DIVORCE EFFECTS'AND CAUSALITY IN THE SOCIAL SCIENCES 45 10-15 years of the top journals in the subject will This stance may be defensible when we are bear this out. concerned only with identifying those explanatory Though rarely explicitly discussed,the method-factors -such as social class,educational attain- ology implicitly used to identify causes in ment,region of residence,and so forth -that have sociology and demography is an unsystematic form macro-level significance.But in the divorce effects' of what Goldthorpe (this volume)terms 'robust issue,as in many others,interest is not restricted in dependence.In this scheme of things,it is supposed this way.It is precisely the individual-level variation that a correlate is a cause if it is associated with an in outcomes that is at issue.Nevertheless,when outcome,precedes it in time,and continues to be reported,R2or pseudo-R2 values are typically low, associated,to some degree,when those potential of the order of 0.0-0.3.18 The low pseudo R2 values confounding factors currently thought relevant are suggest that we are very far from being able to pre- controlled for.As Freedman (1985,1991,1997)and dict which children will experience adverse Goldthorpe (1998)argue,these conditions are outcomes.This being so,there are many unmea- inadequate for causal inference.One would need, sured predictors of these outcomes missing from at a minimum,a behavioural theory whose assump- the models typically presented.As a result,the pos- tions had been verified,and an investigation of the sibility remains that parental divorce is no more than functional form linking the various elements of the a proxy for unmeasured,or poorly measured,prior theory.Consider Tukey's criterion that 'The estab- factors.An alternative possibility is that the out- lishment of a causal relation always requires two comes themselves are to a substantial degree elements,one empirical,the other theoretical.The random occurrences,with a large proportion of empirical observed regularity or experimental result error variance.If so,this remains to be demon- has to be such that its occurrence is theoretically strated,though distinguishing between error impossible unless"A caused B"'(Tukey,1986b:variance and variance attributable to unmeasured 208).This is a strong criterion,and it certainly is factors could be far from straightforward.There not met in the evidence on'divorce effects'and prob-are two further reasons for supposing that un- ably not in any of the many thousands of causal measured variables are likely to be an important claims that are scattered through the pages of our component of the variance unaccounted for in embryonic science.Theoretical ideas are sometimes divorce effects'studies.First,as we saw earlier, used in the divorce effects'literature,but their most of the factors included in regression analyses assumptions are never checked and the correct of outcomes in 'divorce effects'studies are fairly form for expressing the relation among the explana- crude socio-economic factors of the kind used in tory variables and their relation to the outcome traditional demographic differentials research. variable is never examined. They are not the kinds of variables that would be expected to account for individual differences in outcome nor for individual differences in divorce- The 'Divorce Effects'Question is about Individual Differences propensity.Second,many 'divorce effects'studies search for an association between divorce and out- Individual differences are crucial to this issue,and come variables rather than seeking the predictors are rarely if ever touched on in the sociological and of those outcomes.Data sources are chosen specifi- demographic literature.Sociology and demography cally for that purpose rather than to understand the are concerned with,as Rutter (1982)puts it,ques- outcomes themselves.19 tions of 'how many?'rather than 'who?'Our stock Alongside our inability to account for the out- in trade is the level or frequency ofa phenomenon or comes thought to be linked with parental divorce, behaviour rather than identifying precisely who will we cannot account for divorce either.Statistical experience or display it (see also Lieberson,1985). models of divorce account for little of the variance Accordingly,it is often claimed that low values of in divorce risk (Murphy,1985;Bracheretal.,1993).If R2 or analogous measures of goodness of fitl are to we do not know how and why parental divorce be expected when modelling demographic and occurs,we cannot claim with assurance that it is a social behaviour and are not a cause for criticism. cause of later outcomes.20 It is unlikely that divorce%:>%-     /     /     5                               K  #   *  2  4 !                                                         + F #%&'- %&&% %&&.*  K  #%&&'*              1                              "         ) 5 "4    25                        5                         22+   7444 #5 " %&'EA 9:'* 5                2 3 4                            5         2 3 4           "                            ='  6/ 9    <  '  !  3                               I  #%&'9*        2 @4    2 @4 1   "                          #  J %&'-* +          I9          %.                      5               2 4   >                    >     7   2 3 4                  !                <     I9   I9           : :>: = %' 5    I9                       5                      +                              +                                !                                      5                "               2 3 4   F                    2 3 4            "      3   5    "               3        3       2 3 4                   "       
                      %& +                 "                              " #6  %&'-( 7   %&&=* !    "                         9: !  "