Disability and pay: a decomposition of the pay gaps of disabled men in the UK

1
Disability and pay a decomposition of the pay
gaps of disabled men in the UK

• Simonetta Longhi, Cheti Nicoletti and Lucinda
  Platt
• ISER, University of Essex
• Cambridge September 2009

2
Background

• Disabled employees experience a major deficit in
  pay, compared to non-disabled around 11 for men
  (a difference of c.1.30 per hour) and 22 for
  women.
• (Compare though to c.16 for non-disabled women
  and 21-23 for Pakistani and Bangladeshi men)
• Concern to measure the extent to which disabled
  people face employment discrimination and
  whether that is changing (including in response
  to legislation)
• DDA aimed to address discrimination against
  disabled people in employment more energetically
  than before
• Employment discrimination can be at point of
  employment entry or within the labour market e.g.
  in pay.
• But differences in pay among employed can stem
  from differences in qualifications, in types of
  occupation, and in productivity
• Also vary substantially in average personal and
  employment characteristics compared to the
  overall labour force older, less well qualified,
  higher rates of part-time work etc. regional
  concentration, also some occupational segregation

3
Addressing pay gaps

• Traditional approach to estimating discrimination
  in pay (e.g. for women, ethnic minorities)
• Decompose pay into the part explained by
  differences in characteristics and the residual
  unexplained part.
• Attribute residual fully to discrimination, or
• be more cautious residual includes
  discrimination plus unmeasured characteristics of
  relevance but still regarded prima facie as
  evidence of discrimination.
• But in application to disabled persons pay gaps
  there are both conceptual and methodological
  problems.

4
Conceptual / methodological issues

• Disability different from sex
• Issues around productivity
• Issues about who is disabled who is protected
  by legislation
• Should we also be concerned about differences in
  explained part?
• Oaxaca popular but
• when groups compared different can end up with
  out of sample estimation
• focus on mean but other parts of the
  distribution may be very relevant
• Weighting decomposition approaches more robust
  and can explore different points of distribution
  but dont give detailed decomposition

5
Definition of disability

• For the definition adopted by the DDA, disability
  is defined as long term illness limiting daily
  activities.
• Also possible to examine those with long-term
  illness which doesnt limit activities (not
  covered by Act)
• Previous research in the UK has used long-term
  illness alone to define disability and work
  limitations to define differences in productivity

6
Measuring limits on productivity

• Condition limits amount of work
• Condition limits kind of work
• Co-morbidities (proxy for severity)
• Time off for sickness in any of the weeks
  preceding an interview, versus no time off in any
  of the weeks preceding interview (utilises all
  interviews per individual not just one wave)
• Added sequentially to evaluate impact on pay gap

7
Regression based decomposition

• Oaxaca decomposition (see Blinder, 1973 Oaxaca,
  1973) used to explain mean differences using
  linear regression models
• Advantage it allows for a detailed decomposition
  of the pay gap
• Disadvantage it can produce unreliable results
  if the linearity assumption is too restrictive
  and if the covariates for the two groups do not
  have common support so that the counterfactual
  mean estimation is based on out of the sample
  predictions (see Barsky et al 2002) and Nopo
  (2008).

8
Weighting based decomposition (DiNardo et al 1996)

• Using binary model to predict the probability of
  belonging to a particular group (propensity
  score) to compute weights .
• Counterfactual mean or quantiles are estimated by
  using the weights to equalize the distribution of
  the characteristics between groups with different
  personality traits
• Advantage it does not impose a linearity
  assumption between log pay and covariates and
  does not require a common support for the
  explanatory variables but only for the propensity
  score
• Disadvantage it does not allow for a detailed
  decomposition of the pay gap

9
Combined weighting and regression decomposition

• Weighted estimation of linear regression (for the
  mean pay decomposition) and unconditional
  quantile regression (for the quantile differences
  decomposition) with weights based on the
  propensity score (predicted probability) of
  having high rather than low levels of a
  personality trait.
• Advantage 1 This estimation is consistent if
  either the weights (i.e. the binary model) are
  correctly estimated or the regression models are
  correctly specified.
• Advantage 2 The closeness of the generalized
  Oaxaca decomposition and combined decomposition
  results tells us the confidence with which we
  can use the detailed results for the contribution
  of different characteristics deriving from the
  generalized Oaxaca decomposition
• Note that generalized Oaxaca can be applied to
  decompose quantile differences (Firpo et al 2007)
  using unconditional quantile regressions (see
  Firpo et al 2009). It is similar to the Oaxaca
  method except for the fact that the dependent
  variable is given by the recentered influence
  function

10
Contribution of this paper

• More precise definition of disability
• Also looks at non-disabled with a long term
  health condition
• Better operationalisation of productivity in
  stages and
• Differentiate
• where those not limited in productivity are
  similar to non-disabled
• Where characteristics mop-up the pay gap
• Where residual gap which is not accounted for by
  characteristic
• Distinction between types of disability where
  discrimination may be differentially associated
  with type
• physical long-term conditions and long-term
  mental health conditions
• Decompose pay gaps across the distribution of pay
• Produce robust estimates of explained and
  unexplained components using combined regression
  and weighting approach
• Consider explained as well as unexplained
  components

11
Data UK Labour Force Survey 1997-2008

• Quarterly survey, semi panel (respondents
  followed for five waves), nationally
  representative unclustered probability sample of
  c. 50,000 households per quarter, with
  information on responding adults. Earnings
  information collected in waves 15
• We use 47 quarters, wave 1 responses to produce a
  sample of men aged between 23 and 64, living
  in the UK and in paid employment (excluding
  self-employed). We restrict our sample to those
  who are White British and UK born. Our total
  sample is 120,835 cases
• Compare pysically and mentally disabled and those
  with a physical/mental non-activity limiting
  long-term health condition, according to whether
  work-limited, severity of condition, and lack of
  sickness absences, with those with no long-term
  health condition.
• Log hourly wage (from pay and hours information)
• Wage determinants age age squared, job tenure
  and square education level, part-time job,
  private sector, firm size, region, occupation
• Logit (for weighting by propensity to belong to
  group) also includes dummies for marital status
  and children (lt5 and 5-15)

12
Summary of groups analysed

• Non activity limiting long term physical health
  condition
• Physically disabled (activity limiting condition)
• Non activity limiting long term mental health
  condition
• Mentally disabled (activity limiting condition)
• Reference group no long term health condition
• Within 1-4, look at all and then successive
  subsets of those
• Where the condition doesnt limit the amount of
  work
• (a)where the condition doesnt limit the kind of
  work
• (b)no comorbidities
• (c)no days off sick in any waves observed

13
Rates of Disability

• In the population aged 16 and over, 64.7 percent
  of people do not have any long term health
  condition
• 15 percent have a along term health condition
  that does not limit activity
• the remaining 20.3 percent have a long term
  condition that also limits activity (disability).
  
• Among those with non activity limiting long term
  health condition, 84.3 percent have a physical
  disability as their main health problem, while
  for 3.5 percent the main health problem is a
  mental condition.
• Among those disabled, for 76 percent the main
  condition is a physical condition, while for 9.1
  percent the main condition is a mental health
  problem.
• Among those with a long term physical condition,
  the condition limits activity for 33.6 of cases,
  it limits the amount of work for 23.4 of cases
  and it limits the kind of work for 36.1 of cases
• Among those with a long term mental health
  problem, the condition limits activity for 55.0
  of cases, it limits amount of work for 38.5 of
  cases and it limits the kind of work for 53.0 of
  cases

14
Results decomposition at the mean 1. Non
activity limiting physical condition
Mean Gap Composition effect (Combined) Residual effect (Combined) Composition effect (Oaxaca)
1a) All -0.050 -0.023 -0.026 -0.020
1b) 1a does not affect amount of work -0.030 -0.010 -0.020 -0.008
1c) 1b does not affect kind of work -0.012 0.001 -0.013 0.005
1d) 1c no other conditions -0.008 0.005 -0.012 0.007
1e) 1d no days of sickness leave 0.003 0.000 0.003 0.005
15
Results decomposition at the mean 2. Physical
disability
Mean Gap Composition effect (Combined) Residual effect (Combined) Composition effect (Oaxaca)
2a) All -0.141 -0.061 -0.080 -0.058
2b) 2a does not affect amount of work -0.051 -0.009 -0.042 -0.008
2c) 2b does not affect kind of work -0.018 0.013 -0.031 0.012
2d) 2c no other conditions -0.003 0.021 -0.024 0.019
2e) 2d no days of sickness leave 0.005 0.020 -0.014 0.024
16
Results decomposition at the mean 3. Non
activity limiting mental health condition
Mean Gap Composition effect (Combined) Residual effect (Combined) Composition effect (Oaxaca)
3a) All -0.131 -0.084 -0.047 -0.076
3b) 3a does not affect amount of work -0.103 -0.062 -0.041 -0.063
3c) 3b does not affect kind of work -0.067 -0.032 -0.034 -0.033
3d) 3c no other conditions -0.054 -0.021 -0.033 -0.017
3e) 3d no days of sickness leave -0.001 0.001 -0.002 -0.001
17
Results decomposition at the mean 4. Mental
disability
Mean Gap Composition effect (Combined) Residual effect (Combined) Composition effect (Oaxaca)
4a) All -0.297 -0.130 -0.168 -0.093
4b) 4a does not affect amount of work -0.184 -0.071 -0.113 -0.053
4c) 4b does not affect kind of work -0.151 -0.044 -0.108 -0.051
4d) 4c no other conditions -0.166 -0.062 -0.105 -0.052
4e) 4d no days of sickness leave -0.164 -0.145 -0.019 -0.141
18
Decomposition across the pay distribution
physically disabled - all
Gap Composition effect (Combined) Residual effect (Combined) Composition effect (Oaxaca)
10th percentile -0.117 -0.040 -0.077 -0.029
25th percentile -0.125 -0.066 -0.059 -0.065
50th percentile -0.135 -0.085 -0.050 -0.080
75th percentile -0.137 -0.062 -0.075 -0.065
90th percentile -0.140 -0.050 -0.089 -0.048
19
Decomposition across the pay distribution
physically disabled no productivity limitations
Gap Composition effect (Combined) Residual effect (Combined) Composition effect (Oaxaca)
10th percentile 0.003 0.013 -0.010 0.018
25th percentile 0.000 0.019 -0.020 0.016
50th percentile 0.004 0.005 -0.001 0.012
75th percentile 0.011 0.020 -0.009 0.018
90th percentile -0.007 0.022 -0.030 0.033
20
Decomposition across the pay distribution
mentally disabled - all
Gap Composition effect (Combined) Residual effect (Combined) Composition effect (Oaxaca)
10th percentile -0.223 -0.078 -0.145 -0.049
25th percentile -0.244 -0.104 -0.140 -0.097
50th percentile -0.294 -0.151 -0.143 -0.106
75th percentile -0.301 -0.152 -0.149 -0.085
90th percentile -0.242 -0.154 -0.088 -0.087
21
Decomposition across the pay distribution
mentally disabled no productivity limitations
Gap Composition effect (Combined) Residual effect (Combined) Composition effect (Oaxaca)
10th percentile -0.025 -0.026 0.001 -0.084
25th percentile -0.059 -0.054 -0.004 -0.034
50th percentile -0.130 -0.103 -0.027 -0.123
75th percentile -0.283 -0.127 -0.156 -0.171
90th percentile -0.261 -0.101 -0.161 -0.193
22
Detailed decomposition mentally disabled with no
productivity limitations
23
Conclusions (1) the good news

• We find little or no evidence of discrimination
  as most of the gap can be explained in terms of
  reduced productivity of workers with a long term
  illness.
• Those without apparent productivity differences
  are no different in pay or in pay-relevant
  characteristics from non-disabled
• There is no evidence that those who have a
  long-term health condition but do not fall under
  the DDA are subject to discrimination

24
Conclusions (2) But

• For disabled people with a mental condition that
  affects daily activity an unexplained pay gap
  remains, but only at the top of the wage
  distribution.
• For those with a mental health disability where
  the difference at the mean is explained by
  characteristics, the characteristics themselves,
  particularly occupation which plays the largest
  role - may also be shaped by discrimination
• Are those with mental health conditions who are
  relatively well qualified selecting into lower
  paying occupations which accommodate them?
• Approach assumes that less productive workers
  are not also subject to discrimination on account
  of their condition / its severity / its impact on
  their performance, which may be a strong
  assumption to make (they may differ in their
  experience of workplace and employers from those
  with no work-related limitations).

25

• The End
• Comments please!
• or to lplatt_at_essex.ac.uk