European Sociological Revien,Vol.17 No.1,33-57 33 'Divorce Effects'and Causality in the Social Sciencesl Maire Ni Bbrolchain Causality in social science is examined in the context of the question whether children are adversely affected in the long term by parental divorce.The paper identifies first what precise question is at issue.The data and methodology required to examine the question are specified.Studies approach- ing these requirements are examined and it is found that the largely causal conclusions drawn are not justified.To question the causal status ofevidence on the hypothesis is not to imply that it is false.The finding of the present paper is,rather,that the evidence is insufficient to allow the inference that divorce causes long-term adverse effects to the children.More than this,it appears that the sociologi- cal and demographic literature has not,by and large,addressed seriously the difficulties of establishing causal effects in this area.Unjustified causal inference is not unique to the divorce field.Correlational data are routinely interpreted as causal in other areas of social science also.Cau- sal interpretation is a source of difficulty in general in the social sciences and increasingly this is being recognized and discussed.Some reasons for the difficulties involved are suggested and some interim pragmatic approaches to resolving them are proposed. The Problem Divorce,with its supposed long-term conse-Outside the clinical psychiatric literature,academic quences,is a regular fixture in both popular and writing on divorce effects'does not tackle seriously academic discussion of the family in the UK and the problem of establishing causal linkage.The dif- the US,though much less so in continental Europe.ficulties with causal inference that beset this area are It is widely assumed that a causal link between par-common to the social sciences in general (McKim ental divorce and adverse long-term outcomes for and Turner,1997).Social science lacks,at present,a children is well established.In fact,although a statis-scientific toolkit for seeking and validating causal tical association between the two is found commonly,evidence.The issues examined here are,thus,of though not universally (Hullen,1998;Cheese-very general relevance. brough,1999),there is as yet no evidence that the Because divorce is centre-stage in public debate divorce of their parents canses long-term damage to about family values it has a political charge that children.At the same time,evidence against the pro-almost certainly contributes to the current over- position,such as found in Yeung ef al.(1995)and interpretation ofevidence.Divorce has personal sig- Kiernan (1997),is sometimes overlooked.Demo-nificance also.Strong feelings,acknowledged or graphic and sociological discussion of the subject not,can be aroused by the subject since the vast has been marred by over-interpretation of statistical majority of people have,have had,or expect to have associations as causal.Detailed,high-quality evi-extensive personal investment in a family.These dence is needed to justify causal inference,whether features distinguish social,psychological,and relating to divorce or any other hypothesized cause.demographic phenomena in general from the Oxford University Press 2001
! " ! # ! $ ! ! " ! #$ %&&'( ) %&&&* + , #%&&-* #%&&.* " / / 0 1 2 3 4 " " 5 #6 5 %&&.* " " " 5 7 " / / 5 1 8 9::% ; %. -.
34 MAIRE NI BHROLCHAIN study of e.g.the structure of atoms and molecules.In this case,we would not know wby such children But the emotive charge is probably especially fare less well than others. pronounced in relation to family matters.Children's Version B2 seems initially to be the question that response to divorce is embedded in broader social-we might ask if the subject-matter lent itself to political controversy,but the focus of the present experiment,which it clearly does not.But B2 side- paper is on the scientific and professional aspects. steps the true experimental question.An experi- What is the evidence?How has it been interpreted? mental question would be posed in factual rather Is such interpretation valid?If not,why not?Why is than counterfactual terms:Do children in two- evidence on the question over-interpreted?What parent families suffer long-term disadvantage can we do to move up a gear,in a scientific sense? wben their families are subjected to divorce?' Phrased thus,the question is absurd.Divorce is undertaken voluntarily by one or both parents The Question at Issue and cannot be imposed on both parties except in rare circumstances (e.g.the case of the Egypt- The question 'Does parental divorce have adverse ian couple faced with coercion by Islamic law to long-term effects on children?'is less straight-divorce because of what was seen as the apostasy forward than it appears.2 It can be understood in at of the husband).It is impossible to assign least two ways: families at random to a divorce treatment'and a A-Aggregate-level:Does the level of divorce in a 'stable family treatment,and certainly not a society influence the long-term welfare or well- bereavement treatment,holding relevant vari- being of children in that society? ables constant,and examine the 'effects'on the B-Individual-level:Does the divorce of their par- children in the long term.Nevertheless,a weak ents adversely affect the children of the marriage version of this idea probably assails our thinking in the long term? when we attempt to puzzle through the un- By and large,it is the individual-level version of the certainties surrounding the issue.A natural way question that has been the subject of both academic of trying to get an intuitive grasp of the question and public debate although elements of the aggre-is by asking what would happen to the children gate issue may also be influencing thinking and of such-and-such a family with whom we are comment.3 The individual-level question is the pri-acquainted,usually a happy one,if the parents mary focus of this paper and might be specified in were to separate.Such thinking misleads,how- several ways,only one of them correct:+ ever,because divorcing couples are not a B1:In the long term,do children in two-parent random selection of married couples,and so our families fare better than those in one-parent mental experiment will probably be biased.Our families? attempts to think intuitively would be better B2:Womld children in two-parent families suffer served if we were to imagine the hypothetical adverse outcomes in the long term if their par-costs and benefits to cbildren in families at genuine ents were to divorce? risk of divorce,that is,those experiencing the kinds of diff- B3:Would the children of those couples who do culties that usually precede marital breakdown. divorce suffer fewer adverse outcomes in the Question B3 is,in my view,the true question at long term if their parents were to stay together?issue.Strictly speaking it cannot be answered in Question Bl appears to be the one that many com-this form since it requires that the same individuals mentators have in mind.Although the correct follow two mutually exclusive paths through the life- answer to this question appears to have been 'yes,cycle and that the associated outcomes be observed. this does not mean that divorce adversely affects The best initial approximation to the implied com- children,since the question in this form is merely parison is by statistical means.There are two about a correlation.5 Any difference in outcomes components to the comparison:first,specifying may be the result of factors (that give rise to or the outcome measure and second,approximating impede parental divorce that(b)would have been to the randomization that would take place experi- present in any case had the divorce not occurred. mentally.Ideally a comparison should be made of
7 )4 ? @ $ @ ! @ ! @ ? @ ? @ 5 2 3 @4 9 ! A + > + A B @ 7 > ! A 3 @ 7 / B " = 5 AC 7%A ! @ 79A 3 @ 7=A 3 @ D 7% + 24 3 - + 3 #* #* ! " ; 79 " / E 7 79 + A 2 3 / @4 8 " # ! * ! 2 4 2 4 2 4 23 4 < " " 00 + " " 1 " ! " " # ! $ !% D 7= " 5 5 A 0 " !
DIVORCE EFFECTS'AND CAUSALITY IN THE SOCIAL SCIENCES 35 the change in some outcome from before to after 3 Pre-separation predictors of post-separations divorce among the children of divorcing couples outcomes are controlled for,including espe- with change over an equivalent period among the cially any that are joint predictors of separation children of stable marriages.To avoid bias,the pre- and outcome;such predictors can be classified as divorce observation should be made prior to )specific to the child-i.e.child characteristics divorce,and this may be essential in any case because and (specific to the parents or family; some phenomena can only be observed concur- 4.To avoid bias,these predictors are measured rently.The timing of the post-divorce measurement before separation. should be unambiguous -it should relate to the Pre-separation child characteristics are of impor- period after divorce.To approximate to random-tance in general,but particularly in relation to ization,this comparison must be net of any pre-educational outcomes,since it is known that early divorce factors that influence either or both divorce ability and achievement are related to later school propensity and(the risk of)the outcome of interest.achievement and continuation rates.Condition(2) Again,observations on these covariates should be rules out studies in which initial observation of taken before any divorce occurs. children in intact families is at,say,age 12,a second In longitudinal research on children,some out- observation is at age 18,parental divorce is identified comes lend themselves to a formulation in terms of as having occurred some time between 12 and 18, change and others do not.For example,before-and-and the outcome is whether the child had ever after measures can be obtained of school attainment experienced some event-e.g.leaving school- or of emotional adjustment and a measure of change between 12 and 18.In such cases the difficulty is constructed.But others do not fit this mould.Some that the outcome might have preceded separation. behaviours and experiences become possible with These may appear to be stringent conditions though the passage of personal time that were either impos-in fact they should strictly be a great deal more so.A sible or exceedingly rare at an earlier age.Examples full set of conditions would specify that we know the include first sexual experience,first pregnancy, causes of divorce and that measures of these are cohabitation,and so on.Excepting abnormal cir- entered into the equations.There is one further pro- cumstances,such experiences occur only after mising approach via sibling studies that will be sexual maturation.In these cases,the true question considered presently. of interest is whether the rise of an early first preg- nancy etc.is raised by parental divorce beyond what it would otherwise have been.As things The Evidence stand,this risk is unobservable at the individual level.Furthermore,comprehensive predictors of The literature on the subject is in three broad future risk are,as yet,unavailable and may never be groups,although there is some overlap between known.The best we can do,within current metho-them: dological resources,is to make the comparison 1.studies reported in psychology sources,usually controlling in some way for pre-divorce characteris- but not exclusively based on small samples; tics and circumstances that are predictive of the later 2.reports in psychiatric publications,also usually behaviours. small in scale;and In sum,minimal requirements for establishing a 3.studies by demographers and sociologists,pri- link between parental divorce and later outcomes marily analyses of large data-sets. for children,whether positive or negative,that The paper addresses chiefly studies of the third kind would be a candidate for causal investigation are while drawing also on the clinical literature. these: Investigations of the first type,discussed briefly 1.Where possible,a change variable is used,with here,are numerous and in the main weak in meth- measurements taken both before and after the odology.A much cited meta-analysis by Amato and event; Keith (1991)summarized 92 such studies,that are 2 If not,that outcomes relate unambiguously to probably a fair sample of the psychological literature the period after the separation takes place; on 'divorce effects.The methodology of 84 of the 92
5 5 > 5 0 B # " * + " ! F / 7 ! ! + " F " " 5 " . ! " A % ? " ( 9 ! " ( = 8 ' / ( #* > #* ( C 5 8 " ) #9* %9 %' %9 %' > > %9 %' ! G 5 + " 5 5 / A % ( 9 ( = 5 B " ! B " + + #%&&%* 0 &9 2 3 4 5 'C &9
少 MAIRE NI BHROLCHAIN primary studies in this meta-analysis has been exam-question.Some authors acknowledge that causal ined by Jameson(1996)to evaluate the quality of inference may not be justifiable,but most ignore information used in the meta-analysis.The studies this proviso in terminology and discussion.At face were found to be mainly below regional level in value this literature gives the impression that coverage,nearly half did not use probability sam- parental divorce caures children to experience a pling at any stage,most gave no details of response range of long-term disadvantageous outcomes and rate,and most used no statistical controls at all.it is routinely interpreted in this way.Furthermore Finally,the large majority(89 per cent)were cross- this conclusion often appears trouble-free,giving sectional in design,with only one having any ele- rise to few difficulties of method,data,or inference. ment of before and after'comparison.On this My view is that there are major difficulties ofall three evidence,the psychological literature on the subject kinds. has so many shortcomings that it can scarcely sup- Most existing studies make comparisons that are port an inference concerning an association poorly designed,and with inadequate sets of control between parental divorce and children's outcomes, variables.There is much duplication of data-sets. let alone inference as to a causal influence in this Most rely on data that are of poor quality in some regard.9 crucial respects and take no steps to eliminate poten- There are many studies of 'divorce effects'in the tial biases in the data.Regression coefficients are sociological and demographic literature that employ routinely interpreted as representing causal effects. large data-sets.These are the focus of interest here. Little attention is given to individual differences Many are cross-sectional,or use longitudinal data in either in the outcomes of interest or in the propen- a cross-sectional way.Some of the outcomes consid-sity to divorce.Findings at variance with the ered are measures of school attainment or of supposed adverse influence of parental divorce are behavioural or emotional adjustment prior to age sometimes overlooked.Finally,hypothesis testing 16.A large proportion relate to socio-demographic is weak.The divorce effects'literature shares these indicators in adolescence and young adulthood-deficiencies with much of the social science litera- staying on at school beyond the compulsory limit ture.But the problems are seen in high relief in and other measures of educational attainment, relation to divorce because there has been so much early cohabitation or marriage,early parenthood,apparent certainty about the issue in public debate. and the like.Most of these studies use some form To evaluate the evidence relevant to the causal of multiple regression,with child outcomes as connection between parental divorce and long- dependent variables,and with independent vari-term outcomes for children,those large sample stu- ables featuring family structure and anything from dies have been identified that meet the criteria one to many control variables.Measures of family specified in the preceding section:before and after structure differ greatly between studies,ranging studies-those beginning with a sample of child- from cross-sectional parental marital status,parents ren in intact two-parent families-that evaluate ever divorced before a given age,through to time-change in some outcome from before to after the varying indicators of family status.Some look at a divorce or,if not,an outcome that occurs unam- single outcome,others at a range of outcomes.biguously after separation,in which both pre- Where several outcomes are examined they are separation child and parental covariates,measured sometimes closely related.With some exceptions,a prior to separation,are controlled for.There are significant coefficient is usually found for family dis-many other studies of the subject that do not have ruption,and where type of disruption is given,the these characteristics and these have,on the whole, coefficient for divorce tends to be statistically signif-made a useful contribution in establishing a basic icant and that for widowhood not so.This is not so correlation between divorce and outcomes.Studies clear-cut in those few studies using dynamic analysis known to me as of mid-1999 that meet the criteria of family change (Wojtkiewicz,1993;Wu and specified number 10,one of which (Cherlin ef al., Martinson,1993;Wu,1996).With some exceptions,1991)employs two data sources,making 11 analyses regression coefficients are usually taken to reflect a in all.10 The data source on which each analysis is causal effect of family disruption on the outcome in based,together with sample size,children's ages
H #%&&E* 5 F / #'& * 2 4 1 / 4 B & 5 2 3 4 5 6 / %E + > " 6 6 3 " ? ? G G 5 #?/ " 0 %&&=( ? 6 %&&=( ? %&&E* ? G " B 3 " / + F G 6 / G " 6 " 5 6 " I G 3 J 3 F B " F " 5 2 3 4 7 5 A > > 5 / " %&&& %: #) %&&%* " %% %: 5 0 4
DIVORCE EFFECTS'AND CAUSALITY IN THE SOCIAL SCIENCES 37 PID 13pu35 [pue motsrgaq jo pur /Ieuonows IIe)s3unIp (L ae e powenjAa Ieuared (Auq)L Ieuosiad pue Arvuq IIe) (L i pue ad eu s.puyo jo 'ou'uop ( 220 pue M艺 6l.6961 SOON 69615991 SOON 1661 t661 鉴
&' ( " " % & ' ( ) * $ + ( ( ! ( , ( - ) I %&&% ..'% . %E ./ >>>> ( K ) % %&&% %%'=. . %% ./ > . 01 / ' 0 2 3 > ( # * <LM 7M % %&&C %&&- 9::: <) #* %&E-N%&.C %&'% .'EE . %E 9= / ! / 4 2 / 03 " / / / 2 / 03 " / 4 ) 4 2 / 03 " / / / / / 2 03 % / ) ! 2 03 # *
38 MAIRE NI BHROLCHAIN PI4D (apu Jo somseow AlreJ/iuared [wuosiad u)se to) 元 (5661 I LIOoInO 2-2 9 元 aldwes 15601 pawutjuo5[q pue SDN pue s961 sawooino 61/5961 SODN 61961 16610 (Sn)OSN 1861261 (Sn)OSN Apms uay 'epAcd
&' % ' ( ) & * $ + ( ( ! ( , ( - ) ) % %&&- >>> 6 0 2 03 0 2 / 3 # * ) J %%&&- >>> 6 0 # ) %%&&-* %&&. 5 = C E ' >>> 6 0 # ) %%&&-* 7 %&&' %% %9>%E " / " / 4 7 6 4 ) 4 %&..>%&'% > ; > 4 ( K ) %%&&% %% %%>%E / > > " , > ( # *
DIVORCE EFFECTS'AND CAUSALITY IN THE SOCIAL SCIENCES 39 PI42 IOLABq58 Jo sounseaw drysuopeal Jo samsow [euared suqs spP Jo 'oN s,Jamow are inodop pooys Jaie (S's siuaned odwes moqe c0000 Be pue EZII SO-ASIN sn) 盏 -ASIN (sn)so 26619061 (sn)Aprs 2861061 Apms pue 7661 oy pue -166l
&' % ' ( ) & * $ + ( ( ! ( , ( - ) 6 ) %&&- ) #* %&'E %&'' %%9= = CO # - -* 9%>9' - EO #@ . -* 9=>=: 8 " / " / %.9 > " : H"" %&&' 5 E . F =7 C7 ) #* %&'' %&&9 %EC: 9>%: 9=>=: E>%C 9.>=C . / ( , > ( + 6 J %&&% 5 . $ 7 #* %&': %&'9 %::::@ %C>%- %E>%. I # * > > I %? 9<)A < ) # %&-' 7 ) *( <)A < ) #*( <J,)( < J , ) #*( $7A $ 7 #* =F C A - "
40 MAIRE NI BHROLCHAIN iueoyrudis ssep [eos pue aooyp ouou :sAog painseou sowooino (s3100s owoy SuLAeYI o1saaV 也 owooino sneadt-0 uondnsip Al!uej (91ae)tL61 Aq pooo!p aojnos SON SOON SCON SOON Apnis 003566106 S661
& ) " " * ; ! & * ' " % ; + I %&&% & %E 6 # * 6 IA # * 7 ) %%&&% C %% 6 N # * 7A KA 6 N C 7A 6 N KA I N %E %E>9= J 6A FA J 6A FA ) %%&&- %E %E>9= J # 9=* 6A FA 6A FA )J %%&&- %E 9= 6 6 ! .>%E A 6 3 .>%E .
DIVORCE CTS' ND CAUSALTY THE SOCIAI SCIENCES 'ce ie ao A5 pofopduaun A3'ce le auisnoy [epos u! suopeoy![enb qary uorsuaypdwoo Surpea :sAog suonty!enb qioq u uoissadop orxuy -ad pue uondnasip sasto souooino posnjuoo s [uopou pue odui'asua ddeyun Sauoamo 91-II - pue aoJoAlp 455C-0 51512-0 51512-0 uopdns!p AIrej ()L6I Aq paoros!p Aq Paolo!p (11-L saae) uoneedas (+s)8861 Aq paidnas!p Z661 Aq paidnas!p (I/EI3e)(861 B SOASIN SOASIN 16617P/uq SH 1661 ueue DNN Pue Suosy Apnis
& ) " " * ; ! & * ' " % ; %&&. 5 = C E ' & 9C %E>== %%* %&'% # %%>%E* :>- %%>%E A G %%* %&'% # %%>%E* :>- %%>%E 7 9 -O 7 7A KA 7A KA H"" %&&' C E>%C + + ( B > + 6J %&&% 5 . $7 F %&': # %=N%C* %&'9 # %-N%E* :>9 %-N%E 6 2 4A + 6 %? 9F N =)G -P C+ ) ( G -! .>%E 9=P E1 G %:P
42 MAIRE NI BHROLCHAIN and control factors employed are summarized in Table 19 of Kiernan (1997)indicates that parental Table 1.Further features,including a brief resume divorce is not associated with a higher risk of of findings,are given in Table 2. partnership dissolution,once age at first partnership The purpose here is to consider whether these is allowed for.Later tables suggest a higher risk studies provide evidence of a long-term causal effect among men whose parents divorced.13 There are of parental divorce on the outcomes examined. several reports of interaction effects that are of con- Note that these are the best-designed studies cur- siderable potential interest.Elliott and Richards rently available and so comments on their (1991)find that in relation to reading scores at 16 par- methodologyare made in asceptical rather than a cri- ental divorce interacts with social class,with tical spirit.We see first that there is a good deal of children from divorced manual backgrounds duplication of data sources.Of the eleven tabulated,improving their performance relative to age 7 and six are based on the UK National Child Development the opposite occurring in the divorced non-manual Study (NCDS)(1958 birth cohort).The US National group.Chase-Lansdale et al.(1995)find an inter- Survey of Children features twice,as does the Child action between parental divorce during the age- Supplement of the National Longitudinal Survey of range 7-16 and emotional problems at 7,the Youth (NLSY-CS)(US).In two studies,there is con- outcome being mental state at 23.The data suggest siderable redundancy between several outcomes as that among children with behaviour problems at 7, may be seen in Table 2(Ni Bhrolchain efal.1 1994,the divorce group was little different in mental 1995,2000;Cherlinefal.,1995).Six of the studies are health scores at 23 from the intact family group at relatively short term,with an average time between 23.However,among those with few behaviour pro- separation and second observation of 1-3 years- blems at 7 the divorce group had poorer mental Baydar(1998),Cherlinetal.(1991,both data sources),health scores at 23 than those in intact families 7- Astoneand McLanahan(1991),MorrisonandCherlin 16.Finally,Jekielek's(1998)rigorous study reports (1995),Jekielek (1998)-though in some of these an interaction between divorce and parental con- cases the gap may have been as much as 5-6 years.As flict,with children in the most conflictual families it is generally accepted that children areaffected in the at first observation improving at second observation short term by parental divorce,these are not relevant if their parents had divorced in the interim and wor- to the issue of long-term effects,though they report sening if the family had remained intact. some interesting findings.All five of the longer-term studies are based on the UK NCDS data.Thus evi- dence of adverse long-term effects of divorce on Limitations of the Evidence children appears to be based currently on just one data source,the British 1958 birth cohort study.12 Control Variables While the purpose here is not to discuss the find-All of the studies considered here employed a multi- ings of these studies in detail,a few points are variate analysis of some kind.As we see in Table 1,the worthy of particular note.There is some inconsis- most common controls used are socio-demographic tency between the findings of two sets of studies features such as gender of the child,race,age of the (Ni Bhrolchain etal.,1994,1995,2000 and Cherlin et child,region,social class,parental education,and al,1995),in varying analyses of the same or similar number of siblings.A few include more specific par- outcomes using the same data source and similar,ental characteristics such as the mother's age at though not identical,designs.There is one investi-marriage or at the study child's birth.If family back- gation of a range of socio-economic outcomes- ground is covered,it tends to be the family's socio- qualifications,unemployment,and the like-up economic circumstances rather than psychological to age 33,the longest-term study of those examined or social features of the family environment. here.Of the six male and four female outcomes Many variables of primafacie relevance are missing examined in before-and-after fashion only one is from these analyses.Four general categories of statistically significant for each sex(Kiernan,1997).variable tend to be absent:the behaviour and In addition,though not based on a correct before- personal characteristics of the parents,parental and-after design or on a full multivariate analysis, interaction with each other,parent-child
0 5 % F M M 5 9 5 3 = > 7 #%&&'* ) #%&&% * + 6J #%&&%*6 ) #%&&-* H"" #%&&'* > ->E + 3 3 + " > == 1 # %&&.* ! 5 %& #%&&.* " J " %= 5 3 I #%&&%* %E " . )J #%&&-* .>%E . 9= 5 . 3 9= 9= $ . 9= .> %E F H""4 #%&&'* B B 8 + " + 5% + 4 4 ! " 4 6 F A