by the share of the labor force (.8), one can, therefore,
generate a measure of the average skill specialization within that particular
major group (3.1). This calculation can also be done at the lower submajor
level, and the mean of these calculations has been used to get proxy for s.12
The resulting variable has 27 values ranging from 0.4 to 4.7.
Because the theoretical concept of skill speci¬city is a relative variable,
the ¬nal step is to divide the absolute skill specialization measure, s , by the
ISCO measure of the level of skills.13 This gives us a proxy for s /(s + g) that
we will refer to as s 1 . Alternatively, we can divide s by a proxy for peoples™
general skills, g, which gives us a measure for s /g. This alternative mea-
sure is called s 2 . The proxy used for g is the respondent™s highest academic
degree as recorded by the respondent (see Appendix 3.C for details).
The second approach to measuring skill speci¬city is based on the obser-
vation that the probability of moving from any particular job into one that
makes use of a worker™s skills (State I ) is rq for speci¬c skills workers and q
for general skills workers, where r < 1. If we conceive of rq as an element
in the continuum [0, q ], r would then be a measure of the asset speci¬city
of a worker™s skills. At the heart of the concept of job speci¬city is the idea


12 The sensitivity of s to small differences in the number of unit groups assigned to each
higher-level group is greater at lower levels of aggregation, and these differences may not
accurately re¬‚ect differences in skill speci¬city. This source of error is minimized at the
highest level of aggregation. However, the greater variance of the measure at lower levels
of aggregation helps reduce the standard error on the estimated parameter for the skill
variable.
13 Using an absolute measure of s generates results that are downward biased. At the limit, if
the (unknown) correlation between s and g is 1, s will have no effect on preferences. It is,
therefore, important to develop relative measures.

94
Explaining Individual Social Policy Preferences


that outside options are more limited for workers with speci¬c skills than
for workers with general skills.
The 1997 ISSP survey contains a question that precisely taps workers™
assessments of their outside options (ISSP 2000). The question reads as
follows:
If you were looking actively, how easy or dif¬cult do you think it would be for you
to ¬nd an acceptable job?

The respondent could answer “very easy,” “fairly easy,” “neither easy nor
dif¬cult,” “fairly dif¬cult,” and “very dif¬cult.” The key here is that the
dif¬culty of ¬nding an acceptable job is likely to be related to how portable
a person™s skills are. High skill speci¬city means that there are fewer jobs
where these skills are used, and the number of job openings is also likely to
be smaller because asset-speci¬c investments by employers and employees
tend to lengthen tenure and limit turnover. In addition, the probability of
¬nding an appropriate job close to a person™s current residence, which is
also a likely component of what an individual considers “acceptable,” falls
with the number of job openings in a given geographical area.14 Asking
people about the probability of ¬nding an acceptable job is, therefore, likely
to generate answers that are systematically related to a person™s skills. In
the absence of extensive information about individual work histories, and
employment conditions in particular labor market niches, the question is,
therefore, about as good a measure of rq as one could hope for. It is referred
to as s 3 .
There is, however, an ambiguity in the relationship of s 3 to the theo-
retical concept of s. The reason is that we cannot know for sure if peoples™
responses re¬‚ect their absolute level of speci¬c skills or the relative share of
their skills that is speci¬c. To make sure that the skill measure is a rela-
tive measure, as required by the theoretical model, we can divide s 3 by g.
This alternative measure is called s 4 . If s 3 is already a relative measure, we
simply get another relative measure that should also be positively related
to preferences for social spending.
The different skill measures, and their intercorrelations, are listed in
Table 3.1. Not surprisingly, the correlations are higher between measures

14 In a path-breaking analysis, Scheve and Slaughter (2001) argue and show empirically that
home ownership can be treated as a relatively immovable asset that affects people™s prefer-
ences for trade protection. It would be interesting to include an interaction term for home
ownership and the question about the dif¬culty of ¬nding an acceptable job. But residential
status is unfortunately not recorded by ISSP.

95
Political Foundations of Social Policy

Table 3.1. Summary of Independent Skill Variables

Intercorrelations
Variable
Name De¬nition Comment
s1 s2 s3 s4
(Share of ISCO-88 1
s1
level 4 groups)/(Share
of labor force) · ISCO
level of skillsa
(Share of ISCO-88 0.8 1
s2
level 4 groups)/(Share
of labor force) · (Level
of general education)a
s3 b Response to question 0.4 0.5 1 Not clear whether
about dif¬culty of this is a measure of
¬nding an acceptable absolute or relative
job skills
s3 · (Level of general
s4 b 0.7 0.6 0.6 1 Assumes that s3
education) measures absolute
skills (though s4
will always be a
relative measure)
a Shares are calculated at both the ¬rst and second ISCO-88 level and then averaged.
b The number of categories on s3 and s4 have been reduced to the same number as on s1 and
s2 before calculating the intercorrelations.


using either the survey question or the ISCO classi¬cation. The lowest
correlations are between s 3 and s 1 or s 2 . To some extent, this may re¬‚ect
that s 3 is an absolute rather than a relative measure, but the main reason is
simply that s 3 is in¬‚uenced by a number of factors (such as how much people
like their current co-workers) that are unrelated to either skills or social
policy preferences. These factors will wash out in the regression, but they
reduce the correlation with the other measures. To facilitate comparison of
the effects of the different variables in the subsequent regression analysis,
all proxies for s have been divided by their standard deviations.
One ¬nal methodological issue needs to be addressed. Because the ques-
tion used as the basis for s 3 and s 4 was asked only in the 1997 survey, whereas
all the questions about spending were asked only in the 1996 survey, it was
necessary to “translate” the 1997 information on s 3 so it could be used in
the 1996 survey (s 4 can be always be calculated from s 3 ). For this purpose,
averages for s 3 were calculated at the three-digit ISCO-88 level in the 1997
survey and then assigned to individuals in the 1996 survey based on their
96
Explaining Individual Social Policy Preferences


three-digit ISCO classi¬cation in that survey.15 Because the classi¬cation
of occupations is motivated by the skills required in these occupations, it is
reasonable to expect that original information about s is preserved to a con-
siderable extent in this translation. Moreover, because the 1996 and 1997
samples are drawn from the same populations,16 it is shown in Appendix
3.D that s 3 , averaged by ISCO level 3 groups, is an unbiased estimator for
the original variable.
In addition to the skill variables (s 1 “ s 4 ), self-reported pretax and transfer
income was used as a proxy for y (converted into dollars at 1996 exchange
rates). Gender also enters as a key variable in the regression for the reasons
outlined previously. In addition, the regression includes the following set
of controls:

Age. Older workers are likely to be more concerned with job security
and income than younger workers because their time to retirement is
shorter and their ability to ¬nd new employment is likely to be more
limited.
Union membership. Because one of the main functions of unions is to in-
sure their members against labor market risks, it is reasonable to expect
that union members are particularly concerned with social protection
(see, for example, Korpi 1989).
Part-time employment. Part-time employees are often in vulnerable labor
market positions, which may cause particular concern for job secu-
rity and income protection. On the other hand, part-time employees
depend more on ¬‚exible labor markets to generate nonstandard jobs,
which suggests a countervailing effect.
Nonemployed. Esping-Andersen (1999) has argued that some outsider
groups may share an interest in social and economic policies that
maximize their ability to enter employment. But this is an extremely
heterogeneous group that may not have common policy preferences.
We need to include the variable to control for the possibility that the
nonemployed have very different attitudes than the employed.
Unemployed. The expectation is obviously that the unemployed, relying
as they do on transfers, will support high levels of income protection.

15 There are 116 unique groups at the three-digit level. The more ¬ne-grained four-digit
level is not available for some countries and contains a large number of empty categories
where it is.
16 With the minor quali¬cation that those who turned 18 between the 1996 and 1997 surveys
were not part of the population in the former survey.

97
Political Foundations of Social Policy


Self-employment. The self-employed are expected to favor free markets
and low levels of social protection because they depend on ¬‚exible
labor markets and often on relatively low-paid workers.
Information. It is conceivable that better information about the economy
yields particular views on the desirability of social spending. There
was an intense public debate about the proper role of the state in the
1990s, and it could be argued that better informed people may re¬‚ect
the predominant view in this debate, which tended to see cut-backs
as necessary on ef¬ciency grounds (corresponding to a higher cost
of distortionary taxation in the model). Information is measured by
respondents™ subjective understanding of politics (see Appendix 3.C
for details).
Left“right position. Attitudes to social protection may in part be a re¬‚ec-
tion of people™s ideological predispositions, or perhaps the socializing
effects of political parties.17 This possibility is controlled for by in-
cluding positions on a left“right scale based on the respondent™s de-
clared support for parties that are ranked from far left to far right (see
Appendix 3.C for details).
National unemployment. Although the theory implies that individuals dis-
count cyclical unemployment, it has been suggested that such unem-
ployment could have an impact on individual-level social preferences.
Testing this assumption requires a multilevel modeling procedure,
with countries as level 1 and individuals as level 2. Collapsing both
levels into a single equation (as shown in Appendix 3.E) implies the
inclusion of the product variables U j · yi j and U j · s i j in the regres-
sion model, where U j is the rate of unemployment in country j (see
Appendix 3.C for details on measurement).


3.2.3. Findings
The regression model in Equation (3.15) was estimated on all countries si-
multaneously (technically speaking as a single-stage multilevel procedure to
incorporate the possible impact of national macroeconomic conditions).18
To cope with problems of missing observations, a multiple imputation tech-
nique developed by Gary King and his associates was used (see Honaker

17 Note that party support may in part be endogenous to skills. If so, the effect of skills will
be underestimated by the parameter for s.
18 All data analysis was done using Stata 6.0 for Windows.

98
Explaining Individual Social Policy Preferences


et al. 1999). This strategy is superior to the traditional approach of “list-
wise deletion,” which is both inef¬cient and potentially biased (King et al.
2001).19 The following presentation is divided into a section with the key
results and a section that tests the robustness of these results and discusses
potential objections to the way the results are being interpreted.

The Basic Results To give a sense of the central tendency of the estimates,
Table 3.2 shows the results from a pooled analysis, including a full set of
country dummies. Because the Italian survey was conducted in 1990 and
lacks information on several of the control variables, it was not included in
the calculation of these pooled results. In the next section, the results for
Italy are shown to be consistent with those presented in Table 3.2.
The model in column (1) uses the average of the four measures of skills,
called s composite , as a summary variable for skill composition. The next four
columns show the results for each of the component measures (s 1 “ s 4 ).
Model (6) is identical to (1) except that the regression now includes union
membership as an independent variable. Because union membership was
not recorded in Australia, the estimation of model (6) excludes this country.
In interpreting the results, ¬rst note that the parameters for income, y,
and the four measures of skill, s 1 “ s 4 , are in the predicted direction and
highly statistically signi¬cant. The negative effect of income implies that
people™s risk-aversion is not suf¬ciently high to make their demand for
transfers rise with income. Technically speaking, RRA < s g/(s g ’ w/2),
which means that the Meltzer-Richard redistribution logic dominates the
insurance logic. As expected, the relationship is little affected by differences
in national unemployment rates, despite considerable variation in unem-
ployment in the survey year. Thus, a one standard deviation increase in
unemployment would only change the parameter on y from .0033 to .0038.
Yet, for my purposes, the key ¬nding is the positive effect of speci¬c skills
on preferences for spending (which implies that RRA > 0). Each of the four
(standardized) skill variables is associated with signi¬cantly higher support
for spending, and three of the four measures exhibit similar magnitudes of
effects. Again, these relationships hold for all levels of unemployment as can
be seen from the negligible parameters for Uj · s ij .20 The parameter for s 3 is

19 In practice, however, the results are very similar to those obtained by using listwise deletion.
The effects of the theoretical variables tend to be slightly stronger when listwise deletion
is used, but the standard errors are also larger.
20 For example, a one standard deviation increase in U j only reduces the effect of s composite
from .23 to .22.

99
Table 3.2. Support for Social Spending among the Publics of Ten OECD Countries, 1996
(standard errors in parentheses)

Dependent Variable: Support for Social Spendinga
(4)b (5)b (6)c
(1) (2) (3)
’0.0033—— ’0.0036—— ’0.0038—— ’0.0044—— ’0.0035—— ’0.0036——
Income
(0.0002) (0.0002) (0.0002) (0.0002) (0.0002) (0.0002)
0.233—— 0.219——
’ ’ ’ ’
scomposite
(0.014) (0.013)
0.148——
’ ’ ’ ’ ’
(0.010)
s1
0.150——
’ ’ ’ ’ ’
(0.010)
s2
0.105——
’ ’ ’ ’ ’
s3
(0.013)
0.218——
’ ’ ’ ’ ’
s4
(0.014)
0.0029—— 0.0043—— 0.0034—— 0.0042—— 0.0018—— 0.0027——
Age
(0.0006) (0.0006) (0.0005) (0.0006) (0.0006) (0.0006)
0.215—— 0.208—— 0.205—— 0.124—— 0.148—— 0.198——
Gender
(female) (0.018) (0.018) (0.018) (0.019) (0.019) (0.019)
0.185——
’ ’ ’ ’ ’
Union
membership (0.023)
’0.076—
’0.029 ’0.041 ’0.033 ’0.058 ’0.031
Part-time
employment (0.028) (0.028) (0.028) (0.031) (0.031) (0.029)
0.293—— 0.313—— 0.311—— 0.320—— 0.309—— 0.325——
Unemployed
(0.041) (0.041) (0.042) (0.047) (0.046) (0.043)
——
’0.081—— —— ——
’0.074——
Nonemployed ’0.079 ’0.086 ’0.080 ’0.038
(0.025) (0.025) (0.025) (0.026) (0.026) (0.026)
——
’0.235—— —— ——
’0.221—— ’0.184——
Self-employed ’0.232 ’0.250 ’0.222
(0.029) (0.028) (0.028) (0.028) (0.029) (0.027)
——
’0.045—— —— ——
’0.050—— ’0.043——
’0.041 ’0.047 ’0.069
Informed
(0.008) (0.008) (0.008) (0.010) (0.010) (0.009)
——
’0.051—— —— ——
’0.047—— ’0.041——
’0.050 ’0.050 ’0.047
L“R party
support (0.004) (0.004) (0.004) (0.005) (0.005) (0.005)
’0.0002—— ’0.0002—— ’0.0002—— ’0.0003—— ’0.0004—— ’0.0003——
Uj · yij
(0.0001) (0.0001) (0.0001) (0.0001) (0.0001) (0.0001)
’0.012— ’0.012—
Uj · s ij ’0.008 ’0.004 ’0.002 ’0.008
(0.005) (0.004) (0.004) (0.005) (0.005) (0.004)
Adjusted 0.21 0.20 0.20 0.18 0.20 0.22
R-squared
14,101 14,101 14,101 10,956 10,956 11,950
N
— Signi¬cant at the .05 level; —— signi¬cant at the .01 level.
a All regressions included a full set of country dummies (not shown).
b Excludes Australia, Ireland, and Italy for which data are not available.
c Excludes Australia for which union membership data are not available.


100
Explaining Individual Social Policy Preferences


lower than for the other measures, but this is not entirely unexpected given
that this variable may capture absolute rather than relative endowments of
speci¬c skills (or a combination of absolute and relative endowments). For
all correlations between s and g that are greater than ’1, absolute measures
of s will yield lower parameter estimates than relative measures.21
Considering the very different approaches to measuring skills, it is reas-
suring that the results are consistent across de¬nitions. Yet, statistically sig-
ni¬cant effects do not necessarily imply large substantive effects. Table 3.3,
therefore, shows the estimated portion of the explained variance accounted
for by each of the independent variables, as well as the impact on preferences
of a one standard deviation change in each of the independent variables.
The estimates are based on the results of model (6) in Table 3.2, which
includes all the relevant variables.
Although it is not possible to attribute precisely the proportion of ex-
plained variance to each of the independent variables, it is possible to calcu-
late likely ranges. The upper bounds of these ranges are found by recording
the increase in explained variance (measured as a percentage of the total
explained variance) when a variable is included as the ¬rst predictor (apart
from the country dummies). This number encompasses every direct, indi-
rect, and spurious effect of the variable. The lower bounds are calculated
as the increase in explained variance (as a percent of the total explained
variance) when a variable is entered as the last predictor. This procedure
eliminates all hypothesized individual-level spurious effects of the variable
but also discounts all possible indirect effects. The true explanatory power
of any variable is likely to be somewhere between these bounds.
Using this method, Table 3.3 shows that income and skills are unam-
biguously the most important variables in explaining social policy prefer-
ences among the ones included in this analysis. Thus, income accounts for
between 11 and 51 percent of the total explained variance, whereas skills
account for between 26 and 38 percent. Jointly, income and skills capture
between 38 and 73 percent of the explained variance, with the rest accounted
for by the controls.
The key role of income and skills in explaining social policy preferences
is con¬rmed when we consider the impact of a one-standard deviation
change in these variables (column 3). A standard deviation change in either
variable is associated with about 20 percent of a standard deviation change

21 In fact, the correlation between s 3 and a measure of g based on general education is close
to 0 in the data, which implies that an estimated effect of s 3 is half the “true” effect of skills.

101
Political Foundations of Social Policy

Table 3.3. Estimates of the Magnitude of the Effects of Independent Variables

Proportion of Impact of a 1 Standard
Explained Variancea Deviation Changed
Lower Boundb Upper Boundc 95% Con¬dence Interval
’0.22 ’0.19
Income 11 51
26 38 0.19 0.22
scomposite
Age 1 2 0.03 0.05
Gender (female) 6 17 0.09 0.11
Union membership 1 4 0.07 0.09
’0.02 ’0.00
Part-time employment 0 0
Unemployed 3 9 0.06 0.08
’0.03 ’0.01
Nonemployed 0 8
’0.07 ’0.05
Self-employed 2 8
’0.05
Informed 1 8 0.04
’0.09 ’0.07
L“R party support 5 5
Uj · yij 1 13
Uj · sij 0 8
Income and scomposite 38 73 0.38 0.44
All controls combined 27 52 0.36 0.47
a Increase in explained variance by each variable as proportion of the total explained variance
of all (nondummy) variables (based on model (6) in Table 3.2).
b Increase in explained variance (compared to model with only country dummies) when each
variable is included as the last variable.
c Increase in explained variance when a variable is included as the ¬rst variable.
d The change in support for social spending (measured in standard deviations) as a result of a
one standard deviation increase in each of the independent variables (in the cases of income
and scomposite , unemployment is kept at its mean). The last two rows assume changes in
the independent variables that raise support for spending (and take into account that some
combinations of the employment variables are impossible).


in preferences (since the dependent variable is standardized, the recorded
effects can be interpreted directly in terms of standard deviations). Together,
the impact of income and skills is as great as the joint effect of a standard
deviation change in all controls simultaneously. Note also that the effects
of both variables are estimated very precisely, varying in a narrow range
between (’)0.19 and (’)0.22 (95-percent con¬dence interval).
The results for the controls also generally con¬rm the expectations. Indi-
viduals who are particularly exposed to labor market risks “ the unemployed,
women, and older workers “ are more favorably disposed to increasing so-
cial spending than others. The same is the case for union members, whereas
the self-employed are more likely to oppose social spending. Those who
consider themselves well informed about politics are also more likely to
102
Explaining Individual Social Policy Preferences


oppose spending, perhaps re¬‚ecting a political reality at the time that was
hostile to the welfare state. Supporters of right parties, not surprisingly,
also express less support for social spending than supporters of left parties.
Finally, note that the attitudes of part-time employees and those outside
the labor market are indistinct from the attitudes of others. These groups
are evidently too heterogenerous to share any common interest in social
policies.
Table 3.4 shows the results for each spending area separately.22 They
are quite similar across issue areas for the theoretical variables, although
the effects are somewhat stronger for employment protection.23 From the
perspective of the insurance model, as de¬ned previously, it is notable that
the negative effect of income is just as strong for unemployment protection
where transfers go only to those out of employment. This result holds even
when we exclude high-income earners, casting some doubt on the Moene-
Wallerstein argument that egalitarian societies, all else being equal, spend
more on social insurance than inegalitarian ones because the median voter™s
income is higher (Moene and Wallerstein 2001). Higher income appears
to be always linked to preferences for lower spending.
The most notable differences across spending areas are, for the most
part, easy to explain. Thus, it is no surprise that unemployed are far more
concerned with unemployment protection than any other policy area. Like-
wise, it is pretty obvious why older people are particularly keen to raise
pensions (see the effect of age on the pension variable), and it is perhaps
also understandable that nonemployed are less enthusiastic about doing this
given that pensions are, for the most part, linked to employment. It is per-
haps more puzzling that those who consider themselves well informed are
particularly opposed to employment protection. However, recall that the
survey question referred to protection of jobs in declining industries, which
might be perceived by the well informed as particularly damaging to overall
economic ef¬ciency. A more intriguing result is that the effect of skills on
support for employment protection is weaker in countries with high unem-
ployment. Yet, this is the only policy area where the unemployment rate has
this effect, and because there are so many potentially confounding variables
at the national level, we should perhaps not attach too much weight to this

22 Each variable ranges between 1 and 5, and all are de¬ned so that higher values means
greater support for protection.
23 This is partly, though not fully, explained by more variation in the answers to the job
protection question. The standard deviation for this item is 1.15, whereas for the others it
is .94 (unemployment protection), .83 (health insurance), and .81 (pensions).

103
Political Foundations of Social Policy

Table 3.4. Support for Spending in Four Areas of Social Protection (standard errors
in parentheses)

Dependent Variablea
Wage Protection
Employment Unemployment Health
Protection Protection Insurance Pensions
’0.0033—— ’0.0019—— ’0.0017—— ’0.0018——
Income
(0.0003) (0.0002) (0.0002) (0.0002)
0.248—— 0.144—— 0.079—— 0.138——
s composite
(0.015) (0.016) (0.012) (0.012)
’0.0013— 0.0027—— 0.0011— 0.0051——
Age
(0.0007) (0.0005) (0.0005) (0.0005)
0.269—— 0.072—— 0.134—— 0.088——
Gender
(female) (0.021) (0.018) (0.016) (0.015)
’0.047—
’0.028 ’0.027
Part-time 0.027
employed (0.033) (0.027) (0.024) (0.023)
0.120— 0.536——
Unemployed 0.043 0.066
(0.048) (0.040) (0.037) (0.036)
’0.063— ’0.070—— ’0.061——
’0.014
Nonemployed
(0.029) (0.022) (0.021) (0.021)
’0.191—— ’0.273—— ’0.060— ’0.084——
Self-employed
(0.029) (0.024) (0.027) (0.025)
’0.068—— ’0.016—— ’0.020——
’0.004
Informed
(0.010) (0.008) (0.007) (0.007)
’0.043—— ’0.040—— ’0.029—— ’0.019——
L“R party support
(0.005) (0.004) (0.004) (0.004)
’0.0003—— ’0.0002——
Uj · yij ’0.0001 ’0.0001
(0.0001) (0.0001) (0.0001) (0.0001)
’0.016——
Uj · sij ’0.005 ’0.002
0.004
(0.006) (0.005) (0.004) (0.004)
Adjusted R-squared .15 .16 .13 .14
14,101 14,101 14,101 14,101
N
— Signi¬cant at the .05 level; —— signi¬cant at the .01 level.
a All regressions included a full set of country dummies (not shown).



result. Overall, the ¬ndings by area are intuitive and consistent with the
overall argument.


Gender Differences Gender stands out among the control variables. It ac-
counts for between 8 and 17 percent of the total explained variance, and it
104
Explaining Individual Social Policy Preferences


has the greatest impact after income and skills. As argued earlier, the likely
reason is that women require more protection than men in comparable jobs
because they need to be able to leave the labor market for the purpose of
child rearing and then return to it. This suggests that the gap in gender
preferences depends on the extent to which women participate in the labor
market. Because the welfare of nonworking women relies more on the in-
come of men than does the welfare of working women, nonworking women
have a stronger incentive to support policies that raise the take-home pay
of males. Nonworking women will still care about their outside options,
but policies that reduce the relative wage of men also reduce the income of
families where the woman does not work.
To explore this possibility, the ¬rst column in Table 3.5 adds labor force
participation and a term for the interaction between gender and labor force
participation. Labor force participation is coded 1 for those who are full-
time employed, 0.5 for part-time employed, and 0 for those who are less
than part-time employed or outside the labor market. As a consequence,
two of the employment variables (part-time and nonemployment) from
Table 3.4 had to be dropped.
Adding the interaction term creates some problems of collinearity, with
68 percent of the variance in this variable explained by its constituent terms.
Still, the results are statistically signi¬cant and make sense in terms of the
theoretical argument. Thus, when a woman is not working (the value on the
labor force participation variable is zero), her predicted support for more
social protection is .16 higher than men™s, whereas it is .24 higher than men™s
when she is working full-time. In other words, women are more likely to
share policy preferences with men when they are not working. This pattern
is the same across categories of protection.
Columns 2 and 4 of Table 3.5 look at two predicted indirect effects of
gender. Column 2 uses income as the dependent variable and shows that
women, not surprisingly, earn less than men. On average, they make about
$400 less than men (in 1996), and this ¬gure roughly doubles if we also
take into account that women participate less in the labor market. Because
lower income translates into greater support for social spending, the effect
of gender on preferences is obviously magni¬ed by income.24


24 The regression uses education, instead of skill speci¬city, to capture the effects of past skill
investment on income. In principle, we should be using total skills, but the variable that
can be derived from the ISCO classi¬cations is very crude. At any rate, it does not matter
much for the effect of gender.

105
Political Foundations of Social Policy

Table 3.5. Gender Effects on Preferences, Income, and Skills (standard errors
in parentheses)

Dependent Variablea
Support for Social
Protection Income Skill Speci¬city
’0.003——— ’ ’
Income
(0.000)
0.243——— ’ ’
s composite
(0.012)
0.179——— ’20.00——— ’0.266———
Gender (female)
(0.026) (0.73) (0.015)
57.83——— ’0.219———
Labor force 0.044
participation (0.028) (0.96) (0.017)
0.071——
Gender — labor force ’ ’
participation (0.034)
11.304———
’ ’
Education
(0.301)
0.003——— 0.764——— 0.005———
Age
(0.001) (0.025) (0.000)
0.309——— ’52.20——— 0.410———
Unemployed
(0.039) (1.71) (0.037)
’0.224——— ’0.059——
’1.85
Self-employed
(0.023) (1.28) (0.022)
0.039——— ’ ’
Informed
(0.008)
’0.050——— ’ ’
L“R party support
(0.004)
’0.0002———
Uj · yij ’ ’
(0.0001)
’0.009——
Uj · sij ’ ’
(0.004)
Adjusted R-squared .15 .16 .13
14,101 14,101 14,101
N
——— Signi¬cant at the .05 level; —— signi¬cant at the .01 level.
a All regressions included a full set of country dummies (not shown).



On the other hand, about half of this effect is canceled out by another
indirect effect: the lower propensity of women to invest in speci¬c skills
(column 2). The skill speci¬city variable is 0.27 standard deviations lower for
women than it is for men. Because women know that they are likely to leave
their jobs before they can reap the full returns on speci¬c skill investments,
106
Explaining Individual Social Policy Preferences


they are dissuaded from making such investments in the ¬rst place. Put
differently, because women have a comparative advantage of investing in
general skills, they will tend to specialize in these. This provides micro-
level support to the interpretation of the macro-level data on occupational
segregation presented in Chapter 1, and it explains why almost half the
effect of gender on preferences disappears if the skill variable is removed
from the regression.


Robustness Tests In this section, the robustness of the results are tested and
some potential objections to the interpretation of the results are addressed.
It can ¬rst be noted that the ¬ndings for y and s stand up to any combination
of the controls included previously, and they are robust to the inclusion of
any other variable used in the survey, hereunder region, public sector em-
ployment, urbanization, and supervisory position “ in any combination.25
Even though income and skills are powerful explanatory variables in the
pooled analysis, pooling can disguise considerable cross-national variation
in the strength of the results, and sometimes estimated parameters can even
reverse in particular cases. In addition, pooling usually yields exaggerated
t-scores compared to those found for individual countries.26 The regres-
sions for each of the eleven countries were, therefore, run individually. The
results for the theoretical variables are shown in Table 3.6.
Note that every regression yields results that are consistent with the
pooled analysis, with each of the sixty parameters recording the correct
sign and most being signi¬cant at the .01 level or better. The composite
skill variable is always signi¬cant at a .01 level or better, and for nine of
the eleven countries the parameter estimates for s vary in a fairly narrow
range between 0.16 and 0.29 (the parameter in the pooled analysis is .23).
Only Ireland and Italy fall slightly out of the pattern with parameters just
below .12. Yet, the effects for these countries are still statistically highly
signi¬cant, and it should be noted that s composite in both cases are based on
only two proxies for s. In the case of Italy, these proxies also use a crude


25 None of these variables were used in every survey, so instead of cluttering the presentation
with several additional columns, these variables are out of the main analysis.
26 The reason is that the standard error has the form (Standard error of equation
error)/(Standard error of variable). Because the denominator is the square root of the
sum of squares of the explanatory variable divided by N, this normally increases with N
since a squared term is added on the top and 1 is added to the bottom (though it does not
have to be so).

107
Political Foundations of Social Policy

Table 3.6. Income, Skills, and Support for Social Spending in Eleven OECD Countries
(t-scores in parentheses)a

Incomeb scomposite s1 s2 s3 s4 N
’0.0030—— 0.156—— 0.129—— 0.127—— n.ac
Australia n.a 2151
(’0.0004) (0.027) (0.023) (0.026)
’0.0029—— 0.219—— 0.105—— 0.121—— 0.135—— 0.181——
Britain 989
(’0.0006) (0.042) (0.026) (0.027) (0.039) (0.046)
’0.0054—— 0.219—— 0.102— 0.140—— 0.087— 0.220——
Canada 1182
(’0.0008) (0.053) (0.042) (0.045) (0.041) (0.047)
’0.0055—— 0.235—— 0.158—— 0.147—— 0.097—— 0.111——
France 1312
(’0.0007) (0.038) (0.035) (0.028) (0.037) (0.026)
’0.0027—— 0.255—— 0.182—— 0.155—— 0.115—— 0.212——
Germany 2361
(’0.0007) (0.030) (0.028) (0.023) (0.035) (0.028)
’0.0030—— 0.116—— 0.092—— 0.122——
Ireland n.a n.a 994
(’0.0008) (0.028) (0.029) (0.027)
0.105—— 0.104— 0.095——
’0.0021
Italy n.a n.a 983
(’0.0016) (0.037) (0.040) (0.034)
’0.0026—— 0.257—— 0.139—— 0.165—— 0.076—— 0.250——
Norway 1344
(’0.0006) (0.033) (0.024) (0.027) (0.027) (0.030)
’0.0049—— 0.176—— 0.125—— 0.094—— 0.134—— 0.131——
New 1198
Zealand (’0.0007) (0.038) (0.030) (0.030) (0.041) (0.034)
’0.0055—— 0.282—— 0.130—— 0.140—— 0.156—— 0.278——
Sweden 1238
(’0.0010) (0.039) (0.030) (0.028) (0.033) (0.030)
’0.0024— 0.294—— 0.192—— 0.189—— 0.272——
United 0.042 1332
States (’0.0010) (0.052) (0.030) (0.037) (0.036) (0.058)
— Signi¬cant at the .05 level; —— signi¬cant at the .01 level.
a All regressions included the same set of controls as in Table 3.2, column (1).
b The effect of income is only shown for scomposite but varies little across the four measures of s.
c n.a.: Data not available to estimate this parameter.



occupational variable that maps rather poorly onto ISCO-88, potentially
diluting the skill distinctions between categories.
As in the case of the pooled analysis, it should be noted that the results
for s3 are somewhat weaker across all cases than for the other skill measures,
but only in one instance (the United States) is the result statistically insignif-
icant. Given the variety of countries and the differences in measurements,
the combination of results provides clear support for the theory.
Another objection that can be raised to the ¬ndings for skills is that they
may in part be capturing an ideological aversion to government spending
among those with higher education. Two of the measures of s have formal


108
Explaining Individual Social Policy Preferences


education in the denominator, and the other two implicitly assume that such
skills are part of the denominator. In quantitative terms, general education
accounts for roughly one third of the variance in s composite . It is, therefore,
conceivable that the proxies for skills may in part capture an ideological
effect of higher education. For example, much of the economic theory
taught to university students during the 1990s emphasized the ef¬ciency of
free markets over state intervention.27
To some extent, I have already controlled for this possibility by includ-
ing variables for people™s assessment of their own level of information, as
well as their support for parties on the left“right scale. If the highly ed-
ucated consider themselves better informed about the costs of generous
social spending, this is likely to show up in the variable measuring infor-
mation. Likewise, those ideologically committed to a small welfare state
are presumably more likely to support right parties. The fact that a large
effect of skills persists after control for these variables suggests that the
conception of skills as assets is correct.
But there may still be unmeasured aspects of formal education that some-
how confound the effects of the skill variable. One way to address this issue
is simply to include general education as a separate variable. In that way,
the effect of s composite will only pick up the effects of speci¬c skills. In this
setup, one should expect formal education to have the opposite effect of the
speci¬c skills variable, and the separating out of general skills will necessar-
ily weaken the effect of the original variable if general education is indeed a
measure of general skills. However, we can be certain that whatever effect
remains of s , it cannot be attributed to general education.
The ¬rst column of Table 3.7 shows the results of reestimating model
(1) in Table 3.2, using formal education as a separate independent variable.
Formal education has a strong negative effect on support for social protec-
tion. This is consistent with the skill asset argument. But more importantly,
the parameter on the speci¬c skills variable remains positive and statistically
signi¬cant. Not surprisingly, the effect of s falls from 0.23 to 0.14, but this
is still a very considerable impact. Even if one were to discount the effect
of general education as a measure of general skills completely, the results
lend unambiguous support to the argument.


27 In terms of the formal model this can be captured by different assessments of the distor-
tionary effects of taxation.



109
Political Foundations of Social Policy

Table 3.7. Formal Education and Support for Two Types of Spending in Ten
OECD Countries, 1996 (t-scores in parentheses)

Support for Support for Postmaterialist
Social Spending Spending
’0.105—— 0.130—— ’
Formal education
(0.008) (0.007)
0.143—— ’0.097——

scomposite
(0.015) (0.014)
’0.0027—— ’0.0004
Income 0.0004
(0.0002) (0.0002) (0.0002)
0.0015— ’0.0064—— ’0.009——
Age
(0.0006) (0.0006) (0.001)
0.203—— 0.092—— 0.087——
Gender (female)
(0.018) (0.018) (0.019)
0.104—— 0.112——
’0.025
Part-time
employment (0.028) (0.029) (0.030)
0.303—— 0.085— 0.089—
Unemployed
(0.040) (0.043) (0.044)
’0.067—— 0.077—— 0.093——
Nonemployed
(0.025) (0.024) (0.025)
’0.243——
Self-employed 0.021 0.011
(0.028) (0.025) (0.026)
’0.031—— 0.069—— ’0.083——
Informed
(0.008) (0.009) (0.009)
’0.049—— ’0.060—— ’0.059——
L“R party support
(0.004) (0.005) (0.005)
Uj · yij ’0.0002 0.0000 0.0001
(0.0001) (0.0001) (0.0001)
Uj · sij ’0.006 0.008 0.007
(0.005) (0.005) (0.006)
Adjusted R-squared 0.22 0.09 0.07
14,101 14,101 14,101
N
— Signi¬cant at the .05 level; —— signi¬cant at the .01 level.
Note: Regressions included a full set of country dummies.




Yet, results for the postmaterialist spending index explained earlier sug-
gest that it would be a mistake to treat general education as a proxy for
unmeasured ideological effects. Surely, if highly educated individuals be-
lieve in the ef¬ciency of free markets and the waste of government spend-
ing, they should also oppose public spending on the environment, culture,


110
Explaining Individual Social Policy Preferences


and the arts.28 But the exact opposite is true as shown in column (2) of
Table 3.7. People with high general education are much more likely to sup-
port government spending on these areas than others. Conversely, if we
use the composite measure of speci¬c skills (column 3), the effect of skills
is reversed: Speci¬c skill workers want less postmaterialist spending, even
though they support more social spending. Evidently people prefer gov-
ernment spending in areas that are particularly conducive to their personal
welfare. General skills workers demand little social protection but are en-
thusiastic consumers of a clean environment and state-subsidized culture.
Speci¬c skill workers are deeply concerned with social protection but are
not enthusiastic about state subsidization of environmental causes and the
arts. There is no blanket support for, or opposition to, government spending
among any particular group of workers.


3.3. Conclusions
Because a substantial portion of both national and personal income can be
attributed to human capital, broadly conceived, it is not surprising that the
asset speci¬city of this capital matters a great deal for the amount of social
insurance demanded by individual workers. Like physical capital, human
capital can be more or less mobile, and workers who have made heavy
investments in asset-speci¬c skills stand a greater risk of losing a substantial
portion of their income than workers who have invested in portable skills.
For this reason, speci¬c skill workers have a greater incentive to support
policies and institutions that protect their jobs and income.
Because social protection tends to bene¬t low-income people more than
high-income people, position in the income distribution also divides public
opinion. However, at any given level of income, workers with speci¬c skills
are more inclined to support high levels of protection than workers with
general skills. This may help us understand cross-national variance in social
protection because, as explained in Chapter 1, the pro¬le of skills varies de-
pending on the structure of the educational system. If these differences are

28 It is true that “the environment” may be conceived as a collective good improving overall
welfare (it is a little harder to argue this with respect to subsidies to the ¬ne arts), but
by the same token social protection may be conceived as welfare-improving insurance.
The point is not that the highly educated are more informed about what is “good” and
“bad” spending, that is already controlled for, but that they may have internalized a general
aversion to government spending through their educational experience.



111
Political Foundations of Social Policy


re¬‚ected in the political preferences of electorates, and if political parties
adopt policies to attract the support of voters, it suggests a new explana-
tion of the welfare state based on differences in national skill pro¬les. As
explained in Chapter 4, however, the translation of policy preferences into
policies is not direct. Even if the median voter is pivotal in electoral compe-
tition, the long-term social preferences of the median will not necessarily
be re¬‚ected in policies. Nevertheless, having a theory of preferences, such
as the one presented in this chapter, is the ¬rst step toward explaining
cross-national variance in social policy.
The model also suggests a solution to the long-standing puzzle that in-
come equality is linked to higher social spending when comparing across
countries. As we know from Chapter 1, vocational training activity is
strongly positively related to pretax income equality (the correlation co-
ef¬cient is .73 using d9/d1 earnings ratios), and if a speci¬c skill structure
is simultaneously linked to more spending, as suggested by the model and
evidence presented in this chapter, it follows that income equality and so-
cial spending will go hand in hand. In the pure Meltzer-Richard model,
this is ruled out because the pressure for redistribution is always greatest in
countries with the most skewed distribution of income.




Appendix 3.A. Mathematical Proofs

Derivation of Results (3.12) and (3.13) in Model III
The choice of the optimal R requires that
VR ≥ 0 ” β · u (g) · 2g/w = γ · u (R)
Totally differentiating both sides we get

· [g · u (g) + u (g)]
dR w
=
2g 2
dg
β· · u (g) + γ · u (R)
w
Because the denominator is negative,
dR
>0 [gu (g) + u (g)] < 0
iff
dg
112
Explaining Individual Social Policy Preferences


which implies
gu (g)
RRA(g) ≡ ’ >1
u (g)
where RRA(x) is the Arrow-Pratt de¬nition of relative risk aversion de¬ned
at c = x. The inequality conditions speci¬ed in (3.12) and (3.13) follow
directly.

Proof for Result I in Model IV
1. Note ¬rst that t = 1 maximizes t/(1 + t) when 0 ¤ t ¤ 1. Also, if t =
1, R = w/2.
2. From (3.6) the necessary condition for optimal R is
2s g 2g
± · u (s g) · 1 ’ + β · u (g) · 1 ’ + γ · u (R) ≥ 0
w w
(A3.1)

If R = w/2, s g = g = R; hence, the maximum combination of s g and g at
which R = w/2, assuming it exists, requires that this condition holds with
equality and that u (s g) = u (g) = u (R). These conditions imply directly
that
w w
± · s g + β · g = (± + β + γ ) · =y=
2 2


Proof for Results II and III of Model IV
The necessary condition for optimal choice of R is VR (R, s , g) = 0. This
is given by (A3.1).
Totally differentiating VR gives
2s g 2R 2
±· u (s g) · ’1 · 1’ · g + u (s g) · g · · ds
w w w
2s g 2R 2
+ ± · u (s g) · ’1 · 1’ · s + u (s g) · s · · dg
w w w
2g 2R 2
+ β · u (g) · ’1 · 1’ + u (g) · · dg
w w w
2 2
2s g 2g
= ± · u (s g) · ’1 + β · u (g) · ’1 + γ · u (R) · d R
w w
(A3.2)
113
Political Foundations of Social Policy


Note: (1) The term in curly brackets on the right-hand side, which
will be called B, is negative. (2) We can write (s g ’ w/2) · (1 ’ 2R/w) =
s g ’ w/2. And (3)

[u (s g) · (s g ’ w/2) + u (s g)]
s g ’ w/2
= u (s g) · 1 ’ RRA · ≡ u (s g) · L(s g) (A3.3)
sg

So (A3.2) can be written as

u (s g) · L(s g) · ± · g · d s + u (s g) · L(s g) · ± · s · d g + u (g) · L(g) · βd g
= (w/2) · B · d R (A3.4)

Since d y = ± · g · d s + ± · s · d g + β · d g, we can further rewrite (A3.4) as

2± · β · g u (s g) · L(s g) ’ u (g) · L(g)
dR = · · ds
±s + β
wB
u (s g) · L(s g) · ±s + u (g) · L(g) · β
2
+ · · dy (A3.5)
±s + β
wB

To prove Results III and IV, note that in terms of (A3.5) ‚ R/‚ y = d R/d y
and ‚ R/‚s = d R/d s. First, it is shown that L(s g) < L(g). From the def-
inition in (A3.3), this follows if s > 1 “ as is the case apart from purely
general skills “ and if RRA > 0. Result III is that sgn ‚ R/‚ y < 0 if RRA <
s g/(s g ’ w/2). Since B < 0, L(s g) < L(g) and u (x) > 0, this follows
from (A3.5) if L(s g) > 0. This requires that RRA < s g/(s g ’ w/2). This
is a suf¬cient condition: A necessary and suf¬cient condition is that the
numerator of the second term in square brackets on the right-hand side of
(A3.5) is positive.
Result IV is that sgn ‚ R/‚s > 0. Since B < 0, this requires that the nu-
merator in the ¬rst square bracket on the right-hand side of (A3.5) is
negative. Since u (s g) < u (g) from diminishing marginal utility, a suf¬-
cient condition is that L(s g) < L(g), which is true so long as RRA > 0
and s > 1. So Result IV follows from the existence of risk aversion and
speci¬c skills.




114
Explaining Individual Social Policy Preferences


Appendix 3.B. Deriving the Estimating Equation
In this appendix, it is demonstrated that the estimating equation used in the
analysis

R =k+b ·y +c ·s (A3.6)

is equal to

‚R ‚R
R=k+ ·y+ ·s (A3.7)
‚y ‚s

where (A3.7) is a ¬rst-order Taylor expansion of VR (R, s , g) = 0 and y =
± · s · g + β · g evaluated around (R, s , g) = ( R, s , g) ≡ x.

Proof. The ¬rst-order Taylor expansion of VR is given by

VR,g
VR,s
R=K+ s+ (A3.8)
g
VR,R VR,R

In terms of (3.5A),

u (s g) · L(s g) · ± · g
VR,s (x)
= (A3.9.1)
(w/2) · B
VR,R (x)

and
u (s g) · L(s g) · ± · s + u (g) · L(g) · β
VR,g (x)
= (A3.9.2)
w
VR,R (x) ·B
2
The ¬rst-order Taylor expansion of y is

y = k(x) + [± · s + β] · g + [± · g] · s (A3.10)

Rewrite (A3.10) as

y ’ k(x) ’ ±g
g=
±s + β

and substitute into (A3.8), using (A3.9.1) and (A3.9.2). This yields (A3.7).




115
Political Foundations of Social Policy


Appendix 3.C. Detailed Information about Variables

Dependent Variables
The spending variable, R, is based on four issue items in the ISSP surveys.
The ¬rst three are based on the following question:
Listed below are various areas of government spending. Please show whether you
would like to see more or less government spending in each area. Remember that if
you say ˜much more™, it might require a tax increase to pay for it. The respondent
is then presented with the different spending areas (unemployment, health, retire-
ment) and the following range of possible responses: 1. Spend much more; 2. Spend
more; 3. Spend the same as now; 4. Spend less; 5. Spend much less; 8. Can™t choose,
don™t know.

The fourth variable is based on the following question:
Here are some things the government might do for the economy. Please show which
actions you are in favor of and which you are against. Please tick one box in each line.
One of the actions is: Support for declining industries to protect jobs: 1. Strongly
in favor of; 2. In favor of; 3. Neither in favor of nor against; 4. Against; 5. Strongly
against; 8. Can™t choose, don™t know; 9. NA, refused.



Independent Variables
s1 and s2 In some countries individuals were classi¬ed using an earlier
version of ISCO (ISCO-68). However, these classi¬cations can be trans-
lated into ISCO-88 with considerable consistency using a coding scheme
developed by Harry Ganzeboom at Utrecht University (see Ganzeboom
and Treiman 1996 and www.fss.uu.nl/soc/hg/ismf for details). The Swedish
occupational classi¬cation is based on an amended version of an older edi-

tion of ISCO. Statistiska Centralbyran (Statistics Sweden) provided us with
a conversion table to translate these codes into ISCO-88 in a reasonably
consistent manner. Britain uses its own national classi¬cation system, but it
is closely related to ISCO-88 and likewise uses skills as the basis for the clas-
si¬cation. The British translation codes were received from U.K. National
Statistics. The only problematic case is Italy, where the few broad categories
used in the 1996 ISSP survey are completely unrelated to the ISCO-88
categories. Instead, I went back to an earlier 1990 ISSP study (ISSP 1993),
which contains a somewhat more detailed occupational variable for Italy.
Using this variable in conjunction with information on educational levels


116
Explaining Individual Social Policy Preferences


enabled us to map the Italian codes to the one-digit ISCO-88 level in a
fairly consistent manner. Yet, because of the lack of direct correspondence,
the results for Italy must be viewed with caution.

General Skills (used in the denominator of s 2 and s 4 ) The variable is used
as a proxy for g and has ¬ve levels: 1, still at school; 2, incomplete primary; 3,
completed primary degree or lower; 4, incomplete secondary; 5, completed
secondary; 6, incomplete and completed semi-higher degree, or incomplete
university degree; and 7, completed university degree. Alternatively, one

<<

. 4
( 11)



>>