44 MAIRE NI BHROLCHAIN their children will be better off,or no worse off,if circumstances and that there is therefore considerable potential they divorce.In this respect alone,they are likely to differ,for bias. prior to divorce,from couples wbo either never consider divorce Obtaining good information on pre-separation or consider the possibility and stay togetber.These pre-existing family conflict and dysfunction is likely to be diffi- differences may contribute to later outcomes,regardless of cult.A crucial first question vis-a-vis such measures whether the couple divorces.This general possibility is is their validity:how well they measure what they are not excluded by existing studies. supposed to measure.It is clearly very sensible that An association between divorce and adverse out- existing data resources should be exploited in initial comes can be explained in other ways also that research on the 'divorce effects'issue.But when it cannot yet be excluded and are not usually consid- comes to causal inference on any issue,but particu- ered in large-sample studies.One possibility is larly a sensitive one,more acute awareness is needed reverse causation -that children with difficulties of the importance of measurement,data quality,and may contribute to marital instability (Corman and the limitations of existing sources. Kaestner,1992).Other possibilities include shared genetic factors between parents and children (McGue and Lykken,1992;Rutter,1999),or that Regression Coefficients are Interpreted situational aspects of the family,such as prolonged as CausalEffects' experience of unemployment,that may be difficult to capture in general-purpose surveys,have an Let us suppose that we have a statistical model in the adverse long-term impact on family relationships, form e.g.of a logistic regression of some outcome including parental interaction with children(for evi- on a range of covariates,in a before-and-after analy- dence of unemployment as a risk factor of divorce sis designed as suggested earlier.All variables see Bracheretal.,1993). thought to be relevant are present and the quality of measurement is high.Suppose further that the regression is coherent in a statistical and substantive Measurement:Data Quality and Bias sense.A linear model is appropriate,no conse- quence of parental divorce is included among the Several of the studies examined use outcome indica- control variables,no factor mediating divorce and tors that are rated by parents and/or teachers,both outcome is in the model nor is any cause ofor result before and after separation(the NCDS behaviour of any such mediating variable.Is this enough to ratings at 7,11,and 16 and the behaviour scale used allow causal inference?The answer is no.Such a in the NLSY-CS).Such ratings are potentially sub-regression coefficient expresses no more than a net ject to bias,both because they are not blind to the statistical association between parental divorce and family history of the child and because the parent's the outcome.Regression coefficients are merely cor- perceptions may alter after divorce.Blind measure-relational data and are as subject to the maxim that ment is needed to avoid potential bias due to 'correlation does not imply causation'as any correla- preconceived ideas about the relation between par-tion coefficient. ental divorce and children's later welfare(Santrock None of the studies available at present approxi- and Tracey,1978;Amato,1991).In medical research,mate this ideal.Nevertheless,with a few exceptions, double-blind trials are recognized as an essential causal inferences are made routinely and regression precaution in evaluating a treatment.Neither the coefficients for divorce and other independent vari- patient nor the physician making the assessment of ables are referred to as 'effects,with the clear sense of e.g.change in a patient's condition is aware which causal effects,of parental divorce etc.on the out- treatment the patient received.Elaborate care is comes analysed.The format of regression analysis taken to eliminate the influences of prior expecta-no doubt gives the illusion of prediction.The fiction tions on the measurement of outcome.Such of causal inference from regression coefficients is precautions appear not to be adopted in divorce not confined to the'divorce effects'issue.It is shared effects'research.This is so in spite of the widespread belief with almost all of present day social demography that cbildren from broken bomes'suffer because of their family and empirical sociology.A glance through the last      3    3    +     "    !  ,"    "                   %&  # ,        "           5             +                                    1        >    G             #)    %&&9* 1                 #6K   J"" %&&9( I  %&&&*                      G                             #       "      7   %&&=* " 1 ' 9  :                   N          #  <)
      . %%  %E         <J,)*           /                        4         7                            4    # "  5 %&.'( +  %&&%* !       0              <        "           4                 "     B                         2 3 4  &               -!  .,                       % 1         B      "   G   +       Q        A              !                       2 3 4   7                                                  7
!   <  
 =6/ J                                             +                                         +                                                   !         @ 5         G                      I G             /     2       4    G <             <                G             23 4        3              5               5         G       2 3 4   !              +