Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

Embed Size (px)

Citation preview

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    1/13

    1

    Section 2.3 no. 64

    QUANTITATIVE CROSS-NATIONAL RESEARCH METHODS

    Gosta Esping-Andersen

    CPISUniversitat Pompeu Fabra

    and

    Adam Przeworski

    Department of Political Science

    New York University

    Trade-offs in Quantitative Comparisons

    Quantitative nation comparisons pose inevitable trade-offs. One is that much of the

    contextual reality of individual nations is sacrificed for the sake of broader

    generalization. We fail to capture the uniqueness that defines a nations culture,

    historical heritage, and endemic logic. The interpretation of a variable may, indeed, only

    be possible when its is studied contextually (Ragin, 1987; and Lieberson, 1991).

    Boolean analysis, as Ragin argues, helps overcome this dilemma. It has advantages,

    such as its ability to build conjunctural models with very few cases, and its ability to

    analyze non-events. But it needs to be guided by strong theory and substantial

    knowledge, its applicability is limited to relatively few cases, and it may be too biased

    in favor of non-additive, conjunctural models. For an empirical application, see Ragin(1994). See also Section 2.3, no. 72.

    A second trade-off has to do with the often limited number of observations available,

    especially in studies of advanced (OECD) democracies where the N rarely exceeds 25.

    In broader World comparisons, however, the N approaches 200. Many attempt to

    supplement few nations with over-time data, as in the case of pooled cross sectional

    and time series analyses. As ever longer data series for individual countries become

    available, the small N problem will gradually diminish. A 20 year time series for 20

    countries yields 400 observations; a 20 year time series for 200 nations yields 4000.

    The most serious small N problem, then, occurs where for theoretical or other reasons,the natural universe is limited. Small samples do not necessarily pose problems of

    statistical inference. Strong theory may not require many observations. They do,

    however, limit what statistical tools can be brought to bear. Fearon (1991 ) argues that

    the smaller the N, the greater the need to make counterfactuals explicit in other words,

    to tighten the theoretical formulation about the precise conditions that will (or will not)

    produce an outcome. Vague theory implies uncertainty about the number of contending

    explanations (which can grow very large) and about their mutual relationship within a

    causal order. The consequence is possible multi-collinearity and, more generally, large

    error terms all of which can only be managed by having more observations. Hence, it

    is likely that estimations will yield non-robust results. The less ambiguous the theory,

    the less the need for large Ns. Small Ns are therefore an especially acute problem in

    disciplines without a firm theoretical architecture, like Sociology or Political Science.

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    2/13

    2

    Western (1998a) and Western and Jackman (1994) argue that Bayesian estimation, by

    allowing for greater uncertainty, produces more robust results when Ns are few, theory

    vague, and explanations are many. For a critique, see Firebaugh (1995); for an

    empirical application, contrasting conventional regression with Bayesian estimation, see

    Western (1996).

    Small Ns can even hold advantages, because a limited nation-sample implies that the

    researcher can far more easily gain maximum advantage from diagnostic scrutiny of the

    residual plots. A set of systematic outliers, for example, can be easily identified in terms

    of their nationhood, and any good comparativist should be able to pin down what

    variable(s) drives their deviance from the fitted regression line. Diagnostics on small

    samples can approach the advantages of in-depth case studies.

    There are, nonetheless, problems inherent in quantitative national comparisons that are

    both more generic and potentially more serious than the small N problem. The

    remainder of our presentation will focus on qualitative and limited dependent variables,

    on selection bias, on lack of independence between observations, and on the problem of

    endogeneity of variables.

    Qualitative and Limited Dependent Variables

    Many dependent variables in cross-national research are either qualitative (assuming

    discrete values) or limited (they assume continues values within some range). The

    choice whether to measure a variable in a qualitative or continuous way is often

    controversial. Bollen and Jackman (1989: 612), for example, argue that difficulties in

    classiying some political regimes speak in favor of using continuous scales becauseDichotomizing democracy blurs distinctions between borderline cases. In contrast,

    Przeworski et.al. (in press) prefer to treat regimes dichotomously or mulitnomially.

    Yet, whether the dependent variable is qualitative or limited, the consequences are the

    same, namely, the need to use non-linear models. Whether we give political regimes the

    values of 0-1 (as do Przeworski et.al, op.cit) or 1-100 (as does Bollen, 1980), it remains

    that the value on the dependent variable cannot exceed its maximum when the

    independent variable(s) tend to infinity, and it cannot fall below its minimum when

    these variables tend to infinity. Linear models, at best, can provide an approximation

    within some range of the independent variables.

    The standard model when the dependent variable is multinomial is

    Pr (Y=j X=x) = F(x),

    where j= 0,1,,J-1 and F is the cummulative distribution function. Since such models

    are treated by any standard textbook (such as Greene, 1997), we need not present them

    here. Such models can be applied to panel data unless the number of repeated

    observations is large.

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    3/13

    3

    Event history analysis is one particular class of non-linear models applied in cross-

    national research. 1 In such models the dependent variable is an event, such as a

    revolution, regime transition, or a policy adoption. The general model is

    Pr [Y(t+dt)=j] = f[y(t), x(t) ], dt 0

    Most often, such models can be conveniently estimated as

    log S (t) = log [1-F(t)],

    where S (t) is the survival function, or the probability that an event lasts beyond time t,

    and F (t) is the cdf. 2 See Section 2.3, no. 53.

    For dichotomous dependent variables, logit and probit give very similar results. For

    multinomial variables, it is often assumed that the errors are independent across the

    values of the dependent variable, which leads to a logit specification. But this implies a

    strong assumption, namely the irrelevance of independent alternatives, that rarely holds

    in practice (Schmertmann, 1994). Multinomial probit, in turn, requires computing

    multiple integrals, which was until recently computationally expensive. Alternatives

    would be semi- and non-parametric methods.

    The distributions which are commonly used in estimating survival models include the

    exponential, Weibul, logistic, and Poisson distributions. The difficulty here is that it is

    very difficult to statistically distinguish between such distributions (with the exception

    of the Weibull and exponential). The Poisson distribution should be favored when the

    events are rare.

    Methods for studying qualitative dependent variables are now standard textbook fare,

    but what warrants emphasis is that the traditional distinction between qualitative and

    quantitative research is becoming increasingly obsolete. Phenomena such as social

    revolutions may be rare, but this just means they occur with a low probability. They

    can, nevertheless, be studied systematically using maximum likelihood methods.

    The Problem of Selection Bias

    In comparative research we are often interested in the effect of some systemic feature,

    institution, or policy on some outcome. Examples include the effect of labor marketinstitutions on unemployment, the effect of political regimes on economic growth, or

    the effect of electoral rules on the number of political parties.

    In such cases, the question is how some feature of X affects some outcomes Y in the

    presence of conditions Z, or PR (Y | X,Z), where the hypothesis is that y = f(x). Thegeneric problem is how to isolate the effects of X and Z on Y.

    If sampling is exogenous (i.e. the probability of observing any y is independent of z),

    the pr (Y=y |z) = pr (y), and standard statistical methods can be used. But if pr (Y= y |z)

    1 For an example, see Usui (1994)2 Again, we refer to standard textbooks for details.

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    4/13

    4

    is different under different conditions Z, sampling is endogenous and the assumptions of

    the standard statistical model are violated. For a detailed treament, see Section 2.3, no.

    87.

    To exemplify the problem: suppose we want to know the impact of political regimes on

    economic growth. We observe Chile in 1985 (Z), which was a dictatorship (X), inwhich case per capita income declined at the rate of Y= -2.26. To assess the effect of

    political regimes, we need to know what would have been the rate of growth (Y) in

    Chile had it been a democracy. This cannot be answered with the available observations

    because Chile in 1985 was indeed a dictatorship. The comparativist in such a case

    would proceed quasi-experimentally, and look for a case that matches Chile in all

    aspects except authoritarianism. In this way one would compare an authoritarian with a

    democratic Chile.

    But what if Chiles status as a dictatorship in 1985 was due to some factors that also

    affected its economic performance? This could be because growth itself influences the

    survival of regimes; a countrys level of development may be associated both with

    regime selection and growth; some unobservable factors (say enlightened leadership

    or, perhaps, measurement error) may be common to both variables. Whatever the

    underlying selection mechanism, the basic consequence is that there will be cases

    without a match, and this implies that our inferences will be biased.

    The problem raised by non-random selection is how to make inferences from what we

    observe to what we do not. It is a problem of identification (Manski, 1995). For what do

    we know when we observe Chile as a dictatorship in 1985? We observe the fact that,

    given conditions Z, the Chilean regime (X) is a dictatorship. We also know its rate of

    economic growth, given that it is a dictatorship under conditions Z. We do not observe,however, its rate of growth as a democracy under conditions Z. The issue is one of

    counterfactual observations: the rate of growth that an observation with Z = zi ,

    observed under dictatorship, would have had under democracy. We do not know what

    this value is and, hence, we face under-identification.

    What we need to know are two distributions, P(Y=y | x=j, Z) and P(Y=y | x=k, Z), andtheir expected values, where we now think of the possible values of X more generally as

    j,k = 0,1,2,,N. The first is the distribution of Y in the entire population if all cases

    were observed as dictatorships; the second is the distribution of Y in the entire

    population if all cases were observed as democracies. Clearly, if we want to know the

    impact on y of being in states x, we need to determine the difference between these twodistributions, conditional on z.

    It may seem strange to refer to the population as including unobserved cases. But in

    order to compare the effects of states x on the performance of an individual case

    characterizedby a zi , we must allow the case to be potentially observable under the full

    range of X. 3 We observed this case as X = j, but must also imagine the possibility that

    it would have been X j. We must think in terms of a super-population consisting ofa continuum of potential cases. The actual sample is then regarded as having been

    drawn by nature from this super-population (Pudney, 1989: 45).

    3 Note that in our example X assumes only two values. But, generically, this is not necessary.

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    5/13

    5

    If the actual population is an exogenous sample of the potential one, any multivariate

    combination (xi, zi) can be drawn with a positive probability. Such exogenous sampling

    allows us to isolate the effect of x on y given z, that is, to control for z the effects of x

    on y. If the sample is exogenous, we can match pairs (x i = j, zi) and (xi = k, zi). But if

    sampling is endogenous, there will be some cases without a match: some of the x i s

    will not have support. Where we have non-random selection, quasi-experimentalcomparisons fail regardless of the number of observations.

    The methods to correct for selection bias consist of constructing counterfactuals, that is,

    of filling the unobserved supports of the distribution Y for all X. Following Heckman

    (1979), when sampling is not random, the regressions y = f(x,z) suffer from omitted

    variable bias. When such regressions are estimated on the basis of the observed sample,

    the variable that is omitted is the expected value of the error in the equation that

    specifies how the observations are selected into the sample:

    Pr (x=j) = F(Z) + u.

    If E (uj | x=j) 0, then

    E (uj| x) = juE (u | x),

    where ju is a regression coefficient of uj on u.

    Finally, E (u | x=j) = j, where the j s are the inverse Mill ratios, or the hazard rates.

    Since we know that Yj is observed when x=j, we can write the expected values in the

    observed sample as

    E (yj | x=j, z) = zj + juj .

    Note that if this equation is estimated on the basis of the observed sample, the variable

    j is omitted from the specification.

    We can now see why controlling for the variables that enter both into selection and

    outcome equations may indeed worsen the selection bias (Achen, 1986). Following

    Heckman (1988), we distinguish first between selection on observables and on

    unobservables. Selection on observables occurs when the expected covariance E (uju

    |

    z) 0, but once the observed variables Z are controlled for it vanishes, so that E (uju |z) = 0. Selection is on unobservables when E (uju) 0 and E (uju | z) 0, which meansthat controlling for the factors observed by the investigator does not remove the

    covariance between the errors in the outcome and the selection equations. Now note

    that the regression coefficient ju = cov (uju)/var (uj). If selection is on unobservables,controlling for some variable x in the outcome equation may reduce the error variance uj

    without equally reducing the covariance uju. Hence, the coefficient on the omitted

    variable will be larger and the bias will be exacerbated.

    In short, the expected values of the observed cases will be biased because they covary

    with the variable which determines which cases are observed. If selection is exclusively

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    6/13

    6

    on observables, this bias can be corrected by traditional controlling techniques. But if it

    is on unobservables, such controls only worsen the bias.

    Correcting for selection bias is not uncontroversial. It has been found that corrections

    for selection are not robust, but are highly sensitive to relatively minor changes in

    assumptions about distributions (Goldberger, 1983). Others have found that someestimation methods fail to correct for this bias and may even exacerbate it (Stolzenberg

    and Relles, 1990). As Heckman (1988: 7) argues, the quandary we face is that different

    methods of correcting for selection bias are robust if there is nobias to begin with; if

    there is, there is no guarantee that the methods are robust.

    The logic of the problem is similar whether we study large or small Ns (Fearon, 1991).

    When de Toqueville concluded that revolutions do not bring about social change, the

    reason might be that they occur only in countries where it is difficult to change society.

    Even when N=1, the issue of selection bias does not disappear: The French revolution

    in 1789 may have been caused by the same conditions that made social change so

    difficult. It is possible that a revolution in a country where social relations are easier to

    change would have provoked change. But then a revolution would not have been

    necessary.

    Comparativists who conduct case studies cannot benefit from statistical distributions to

    generate the counterfactuals. Yet, as in all cases of comparison the problem remains and

    it should, therefore, be standard practice to ask counterfactual questions of ones case or

    cases.

    The Problem of Independence

    Statistical inference must assume that the observations on a variable are independent

    one of the other. Is country A s performance truly independent of what happens in

    country B? Is what happens at t+1 independent of events in t? Usually not, and this

    implies the need for corrective procedures. In most cases, however, rigorous correction

    will entail that the de facto N (nations or years) diminishes; in some instances,

    statistical dependency cannot be resolved at all.

    Cross-sectional analysis almost invariably assumes that nations and their properties (say

    budgets or institutions) are independent one of the other. This should not be assumed.

    We know that the Scandinavian countries have a similar and shared history, deliberatelylearning from each other through centuries, thus creating similar institutions and path

    dependencies. The same goes for Austria and Germany, for Belgium and the

    Netherlands and, arguably, for all the Anglosaxon nations. The issue is captured in

    Castles (1993) families of nations. World samples have a similar problem: Japans

    long hegemony in East Asia will have influenced Korean society; Confucianism has had

    a pervasive influence throughout the region. Similar stories are easily told for Latin

    America and Africa. If the World is a set of nation clusters, the real N is not 20-odd

    OECD countries or 150-odd World nations.

    When nations form families, but are treated as if they were all unique and independent,

    we are likely to get biased coefficients and, very probably, unequal error variance

    (heteroskadisticity). Sweden alone will drive the regression line in just about any

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    7/13

    7

    welfare state analysis, and when also Denmark and Norway are treated as discrete

    observations, the bias is multiplied in so far as all three in reality form part of the same

    political economy (Scandinavia). Diffusion effects that operate between members of a

    nation-cluster can also result in heteroskadistic disturbance in the cross-section. Such

    can be corrected by, for example, adding a variable that captures the common

    underlying property that drives the disturbance (say, a dummy for being Scandinavia)but, again, this correction absorbs precious degrees of freedom in a small N study and,

    substantively, amounts to reducing the three nations to one observation. One attempt to

    estimate comparative models in which it is presumed that nations cluster can be found

    in Esping-Andersen (1999).

    Lack of independence in a time-series is normally taken for granted, since this years

    budget or election outcome is almost inevitably related to last years budget or the

    previous election. The standard assumption is a first-order (AR1) serial correlation. In

    comparative research virtually all time-series applications are pooled with cross-

    sections. But, where Ns are very small, one may as well simply compare across

    individual time-series estimations, as do Abraham and Hausman (1994), or Esping-

    Andersen and Sonnberger (1991). Time-series are meant to capture historical process.

    Yet as Isaac and Griffin (1989) argue, they easily end up being a-historical.

    Pooling cross-sectional with time series data (panel regressions) has become very

    widespread, especially in studies of the limited group of advanced (OECD) societies. In

    many cases, the panel design is chiefly cross-sectional (more nations than years), as

    exemplified by Alvarez et.al. (1991), and Iversen and Wren (1998); others are

    temporally dominated (as in the case of Hicks, 1994a; or Hicks and Swank, 1992; for a

    discussion, see Stimson, 1985). Panel models are especially problematic because they

    can contain simultaneous diachronic and spatial interdependence and, worse, the twomay interact. The standard method for correcting contemporaneous error correlation

    (GLS) applies only where the ts well exceed nations (which is rare). The consequence

    is that t-statistics are overestimated, errors underestimated, and the results may therefore

    not be robust (Hicks, 1994b; Beck and Katz ,1995) . The Beck and Katz (1995)

    procedure, can correct for temporal and cross-sectional dependency one at a time, but if

    the two interact, no solution exists. See also Beck et.al (1998) for an application to

    maximum likelihood estimation.

    Panel models can be based on two types of theoretical justification. One is that events

    or shocks occur over time that affect the cross-sectional variance. There is, for

    example, a huge recent literature on the impact of labor market rigidities onunemployment: regulations vary across nations but also across time because of de-

    regulatory legislation (see for example Nickell, 1997). De-regulation in a country

    should produce a break in its time series, and the auto-correlation element will be split

    into the years preceding and following the break. The second justification, not often

    exploited, is to interpret autocorrelation as an expression of institutional or policy path

    dependency. In this instant, the rho must be treated as a variable. The problem, of

    course, is that the rho is likely to combine theoretically relevant information as well as

    unknown residual autocorrelation. The researcher can accordingly not avoid including a

    variable that explicitly measures path dependency.

    If we insist on faithful adherance to the real World, panel regressions will require so

    much correction against dependency that the hard-won additional degrees of freedom

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    8/13

    8

    that come with a time-series are easily eaten up. And how many can truthfully claim

    that time and country dependencies do not interact? Indeed, most sensible

    comparativists would assume they do: if nations form part of families it should also be

    the case that the timing of their shocks, events, or policies is interdependent. Such

    intractable problems are certainly much more severe in small-N comparisons, and this is

    reflected in the prevailing lack of robustness that is endemic in the OECD arealiterature. In Beck and Katz (1995) re-estimations of the Hicks and Swank (1992)

    study, to give an example, several key variables turned out insignificant. A marginal

    difference in measurement, the inclusion or exclusion of one country, the addition or

    subtraction of a year here or there, or the substitution of one variable for another, can

    change the entire model.

    One alternative is to construct multi-level models which explicitly take into account the

    possibility that nations may cluster (for an overview, see Goldstein, 1987).

    Nieuwbeerta and Ultee (1999) have, for example, estimated a three level (nation, time,

    and individual) model of the impact of class on party choice within the context of

    nations social mobility structure. A second alternative, in particular when the

    dependent variable is categorical, is to exploit the advantages of event history analysis.

    But, here the time series needs to be quite long considering that theoretically interesting

    events, such as revolutions, democratization, or even welfare reforms, are far between.

    In the event history context, analytical priority is usually given to temporal change,

    which brings it much closer to traditional time series analysis. But rather than having to

    manipulate autocorrelation, time sequencing (states and events) is actively modelled

    and thus gains analytic status. For an application to nation comparisons, see Strang

    (1994), Usui (1994) and, especially Western (1998b), which also can stand as an

    exemplar of how to minimize the interdependency problem. See also Section 2.3, no.

    53.

    There are two particular cases where the lack of independence among observations

    simply prohibits adequate estimation. The first, noted above, occurs when time and

    nation dependencies interact. The second is when globalization penetrates all nations

    and when many nations (such as the European Union) become subsumed under

    identical constraints. In this instance, existing cross-national correlations will strenthen

    and we may, indeed be moving towards an N=1. Of course, global shocks or European

    Union membership do not necessarily produce similar effects on the dependent variable

    across nations or time. If nations institutional filters differ, so will most likely the

    impact of a global shock on, say, national unemployment rates. Here we would specify

    interaction effects, but that would be impossible in a pure cross-section, and extremelydifficult in a time series, unless we already know how the lag structure will differ

    according to institutional variation.

    4. The Endogeneity Problem

    All probabilistic statistics require conditional independence, namely that the values of

    the predictor variables are assigned independently of the dependent variable. The basic

    problem of endogeneity occurs when the explanans (X) may be influenced by the

    explanandum (Y) or both may be jointly influenced by an unmeasured third. The

    endogeneity problem is one aspect of the broader question of selection bias discussed

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    9/13

    9

    earlier. General overviews of the endogeneity problem can be found in Manski (1995)

    and King, Keohane and Verba (1994). See also Section 2.3., no. 88.

    The endogeneity issue has been intensely debated within the economic growth literature

    in terms of the causal relationship between technology and growth. But it applies

    equally to many fields. For example, comparativists often argue that left power explainswelfare state development. But are we certain that left power, itself, is not a function of

    strong welfare states? Or, equally likely, are both large welfare states andleft power

    just two faces of the same coin, different manifestations of one underlying, yet un-

    defined, phenomenon? Would Sweden have had the same welfare state even without its

    legendary social democratic tradition? Perhaps, if Swedens cultural past

    overdetermines its unique kind of social democracy and social policy. If this kind of

    endogeneity exists, the true X for Sweden is not left power but a full list of all that is

    Sweden. The vector of the Xs becomes a list of all that is nationally unique.

    The endogeneity problem becomes easily intractable in quantitative cross-national

    research because we observe variables ( Ys andXs) that represent part of the reality of

    the nations we sample. Our variables are in effect a partial reflection of the society

    under study, and the meaning of a variable score for one nation may not be metrically

    equivalent to that of another a weighted left cabinet score of 35 for Denmark and 40

    for Sweden probably misrepresents the Danish-Swedish difference if, that is, the two

    social democracies are two different beasts.

    A related, and equally problematic, issue arises in the interpretation of coefficient

    estimations. In the simple cross-sectional regression, the coeffient for left power,

    economic development or for coordinated bargaining can only be interpreted in one

    way: as a constant, across-the-board effect irrespective of national context. If Denmarkhad 5 points more left power, its welfare state should match the Swedish (all else held

    constant). In fixed-effects panel regressions, the Xs are assumed to have an identical

    impact on Y, irrespective of country. Can we really accept such assumptions? Probably

    not.

    The reason that we cannot is that it is difficult to assume that variables, say a bargaining

    institution or party power balance, will produce homogenous, monotonically identical,

    effects across nations for the simple reason that they are embedded in a more complex

    reality (called nation) which has also given rise to its version of the dependent variable.

    In this case, the bias remains equally problematic whether we study few or many Ns; it

    will not disappear if we add more nations.

    Quantitative national comparisons rarely, if ever, address such endogeneity problems.

    Studies of labor market or economic performance routinely presume that labor market

    regulations or bargaining centralization are truly exogenous variables, whose effects on

    employment is conditionally identical whether it is Germany, Norway or the United

    States (well-known examples are Calmfors and Driffill, 1988; Hicks and Kenworthy,

    1999). The same goes for welfare state comparisons with their belief that demography

    and left power are fully exogenous and conditionally unitary (for example, Wilensky,

    1975; Pampel and Williamson, 1989).

    There exist several, not necessarily efficient, ways of correcting for the endogeneity

    bias. If we have some incling that the bias comes from variable ommission, the obvious

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    10/13

    10

    correction entails the inclusion of additional controls. An example of this kind was

    Jackmans (1986) argument that Norways North Sea Oil was over-determining the

    results in the Lange and Garrett (1985) study. Small N studies with strong endogeneity

    have little capacity to extend the number of potentially neccessary controls. A second

    approach is to limit endogeneity in X by re-conceptualizing and, most likely, narrowing

    Y. The welfare state literature provides a proto-typical example: aggregate socialexpenditure was increasingly replaced by measures of specific welfare state traits.

    Controls, no matter how many, will however not resolve the problem under conditions

    of strong sampling bias. As discussed above, the best solution in such a situation is to

    concentrate more on the theoretical elaboration of causal relations between variables. If

    we can assume that our estimations are biased because Y affects the values on X, or

    because both are jointly attributable to a third underlying force, thinking in

    counterfactual terms (would Swedens welfare state be the same without Swedens

    social democracy) will force the researcher to identify more precisely the direct or

    derived causal connections (Lieberson, 1987; Fearon, 1991). If bias can be assumed to

    come from the assumption of monotonically homogeneous effects of the X across all

    nations, the researchers attention should concentrate on identifying more precisely the

    conditional mechanisms that are involved in the causal passage from an X to a Y (why

    and how will a 5 point rise in left power make the Danish welfare state converge with

    the Swedish?).

    Bibliography:

    Abraham K G and Hausman S N 1994 Does employment protection inhibit labourmarket flexibility? Pp 59-94 in Blank R, ed. Social Portection versus Economic

    Flexibility. University of Chicago Press, Chicago.

    Achen C H 1986 The Statistical Analysis of Quasi-Experiments. University of

    California Press, Berkeley.

    Alvarez R, Garrett G, and Lange P 1991 Government partisanship, labor organization,

    and macroeconomic performance.APSR, 85: 539-56

    Amemyia T 1985 Advanced Econometrics. Cambridge, Mass: Harvard University

    Beck N and Katz J 1995 What to do (and not to do) with times-series, cross-section

    data.APSR, 89: 634-47

    Beck N , Katz J, and Tucker R 1998 Taking time seriously: time series-cross-section

    analysis with a binary dependent variable.AJPS, 42: 1260-88.

    Blossfeld H P. and Roehwer G. 1995 Techniques of Event History Modelling: New

    Approaches to Causal Analysis. Mahwah: Erlbaum.

    Boeri T 1999 Enforcement of employment security regulations, on-the-job search and

    unemployment duration.European Economic Review, 1: 65-89.

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    11/13

    11

    Bollen K 1980 Issues in the comparative measurement of political democracy.ASR, 45:

    370-90

    Bollen K. And Jackman R W 1989 Democracy, stability, and dichotomies.ASR, 54:

    438-57

    Calmfors L and Driffill J 1988 Bargaining structure, corporatism, and macroeconomic

    performance.Economic Policy, 6: 13-61

    Castles F G 1993 Families of Nations. Aldershot, Dartmouth.

    Esping-Andersen G 1999 Social Foundations of Postindustrial Economies. Oxford

    University Press, Oxford

    Esping-Andersen G and Sonnberger H 1991 The demographics of age in labor market

    management. Pp 227-249 in Myles J and Quadagno J, eds. States, Labor Markets, and

    the Future of Old-Age Policy. Temple University Press, Philadelphia.

    Fearon J 1991 Counterfactuals and hypothesis testing in Political Science. World

    Politics, 43: 169-95

    Firebaugh G 1995 Will Bayesian inference help? A skeptical view. Sociological

    Methodology, 25: 469-472

    Goldberger A 1983 Abnormal selection bias. In Karlin S, Amemyia T, and Goodman L,

    eds. Studies in Econometrics, Time Series, and Multivariate Statistics. Academic Press,

    Stamford, CT.

    Goldstein H 1987 Multilevel Models in Educational and Social Research. Oxford

    University Press, New York

    Greene W H 1997Econometric Analysis. 4th edition. Macmillan, New York

    Heckman J 1988 The microeconomic evaluation of social programs and economic

    institutions. In Chung-Hua Series of Lectures by Invited Eminent Economists no. 14.

    The Institute of Economics, Academia Sinica, Teipei.

    Hicks A. 1994a The social democratic corporatist model of economic performance in

    short and medium term perspective. Pp 189-217 in Janoski T and Hicks A, eds. The

    Comparative Political Economy of the Welfare State. Cambridge University Press,

    Cambridge

    Hicks A 1994b Introduction to pooling. Pp 169-188 in Janoski T and Hicks A, eds.

    Op.cit

    Hicks A. and Kenworthy L. 1999. Cooperation and political economic performance in

    affluent democratic capitalism.AJS, 6: 1631-1672.

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    12/13

    12

    Hicks A and Swank D 1992 Political institutions and welfare spending in industrialized

    democracies.APSR, 86: 658-74.

    Hsiao C. 1986.Analysis of Panel Data. Cambridge University Press.

    Iversen T and Wren A 1998 Equality, employment, and budgetary restraint: thetrilemma of the service economy. World Politics, 50: 507-46.

    Isaac L and Griffin L. 1989. Ahistoricism in time-series analyses of historical process.

    ASR, 54:873-90.

    Jackman R 1986 The politics of economic growth in industrial democracies, 1974-1980:

    Leftist strength or North Sea oil?Journal of Politics, 48: 242-56

    Janoski T and Hicks A. eds 1994. The Comapartive Political Economy of the Welfare

    State. Cambridge University Press.

    King G, Keohane R, and Verba S 1994Designing Social Inquiry. Princeton University

    Press, Princeton, NJ.

    Lange P and Garrett G 1985 The Politics of growth.Journal of Politics, 47: 792-827

    Lieberson S 1991 Small Ns and big conclusions. In Ragin C and Becker, H S eds.,

    What is a Case? New York: Cambridge University Press.

    Lieberson S 1987Making it Count. University of California Press, Berkeley

    Maddala G S 1983Limited-Dependent and Qualitative Variables in Econometrics.

    Cambridge University Press, Cambridge.

    Manski C 1995Identification Problems in the Social Sciences. Harvard University

    Press, Cambridge, Mass.

    Nickell S 1997 Unemployment and labor market rigidities: Europe versus North

    America.Journal of Economic Perspectives, 3: 55-74

    Nieuwbeerta P and Ultee W 1999 Class voting in Western industrialized countries.

    EJPR, 35: 123-160

    Pampel F and Williamson J 1989Age, Class, Politics and the Welfare State. Cambridge

    University Press, Cambridge.

    Przeworski A, Alvarez M R , Cheibub, J A , and Limongi F In press. Democracy and

    Development: Political Regimes and Material Welfare in the World, 1950-1990.

    Cambridge University Press, New York.

    Pudney S 1989Modelling Individual Choice: The Econometrics of Corners, Kinks, and

    Holes. Cambridge University Press, Cambridge.

    Ragin, C. 1987. The Comparative Method. Berkeley: University of California Press

  • 7/30/2019 Qunatative Cross National Research Methods (Esping Andersen G. - Przeworsky a., 2000)

    13/13

    13

    Ragin C 1994 A qualitative comparative analysis of pension systems. Pp. 320-345 in

    Janoski T and Hicks A, eds. The Comparative Political Economy of the Welfare State.

    Cambridge University Press, Cambridge

    Schmertmann C P 1994 Selectivity bias correction models in polychotomous sampleselection models.Journal of Econometrics, 60: 101-32.

    Stimson J 1985 Regression in time and space: a statistical essay.AJPS, 29: 914-47

    Strang, D. 1991. From dependency to sovereignty: an even-history analysis of de-

    colonilization, 1870-1987.ASR, 6: 846-60.

    Usui C 1994 Welfare state development in a world system context. Pp 254-277 in

    Janoski T and Hicks A, op. Cit

    Western, B. 1996 Vague Theory and Model Uncertainty. Sociological Methodology, 26:

    165-192.

    Western B 1998a Causal heterogeneity in comparative research: a Bayesian

    hierarchical modelling approach.AJPS, 42: 1233-1259

    Western B 1998bBetween Class and Market: Postwar Unionization in the Capitalist

    Democracies. Princeton University Press, Princeton, NJ

    Western B and Jackman S 1994 Bayesian inference for comparative research.APSR,

    88: 412-23

    Wilensky H 1975 The Welfare State and Equality. University of California Press,

    Berkeley.