. 8
( 11)


to this. German and Japanese car makers pushed Americans ones to adopt
more ¬‚exible and ef¬cient production methods, while competition from
American high-tech ¬rms triggered massive investments in information
technology everywhere else. But competition and innovation is a general
feature of capitalism and would surely have occurred in large measure even
in the absence of growing trade and capital mobility.

5.4. Conclusion: The Welfare State at the End of Industrial Society
Based on data for sixteen OECD countries over a thirty-¬ve-year period,
there is little evidence that trade, or capital mobility for that matter, has
played an important role in the expansion of the modern welfare state. Cor-
respondingly, there is little support for the idea, proposed by Katzenstein
(1985), that large countries will become more akin to small corporatist wel-
fare states as they grow increasingly exposed to the vagaries of international
markets. On the other hand there is also little evidence that globalization is
a major threat to the welfare state. Instead, what has propelled much of the
expansion of the welfare state since the early 1960s is a dynamic process of
technological progress in manufacturing combined with the saturation of
markets for agricultural and industrial products.
Somewhat paradoxically, therefore, the very process that has decimated
the traditional industrial working class underpins the growth of the welfare
state. However, it does not mean that partisan politics has played no role.
By and large, left governments have been more prone to raise spending
in response to deindustrialization than have right governments. Likewise,
institutions that tend to be associated with “left” politics, namely elaborate
vocational training and proportional representation, have greatly magni¬ed
the spending effects of deindustrialization.
Coping with Risk

Perhaps the most pressing political question ¬‚owing from the results
is what happens when the deindustrialization process comes to a halt.
Figure 5.7, which shows the relationship between the share of the work-
ing age population outside industry and the speed of deindustrialization,
clearly suggests that such a slowdown is occurring. From very different
starting points in the early 1960s (the labeled entries to the left of the
dotted line), all countries have been converging on a “postindustrial” equi-
librium in the 1990s (the labeled entries to the right of the dotted line),
albeit at very different speeds. This process of convergence has been ac-
companied by a slowdown in the pace of deindustrialization (the vertical
axis). If the insecurities associated with past deindustrialization spread de-
mands for compensation and for socialization of risks well into the middle
classes, will the stabilization of the sectoral employment structure reverse
the process?
As already noted, this is unlikely to happen. The costs of retrenchment
are concentrated on vocal constituencies, and most people have simply

Figure 5.7 Convergence toward the service economy, 1960“1995.
Notes: Labeled dots are the ¬rst (1960“3) and last (1992“5) observations in the data set. Early
observations fall to the left of the dotted line, except in the cases of Canada and the United
States, where both observations are to the right of the line. The unlabeled dots represent the
observations of intervening years.

Forces of Change

become too dependent on the welfare state to want its dismantlement
(Pierson 1994, 1996). However, the slowdown is likely to be accompa-
nied by more intense distributive battles between those in secure and those
in insecure labor market positions. If people in secure positions know that
they are highly unlikely to end up in insecure ones (i.e., face low labor
market risks), they have less reason to be solidaristic with those in insecure
positions. This general conclusion stands in sharp contrast to the postma-
terialism thesis advanced by Inglehart (1987, 1990) and others, and it will
be subject to more intense analysis in Chapter 6.


New Tradeoffs, New Policies

This chapter applies the welfare production regime argument to the case of
service employment and tests its economic and political implications. Be-
cause job growth in high-skilled, better-paid services is limited by the size
of the domestic market, whereas growth in low-skilled, low-paid service
jobs is hampered by wage compression and high social costs, the transition
toward a more sheltered postindustrial economy has produced a dif¬cult
tradeoff between equality and employment, mediated only by the willing-
ness of the government to increase public employment or subsidize private
employment at a cost to tax payers.
As Anne Wren and I have argued (Iversen and Wren 1998), governments
initially responded to this “trilemma” in a manner that clearly re¬‚ected par-
tisan preferences and broader institutional conditions. Thus, right govern-
ments in liberal market economies sought to further deregulate labor mar-
kets, whereas governments in coordinated market economies embarked on
policies to either ration work (primarily in countries with independent cen-
tral banks and strong Christian democratic parties) or to increase spending
on public employment (primarily in countries with highly centralized wage
bargaining and strong social democratic parties). The division between vari-
eties of capitalism described in Part I has thus been overlaid by new divisions
that are the result of different political responses to deindustrialization.
This chapter takes a fresh look at the tradeoffs and how governments in
different parts of the world have adapted to them. After a period of intensely
partisan responses, governments are increasingly embarking on strategies “
including deregulation of part-time employment, new educational initia-
tives aimed at upskilling, service trade liberalization, and tax reforms “
that are designed to supplant the trilemma and return to a more virtuous
cycle between social protection, equality, and private sector employment.
Forces of Change

In the ¬nal section of this chapter, I discuss the nature of these new policy
initiatives, including their underlying coalitional dynamics, their record of
success so far, and their prospects for success in the future “ in particular the
possibilities for recreating the linkage between generous social protection,
high wages, and open trade that existed in the golden era of the industrial
economy (outlined in Chapter 2).

6.1. A New Set of Tradeoffs
At the crux of the regime argument applied to services is the notion that
wage structure affects relative prices and employment. High labor costs
for low-skilled labor raise relative prices in low productivity services and
are therefore not conducive to the rise of low-cost, labor-intensive services
(Esping-Andersen 1990; Iversen and Wren 1998). This section examines
the empirical basis for this thesis in detail, focusing ¬rst on prices and then
on employment. While the main aim is to establish the economic tradeoffs
faced by governments in a deindustrializing economy, I note that the analy-
sis provides a solution to one of the most enduring puzzles in international
economics, namely why domestic price levels vary systematically across
countries “ what is known as the purchasing power parity puzzle.

6.1.1. Prices
It is very dif¬cult to compare prices on particular services directly across
countries. Detailed international classi¬cations of services do not exist, and
even very similar services, such as long-distance phone rates, typically ex-
hibit so many subtle product differences that systematic cross-country com-
parisons are all but impossible. The approach I adopt here relies instead
on international comparisons of the prices on comparable baskets of goods
and services, called purchasing power parities (PPP).1 Although these com-
parisons are based on the domestic prices of both manufactures and ser-
vices, because prices are likely to diverge more on sparsely traded services
than on highly traded manufactures, lasting price differentials are likely
to primarily re¬‚ect differences in prices on services. Another advantage of
using these data is that I can plug into an well-established theoretical and
empirical tradition in economics, purchasing power parity theory, with a
clear methodology and an agreed set of empirical puzzles.
1 The baskets are not identical, but the OECD encourages countries to follow the same
standards, and the measures are widely considered reasonably reliable for comparison.

New Tradeoffs, New Policies

The PPP approach to international price comparisons is to ¬rst deter-
mine the exchange rate, usually in dollars, that would buy the same basket
of products in different countries. This PPP exchange rate can then be
compared to the actual exchange rate. If the latter is higher, in dollars, than
the PPP exchange rate, the domestic price level is above the price level
in the comparison country. A simple measure of the degree of over- or
undervaluation at time t is the real exchange rate q:
p t—
qt = e t (6.1)
where e is the nominal (dollar) exchange rate, p is the domestic consumer
price index, and p* is the U.S. consumer price index. The fraction p/p* is
the PPP exchange rate.
According to the law of one price, identical products that can costlessly be
transported from one location to another should have the same price in all
localities. If not, goods-market arbitrage would quickly eliminate any price
discrepancies. Hence, at least over a period of time suf¬ciently long to allow
for price adjustments to shocks, q should equal 1. This is the fundamental
null hypothesis in PPP theory. In the strong version of the theory, q is always
equal to 1. But if prices are sticky, short-term price or exchange rate shocks
will not be immediately eliminated. In the weaker version, therefore, PPP
theory predicts parity only after the period of time it takes for prices to
adapt. Most crucially, perhaps, if nominal wages are set through collective
wage bargaining, price adjustments will be delayed by the frequency with
which wages are renegotiated.
There is a wealth of empirical studies seeking to test the PPP thesis
against data on real exchange rates. The results are mixed. Early studies
found no evidence that deviations from PPP were ever corrected, imply-
ing that real exchange rates could be treated as random walks (Krugman
1978; Adler and Lehmann 1983; Meese and Rogoff 1988). This ¬nding was
disconcerting, if not outright embarrassing, for economists. Subsequent ev-
idence, however, suggests that real exchange rates do eventually revert to
their mean value. Estimates of the time it takes for disturbances to decay
vary, but most imply half-life ranges of 3“6 years (a half-life is the time it
takes for a PPP deviation to decay by 50 percent). Recent examples of this
research include Frankel and Rose (1996), Lothian and Taylor (1996), Oh
(1996), and Papell (1997).
The ¬ndings of mean reversion are comforting for PPP theory, but, as
noted by Rogoff, “a half-life of three to ¬ve years [is] seemingly too long to
Forces of Change

be explained by nominal rigidities” (1996, p. 648). A related, and potentially
more serious, problem is that many of the empirical results leave open the
possibility that real exchange rates revert to means that are persistently
different from PPP, or a real exchange rate of one. Indeed, this is what the
empirical results in this section clearly imply. I argue that these deviations
from parity, the purchasing power parity puzzle, are in large measure the
result of cross-national differences in the level of social protection “ in
particular differences in the politically mediated wage structure.
To see this, ¬rst express Equation (6.1) in natural logarithms:

ln q t = ln e t + ln p t— ’ ln p t (6.2)

If PPP holds in the long run, the right-hand side must revert to zero over
time (the real exchange rate is unity). By implication, deviations from zero
must be temporary, and any disturbance must be followed by a decay pro-
cess. This decay process can be estimated using the following regression

ln q t = ρ ln q t’1 + µt (6.3)

where ρ must be between 0 and 1 for disturbances to decay over time.
Alternatively, by subtracting ln qt’1 on both sides, (6.3) can be written as

ln q t = δ ln q t’1 + µt (6.4)

where δ = (ρ ’ 1) is each period™s decay in the initial deviation from PPP.
For example, if δ = ’.25 it means that disturbances are damped out at
25 percent in each period.
Tests of this model on single currencies in the post“Bretton Woods era
all tend to fail to reject the null hypothesis of a random walk. As Wu sum-
marizes the evidence, “while results obtained by employing long-horizon
data (century plus) offer limited support for the validity of PPP in the long
run, those obtained by using the post“Bretton Woods data remain largely
negative” (Wu 1996, p. 55). To overcome these problems, a number of
recent studies use pooled data for several countries. This method is partic-
ularly relevant to our purposes because it allows us to detect cross-national
differences in deviations from PPP.
If real exchange rates are measured against the same base currency, here
the U.S. dollar, the model to be estimated is simply

ln q i,t = δ ln q i,t’1 + µi,t (6.5)
New Tradeoffs, New Policies

where i indexes countries. Because there is leftover ¬rst-order serial corre-
lation, it is common practice in this literature to include a lagged difference
term.2 With this re¬nement, Table 6.1, column 1, shows the results of es-
timating this equation on annual data for eighteen OECD countries in the
post-Bretton Woods era.
Note that the parameter on the lagged dependent level variable is nega-
tive so that deviations from PPP dampen over time. This process of mean
reversion, however, is rather slow with a half-life of almost 4 years. This
is at the higher end of the range of existing estimates, and it is signi¬-
cantly above the estimates in some studies of OECD currencies for the
post“Bretton Woods period where the half-life tends to be much shorter.
For example, Wu (1996) ¬nds a half-life of 2.3 years, Papell (1997) one of
2.5 years, and Oh (1996) one of between 1 and 2 years. However, this
discrepancy turns out to be the result of a simple difference in model speci-
¬cation. Whereas I have used absolute PPP levels, the practice in the litera-
ture is to measure real exchange rates as deviations from their national means.
This procedure is equivalent to controlling for country-speci¬c effects:
ln q i,t = δ ln q i,t’1 + b i d i + µi,t (6.6)

where di is the dummy variable for country i. Column 2 in Table 6.1 shows
the results of estimating this ¬xed effect model.
Note that there is a notable increase in the explained variance from
the model without dummies, and an F-test unambiguously shows that the
dummies belong in the model. Moreover, the half-life of deviations from
parity is now signi¬cantly reduced to only 1.7 years. This is more consistent
with a sticky price hypothesis than the 4-year half-life in Equation (6.6). It
may still be too long to be fully explained by slowly adjusting wages and
prices, and it is possible that trade costs at the consumer level provide an
important reason for slow price adjustments (see Rogoff 1996; Obstfeld and
Rogoff 2000). But it is quite in agreement with existing results.
The key issue for our purposes, however, is not the speed of price ad-
justments but the fact that real exchange rates in many countries never con-
verge to PPP, as implied by the signi¬cant effects of the country dummies.
To ¬nd out how much, on average, the real exchange rate of a country
is overvalued, we take the inverse of the log value of the parameter for

2 Higher order lags show no signi¬cant effects.

Forces of Change

Table 6.1. Real Exchange Rates for Eighteen OECD Countries, 1973“1997 a

(1) (2) (3) (4) (5)
’0.66—— ’1.65——— ’0.97——
Intercept 0.02 0.00
(3.67) (0.00) (’2.18) (’3.39) (’2.80)
’0.17——— ’0.32——— ’0.34——— ’0.36——— ’0.25———
ln(Real exchange rate)t’1
(’8.12) (’11.62) (’11.73) (’9.87) (’8.48)
0.07—— 0.18——— 0.11———
ln(GDP per capita)t “ “
(2.19) (3.66) (3.06)
d5/d1 ratiot’1 “ “ “
(’0.73) (’4.00)
Australia “ 0.02 0.03 0.01 “
Austria “ 0.02 0.05 0.07 “
Belgium “ 0.02 0.04 0.04 “
Canada “ 0.01 0.02 0.06 “
Denmark “ 0.08 0.10 0.06 “
Finland “ 0.07 0.09 0.09 “
France “ 0.02 0.04 0.05 “
Germany “ 0.05 0.07 0.06 “
Ireland “ 0.04 0.11 “
’0.04 ’0.01
Italy “ 0.03 “
Japan “ 0.07 0.09 0.12 “
Netherlands “ 0.03 0.06 0.07 “
’0.05 ’0.03 ’0.01
New Zealand “ “
Norway “ 0.11 0.13 0.09 “
Sweden “ 0.09 0.11 0.09 “
Switzerland “ 0.09 0.10 0.13 “
’0.02 ’0.00
United Kingdom “ 0.02 “
0.37——— 0.42——— 0.43——— 0.41——— 0.37———
Lagged difference term
(8.60) (10.03) (10.24) (7.64) (6.99)
R-squared .208 .310 .317 .340 .265
Number of observations 450 450 445 284 284
Signi¬cance levels: ——— <.01; —— <.05 (t-scores in parenthesis).
a Exchange rates are measured against the U.S. dollar and expressed in logged differences.

The reference country for the country dummies is the United States.
Sources: Exchange rates: OECD (1999c); GDP per capita: World Bank™s Global Development
Network Growth Database at www.worldbank.org/research/growth/GDNdata.htm (this variable
is constructed from Penn World Table label=”5.6”, Global Development Finance and World
Development Indicators); d5/d1 ratios: OECD Electronic Data Base on Wage Dispersion, undated.

that country™s dummy and subtract 1 (parity) from the result. The long-
run equilibrium value is determined by dividing by ’δ. Using this formula,
the Swedish real exchange rate, for example, turns out to be an average of
31 percent overvalued compared to the U.S. dollar. Hence, a dollar would
New Tradeoffs, New Policies

on average buy 31 percent less in Sweden than in the United States during
the period 1973“97. For comparison, the Swedish real GDP per capita grew
by a mere 9 percent in this period, implying that an elimination of the price
gap between Sweden and the United States would add three times more to
the Swedish GDP than what real economic growth accomplished during
this period.3 Deviations from PPP are clearly not trivial by any reasonable
economic yardstick. But how do we explain them?
In a classic formulation of PPP theory, Balassa and Samuelson propose
one possible solution (Balassa 1964; Samuelson 1964). They hypothesize
that because productivity in traded sectors rises faster than productivity
in nontraded sectors “ in accordance with the Baumol hypothesis “ rich
countries will have higher real exchange rates than poor countries insofar
as competitive labor markets eliminate wage differentials between sectors.
The logic is that lower productivity growth in one sector will result in higher
real labor costs, and therefore higher relative prices. High per capita GDP,
which equals high relative productivity in the traded sector, will therefore
be associated with a high real exchange rate.
Rogoff (1996) shows that this proposition is supported by data covering
both rich and poor countries, but he also shows that per capita income fails
to explain most of the variance among developed countries. The same con-
clusion follows if GDP per capita is included in our previous regression (see
column 3 of Table 6.1). A 1 percent increase in per capita income raises the
long-term real exchange rate by about 0.2 of a percentage point, although
this estimate is not very accurate (the effect is statistically different from
zero only at a .05 signi¬cance level). More importantly, the country-speci¬c
effects are not much affected by the inclusion of per capita income. The
correlation coef¬cient between the parameters on the country dummies
before and after inclusion of per capita income is 0.98, and the mean pre-
dicted currency overvaluation in fact rises from 11 to 17 percent (in the case
of Sweden it is now 36 percent).
To explain the variance among developed countries, we need to drop
the assumption of perfectly competitive labor markets. If skills are partly
¬rm, industry, or sector speci¬c, or if there are market imperfections, then
wages for workers at similar skill levels can vary considerably between traded
and nontraded sectors (as well as within these). Yet, such productivity-based
wage differentials do not necessarily affect workers in the same sector to the

3 In fairness, though, it should be mentioned that 1993 was a terrible year for the Swedish
economy. Using 1994 instead shows a 23 percent increase in real per capita.

Forces of Change

same extent. Some may be more unionized than others, while some will be in
product market segments with price-inelastic demands that make it easier
to externalize higher wage costs. Still others will enjoy monopoly power
through accreditation, certi¬cation, and other measures that may restrict
labor supply. In any of these scenarios the most important determinant of
the price effects of differential productivity growth is the extent of intra- and
inter-industry wage-coordination. And we know that coordinated wage-
setting institutions tend to raise relative wages for low-paid workers, which
tend to be concentrated in certain nontraded services.
It should be noted that this logic does not require wages to be institution-
ally coordinated across sectors. The relative price effect will occur whenever
the wage leaders in low-productivity, nontraded industries, based on their
labor market power, are able to keep up with wages in traded sectors, and
wages within the sector are tightly coupled as a result of intra-industry wage
coordination. Under these conditions, the relative prices of services pro-
duced with low-paid, low-productivity labor will be higher. That said, it is
generally true that the higher the level of coordination, the more wage com-
pression across sectors there is, and the greater the relative price effect is.
The effect of wage compression is illustrated in Figure 6.1, which shows
the relationship between earnings dispersion and the percentage overval-
uation of countries™ real exchange rates (using the procedure described
previously in the Swedish example), controlling for per capita income. The
dispersion measure is OECD™s ¬gures for the earnings of a worker with the
median earnings relative to a worker with earnings in the bottom decile
(d5/d1 ratios).4
Note that the relationship is in the predicted direction and moderately
strong (r = .56). For example, the three egalitarian Scandinavian coun-
tries have signi¬cantly “overvalued” real exchange rates, whereas three
inegalitarian countries “ Britain, Canada, and the United States “ have rel-
atively undervalued currencies. Italy is clearly an outlier, exhibiting a com-
pressed wage structure but also a relatively “cheap” currency. Unlike other
¬gures, however, OECD™s estimate of Italian wage compression varies. In
the 1997 OECD Employment Outlook, the average d5/d1 ratio reported
for Italy is 1.8, whereas the average in OECD™s Electronic Data Base on

4 The numbers are averages for the period from 1979, when data starts in most countries, to
the early 1990s. I used d5/d1 rather than d10/d1 ratios because dispersion at the lower end of
the earnings distribution is more pertinent to the argument than earnings for high-income

New Tradeoffs, New Policies


Overvaluation of currency

Denmark Switzerland
Finland Japan

20 Netherlands
Belgium France
United States
New Zealand
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
d5/d1 wage ratios

Figure 6.1 Wage dispersion and average currency overvaluation, 1993“1997.
Note: Figures for overvaluation are based on the estimated parameters for the country dum-
mies model, Table 6.1, column 3.
Source: OECD, Electronic Data Base on Wage Dispersion, undated.

Wage Dispersion (undated) is 1.4. Using the number in the printed version
eliminates Italy as an outlier and raises the correlation coef¬cient between
dispersion and real exchange rates to .70. If the electronic ¬gure is more
accurate, a plausible explanation may center on the many and large deval-
uations of the Italian Lira in the post-Bretton Woods period.
If we include d5/d1 ratios directly in the regression equation (column 4
in Table 6.1), we can get a more precise estimate of the price effects of
the wage structure.5 Because we only have dispersion data for a subset
of the original period from 1973 to 1997, the number of observations is
signi¬cantly reduced. In some cases, data are only available for part of the
1980s, and in one case, Ireland, there is only a single observation (for 1993).
Another problem is multicollinearity: The country dummies explain nearly
97 percent of the variance in wage dispersion, making it dif¬cult to obtain
statistically signi¬cant results for the dispersion variable. This is evident if
we take the dummies out of the equation (column 5) since the parameter
for wage dispersion then becomes highly signi¬cant.

5 I am using the electronic data here and throughout the rest of this chapter.

Forces of Change

Removing the dummies also slightly increases the effect of dispersion,
but the parameters are very similar in both regressions. Because we are
con¬dent about the relationship between wage structure and real exchange
rates, we can therefore get a good sense of the substantive impact from ei-
ther estimate. It turns out that the immediate (one year) effect of a 1 percent
increase in dispersion is to reduce prices by 0.15 percent (using the ¬xed
effect model). After 2 years, the effect is 0.21 percent, and in the long run it
approaches 0.41 percent. If we converted the dependent variable into per-
centage over- or undervaluation, the long-term parameter on the dispersion
variable would be ’52, which is nearly identical to the slope on the regres-
sion line in Figure 6.1 (’53). The results of the multivariate pooled times-
series analysis thus con¬rm the simple bivariate cross-sectional analysis.
A nice check on these results is to make use of the so-called Big Mac Index
developed by The Economist. The Big Mac PPP is the exchange rate that
would mean that a McDonald Big Mac hamburger, presumably an almost
identical product wherever it is sold, costs the same in that country as in
the United States. Comparing actual exchange rates with PPPs indicates
whether a currency is under- or overvalued, or, alternatively, whether a Big
Mac is cheap or expensive. And because Big Macs are not internationally
traded, we have a direct measure of prices in a nontraded industry.
The index is available only for a subset of the OECD countries, and only
for 14 years. Still, relating d5/d1 ratios to average currency overvaluation
by this measure (Figure 6.2) produces a very similar pattern to that in
Figure 6.1. The most notable difference is that the range of under- and
overvaluation is almost twice that for the basket of goods and services used
to calculate regular PPP ratios. The likely reason is that the price of a Big
Mac is determined primarily by the cost of nontraded inputs, especially low-
skilled labor, whereas the regular PPP basket of goods and services contain
both traded and nontraded components as well as labor inputs across the
wage scale.
To take an illustrative example on the regression line, the average overval-
uation of the Swedish currency compared to the Big Mac PPP is 50 percent,
whereas the overvaluation of the currency compared to the basket PPP is
about 35 percent in the same period. We can explain the Big Mac overval-
uation directly in terms of differential wages. Assuming that McDonald™s
workers™ wages are on average in the bottom twentieth percentile of the
earnings distribution in both countries, which seems reasonable, OECD
wage data show that a Swedish worker is earning 52 percent more than
an American worker during the period for which both wage and Big Mac
New Tradeoffs, New Policies

80 Switzerland

Big Mac overvaluation

40 France
Italy Germany UK

United States
New Zealand
-20 Australia

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
d5/d1 wage ratios

Figure 6.2 Wage dispersion and the relative price of Big Macs, 1988“2000.
Sources: Big Mac Index: The Economist, 6.9.86, 17.1.87, 2.4.88, 15.4.89, 8.5.90, 13.4.91,
18.4.92, 17.4.93, 9.4.94, 15.4.95, 27.4.96, 12.4.97, 11.4.98, 3.4.99, 29.4.00. d5/d1 ratios:
OECD, Electronic Data Base on Wage Dispersion, undated.

price data are available (at the bottom tenth percentile of the distribution,
the ¬gure is 73 percent). This is almost exactly equivalent to the surcharge
for a Big Mac in Sweden compared to that in the United States. More-
over, this gap in the wages of low-paid workers is not simply a re¬‚ection of
higher average Swedish wages (which is essentially the Balassa-Samuelson
hypothesis) “ the mean wage in Sweden is almost equivalent to that in the
United States during this period (in fact it is 6 percent higher). Swedes pay
more for their Big Macs than Americans primarily because Swedish wages
are more compressed.6
From this analysis, it seems safe to conclude that permanent differences
in the price level of different countries “ or deviations from PPP “ are
in large measure attributable to a combination of lower productivity in
nontraded services and institutionally mediated wage compression. After
these effects have been taken into account, the real exchange rate data are
fully compatible with the thesis that prices on traded goods converge to PPP

6 Of course, there are inputs other than labor that matter for hamburger prices. But many
of these are traded and therefore less likely to account for cross-national differences, and
about three quarters of all costs in the hotel and restaurant business are due to labor costs.

Forces of Change

after a lag that can be explained by a combination of sticky prices/wages
and trade imperfections.

6.1.2. Employment
When relative prices of an industry rise, all else being equal, production
and employment fall “ assuming, of course, that prices and employment
are market determined.7 The second effect of wage compression is, there-
fore, an employment effect. The magnitude of this effect, however, varies
across sectors. As discussed previously, intraoccupational compression in
the heavily traded manufacturing sector primarily has the effect of shift-
ing the comparative advantage of countries toward higher value added and
skill-intensive activities without reducing employment. As long as unions
are fairly encompassing, they have an incentive to set general wages at a
level consistent with maintaining international competitiveness.
In nontraded sectors, by contrast, because specialization has not evolved
to the same extent, wage and other forms of social protection lead to ris-
ing relative prices and reduced employment growth if services are market
supplied (Iversen and Wren 1998; Scharpf 2000; Manow 2002). The mag-
nitude of this effect, however, is conditioned by several factors. First, and
rather obviously, the greater the capacity of an industry to increase pro-
ductivity is, the smaller the relative price effect is, and the smaller the
employment effect is. Second, the lower the price elasticity of demand
is, the easier it is for ¬rms to externalize the costs of higher real wages
and not cut employment. This tends to be true in skill-intensive services
such as consulting and medicine. Conversely, the more an industry re-
lies on low-skilled labor to produce easily substitutable products, and the
higher the proportion of total costs is that goes to wages, the more sensitive
the industry will be to wage compression. Consequently, companies in an
industry that is lagging in productivity, faces price-elastic demand, relies
heavily on low-paid labor, and engages in production where labor costs
constitute a high share of total costs will be particularly vulnerable to wage
With these distinctions in mind, Table 6.2 compares three private ser-
vice sectors in terms of productivity growth, share of labor costs in total
costs, and relative wages. Together these services account for the entire

7 I consider the possibility of nonmarket provision of services later in this chapter, where the
cost of provision is covered through taxation.

New Tradeoffs, New Policies

Table 6.2. Productivity, Relative Earnings, Labor Shares, and Private Employment in
Four Sectors

Relative Labor Change in
Wagesb Sharesc Employmentd
Manufacturing (Isic 3) 3.1 100 0.70
Business services (Isic 8) 0.0 106 0.41 2.7
Wholesale and retail trade (Isic 6) 1.2 82 0.73 2.6
Social and personal services (Isic 9) 70 0.69 4.7
a Average growth in total factor productivity 1970“95.
b Average sector wages as a percent of manufacturing wages.
c Average share of labor compensation in value added.
d Change in employment as a share of the working-age population (population-weighted).

Source: Calculated from OECD (1999a).

rise in private employment between 1970 and 1995.8 Manufacturing em-
ployment, which has everywhere declined, is used as a reference for the
It is apparent from these data that business services (Isic 8) are quite
different from other services. Although productivity is stagnant, relative
wages are much higher than in other services “ presumably a result of more
skill-intensive production “ and the factor share of labor in total output is
notably smaller than in other sectors. The latter is likely a re¬‚ection of the
importance of proprietary standards and intellectual capital that simulta-
neously raise capital intensity and the level of wages. The wholesale and
retail sector (Isic 6), and social and personal services (Isic 9), by contrast, are
both relatively low-pay sectors with high labor shares and low productivity.
Social and personal services are particularly dependent on low-paid labor
and record the lowest capacity for productivity growth. On the other hand,
wholesale and retail trade is the most labor-intensive sector. Unfortunately
it is not possible to get any systematically comparable indicators for price
elasticity. Yet, one would expect that the more knowledge-intensive char-
acter of business services makes demand less price elastic than in the other
two sectors where price competition tends to be ¬erce.
To explore the linkage between wage structure and private employment,
the four panels in Figure 6.3 show the relationship between dispersion of
industry earnings and annual change in private employment (as a percent

8 I ignore a small sector of transport and storage services where employment has declined
slightly (0.14 percent of the working-age population).

(b) Finance, insurance, real estate, and business services

(a) Manufacturing 0.6
0.4 0.5
-0.1 0.2
-0.6 -0.1

Change in employment (annualized)

Change in employment (annualized)
0.1 0.15 0.2 0.25 0.3 0.35 0.1 0.15 0.2 0.25 0.3 0.35
Dispersion of earnings Dispersion of earnings

(c) Wholesale, retail, restaurants, and hotels (d) Community, social, and personal services




0.2 0.2

0.1 0.1

0 0

-0.1 -0.1

-0.2 -0.2

Change in employment (annualized)
-0.3 -0.3
Change in employment (annualized)

-0.4 -0.4
0.1 0.15 0.2 0.25 0.3 0.35 0.1 0.15 0.2 0.25 0.3 0.35
Dispersion of earnings
Dispersion of earnings

Figure 6.3 Dispersion of earnings and employment growth in four economic sectors.
New Tradeoffs, New Policies

of the working-age population) by sector. As in Iversen and Wren (1998),
I use ¬gures for industry dispersion in wages, rather than d5/d1 ratios,
because these are available for the same countries and years as the employ-
ment data. Yet, as I show in the subsequent statistical analysis, the pattern
is similar using d5/d1 ratios.
Because the yearly changes (small dots) are very volatile, the graphs also
use 5-year averages (squares) as well as whole-period averages (triangles).
Note that for two sectors, manufacturing and business services, there is
virtually no relationship between wage dispersion and employment growth.
The only difference between these sectors is that manufacturing has experi-
enced a dramatic drop in employment, while producer services have been a
net gainer (as indicated by the averages on the y-axis). Yet, while the growth
in the number of jobs in business services has been rapid, in no country
except the United States has the gain in employment in business services
been suf¬cient to compensate for the loss in manufacturing.
The sectors that account for most of the divergence in private sector
employment performance across countries are wholesale and retail trade
(Isic 6), and especially social and personal services (Isic 9). In both sectors,
greater dispersion is associated with more rapid growth in employment.
This is particularly notable in social and personal services, which has been
the main engine of employment across the OECD area during the past three
decades. The pattern is consistent with the expectations ¬rst explained in
Iversen and Wren (1998). Thus, employment in the slow-productivity and
low-wage sectors exhibits a close association with wage compression, while
in other sectors it does not.
Table 6.3 shows the magnitudes of the differences in the relationships
between wage structure and private employment performance, using slopes
of the regression lines in Figure 6.3 and correlation coef¬cients at different
levels of aggregation as statistics.9 Con¬rming a visual inspection of the
graphs, there is no relationship between dispersion and employment in
manufacturing or in business services “ the latter might even be slightly
negative “ whereas there is a clear positive relationship in the wholesale
and retail sector and especially in social and personal services.
The relationship between dispersion and employment is likewise strong
if we combine the latter two sectors (Isic 6 and 9), which we may simply
call consumer services. At the level of whole-period averages, the bivariate

9 One large outlier, Canada in 1989, is left out here and in the subsequent analysis. The
reasons for this one outlier (out of 331) is unclear.

Forces of Change

Table 6.3. The Bivariate Relationship between Dispersion of Earnings and
Employment Growth in Three Service Sectors

Periodization Sector Slope Coef¬cient
Annual (14 — (1) Manufacturing 0.31 0.04
25 obs.) (2) Finance, insurance, and 0.07
business services
(3) Wholesale, retail, 0.82 0.18
restaurants, and hotels
(4) Social and personal 1.84 0.48
(5) Consumer services (3+4) 2.66 0.40
5-year averages (1) Manufacturing 0.32 0.06
(14 — 5 obs.) ’0.18
(2) Finance, insurance, and 0.08
business services
(3) Wholesale, retail, 0.92 0.28
restaurants, and hotels
(4) Social and personal 1.88 0.67
(5) Consumer services (3+4) 2.80 0.56

Whole period (1) Manufacturing 0.07 0.03
(14 obs.) (2) Finance, insurance, and 0.24
business services
(3) Wholesale, retail, 0.74
restaurants, and hotels
(4) Social and personal 0.90
(5) Consumer services (3+4) 0.90
— Statistically signi¬cant at a .01 level (only pertains to whole period slopes since
the observations in the time series data are not statistically independent).
Source: Calculated from OECD (1999a).

relationship is almost perfect (r = .9) whether we consider social and per-
sonal services separately or look at consumer services as a whole.
But bivariate correlations can be misleading, especially in data that ex-
hibit both cross-sectional and cross-time variance, and other factors surely
play a role. First, it has been argued that restrictive macroeconomic con-
ditions in many European countries during the 1980s and early 1990s hurt
employment performance in those countries (Scharpf 1991; Soskice 2000;
Hall and Soskice 2001; Manow 2002). Furthermore, as argued by Scharpf

New Tradeoffs, New Policies

(1991), Hall and Franzese (1998), Iversen (1999) and others, equilibrium
unemployment can vary across countries depending on the particular setup
of macroeconomic institutions. We should therefore expect better unem-
ployment performance to be associated with better employment perfor-
mance. However, this does not explain why cross-national employment per-
formance varies across sectors, as is clearly the case.
This intersectoral variance may potentially be explained with reference
to the effects of trade with less-developed countries. As argued by Wood
(1994) and Leamer (1996), LDC trade puts downward pressure on wages
in those industries with the highest concentration of low-skilled workers.
If low-wage competition simultaneously causes deindustrialization and mi-
gration of workers to services, then LDC trade would confound the rela-
tionship between earnings dispersion and employment. In particular, we
should expect LDC trade to be positively related to employment in low-
paid services. Recall, however, that we found no evidence in Chapter 5 that
LDC trade leads to deindustrialization, so there is reason for some ex ante
skepticism about the empirical importance of this argument.
In contrast to the purportedly strong effect of LDC trade, the analysis
in Chapter 5 found clear evidence that per capita income affects service
employment, and for good reasons. Because demand for services, in accor-
dance with Engel™s law, tends to be income elastic (the higher the level of
income is, the greater the demand is for services), we should expect higher
levels of service employment in high-income countries, and rising levels
of employment in countries experiencing growth in incomes. As with the
macroeconomic argument, however, it is less clear that this helps us explain
cross-national differences in the performance of particular service sectors.
An argument that points both to the supply and demand side concerns
the in¬‚ux of women into the labor market. Female labor force participation
increases the supply of labor; at the same time, it raises demand for social
and personal services such as daycare and food services (see Huber and
Stephens 2000). The effect is dissipated, however, to the extent that these
services (most obviously daycare) are provided through the public sector.
Indeed, if such services are paid for through higher taxes that raise the
costs of privately provided services, the net effect of female labor force
participation on private sector employment would not be clear.
This leads us to an argument about taxation that has been recently pro-
posed by Fritz Scharpf (Scharpf 2000). Scharpf points out that the tax
wedge, which is the added costs of labor attributable to taxation, can hurt

Forces of Change

employment in cost-sensitive services if the tax code does not make exemp-
tions for low-paid workers. In particular, taxes raised through employer
social security contributions and taxation on consumption tend to be regres-
sive compared to taxes on income, and regressive taxes place a particularly
high burden on low-wage sectors. Note that this argument is a subspecies
of the general wage structure argument, since what ultimately matters to
employers is the total wage bill. Yet it important because it pinpoints a
mechanism, the structure of taxation, that can be manipulated politically
without raising inequality.10
A related issue concerns the use of tax revenues. While some revenues
are returned to the private sector in the form of reimbursements for, say,
medical expenses, other revenues are used to employ people to provide
the services directly. Government provision of services varies greatly across
countries, and such provision may have crowded out private services in
some countries, while privatization of public services in other countries
may have “in¬‚ated” private employment performance. This argument is
only unambiguously relevant for social and personal services because the
government is not involved to any signi¬cant degree in the provision of
other services, but it is precisely in social and personal services that we ¬nd
the strongest relationship between earnings structure and employment.
To test these alternative hypotheses, I carried out a regression analysis
using changes in (private sector) consumer service employment as a per-
centage of the working age population, e, as the dependent variable. These
services account for between 67 percent (country weighted) and 72 per-
cent (population weighted) of the growth in total service employment be-
tween 1970 and 1996, and they account for almost the entire cross-national
variance in private service employment performance (because employment
growth varies more across countries in consumer services than in producer
In addition to the wage dispersion variable, I included the following set
of controls designed to capture the preceding arguments:
1. Standardized unemployment rates to control for macroeconomic
conditions (OECD, Economic Outlook, various years)

10 Yet, one has to be careful in predicting the effects of altering the tax structure because a
shift in the tax burden from employer contribution to income taxation may be accompanied
by higher wage demands by low-paid unions. Lower social security contributions raise the
capacity of employers to pay higher wages, which indirectly confers more bargaining power
on unions.

New Tradeoffs, New Policies

2. Per capita income to capture income elasticity effects (Heston,
Summers, and Aten, 2002, Penn World Table 5.6).
3. Trade with less-developed countries as a percentage of GDP to con-
trol for the effect of low wage competition (IMF, Direction of Trade
Statistics Yearbook, various years),
4. Female labor force participation as a percentage of the working-age
population to control for the supply-and-demand side effects of such
participation (OECD, Labour Force Statistics, various years),
5. Consumption taxes and social security contributions as a percentage
of GDP to control for the tax wedge thesis (OECD 2002).
6. Government employment as a percentage of the working-age popu-
lation to control for the possible crowding-out effects of public em-
ployment (OECD, Labour Force Statistics, various years)
In addition, to hedge against arguments about country-speci¬c conditions
that are not captured by these variables, the regression includes a full set of
country dummies.11 Using the same error correction setup as in the case of
prices, the model is
e i,t = δe i,t’1 + b i di + b v xi,t’1,v + xi,t,v + µi,t (6.7)
i=1 v=1 v=1

where di refers to the dummy variable for country i, and xt,v is the vth
independent variable at time t (entered both as changes and as lagged levels),
and is the ¬rst difference operator.
The results of the analysis are shown in Table 6.4, which, for purposes
of comparison, also includes the results for business services (the ¬rst col-
umn). Supporting the ¬ndings in Iversen and Wren (1998), wage disper-
sion is strongly positively related to employment in consumer services, but
not in business services (compare the ¬rst two columns). The precision
of the parameter estimate for consumer services is high and statistically
different from zero.12 In substantive terms, the ¬rst-year effect of a 1-
standard-deviation increase in wage dispersion is to increase employment by
0.15 percent of the working-age population. After 3 years, the effect is 0.45

11 An F-test also indicates that these dummies belong in the model.
12 The setup with a full set of dummies in fact gives a very conservative estimate of the statistical
signi¬cance level. As in the price equations, we have a fairly severe case of multicollinearity
between wage dispersion and the country dummies “ the latter account for 86 percent of
the variance in the former “ and if the dummies are removed, the signi¬cance level of the
wage dispersion variable improves notably.

Forces of Change

Table 6.4. The Determinants of Private Service Sector Employment Growth

Consumer Services (Isic 6 + Isic 9)a
(Isic 8)a
’0.85——— ’0.92———
’0.14 ’0.34
(’1.01) (’3.95) (’3.34) (’1.11)
3.07——— 3.08——— 2.22———
Earnings dispersiont’1 0.40
(1.29) (3.93) (3.96) (3.00)
Earnings dispersiont 2.89 3.03 2.48
(’0.79) (1.47) (1.55) (1.33)
’0.01——— ’0.03——— ’0.03——— ’0.04———
(’3.21) (’5.77) (’5.03) (’6.14)
’0.06——— ’0.12——— ’0.12——— ’0.12———
(’6.18) (’7.57) (’7.52) (’7.18)
0.07——— 0.06——— 0.10———
GDP per capitat’1 0.00
(0.20) (2.88) (2.73) (4.21)
0.11——— 0.36——— 0.37——— 0.41———
GDP per capitat
(3.33) (5.34) (5.43) (6.03)
0.03——— 0.03——— 0.03———
LDC tradet’1 0.00
(0.99) (3.55) (3.87) (3.51)
LDC tradet 0.01 0.01 0.00
(’0.54) (0.68) (1.18) (0.42)
0.01——— 0.01—
Female LF participationt’1 0.00 0.00
(3.50) (0.54) (0.71) (1.94)
0.02——— 0.05——— 0.05——— 0.05———
Female LF participationt
(3.82) (4.52) (4.19) (3.75)
Taxationt’1 “ “ 0.03
(0.73) (1.96)
0.01— 0.04——
Taxationt’1 “ “
(1.78) (2.28)
Public employmentt’1 “ “ “
Public employmentt “ “ “
’0.03—— ’0.04—— ’0.04—— ’0.07———
Lagged dependent level
(’2.18) (’2.12) (’2.09) (’3.81)
Adjusted R-squared .42 .57 .57 .59
Number of observations 331 331 331 331
Signi¬cance levels: ——— < .01; —— < .05; — < .10 (t-scores in parentheses).
a All regressions were estimated with a full set of country dummies using panel cor-

rected standard errors.

New Tradeoffs, New Policies

percent or about 0.1 standard deviation, and in the long run a 1-standard-
deviation increase in dispersion is associated with a 0.8-standard-deviation
increase in employment, or 4.3 percent of the working-age population.13
These estimates, however, must be treated with caution. The data exhibit
strong path dependency, and that makes long-term equilibrium predictions
unreliable. I am capturing a process, service sector expansion, as it unfolds,
and it is risky business to extrapolate the exact long-term equilibrium condi-
tions.14 I return to this issue later. Suf¬ce it to say here that the dispersion
variable exhibits a statistically signi¬cant positive effect on employment.
The long-run magnitude of the effect appears to be large, but the exact size
is dif¬cult to pin down.
Turning to the effect of per capita income, it is as unambiguous as it is
unsurprising. Just as rising income during the latter part of the nineteenth
century fueled demand for manufactured goods, so does rising income today
spur demand for, and employment in, services. As we saw in Chapter 5,
the combined effect of productivity growth in manufacturing and shifts in
consumption patterns led to deindustrialization and simultaneously created
the structural conditions for the rise of services.
Unemployment also has the expected effect on employment, where the
permanent effect can be interpreted as an impact of institutions while the
transitory effect can be interpreted as either the result of the business cycle
or of countercyclical macroeconomic policies. Of course, one cannot cleanly
separate out what is the cause and effect here, because both employment
and unemployment are determined by the same underlying institutions
and policies, but it seems safe to conclude that macroeconomic conditions
do signi¬cantly affect employment performance “ a result that is antici-
pated by Scharpf (1991), Hall and Franzese (1998), Iversen (1999), Soskice
(2000), Manow (2002), and others. This is important to keep in mind when
interpreting the performance of particular countries because some have
bene¬ted signi¬cantly from propitious macroeconomic conditions.
LDC trade appears to have a positive effect on service employment as ex-
pected. However, because there are off-setting indirect effects, LDC trade

13 The results are substantively similar if we run the regressions on Isic 6 and Isic 9 separately
(somewhat weaker for the former and somewhat stronger for the latter). I keep the two
together for presentational economy.
14 The task is a bit like predicting the cruising altitude of an airplane from data on speed and
altitude during the ¬rst few minutes of ¬‚ight. We know that the plane is increasing altitude
at a decreasing rate, and this makes it possible to project the ultimate cruising altitude, but
there are no actual observations of the steady state.

Forces of Change

does not in fact have any net effect on employment, and without the other
variables in the model, it exhibits only a small and insigni¬cant relation-
ship with employment. It is unclear exactly how to interpret these results “
especially why LDC trade is associated with a less dispersed wage structure “
but one is well advised not to draw strong conclusions about the effects of
LDC trade from the direct effect. As I show later in this chapter, there is
also no cross-national evidence that countries more exposed to LDC trade
have higher service employment rates than other countries.
Female labor force participation does not appear to have a signi¬cant
effect on private sector service employment. But, as argued previously, this
may in part be because many of the services most directly associated with
the demand side of female labor force participation, in particular daycare,
are often provided through the state. Correspondingly, when public sector
employment is controlled for (last column of Table 6.4), the effect of female
labor force participation increases (though it is still only borderline signi¬-
cant). Another part of the story is that much of the demand effect of female
labor force participation goes through per capita income (the relationship
between participation and income is positive). The high income elasticity
of services is therefore partly a result of the fact that a substantial portion
of the new demand for services has come from women entering into the
labor market (Huber and Stephens 2000).
As expected, government employment has a statistically signi¬cant and
permanent effect on private consumer service employment.15 A 1 percent
rise in public employment is associated with about an 0.8 percent decline
in long-term private employment. It is not clear, however, how to interpret
this effect because it is likely that a substantial portion results from reversed
causation. If private service employment is growing slowly, governments
will be under greater pressure to employ people in the public sector. As I
argue at length in the next section, this direction of causality is likely to have
been particularly important in countries where the political left is strong,
and these countries also tend to be the ones with the most compressed wage
structures.16 In addition, it is not the case that those countries with relatively
compressed wage structures, but without large public sectors, exhibit better
private employment performance than those with larger public sectors.

15 As one would have expected, it does not have an effect if added to the equation for business
service employment (b = 0.0007 for the lagged government employment variable).
16 In principle, it is possible to deal with this problem by using recursive models. In practice,
it is very hard to develop good instrumental variables, and it is impossible to get people to
agree on their interpretation.

New Tradeoffs, New Policies

This suggests that public employment is not an important determinant of
private employment, and that the causality therefore runs in the opposite
Regardless of how much of the effect of government employment can
be attributed to reverse causation, the positive effect of the wage disper-
sion remains, although the parameter for the permanent effect is now re-
duced from 3.1 to 2.2. Perhaps tellingly, the bivariate regression reported
in Table 6.3 implies a parameter of 2.6, which is right in the middle of these
values. In addition to reducing the parameter of the dispersion variable,
the inclusion of government employment increases the speed with which
private service employment reaches its equilibrium, and this also reduces
the long-term impact of the independent variables. Thus, the equilibrium
effect of a 1-standard-deviation increase in dispersion is now 1.6 percent of
the working-age population compared to 4.3 percent before. This result,
however, almost certainly underestimates the true effect for the reasons just
To get a better idea of the likely range of effects, Table 6.5 shows the
estimated impact on employment of a 1-standard-deviation increase in dis-
persion, using a range of different model speci¬cations. The short-term
effect is the rise in employment after 5 years; the long-term effect is the
estimated increase in equilibrium employment. Moving down in the table
implies that more variables are being added to the model. Thus, the num-
bers in the ¬rst line are the estimates when only per capita income is used
as a control variable. The last line is the fully speci¬ed model correspond-
ing to the last column in Table 6.4. I have estimated the effects both with
and without controls for time periods (i.e., with and without t ’ 1 time
Note that the short-term effect is fairly stable around 0.4“0.7 percent, re-
gardless of speci¬cation. The long-term predictions are also within a fairly
narrow range of 4“7 percent, except when public sector employment is in-
cluded as a control. Then the effect drops to between 1.4 and 1.6 percent.
As noted earlier, a plausible explanation for this drop is that governments

17 One reviewer also suggested that there might be bias in the results because some private
services are publicly subsidized and therfore not really “private” (exactly where the OECD
draws the line between public and private employment is unclear). Yet, if that were true, it
would cut in the opposite direction of my hypothesis: egalitarian welfare states have more
subsidization, which should weaken the relationship between equality and employment.
Also, in the equations using public employment on the right-hand side, this variable is
probably a good proxy for subsidization.

Forces of Change

Table 6.5. The Effect of Earnings Dispersion under Different Model Speci¬cations

Effect of Earnings Dispersiona
With Country Dummies With Country and Time Dummies

Short Termb Long Termc Short Termb Long Termc
With Control for
GDP per capita and 0.39 0.46 6.22
Unemployment and 0.52 7.13 0.46 6.10
LDC trade and 0.70 5.64 0.61 5.71
Female LF 0.72 4.26 0.61 3.99
participation and
Taxation and 0.73 4.39 0.63 4.16
Public employment 0.49 1.60 0.41 1.37
(1.01) (3.14) (0.70) (2.12)
a The effect of a 1-standard-deviation increase in earnings dispersion.
b Percentage increase in employment as a percentage of the working-age population after
5 years.
c Percentage increase in employment as a percentage of the working-age population in the
long run.
d Equilibrium cannot be calculated because coef¬cient on lagged dependent level variable is
indistinguishable from 0.

in countries with a ¬‚exible, and hence dispersed, wage structure are under
less pressure to increase public employment. Putting public sector em-
ployment on the left-hand side supports this interpretation because wage
dispersion is negatively related to public employment. There are therefore
two interpretations: The ¬rst is that causality runs from wage dispersion
to public employment, in which case the latter should not be included on
the right-hand side. Alternatively, public employment does reduce private
sector employment, in which case some of the effect of wage dispersion
on private sector employment goes through public sector employment. If
we take this indirect effect into account, the total long-term effect of a
1-standard-deviation increase in dispersion is to increase private service
employment by 2.1“3.1 percent (noted in parentheses in the last line of
Table 6.5). Based on these results, it seems safe to conclude that about
2.5 percent is the lower bound on the long-term effect of wage dispersion,
which corresponds to 0.5 standard deviation on the dependent employment
I have saved the discussion of taxation for last because the effects of
this variable are in some respects the most intriguing. As noted previously,

New Tradeoffs, New Policies

Scharpf (2000) has argued persuasively for the importance of the variable,
and he can point to a strong negative cross-national correlation between
service employment and social security and consumption taxes. Yet, in the
error correction model, the effect of the variable is reversed, although it is
only borderline signi¬cant. The issue is not simply that the effect of this
variable goes through wage dispersion, although there is a negative and
statistically signi¬cant effect, but that taxation does not have the expected
effect in any speci¬cation of the model.
This phenomenon (i.e., that a strong and theoretically sensible cross-
sectional correlation disappears in a dynamic model with a lagged depen-
dent variable, or in an error correction model) is well known to statistical
practitioners. In the past, it has been taken to mean that the theoretical
intuition must be wrong, but recently it has become an issue of contention
among methodologists. In a recent paper, Chris Achen argues that a per-
fectly sensible relationship in a simple regression on levels can completely
disappear in a dynamic model even though the relationship does in fact exist
(see Achen 2000). The problem arises when there is strong trending in
the data between the dependent and independent variables, and the lagged
dependent variable soaks up most of the cross-sectional variance. In our
case, there is no doubt that the variables are trending; therefore, this is a
potential problem.
The issue applied to taxation is illustrated in Table 6.6. The variables
in this table are the same as before except that only contemporaneous
levels (not lagged levels plus changes) are used in the regression. Col-
umn 1 shows the bivariate relationship between taxation and employment,
which is strong and negative, as in Scharpf ™s analysis. The effect is re-
duced, but still strong, in the multivariate model in column 2, which in-
cludes all the variables from Table 6.5. The problem, of course, is that
the errors in the model are highly serially correlated, so the t-statistics are
One standard method to deal with serial correlation is to add a lagged
dependent variable as in column 3 (Beck and Katz 1995). This regression,
which also controls for country-speci¬c effects, immediately reveals the
problem. While all the other variables retain at least some of their origi-
nal effect, the parameter estimate for the taxation variable not only loses
signi¬cance but turns positive. This is precisely the situation we had in
the original model, and although the present model is not fully speci¬ed
(because we cannot assume that the parameters on the lagged independent

Forces of Change

Table 6.6. The Effect of Taxation under Different Model Speci¬cations

Employment in Consumer Services
(4) Long-Term
a a a
(1) (2) (3)
Intercept 4.66 2.66 “
(39.33) (1.62) (3.67)
Earnings dispersion “ 15.38 16.68
(4.77) (3.67)
’0.38 ’0.36
Unemployment “
(’9.00) (’7.74)
GDP per capita “ 1.26 1.39
(14.48) (8.11)
LDC trade “ 0.05 0.17
(0.77) (2.84)
Female LF participation “ 0.08 0.04
(3.83) (1.88)
’0.48 ’0.21
Taxation 0.01 0.04
(’16.11) (’7.94) (0.70)
Public employment “
(’9.55) (’6.50)
Lagged “ “
dependent level (45.50)
Adjusted R-squared 0.419 0.789 0.997
Number of observations 359 343 331
Signi¬cance levels: ——— < .01; — < .10 (t-scores in parentheses).
a Estimated with a full set of country dummies using panel corrected standard errors.
b Estimated from the results of LDV regression in column 3.

variables are zero), this is a widely used setup that highlights the problem
at hand.18
The last column of Table 6.6 shows the long-term parameters (b/
(1 ’ .84)) of each variable, and it is instructive to compare these results
to the simple regression on levels in column 2. One would hope that the
long-term parameters, which suggest the levels of employment expected in
equilibrium, are not completely out of whack with the results for the nondy-
namic regression on the levels themselves. And such rough correspondence
does indeed hold for wage dispersion, unemployment, and GDP per capita.

18 The assumption of zero effects of lagged variables is not always appropriate; therefore,
one should not, in my view, reduce a fully speci¬ed error correction model to the form in
Table 6.5 unless the results are consistent. In our case, the results are broadly consistent.

New Tradeoffs, New Policies

On the other hand, the dynamic model (static model) appears to overes-
timate (underestimate) the effect of female labor force participation and
underestimate (overestimate) the effect of LDC trade. But again, the result
that is most inconsistent with the simple regression is the one for taxation.
Here there is not simply a lack of correspondence, but a contradiction.
On the basis of these results, it is tempting to conclude that tax structure,
despite the theoretical arguments to the contrary, really does not matter
for employment. This is the conclusion many methodologists, including
Nathaniel Beck who advocates use of lagged dependent variable (LDV)
models, are inclined to draw.19 Chris Achen, on the other hand, suggests
that the simple regression without a lagged dependent variable may tell an
important substantive story, even if the t-statistics are meaningless. This
controversy will have to be resolved among methodologists, and that will
take time. Here I simply conclude that the effect of tax structure on employ-
ment performance is potentially important, but inconclusive. For the time
being, we need to retain a potential role for the tax wedge in understanding
developments in speci¬c countries.
As a ¬nal check on the robustness of the results for wage dispersion, I
reran the error correction model using d5/d1 ratios in place of the indus-
try dispersion measure (Table 6.7). The drawback is a 43 percent loss of
observations since d5/d1 ratios are only available for a subset of years “ in
several cases only for the 1980s, and in one case (Norway) only for 2 years.
To avoid the elimination of single observations as a result of differencing,
I interpolated values whenever the gap between two values in a series was
3 years or less. This is identical to the procedure followed by Rueda and
Pontusson (2000), who analyze the same earnings data. Out of the total of
202 observations, 14 were added in this manner. Because there was almost
no change in the d5/d1 ratio in the cases of interpolation, and given the
short period of time, the procedure does not appear to be problematic.20
The results are broadly in agreement with the previous ¬ndings, al-
though the effects of wage dispersion are somewhat stronger (and always
statistically signi¬cant at a .01 level or better). A 1-standard-deviation in-
crease in dispersion is associated with an increase in employment (as a
percent of the working-age population) of between 1.1 percent (with con-


. 8
( 11)