Research Report No 2
Northern Ireland Economic Research Centre
In early 1992, as part of the Review of Employment Equality in Northern Ireland, the Central Community Relations Unit commissioned research to provide a detailed profile of the Catholic and Protestant male unemployed. The research was conducted by Dr Anthony Murphy and Mr David Armstrong, who have produced this report on their findings.
In carrying out the Employment Equality Review, it is the policy
of the Central Community Relations Unit to publish commissioned
research, with a view to informing wider discussion of the issues.
The views expressed in the report are, however, the responsibility
of the authors and should not necessarily be regarded as being
endorsed by the Central Community Relations Unit.
We are grateful to the ESRC Data Archive, the Department of Economic
Development, and the Policy, Planning and Research Unit (PPRU)
of the Department of Finance and Personnel for providing the Labour
Force Survey and Continuous Household Survey data used in this
report. They are not responsible for the content of this report.
Any errors or omissions are our responsibility.
Catholic and Protestant men in Northern Ireland have very different
unemployment rates (Table 1.1). In the 1991 Census, the unemployment
rate for Catholic men is about two and a quarter times higher
than the unemployment rate for Protestant men. This unemployment
differential has persisted throughout the 1980s. It has provoked
a lively debate amongst academics and policy makers about the
factors which account for it.
Notes: The 1971 and 1981
Census figures are from Compton (1991). The Protestant figures
refer to all non-Catholic denominations. The figures in brackets
are obtained by allocating those in the religion 'not stated'
category proportionately between Catholics and others.
Explaining the Unemployment Differential
The basic issue is whether or not the difference in unemployment
rates reflects differences in labour market opportunities. According
to some (eg Compton 1991) it is possible to explain a large part
of the unemployment differential in terms of 'structural' factors
such as age, number of children, geography, social class and/or
industry which, it is argued, have little or nothing to do with
differences in opportunities., According to others (Eversley,
1989 or Smith and Chambers, 1991) much of the unemployment differential
is explained by religion or factors highly correlated with religion
and not by differences in the observed characteristics of the
two groups. In this paper we address this question of whether
or not the unemployment differential can be largely explained
by 'structural factors'.
The paper investigates a number of different aspects of Catholic and Protestant male unemployment. This includes not only the incidence of unemployment but also the following topics:
non-participation (economic inactivity and its links with unemployment;This approach is unique in providing a more comprehensive picture of the unemployed since both flows to and from unemployment are examined as well as the stock of unemployed. For example, we find that high Catholic male unemployment rates are due to both higher inflow rates into unemployment and longer durations rather than higher rates of labour turnover.
Use of Econometric Models
Two different household survey datasets are used to look at these topics. These two datasets are larger than those previously used. A wide range of statistical techniques have been used to analyse the data. In addition to examining the data using simple univariate and multivariate techniques, we construct a series of econometric models which help to disentangle the effects of various personal characteristics, including religion, which are correlated with each other.
Such models are very useful. For example our models for the incidence
of unemployment throw light on the claim in Compton (1991) that
the unemployment difference between Catholic and Protestant men
is largely explained by observable factors such as age, geography
etc and not by religion or other factors correlated with religion.
Outline of Paper
The outline of this paper is as follows. The findings of previous model-based research on the subject of Catholic and Protestant unemployment are briefly reviewed in Chapter 2. The Labour Force and Continuous Household Survey sample datasets are discussed in Chapter 3. Models of the incidence of unemployment are presented and discussed in Chapter 4. Non-participation and discouragement are considered in Chapter 5. We turn to the duration of unemployment next in Chapter 6. We examine and model the job search behaviour of the unemployed in Chapter 7. In Chapter 8 we consider flows to and from employment and unemployment. Finally, we provide some conclusions in Chapter 9.
Nearly all the 'technical' econometric material is set out in
Appendices. In the main text, sections containing technical material
are in small type. Some readers may wish to skip over much of
In this Chapter we examine two important studies which attempt to model the incidence or duration of male unemployment by religion in Northern Ireland. The large number of studies which examine, but do not explicitly model, the incidence of unemployment are not reviewed here. These are reviewed in Gallagher (1991) and Whyte (1990).
This chapter concentrates on the work of Compton (1991) and Smith
and Chambers (1991) since these are fairly sophisticated studies
which' reach very different conclusions. Compton (1991) suggests
that over 80% of the difference in unemployment rates between
Catholic and Protestant males is explained by a small number of
factors apart from religion. Smith and Chambers (1991) suggest
that over 60% of the difference in male unemployment rates between
Catholics and Protestants in their data is accounted for by religion
or factors correlated with religion. The remaining 40% is accounted
for by differences in the characteristics of the two groups. This
suggests a wide range of disagreement; between 40% and 80% of
the difference in unemployment rates between Catholic and Protestant
men is attributed to differences in the characteristics of the
Modelling the Incidence of Unemployment By Religion
The major limitation of much research on male unemployment and religion has been the absence of quantifiable models. Without model results it is very difficult to identify the relative importance of the various factors which contribute to the different unemployment rates amongst Catholic and Protestant males. In consequence there is the risk that research effort will focus on issues which are either not important or else not very important.
Another problem with many studies is the difficulty of simultaneously handling the large number of explanatory variables which are available. These explanatory variables simultaneously contribute to the high male Catholic unemployment rate. However, ft is unlikely that they are uncorrelated and do not interact with each other. More sophisticated methods are required to ensure that the effects of the various explanatory variables are correctly disentangled.
The paper of Smith (1987) was a major advance since he used a
logistic regression to model the incidence of unemployment by
religion and included a large number of explanatory variables
in his model. A logistic regression or logit model is the simplest
method of modelling an outcome such as being employed or being
unemployed. A range of personal and other characteristics which
explain the incidence of unemployment may be included as explanatory
variables in a logit model.
Smith and Chambers (1991)
The logit model results in Smith and Chambers (1991) update the results in Smith (1987). They take account of some of the criticisms of the results presented in 1987. However the 1987 and 1990 results are very similar.
Smith and Chambers list common explanations of the unemployment differential between Catholics and Protestants apart from direct or indirect discrimination and the chill factor:
Dual Labour Markets:These explanations are not independent. For example, the dual labour market story is consistent with the adverse industry and/or socioeconomic group stories. The population growth story, generally attributed to Compton, relies on the two labour markets being segregated. It begs the question of why Catholic and Protestant jobs remain segregated. The remaining explanations are readily incorporated into the logit model of the incidence of unemployment which Smith and Chambers construct.
Oddly enough Smith and Chambers do not include the following two explanations in their list:
Security Related Jobs:The important point, however, is that few members of the security forces or those engaged in the black economy are likely to respond to household surveys such as the Labour Force and Continuous Household Surveys. Thus they are unlikely to bias the results of any analysis of household survey datasets to a great extent.
Smith and Chambers use CHS data for 1983 to 1985 to construct a logit model of the incidence of unemployment. Their effective sample consist of over 5,500 economically active males aged 16 and over. They include the following explanatory variables:
Age:They also include some interaction terms. They find that religion is highly significant. Not too many of the other variables are significant which may have something to do with the clustered sample design used to collect the CHS data in 1983 and 1984.
The results in Smith and Chambers (1991) confirm Smith's earlier
results (Smith, 1987) which were challenged by Compton et al (1988),
Cormack and Osborne (1988) and others. Smith and Chambers discuss
the criticisms levelled against the results presented in the 1987
papers. Their discussion of the points raised appears reasonable
to us. Of course, as Wilson (1989) and others have pointed out,
Smith was wrong to claim that he had taken account of all relevant
It is worthwhile discussing some of the criticisms of Smith (1987) in some detail since many of the same criticisms may be applied to the findings in this paper.
(i) Sampling Problems
A clustered and stratified sample design was used to collect the 1983 and 1984 CHS data. The primary sampling units were wards selected from each of three strata - Belfast district council, other areas east of the Bann and west of the Bann. Although clustering increases sampling variability and makes modelling more difficult it certainly does not invalidate any analysis below the strata level eg at TTWA level. In support of this it should be noted that when we use LFS data we obtain similar results using clustered 1985/86 data and unclustered 1990/91 data.
(ii) Measures of Goodness of Fit
Smith and Chambers were criticised for not presenting a measure of the goodness of fit of their model. Various measures of fit have been proposed for binary models such as the logit or probit. However, no single measure dominates unlike the use of R 2 in ordinary regressions (Maddala 1983, Dhrymes 1986, Greene 1993). Various pseudo R 2 measures are available but, with a large cross-sectional dataset, high pseudo R 2 values are extremely rare. In these circumstances various measures of fit and mis-specification tests should be presented along with the model estimates.
(iii) Educational Qualifications
Differences between Catholics and Protestants in the mix of subjects studied at school or college may be part of the explanation of the unemployment differential. In the LFS and CHS we have little or no data on subject mix. However, there is no evidence from international studies that subject mix has a large effect on the incidence of unemployment. Accordingly we do not believe that subject mix has anything like as large an effect on the incidence of unemployment as the level of qualifications. The effect of subject mix is likely to be quite small. Miller et al (1990) examine the relationship between degree subject and earnings for graduates. However, as Gallagher (1991) notes this evidence deals with a narrow sector of the labour market and with earnings rather than unemployment. Murphy and Shuttleworth (1994) find that, ceteris paribus, the number and level of qualifications of school leavers and not subject choice affects the incidence of unemployment amongst Catholic and other school leavers.
(iv) Population Growth
There are really two separate stories in the arguments about the effects of population growth on the incidence of unemployment. The micro-economic story has to do with the benefit trap. It is assumed that, ceteris paribus, men with large families have very similar incomes in and out of work and so are more likely to be unemployed. Since Catholic families are larger, Catholics are more likely to be unemployed. Miller and Osborne (1983) reanalysed a large cohort study of the unemployed in Northern Ireland and did not find strong support for large benefit trap effects. In any case our reduced form econometric models include many of the factors leading to high replacement ratios such as the number of children, qualifications, others unemployed in the household, etc. These variables should pick up benefit trap effects.
The macro-economic story, which is associated with Compton (1981, 1991), has to do with the effect of differences in labour force growth on the unemployment differential. This argument cannot really be dealt with using micro household data. However, it should be noted that Compton's examples are based on unrealistic and rather extreme assumptions. For example, he assumes that existing jobs always remain segregated since there is no turnover in these jobs. Smith (1987), Eversley (1989) and Smith and Chambers (1991) all point out the limitations of his examples.
However, possibly the best approach is to construct other more realistic examples or models of the unemployment differential in the presence of differences in labour force growth. In Appendix 2.1 we present a simple model which is used to examine the effect of differential labour force growth on the unemployment differential. Our model is based on what we consider to be plausible assumptions about, for example, separation and engagement rates. However, our model gives results which are the opposite of Compton's results. In particular, our model suggests that the contribution of differences in labour force growth to the unemployment differential is likely to be quite small.
(v) Security Related and Black Economy Jobs
Very few individuals in security related jobs appear in household
surveys such as the Continuous Household and Labour Force Surveys.
Of course some individuals in security related jobs may conceal
their occupations by stating that they are civil servants. There
is little one can do about this. Individuals engaged in the black
economy are also very unlikely to respond to household surveys.
Thus it is unlikely that security related jobs, which are predominantly
held by Protestants, or black economy jobs, which some argue are
more common in Catholic areas, bias the modelling results to a
large extent. However, care must be taken when applying our model
results to aggregate data. Smith and Chambers (1991) discuss the
likely effect of security related jobs on the aggregate unemployment
differential. Their calculations are very crude and their results,
as they correctly point out, are over-estimates of the effects
of security related jobs on the unemployment differential. In
Appendix 2.2 we examine the contribution of security related jobs
to the unemployment differential. Our results suggest that removing
the security related jobs effect would reduce the unemployment
differential by between 10% and 15%.
Compton (1981, 1991) uses Census data and standardization techniques to decompose the difference in unemployment rates between Catholic and Protestant males. For example in Table 3.5 in his 1991 paper, Compton used 1981 Census data to calculate a predicted Catholic male unemployment rate assuming Catholics had the same structure (in terms of age, geography, social class and/or industry) as Protestants. He finds that industry alone explains 65% of the unemployment difference while industry and geography explain 78% of the unemployment difference. Compton finds that age only explains 9% of the unemployment differential. He suggests that this is because the age structure is a poor proxy for labour force growth.
We wish to discuss four important issues regarding Compton's approach:
The second issue is important. There is no unique reference or standard group. The use of an alternative reference group is likely to significantly reduce the explanatory power of Compton's procedure.
(i) Meaning of Standardization Procedure
It is unclear what it means to assume that Catholic males have the same structure as Protestant males since social class or industry are hardly structural. They are just as likely to reflect differences in labour market outcomes as to cause such differences.
Compton discusses this issue at length. He points out that, although
his results provide statistical insights, as explanations of the
unemployment differential "they should be approached with
However, the real problem with Compton's standardization procedure
is that ft is not unique. In Appendix 2.3 we show what is going
on. Compton calculates a "predicted" Catholic unemployment
rate assuming Catholics have the same characteristics as Protestants.
An alternative predicted unemployment rate is obtained when Catholics
and Protestants differ in characteristics but, for any given set
of characteristics, Catholics are assumed to face the same unemployment
rate. This alternative predicted Catholic rate is equally valid
yet it "explains" a much lower share of the unemployment
differential. Other choices yield different results. Compton's
choice of reference base appears to be the one which "explains"
the largest share of the unemployed differential.
As Compton briefly notes, his procedure is not invariant to the
level of dis-aggregation. Under plausible assumptions, greater
dis-aggregation always leads to greater explanatory power.
In principle the multiple standardization procedure used by Compton may be extended to include any number of explanatory variables. In practice only a small number of explanatory variables are used. As a result relevant explanatory variables are omitted. The effect of omitting these variables is not considered. Since the omitted relevant variables are very unlikely to be uncorrelated with the included variables, biased results are very likely.
The basic point about all multiple standardization procedures
is that they say nothing about causality; they are merely accounting
In this chapter we have reviewed two studies which model differences in the incidence or duration of unemployment between Catholic and Protestant men. The two opposing views of the causes of the unemployment differential are represented by the work of Compton (1981, 1990) on the one hand and Smith (1987) and Smith and Chambers (1991) on the other hand.
The results reported in Smith (1987) were an important element
of the 1987 Standing Advisory Commission on Human Rights report
on fair employment which led to the 1989 Fair Employment Act.
It had a high profile and as a result Smith (1987) came in for
a lot of criticism. In the case of his logit model results on
the incidence of unemployment, much of this criticism was undeserved.
Compton's work has come in for a good deal less criticism. We
set out some important new criticisms of Compton's methodology.
However, to date, no real attempt has been made to nest Compton's
and Smith's work in a common framework, in order to understand
why they reach different conclusions and to identify the valid
elements in the two approaches.
This chapter describes the two household survey datasets used
in the analysis. The first dataset consists of four years Labour
Force Survey (LFS) data. The four years are 1985, 19869 1990 and
1991. Religion data were not collected in the LFS between 1987
and 1989. The second dataset consists of four years Continuous
Household Survey (CHS) data. The four years are from 1986 to 1989.
Since the sample size in these datasets in any one year is quite
small, the data from each of the years are pooled.
Labour Force Survey
The LFS is an annual survey carried out each spring by interviewing the adult members of around 4,000 households about their personal circumstances and work. It is the largest regular household survey in Northern Ireland and provides a rich source of information about the labour force using internationally agreed definitions.
The LFS provides a comprehensive picture of the economic activity of the private household population. It does this by classifying all adult survey respondents into three main activities, namely employment, ILO unemployment and economic inactivity, according to their circumstances in the week prior to the interview. In the LFS, the employed, the ILO unemployed and the economically inactive are defined as follows:
Employed:The sample design which was used to collect the LFS data changed in 1987. Before 1987 a clustered sample design was used. This meant that the primary sampling units were wards selected from each of three strata (Belfast district council, other areas East of the Bann and West of the Bann). Addresses were randomly selected within the chosen wards. From 1987 onwards, the clustered sample design was not used. Instead, addresses are randomly selected within three sample strata (Belfast DC, East and West). This change has improved the sample design.
The LFS Sample
The sample used in this report consists of economically active males aged 20 to 59 with a known religion (Catholic, Protestant, other religion and no religion). The age range 20 to 59 was chosen to avoid extensive modelling of participation in education and training schemes and retirement decisions.
The LFS data were recoded to ensure consistency of definitions. The sample data are unweighted and non-respondents to the economic activity and religion questions are excluded. Proxy information was obtained from 54.4% of the sample.
Some details of the LFS sample are set out in Table 3.1. Each
year the LFS sample contains about 2,800 co-operating males aged
20-59 with known religion. This gives a total sample of just under
11 300. The religious composition of this sample is 40.2% Catholic,
56.1% Protestant, 1% other religion and 2.7% no religion.
The Continuous Household Survey (CHS) is similar to the General Household Survey in Britain. The CHS is a continuous survey carried out by interviewing the adult members of around 3,000 households each year about their social and economic conditions.
Since 1985, an unclustered sample design has been used to collect the CHS. This means that the addresses for the CHS are randomly sampled from each of three regional strata (Belfast, East and West). Unlike the LFS, there are separate questionnaires for those interviewed face-to-face and those for whom only proxy information was available. The range of information available for proxies is limited. In particular, no details of highest educational or vocational qualifications are collected. The questions on economic activity in the CHS are generally less comprehensive than those in the LFS. In the CHS, economic activity is classified as follows:
The CHS Sample
Some details of the CHS sample are set out in Table 3.2. The sample contains about 7,600 co-operating males aged 20 to 59 with known religion and economic activity. Data from full or partial interviews are available for 86.6% of these. Only proxy information is available for the remaining 13.4%. Excluding proxies, the sample size is about 6,600. The sample, which excludes proxies, is generally used for the detailed statistical analysis since the proxy responses do not contain information on highest qualifications.
Religion data were missing for some individuals. Where an individual
did not answer the religion question or only proxy information
was available, religion was imputed using the religion of the
head of the household or his/her spouse. The religious composition
of the sample is 38% Catholic, 58.9% Protestant, 0.6% other religion
and 2.5% no religion.
In public debate a great deal of attention is paid to the unemployment
differential ie the ratio of Catholic to non-Catholic unemployment
rates. The average unemployment differential in the LFS and CHS
samples is 2.5. The 95% confidence interval for this differential
ranges from 2.3 to 2.7. This means that we can be 95% confident
that the unemployment differential in the CHS sample lies between
2.3 and 2.7.
Economic Activity in the CHS and LFS Samples
Tables 3.3 and 3.4 show different aspects of the economic activity of Catholics and non-Catholics in the LFS sample and Tables 3.5 and 3.6 show different aspects of the economic activity of Catholics and non-Catholics in the CHS sample.
In the LFS the Catholic and non-Catholic unemployment rates are
25.7% and 10.4% respectively. The corresponding participation
rates are 85.8% and 92.2%. The same pattern shows up in the CHS
data. Catholic men are significantly more likely to be both unemployed
and inactive. Unemployment rates are significantly lower and participation
rates significantly higher in the LFS than in the CHS which reflects
the stricter definition of unemployment used in the CHS.
UNEMPLOYMENT AND NON-PARTICIPATION RATES BY AGE
UNEMPLOYMENT AND INACTIVITY RATES BY AGE
 It is not possible to construct a measure of ILO unemployment using the CHS data. However, this is not a major problem since, when we model the two measures of unemployment using LFS data, we obtain fairly similar results.
 Details of how the confidence intervals are calculated are given in Appendix 3.2. In the LFS sample, the unemployment differential fell from 2.6 in 19U/86 to 2.3 in 1990/91. However a clustered sample design was used to collect the data in 1985 and 19%. Because of this we cannot say that the fall from 2.6 to 2.3 is significant from a statistical point of view. This is discussed in more detail in Appendix 3.1