1
Disability Onset, Disability Exit, and Welfare Benefit
Receipt
Melanie K. Jones
Cardiff Business School
Cardiff University
Tel: +44 (0)2920875079
Email: [email protected]
&
Duncan McVicar
Queen’s Management School
Queen’s University Belfast
Tel: +44 (0)2890974809
Email: [email protected]
Preliminary Version: June 2016
Please do not quote
JEL: H51, H53, I38
Keywords: disability, disability onset, disability exit, welfare benefits, disability insurance,
local labour force survey, propensity score matching
Acknowledgements
This work is based on data from the Annual Population Survey which is produced by the
ONS and is accessed via special licence from the UK Data Archive, University of Essex,
Colchester. The usual disclaimer applies.
2
Abstract
There is consensus in the dynamics of disability and labour market outcomes literature that
disability onset leads to a decline in employment and earnings and an increase in applications
for and receipt of disability insurance. But important questions remain. How does disability
onset impact on welfare benefit receipt beyond disability insurance? How does disability exit
impact on benefit receipt? To what extent do these relationships vary for individuals with
different socio-economic characteristics, in different labour-market contexts, and under
different disability benefit regimes? This paper addresses these questions exploiting rarely-
used longitudinal data constructed from the UK Local Labour Force Survey, combining
propensity score matching with difference-in-differences methods to draw plausibly causal
inferences under explicit assumptions. Disability onset is shown to (robustly) increase receipt
of sickness and disability benefits within one year (by between 3 and 8 percentage points),
although the impact on non-sickness benefits is small and non-robust. Onset effects are larger
for individuals with lower qualification levels, for those experiencing onset of disability
linked to mental health conditions or more severe disability onset, and for those under a
disability benefit regime with less rigorous screening and conditionality. Evidence on
disability exit effects is more mixed: some estimates suggest negative impacts on benefit
receipt, but this conclusion is fragile to different assumptions about diverging prior trends.
Arguably the most convincing exit estimates presented here – comparing those who report
disability exit with those who report disability exit one year later – suggest exit does not
impact on receipt of sickness and disability benefits two years on. There are few clear
differences in estimated disability exit effects by individual characteristics or labour market
context. The apparent asymmetry between benefit receipt impacts of disability onset and
disability exit provides further (indirect) evidence that even temporary disability can have
long lasting economic effects.
3
1. Introduction
Disability benefits provide an essential safety net for many people of working age whose
health means that they are unable to carry out paid work. On the other hand disability benefits
may themselves contribute to low participation and employment rates among people with
disability (e.g. Parsons, 1980; Haveman and Wolfe, 1984; Bound, 1989; Autor and Duggan,
2003; Maestas et al., 2013). High and/or growing disability benefit recipiency rates in many
countries have also led to concerns about their fiscal sustainability (e.g. Autor and Duggan,
2006; Burkhauser et al., 2014).
The more complete our understanding of the relationships between disability, benefit receipt
and employment, the better able we will be to design effective policy concerning these trade-
offs. This paper addresses part of this wider picture concerning the relationship between
disability and benefit receipt. Specifically, in a parallel with the strand of the literature that
goes beyond cross-sectional analysis to examine the dynamic relationships between disability
and labour market outcomes (e.g. Charles, 2003; Jenkins and Rigg, 2004; Mok et al., 2008;
Garcia-Gomez, 2011; Meyer and Mok, 2013; Singleton, 2014; Polidano and Vu, 2015), this
paper uses previously unexploited British longitudinal data to examine the dynamic
relationships between disability and benefit receipt. Like this earlier literature, our focus on
dynamics is motivated by the dynamic nature of both disability (people flow into (experience
onset of) and out of (experience exit from) disability) and benefit receipt, leading to the
question: How does benefit receipt evolve with respect to changes in disability status? We
know time matters. For example, international experience of disability benefit reform
suggests that it is in the period immediately following disability onset that there is most scope
to constrain the growth of disability benefit rolls (by limiting inflows), and to support
employment among people with disability (Burkhasuer et al., 2014). Exploiting longitudinal
data also gives us additional tools for dealing with the endogeneity of self-reported disability.
We begin by examining the impact of disability onset on a set of four benefit receipt
outcomes ranging from the main income-replacement disability benefit – the UK version of
Disability Insurance (DI) known as Employment and Support Allowance (ESA) – to receipt
of any non-universal welfare payment. In doing so we build on the earlier work of Jenkins
and Rigg (2004) which finds a positive impact of disability onset on income from own
disability benefits and from other welfare benefits at the household level in both the onset
4
year and the year after onset using British Household Panel Study (BHPS) data from 1991-
1998. In particular the period we study (2004-2012) follows or spans major reforms to
disability benefits in 1995 and 2008 and to unemployment benefits in 1996 which all
impacted disability benefit rolls (e.g. see McVicar, 2008; Banks et al., 2015). For the US
Singleton (2014) finds a positive impact of disability onset on DI receipt in the onset year and
the seven subsequent years using data from the Survey of Income and Program Participation.
Polidano and Vu (2015) also find a positive impact of disability onset on receipt of any
income replacement welfare payment in the onset year and the four subsequent years using
Household Income and Labour Dynamics in Australia (HILDA) survey data. Looking beyond
DI receipt or catchall measures of benefit receipt is important given that many working age
people with disability are not in receipt of disability benefits but may be in receipt of other
benefits (e.g. for the US see Meyer and Mok, 2013)1, and because flows onto and off
disability benefits are not only from and to employment but also from and to other benefit
payments including unemployment insurance (e.g. for the UK see Sissons et al., 2011; Beatty
and Fothergill, 2015). Taken together this suggests even the sign of disability onset effects on
receipt of welfare payments other than sickness and disability benefits is uncertain ex ante.
Our second contribution is that we are able to examine heterogeneous impacts of disability
onset on benefit outcomes across several dimensions because we have a larger sample, with
more disability onsets, than is typical in this literature. To date what we know in this regard is
limited to differences in DI application and receipt by severity of disability onset from
Singleton (2014) (those experiencing onset of work-preventing disability are much more
likely to apply for and receive DI than those experiencing onset of work-limiting disability)
and differences in receipt of any Income Support payment by broad education level from
Polidano and Vu (2015) (those with no qualifications are much more likely to receive Income
Support than those with vocational or higher-level qualifications). We re-examine both these
dimensions here using British data. There is more existing evidence on heterogeneous
impacts of disability onset on employment and other labour market outcomes. Polidano and
Vu (2015) finds variation by pre-onset employment status (impacts on employment are due
more to reduced inflows than to increased outflows from work) and by education level (larger
impacts for lower educated individuals). Exploiting the same Local Labour Force Survey
(LLFS) data we use here, Jones et al. (2013) finds stronger employment effects of disability
1Meyer and Mok (2013) present descriptive data from the Panel Study of Income Dynamics on recipiency rates
for various welfare benefit payments, including DI, between six and ten years after disability onset. They do not
present estimates of the impact of disability onset on these outcomes, however.
5
onset for men, older individuals, those with more severe disability (proxied by multiple
conditions), but little difference by education level. Again we examine differences in
disability onset effects on benefit receipt along all these dimensions. A cross-country study
by Garcia-Gomez (2011) also suggests there are bigger employment impacts of disability
onset and negative health shocks in countries where disability (and other) benefits are more
generous, where they are conditioned on not working, and where employers do not have to
meet disability employment quotas. We too consider the institutional context for disability
onset, but rather than a cross-country approach we do so by examining variation in outcomes
within Britain either side of a major disability benefit reform introduced in 2008. Finally,
given evidence that disability rolls tend to be higher where and when labour markets are
weaker (e.g. Black et al., 2002; Autor and Duggan, 2003; McVicar, 2006), we also examine
whether the impacts of disability onset vary with local unemployment rates.
Our third contribution is to explicitly examine the impact of disability exit – no longer
reporting a disability – on benefit recipiency. This has been mostly overlooked by the
dynamics of disability literature, despite the fact that for many people disability is temporary
not permanent (e.g. see Burchardt, 2000; Meyer and Mok, 2013). One factor that may have
contributed to this is the perception that disability benefits are essentially an absorbing state
until either death or state pension age is reached, although this appears less the case in some
countries than others (see OECD, 2010). Further, even temporary disability can have long-
lasting effects on labour market outcomes through impacts on human capital accumulation
and through state dependence (e.g. Charles, 2003; Mok et al., 2008; Oguzoglu, 2012a; Meyer
and Mok, 2013). Note that an exception to the dearth of studies on disability exit effects is
Jones et al. (2013) which finds no overall impact of disability exit on employment, but a
small positive impact for women and young people. Here we build on this to examine
disability exit impacts on our range of benefit recipiency measures. As in the case of
disability onset, we also explore the extent to which disability exit effects vary across various
dimensions including gender, age, severity, type of impairment, and labour market context,
although smaller sample sizes for disability exits limits what we learn in this case.
The remainder of this paper is set out as follows. The following section provides a brief
overview of the British welfare system, and in particular disability benefits, pre and post
2008. Section 3 describes the LLFS data we use to estimate disability onset and exit impacts.
Section 4 sets out our approach to estimation and discusses identification. Sections 5 and 6
present and discuss the resulting estimates of disability onset and disability exit impacts,
6
respectively. Section 7 concludes. The supplementary appendix presents additional data
details and the results and brief discussion of sensitivity analysis.
2. Disability and Other Benefits in Britain 2004-2012
Our data cover the period 2004-2012, and in this section we briefly describe the working-age
welfare system in Britain in place at that time, with particular emphasis on welfare payments
for people with disability. The main disability-related benefits in Britain are earnings
replacement benefits and additional costs benefits. In keeping with the disability and labour
supply literature, we concentrate primarily on earnings-replacement benefits here. These are
also the benefits that were most extensively reformed during the 2004-2012 period. The main
additional cost benefit throughout the period was called Disability Living Allowance (DLA),
and this only began to be reformed – with its gradual replacement by Personal Independence
Payments – subsequent to the period of interest here. The working-age recipiency rate for
DLA rose slowly but steadily over the period, from around 4% in 2004 to around 4.5% in
2012. Other major working-age benefit types that are covered by our broadest measure of
welfare recipiency – see Section 3 – include Jobseeker’s Allowance (JSA) (unemployment
benefit), Income Support (means tested social assistance, e.g. for single parents) and Housing
Benefit (for those with low income to help with housing costs). For further details on these
payments see Browne and Hood (2012).
From 2004 until the 27th
October 2008 the main earnings replacement disability benefit for
those unable to work on grounds of disability was called Incapacity Benefit (IB). This was a
contributory benefit, i.e. eligibility required a work history, and was (mostly) not subject to
means-testing. Incapacity for work was determined by government doctors by means of a
Personal Capability Assessment (PCA). IB was paid at one of three flat rates depending on
the length of time the individual had been unable to work: a short term lower rate for the first
28 weeks, a short term higher rate for the next 24 weeks, and a higher long-term rate
subsequently. Those who became sick or disabled while in work were generally ineligible for
IB during the first 28 weeks of a spell out of work and instead could claim Statutory Sick Pay
(SSP), for which employers were responsible.2 Those unable to meet the contributions based
eligibility criteria for IB were potentially eligible for Severe Disablement Allowance (SDA)
2 Note that unlike for DI in the US there is no mandatory waiting period for eligibility for IB (or for SSP for
those in work at the time of disability onset).
7
(although no new claims for SDA were granted after 2001) or to have their National
Insurance credits – contributions towards the state pension – paid. ‘Credits only’ claimants
usually also received Income Support, often with a ‘disability premium’. Recipients of IB,
SDA and ‘Credits Only’ (but not SSP) were collectively referred to as incapacity benefits
claimants – note the practical equivalence in British welfare-speak between claiming and
receiving benefits, as opposed to applying for benefits – and made up the incapacity benefit
roll, which stood at around 6.7% in 2004, having hovered between 6% and 7% since the mid-
1990s.
From 2003 a new set of work-first reforms called Pathways to Work (PtW), aimed at slowing
the inflow to IB and boosting outflows for those having recently joined the roll, was
gradually rolled-out. It made movement onto the IB program (including credits only)
conditional on attendance at work-focused interviews, with the aim of steering at least some
recipients into employment support services and ultimately back into the labour market. It
also introduced a ‘back to work’ bonus payment, provided additional in-work condition-
management health support for those returning to employment from IB, and brought PCAs
forward so they took place three months into the IB claim rather than six months into the
claim. Evaluation evidence on the impacts of PtW has been mixed (see Adam et al., 2010;
National Audit Office, 2010), although the IB claimant rate fell steadily between 2004 and
2008 to around 6.2%. The unemployment rate hovered around 5% over this period.
In 2008, ESA replaced IB (and credits only IB) as the main earnings-replacement disability
benefit for new applicants. This new program of insurance-based benefit for those with
sufficient work history and means-tested social assistance benefit for those without sufficient
work history included a new tougher Work Capability Assessment (WCA), with fewer
exemptions, in place of the existing PCA. The requirement to attend work-focused interviews
introduced under PtW was extended into a requirement to engage in work-related activity for
all but the most severely disabled, linked explicitly to payments, with around one quarter of
the existing benefit payment made conditional upon compliance. There was also no longer a
higher rate of payment for longer-duration claims. Further, from April 2011 existing IB
recipients started to be reassessed under the new ESA eligibility criteria, although this
process was still not complete at the time of writing. Many were judged ineligible as a result
of medical re-screening under the stricter WCA, although some have since successfully
appealed these decisions (see Department for Work and Pensions Quarterly Official Statistics
Bulletin December 2014). Disability recipiency rates continued to fall slowly over the years
8
from 2008-2012, reaching around 6% in 2012. Concurrently, the Great Recession led to a
rapid increase in the unemployment rate (and in the JSA claimant rate), rising to around 8%
by the second quarter of 2009, where it remained through to 2012. The fact that disability
recipiency rates did not increase during or in the years following the downturn, unlike in
earlier downturns, suggests the reforms outlined above may have impacted significantly on
flows onto and off IB/ESA. For more details of this reform and early estimates of its impacts
see Banks et al. (2015).
3. Data
The LLFS data have a number of desirable properties for our purposes. In particular, we are
able to draw on a large sample with sufficient numbers of disability onsets and exits to enable
examination of heterogeneous effects. The data also span a major disability benefit reform –
the switch from IB to ESA from 2008 for new applicants – which allows us to examine the
impacts of disability onset on benefit receipt under different benefit regimes within the same
country. The trade-offs are that the LLFS offers a relatively short longitudinal dimension with
respondents observed for a maximum of four years (so we cannot directly examine longer
term impacts of disability onset/exit), and that the longitudinal sample we end up using is not
fully representative of the wider working age population. The latter should be borne in mind
when drawing conclusions from the analysis presented here.
The LLFS is part of the Annual Population Survey (APS) which also contains observations
from the main Quarterly Labour Force Survey (QLFS) and the APS boost. Special Licence
LLFS data from the January to December APS are pooled from 2004 to 2012. We focus on
the LLFS because individuals are retained for four years with a 25% rotational panel element,
rather than the standard five quarters for the QLFS. We follow Jones et al. (2013) and use the
system variables, employed in the construction of the LLFS, to undertake a matching process
of individuals across time to construct a panel version of the LLFS. The LLFS covers Great
Britain but, as discussed by Jones et al. (2013), since it was designed to boost the sample size
of the main QLFS it is not geographically representative, although this has a limited effect on
the sample composition in terms of personal characteristics, albeit there is a slightly higher
proportion reporting disability and benefit receipt, and lower proportions reporting
employment than in the full APS sample pooled over the same period (see Table A1 in the
appendix).
9
We restrict our analysis to respondents who provide valid information at four consecutive
waves between 2004 and 2012, creating a balanced panel, and who are of working age
throughout. We are left with a maximum sample of 49,071 individuals (196,284 person year
observations). Note that the LFS is not primarily designed as a panel survey and it is the
address rather than the individual that is traced across time. As a consequence observations in
the LLFS panel are restricted to households that did not move address and individuals who
remained resident within these households for four consecutive years. Our sample therefore
excludes individuals who experience disability onset/exit which is associated with selective
residential mobility, e.g. for formal or informal care purposes (Norman et al., 2005). These
are likely to be the most severe onsets or greatest recoveries. However, since migration more
generally is dominated by young and healthy individuals, attrition increases the prevalence of
disability in the LLFS panel relative to the unrestricted pooled LLFS by about two percentage
points. Overall, compared to the full APS sample, the balanced LLFS panel has slightly
higher rates of disability and disability benefit claiming, lower rates of non-sickness benefit
claiming, is slightly older, has fewer full-time students and singles, and more renters, in
addition to the aforementioned differences in geographical coverage (see Table A1).
Disability
The LLFS contains self-reported information on two alternative definitions of disability. Both
require a positive answer to an initial question on long-term health: “Do you have any health
problems or disabilities that you expect will last for more than a year?” This is then followed
by a series of questions: (i) “Does this health problem affect the kind of paid work you might
do? Does this health problem affect the amount of paid work you might do?” (ii) “Do these
health problems or disabilities, when taken singly or together, substantially limit your ability
to carry out normal day to day activities? If you are receiving medication or treatment,
please consider what the situation would be without the medication or treatment.” which
refer to work-limiting (WL) and Disability Discrimination Act (DDA) definitions of
disability, respectively (see Jones et al., 2006 for details). Individuals answering ‘no’ to the
first question on long-term health, or those answering ‘yes’ to the first question but ‘no’ to
the two follow-up questions, are classed here as non-disabled. The prevalence of WL and
DDA disability in the balanced panel is 17.48% and 18.11% respectively, with considerable
overlap. In what follows we present separate analysis for each definition, reporting analysis
10
using the DDA definition in the main text for reasons which we set out below and analysis
using the WL definition in the Appendix.3
In relying on self-reported disability we acknowledge the possibility that measurement error
may bias estimated onset and exit effects towards zero. By focussing on consistent patterns of
disability reporting facilitated by longitudinal data, however, we reduce the scope for
measurement error compared to cross-sectional approaches. Specifically, we follow Jenkins
and Rigg (2004), Polidano and Vu (2015) and others by using two-period measures of
disability onset and exit where we define the onset group as those who experience two
periods reporting no disability followed by two periods of reporting disability (0011), and
those who experience consistent exit as those who report two periods of disability followed
by two periods of no disability (1100). For each we specify two alternative control groups.
Our first control group for onsetters are those who are continuously non-disabled (0000), i.e.
those at risk of onset who do not report onset. Similarly, for the exit group our first control
group are continuously disabled (1111), i.e. those at risk of exit who do not exit. Our second
control group in each case are those who do experience disability onset but one year later (i.e.
0001) and those who do experience disability exit but one year later (i.e. 1110). Table 1
provides the sample sizes for the various treatment and control groups, separately for the WL
and DDA definitions of disability. Note that using these definitions within our balanced
panel, onset or exit can occur at any time between 2006 and 2011, and there is a broadly
equal distribution of such events across this six-year window. Also note we observe fewer
disability exits than disability onsets.
Another potential issue with self-reported disability is justification bias – for a given degree
of disability, benefit recipients may be more likely than others to report themselves as
disabled – which may impart biases to our estimated onset and exit effects in the opposite
direction. The extent to which this is economically important, however, is not clear. Benitez-
Silva et al. (2004), for example, find that self-reported disability status is an unbiased
indicator of DI eligibility decisions. Bound (1991) suggests that justification bias may even
help to cancel out biases due to measurement error. Meyer and Mok (2013) also argue that
some alternative (more objective) measures may themselves be endogenous and are often too
3 The ONS has recently highlighted a discontinuity in the measures of disability in the LFS between 2009 and
2010. This relates to a minor change in the administration of the questionnaire where “I should now like to ask
you a few questions about your health. These questions will help us estimate the number of people in the
country who have health problems” was added to the survey. It is, however, thought to have increased the
prevalence of disability by about 1.5 percentage points.
11
narrow, for example excluding conditions such as mental illness or pain which have no
physical marker (also see Benitez-Silva et al., 2004). Like Meyer and Mok (2013), we are
constrained by the absence of objective measures of disability in the LLFS which might
otherwise be used in place of, or to instrument, self-reported measures (e.g. see Disney et al.,
2006; Garcia-Gomez et al., 2013). But we do have a choice of self-reported definitions, and
our focus on the DDA definition rather than the WL definition of disability is intended to
reduce the potential for justification bias (see Oguzoglu, 2012b). We discuss potential biases
further in Section 4.
Benefit Receipt
All respondents are initially asked whether, in the reference week, they claimed any State
Benefits or Tax Credits (including State Pension, Allowances, Child Benefits and National
Insurance Credits). Those who respond positively are then asked ‘Which of the following
type of benefit or Tax Credits were you claiming?’ and are given a long list of options,
including ‘Sickness or disability benefits’; ‘Unemployment related benefits’; ‘Income
Support’; ‘State Pension’; ‘Family related benefits (excluding Child Benefit)’; ‘Child
Benefit’; ‘Housing/Council Tax rebate’; ‘Other’.4 We first generate a binary variable for
claiming any of the benefits listed excluding those who report only universal benefits (Child
Benefit, State Pension or both). This is our broadest, ‘any benefit’, measure and is reported
by 14.74% of the sample. We subsequently generate two narrower binary measures for
‘sickness benefit’ and ‘non-sickness benefit’. The latter is equal to 1 for those who report
‘Unemployment related benefits’, ‘Income Support’, ‘Family related benefits (excluding
Child Benefit)’, ‘Housing/Council Tax rebate’, ‘Other’, and is zero otherwise.
Those in receipt of ‘Sickness or disability benefits’ are asked to list which type of benefit
they claim and the responses include: ‘Incapacity Benefit’; ‘Severe Disablement Allowance’;
‘Statutory Sick Pay’; ‘Disability Living Allowance’; ‘Attendance Allowance’; ‘Industrial
Injuries Disablement Benefit’; and (from 2009) ‘Employment and Support Allowance’. We
exclude ‘Invalid Care Allowance’ (also reported) from our measure of sickness benefit since
this is claimed by a carer and not on the basis of own disability. In other words, our ‘sickness
benefit’ measure covers both income-replacement and additional costs benefits (the latter not
conditioned on being out of work). We also use this additional information to create a final
4 ‘Tax Credits’ is also listed among the full set of options but this is excluded from the measures available in the
APS datasets.
12
narrower measure of receipt of IB or ESA which is equal to 1 for those who report receipt of
either benefit but 0 otherwise, and is reported by 4.92% of our sample. The percentage of the
working age population claiming each of the four measures is traced from 2004 to 2012 in
Figure 1. Note the decline in IB and rise in ESA recipients following the 2008 reforms, but
the overall stability of the series for IB or ESA and for sickness benefits, including through
the years of and following the Great Recession. In contrast, note the increase in receipt of
non-sickness benefits (driven primarily by increases in JSA receipt), and as a result in the any
benefit measure, from 2008 to 2009.
Table 2 reports sample proportions in receipt of benefits according to each of the four
measures by wave, split into (DDA) disability onset/exit treatment and control groups as
defined above. Note the stability of these sample proportions throughout for the main control
groups (0000 for onset and 1111 for exit), in contrast to increased (decreased) benefit receipt
for the onset (exit) treatment groups. Also note that for some outcomes benefit receipt rises
(falls) ahead of onset (exit) for the treatment group; although the biggest increase (decrease)
tends to be in the onset (exit) year. The alternative control groups are more similar to the
treatment groups in this respect. We return to this point in the following section. Further, note
that the uptake in IB/ESA for those reporting disability onset is modest; most of those
experiencing disability onset do not receive these benefits in the onset year or the year
following onset. Singleton (2014) similarly shows a relatively modest uptake of DI in the US
for those experiencing WL disability onset. Table 3 reports the proportion of the pooled
2004-2012 LLFS sample in receipt of benefits by employment status. Note that only half (one
quarter) of those describing themselves as DDA disabled receive any non-universal benefit
(IB/ESA). For non-employed DDA disabled – just over half of those reporting themselves as
DDA disabled – the corresponding proportions are 80% and 40% respectively.5
As in the disability case, the fact that these benefit receipt measures are self-reported means
we cannot rule out measurement error in our outcome variables. This seems most likely in
responses to questions about specific benefit types, so may affect the narrower measures
more than the broader ones. Note that the LLFS data for the narrowest IB/ESA measure track
the corresponding administrative data for the actual benefit roll very well over the 2004-2012
period, but at nearly two percentage points lower in all years (see Figure A1). The likeliest
explanation is that IB/ESA Credits Only recipients don’t tick the IB/ESA box because they
5 Tables A2 and A3 present the WL equivalents.
13
don’t consider themselves IB/ESA recipients, although they are counted as such in the
administrative data. We may therefore underestimate both disability onset and exit effects on
IB/ESA receipt. For the broader definitions of benefit receipt, it seems reasonable to assume
that measurement error is random, in which case the magnitude of our estimated onset/exit
impacts will be unaffected although they may lose precision.
Control Variables
Like the QLFS the LLFS contains detailed information on personal and employment-related
characteristics using established definitions, measured consistently over time. In what
follows, we condition on a wide set of explanatory variables measured in wave 1, that is, two
years prior to onset (exit) for the treatment group. Following Polidano and Vu (2015) these
variables include age (age squared), gender, highest educational qualification, region of
residence, marital status, presence of dependent children under 16 in the household,
employment status, full-time student status, housing tenure and the local unemployment rate.
In order to further mitigate potential concerns over non-random selection into disability onset
or exit status, we also condition on benefit status in wave 1 and, given individuals may
experience onset in different years, on year first included in the survey to control for time
period effects.
As discussed above, the LLFS does not contain objective information on health. We are able,
however, to examine the sensitivity of our results not only to using different self-reported
measures of disability, but also to the inclusion of measures of self-reported health as
controls. Specifically, we examine sensitivity to conditioning on reporting a long-term health
condition in wave 1 and to reporting prior long-term health problems in wave 1. The relevant
question for the latter is: ‘Have you ever had any health problems or disabilities (apart from
the ones you have just told me about) that have lasted longer than one year?’ Note that the
sample size is reduced when conditioning on prior health because the relevant question is not
asked in proxy interviews.
We are also able to explore heterogeneity in the onset and exit effects by splitting our sample
by gender, age (older and younger defined as above or below age 45), highest qualification
(higher and lower qualifications defined as above or below GCSE grade C), disability type
(main health problem is classified as physical or mental condition at onset), disability severity
(single or multiple health conditions reported at onset), benefit receipt in wave 1, local
economic conditions (proxied by quartiles of the unemployment rate) and onset pre and post
14
ESA. Note that the ESA option, although introduced on 27th
October 2008, was only included
in the questionnaire from 2009 onwards.
4. Approach to Estimation and Identification Issues
We want to estimate the impact of disability onset (exit) on a set of benefit receipt outcomes.
We cannot design a random experiment for this purpose, so we have to use observational data
and rely on econometrics to separate out treatment effects from differences in outcomes due
to selection (observable and unobservable differences between those experiencing onset/exit
and others, some of which may be time-invariant and some of which may be time-varying,
and which are themselves associated with benefit outcomes) and reverse causality (benefit
receipt may impact on disability as well as vice versa). Jenkins and Rigg (2004) showed that
individuals who experience disability onset tend to be more disadvantaged prior to onset than
those at risk of but who do not experience disability onset. The opposite is likely to be the
case for disability exit, i.e. exiters are likely to be less disadvantaged than those at risk of but
who do not experience disability exit. One potential mechanism for reverse causality is that
benefit receipt, or more specifically inactivity associated with benefit receipt, directly leads to
deteriorating health (e.g. Lindeboom and Kerkhofs, 2009; Deb et al., 2011; Coleman and
Dave, 2014). Another is justification bias outlined above: that for a given degree of disability,
benefit recipients may be more likely than non-recipients to report disability in order to
rationalise or justify their benefit status.
Following earlier papers including Garcia-Gomez (2011) and Polidano and Vu (2015), we
start by taking a propensity-score-matching (PSM) approach6 to identify disability onset
(exit) impacts separately from compositional differences between those experiencing onset
(exit) and those at risk but not experiencing onset (exit), under a standard conditional
independence assumption (CIA) (see Rosenbaum and Rubin, 1983). Specifically, we take the
sample of DDA 0011s and 0000s for onset (or 1100s and 1111s for exit) and match exactly
on receipt of any benefit in wave 1 and by year in wave 1 before estimating a probit model
for treatment status (disability onset/exit) regressed on an extensive set of wave 1 observables
as set out in the previous section (also see Table 4). For each individual experiencing onset
(exit) we then find the individual with the most similar probability of experiencing onset
(exit) between wave 2 and 3 given their characteristics but who did not do so. Note that
6 This is implemented using the Stata command psmatch2 (see Leuven and Sianesi, 2003).
15
before matching large and statistically significant differences between the treatment and
control groups exist in each case. After matching these differences in sample means disappear
(see Table 4).7 This also holds when using the WL definition of disability in place of the
DDA definition, and when using the alternative control groups (0001 and 1110) in place of
the standard control groups (see Tables A4-A6). Calculating how the treated individuals’
outcomes differ from their matched partners’ outcomes, and averaging these differences over
all treated individuals, yields our initial estimates of the impact of disability onset (exit) on
those who experience it.8 If the CIA holds this is interpretable as the average treatment effect
on the treated, or ATT. We report these differences separately for each of the two years
preceding onset (exit), for the year coinciding with onset (exit), and for the year following
onset (exit). The timing reflects the short LLFS panel and while we are unable to identify
longer-term trends we capture the period when the impact of onset is most pronounced (Mok
et al., 2008; Meyer and Mok, 2013).
How plausible is the (untestable) CIA in this case? Like Garcia-Gomez (2011) and Polidano
and Vu (2015) we can point to the availability of detailed individual characteristics on which
we match. Nevertheless, unobserved confounding factors are still likely to remain. To the
extent that such unobserved confounders are time-invariant (e.g. preferences, history),
however, we can recover the ATT by either exact matching on wave 1 outcomes (following
Garcia-Gomez, 2011) or by combining PSM with difference-in-differences (DID) (following
Garcia-Gomez et al., 2013; Polidano and Vu, 2015).9 Specifically, we difference the
differences in outcomes between the treatment and control groups between waves 1 and 2
(two and one years prior to onset/exit), waves 1 and 3 (two years prior to and the year of
onset/exit), and waves 1 and 4 (two years prior to and the year following onset/exit). Our
DID-PSM estimate of the ATT between wave 1 and the year of onset (exit) is therefore the
difference in differences between wave 1 and wave 3 outcomes for the matched sample. The
DID between wave 2 and 3 outcomes – not examined by Polidano and Vu (2015) – gives our
DID-PSM estimate of the ATT specifically for the onset (exit) year.10
7 There is one exception in each case: a higher proportion of the onset control group are from Rest of Yorkshire
& Humberside (significant at 95%) and a higher proportion of exit control group are from the East of England
(significant at 90%). 8 The analysis is performed over individuals within the region of common support (that is, where there is
overlap between propensity scores within the treatment and control group), so it’s not quite all treated
individuals. 9 This approach was originally proposed by Heckman et al. (1997).
10 An alternative is to additionally exact match on wave 2 outcomes and examine the difference in wave 3
outcomes.
16
Formally, the DID-PSM estimator can be expressed as:
where n denotes the number of individuals within the treatment group which are each denoted
l, whereas those in the control group are denoted j. The time period pre and post treatment is
denoted here as t’ and t respectively, and )( 'ltlt YY measures the change in benefit receipt
between wave 1 (or wave 2) and subsequent waves for treatment individual l. The change in
outcomes in the matched control is generated by weighting the difference in outcomes across
individuals j )( 'jtjt YY .11
Standard errors are calculated following Abadie and Imbens (2006)
and take into account that the propensity scores are estimated.
The DID-PSM approach outlined above, although it partially relaxes the CIA, rests on the
additional assumption of parallel trends between treatment and control groups. This is also
untestable but standard practice is to examine trends prior to treatment to provide some
indication of the assumption’s reasonableness. Given our definition of onset (exit) and the
four-wave length of the panel, our pre-treatment information is limited to waves 1 and 2 in
each case. Nevertheless we can still learn something from these data. We know from Table 2
that benefit receipt in the treatment groups tends to diverge from benefit receipt in the
(standard) pre-match control groups prior to onset/exit, i.e. between waves 1 and 2. The DID-
PSM estimate for the difference in difference in outcomes between waves 1 and 2 (defined
above) gives the post-match equivalent and, from one perspective, can therefore be
interpreted as a test of diverging prior trends. This interpretation assumes that any divergence
in outcomes prior to onset/exit is not itself part of the treatment effect of onset/exit, but
reflects selection on time-varying confounders (e.g. risk of job loss unrelated to health). For
‘shock’ treatments like acute hospitalizations (Lindeboom et al., 2006; Garcia-Gomez et al.,
2013) or heart attacks (Trevisan and Zantomio 2016) this is the most natural interpretation of
pre-treatment divergence in outcomes. For self-reported disability, however, pre-onset (exit)
divergence in outcomes may also reflect a gradual deterioration (improvement) in health
which occurs prior to reporting disability, in which case divergence in outcomes prior to
11
Although we report estimates based on NN1 matching in the main text, we also test sensitivity to other
commonly-used variants of the PSM matching method (see Table A13). The trade-off in selecting a matching
method is that the more neighbours one uses as matching partners, or the wider the gap in propensity scores one
allows, the more likely it is that the resulting estimate is biased because individuals who are not very similar to
each other are being compared; the fewer neighbours one uses or the smaller the gap in propensity scores one
allows, the less precisely the ATT will be estimated.
1l 1j
jtjtltlt YYjlWYYn
DiD ))(,()(1
''
17
reported disability onset (exit) may be more appropriately interpreted as part of the overall
treatment effect (Meyer and Mok, 2013). This may be even more the case for disability exit if
recovery is more gradual and labour market adjustment occurs prior to rather than after
disability exit (see Jones et al., 2013).
More generally this agnosticism about pre-onset effects is common to both PSM-type papers
(Polidano and Vu, 2015) and the individual fixed effects / event study papers of Mok et al.
(2008), Meyer and Mok (2013), Jones et al. (2013) and Singleton (2014), although this is not
always made explicit. Two exceptions are: Charles (2003), which takes an event study
approach but controls for differences in outcome trends at the individual level, albeit
specified over a long time horizon (10 years pre and post reported onset); and Garcia-Gomez
(2011) which exact matches on onset-year outcomes (see below). We adopt the treatment
effects approach here, but we also follow the event study papers in clearly setting out pre-
onset divergence in outcomes and presenting a range of estimates which recover the ATT
under different assumptions about pre-onset (exit) divergence. Specifically, if we assume that
pre-onset (exit) divergence following wave 1 is itself part of the treatment effect, then the
ATT corresponds to the DID-PSM estimates differenced between waves 1 and 3 (the onset
year) and waves 1 and 4 (the year following onset).12
In contrast if we assume pre-onset (exit)
divergence following wave 1 reflects confounding time-varying heterogeneity, then we have
a choice of two DID-PSM estimates depending on whether we assume the diverging trends
will continue beyond onset (exit) or stop at onset (exit). In the former case we can interpret
the DID-PSM estimate which double differences the difference in outcomes ((wave 3-wave
2)-(wave 2-wave 1)) as a conservative estimate of the ATT under the parallel growth
assumption (see Mora and Reggio, 2012).13
In the latter case the difference in differences
between wave 2 and 3 outcomes gives the ATT for the onset year.14
It is unclear (and is
untestable) which assumption is more reasonable, and we present both estimates in what
follows.15
12
In line with existing studies who find no evidence of pre-onset labour market adjustment more than one year
prior to onset (Charles, 2003, Jones et al., 2013) we use two years prior as our base period to maximise the panel
element of analysis. 13
We assume the trends are linear given we only have two waves of data pre-onset (exit). 14
In both cases corresponding estimates can be recovered for outcomes at wave 4. 15
One possible example of a time-varying confounding factor which could lead to ongoing divergence in
outcomes could be ongoing divergence in local labour markets throughout the 2004-2012 period, given the
association between local labour market tightness and benefit recipiency rates. An example of a time-varying
confounder which would likely not lead to ongoing divergence in outcomes, other than through impacts on
disability, could be job loss unrelated to health.
18
Table 2 shows that the standard parallel trends assumption looks much more defensible for
our alternative control groups for onset (0001) and exit (1110), i.e. those who do experience
onset or exit but one year later than those in the corresponding treatment groups. This is
confirmed by the post-match results presented in the following section. We therefore also
present DID-PSM estimates using these alternative control groups. In this case the DID-PSM
estimate of the ATT for the onset (exit) year is the difference in differences between wave 1
and wave 3 outcomes for the matched sample, which under the assumption of parallel trends
is equivalent to the difference in differences between wave 2 and 3 outcomes. The trade-off
in this case is increased chance of measurement error (classification into the control group is
based on disability onset/exit observed in a single year), that we wash out effects of gradual
deterioration (improvement) in health associated with but occurring in advance of disability
onset (exit), and the fact that it is not meaningful to consider differences in outcomes in wave
4.
Whichever estimator from the above discussion is considered, reverse causality within
onset/exit year remains a potential threat to identification. Specifically, given the annual wave
structure of the LLFS, we know only that onset / exit takes place between wave 2 plus one
day, and wave 3. Similarly, we do not know the benefit inflow date for those not receiving
benefits at wave 2 but receiving benefits at wave 3, and similarly for benefit outflows. If there
is reverse causality, either directly or as a result of justification bias, then if benefit receipt
(exit) precedes disability onset (exit) within year, we would expect estimated treatment
effects to be biased upwards in magnitude. This is not always explicitly discussed in the
papers cited above. An exception is Garcia-Gomez (2011) which uses exact matching on
onset year outcomes and only examines outcomes in the year after onset. The trade-off is that
you learn nothing of onset-year effects, and because you drop those with onset-year effects
the results end up being less generalizable.16
We report estimates from a Garcia-Gomez
(2011) style exact matching approach in the appendix. In the main text, however, we report
estimates under the assumption that any such reverse causality is negligible. In defence of
this assumption we are using a definition of disability (DDA, consistent onset/exit over two
years) which will reduce the scope for justification bias (see Charles 2003 on the latter point)
and we are looking at outcomes over a relatively short duration during which large direct
reverse causality effects are unlikely to have accumulated. Exact matching on benefit receipt
in wave 1 helps to reduce the scope for reverse causality other than that arsing within year.
16
In our application we would also need to match separately for each outcome.
19
Finally, in order to examine whether and how disability onset/exit impacts differently across
different groups of individuals and in different labour market and policy contexts, we split the
sample along a number of different dimensions prior to the matching procedure, using the
standard 0011 and 1100 control groups, and repeat. For conciseness we present just two
estimates in each case: the DID-PSM estimates for wave 3 minus wave 1 and for wave 3
minus wave 2 which include and exclude pre-treatment effects respectively. Likewise we do
not present separate balancing tests or separate discussion of identifying assumptions for the
estimates of heterogeneous effects.
5. The Benefit Recipiency Impacts of Disability Onset
Post-matching estimates of DDA disability onset treatment effects are reported in Table 5.
For each of the benefit measures we first present PSM estimates for the difference between
treatment and control group recipiency rates in each wave. Under the standard CIA, and
assuming negligible reverse causality within year, the wave 3 estimates give the ATTs in the
year of onset and the wave 4 estimates give the ATTs one year on. For all four outcome
measures the wave 3 estimates suggest large positive disability onset effects in the year of
onset, all of which are statistically significant at the 99% level. For IB/ESA and our more
general sickness benefits measures, those experiencing disability onset have wave 3 benefit
recipiency rates of 6.4% and 9.7% respectively, against control group recipiency rates that are
essentially zero or close to zero. For the any benefit measure those experiencing disability
onset have a recipiency rate almost double that for the controls (20.9% versus 11.5%). For
non-sickness benefits the difference (3.7%) is much smaller both in absolute magnitude and
relative to the wave 3 recipiency rate for the control group (10.2%), but still shows that
disability onset is associated with increased flows onto or decreased flows off non-sickness
benefits. Where we have outcome measures in common our findings are qualitatively
consistent with those of Jenkins and Rigg (2004), Singleton (2014) and Polidano and Vu
(2015), with effects of the same order of magnitude. The wave 4 estimates show further flows
onto IB/ESA and more generally onto sickness benefits over the subsequent year, but no
further change in overall benefit recipiency or receipt of non-sickness benefits. In other words
the additional inflows to disability benefits come from within the benefit system rather than
from outside the benefit system. That people with disability in Britain (and elsewhere)
sometimes move on to disability benefits following a spell on other benefits is well known
(e.g. Beatty et al, 2000; Sissons et al., 2011; Beatty and Fothergill, 2015). What we show
20
here, however, is that in the year after the onset year itself, all net flows onto disability
benefits appear to come from existing benefit claimants.
For one of the four outcomes (IB/ESA) there is a statistically significant difference between
the treatment and control groups even in wave 1, which makes us question the standard CIA.
Therefore Table 5 also presents the range of DID-PSM estimates discussed in the previous
section, under the weaker CIA assumption. First consider the difference between wave 1 and
2 recipiency rates. For all four outcome measures the gap between the treatments and controls
grows between waves 1 and 2, although the magnitude of this effect is small and in only one
case – sickness benefits – is this divergence statistically significant at the 95% level. The lack
of evidence for diverging trends suggests we can interpret the DID-PSM estimates
differencing waves 3 and 1 and waves 4 and 1 as ATTs, subject to the weaker CIA and
assuming negligible within-year reverse causality, at least for three of the four benefit
measures. In the light of this we interpret the pre-onset divergence for sickness benefits as
reflecting the effects of declining health in advance of reported disability, e.g. on receipt of
SSP, in which case we can similarly interpret the DID-PSM estimates differencing waves 3
and 1 and waves 4 and 1 as ATTs. For three of the four measures these estimates show large
and highly statistically significant treatment effects in the year of onset, increasing benefit
recipiency by 9.3 percentage points (any benefits), 8 percentage points (sickness benefits) and
5 percentage points (IB/ESA) respectively. The partial exception is for receipt of non-
sickness benefits, on which the impact of disability onset is relatively small (+3.8 percentage
points), although still statistically significant at 99%. The positive sign suggests that net flows
onto non-sickness benefits associated with disability onset exceed flows from non-sickness to
sickness benefits associated with disability onset. The DID-PSM estimates for wave 4-1 are
in line with the straight PSM estimates discussed above.
In terms of robustness, we still get positive, large (for three out of the four measures) and
statistically significant treatment effect estimates if we assume any divergence between
waves 1 and 2 is the result of time-varying confounders that do not drive ongoing divergence
post-onset (DID-PSM wave 3-2), although these estimates are generally smaller in
magnitude. If we assume all prior divergence is driven by confounders and that this
divergence would continue linearly post-onset (DID-PSM waves ((3-2)-(2-1)), then for all
four measures we obtain estimated treatment effects with positive signs, although they are
further reduced in magnitude than is the case under the more lenient assumptions, and only
for IB/ESA is the effect statistically significant at 95%.
21
If instead we compare the onset treatment group to the alternative control group then we can
reject diverging prior trends for all four outcome measures17
, and three of the four DID-PSM
estimates differencing wave 3 and 1 suggest positive, moderately sized or large treatment
effects which are statistically significant at 95% (see Table 6). The exception is non-sickness
benefits, for which the estimated treatment effect is zero according to all estimates, perhaps
casting some doubt on our earlier interpretation that net flows onto non-sickness benefits
associated with disability onset exceed flows from non-sickness to sickness benefits
associated with disability onset. Estimates are also robust to matching method and
additionally matching on prior health (see Tables A13 and A14).
To summarise, the range of estimates under the various assumptions discussed above suggest
that disability onset increases the probability of receipt of any benefit in the year of onset by
between 4.1 and 9.3 percentage points; receipt of any sickness benefit by between 3.0 and 8.0
percentage points; and receipt of IB/ESA by between 3.4 and 5.0 percentage points. Impacts
on non-sickness benefits are smaller and only significant under certain assumptions.
For the WL definition of disability, which is likely to be more closely linked with work-
contingent benefit outcomes, but also more susceptible to justification bias, all four outcome
measures show divergence between waves 1 and 2 which is statistically significant at 95%
(see Table A6). Estimated onset impacts that difference wave 3 and wave 1 outcomes are
larger partly as a result. So too are estimates that difference wave 3 and wave 2 outcomes,
and for three of the four outcomes, the estimates the difference wave ((3-2)-(2-1)) outcomes.
Again the exception is non-sickness benefits for which the DID-PSM estimate identified
under the parallel growth assumption is again zero. There are no such diverging prior trends
when the WL onset group are compared to the alternative WL control group, and in that case
estimated treatment effects are positive and statistically significant for all four benefit receipt
measures (see Table A7).
Finally, interpreting any of the estimates above as an ATT rests on the assumption that within
year reverse causality is negligible. We can get some indication of the reasonableness of this
assumption by following the approach of Garcia-Gomez (2011) and exact matching
additionally on outcomes in the onset year, thereby ensuring that disability onset precedes
any change in benefit status. Table A16 presents the resulting wave 4 PSM estimate for our
17
Again it is sickness benefits for which diverging prior trends looks more difficult to reject, consistent with
deterioration in health occurring in the year prior to reported disability onset.
22
four outcome measures, for both DDA and WL definitions of disability. For all four benefit
receipt measures the estimated onset effect remains positive, large and statistically significant
at 95%. Reverse causality does not drive our estimates.
5.1. Heterogeneous impacts of disability onset
In this section we discuss evidence for heterogeneous effects of disability onset by gender,
qualification level, age, broad disability type (mental/physical), severity (single/multiple
conditions), pre/post-2009, by local labour market unemployment rate quartile, by
employment status, by benefit status, by reporting a long-term health condition, and finally
by reporting a prior long-term health condition. Table 7 presents two estimates in each case:
the DID-PSM estimates for the change in each of the four measures of benefit between waves
1 and 3 and the equivalent estimates for the change in outcomes between waves 2 and 3. The
equivalent estimates using the WL definition of disability are reported in Table A8. The
results are robust to choice of disability definition.
We find big differences in onset effects by qualification level, type of disability, severity of
disability, and whether the individual reports a long term health condition in wave 1. We
discuss each in turn below.
First, consistent with Polidano and Vu (2015), we find a bigger impact of disability onset on
benefit recipiency for those with low qualifications than for those with higher qualifications.
A variety of factors are likely to contribute to this differential effect, including higher benefit
replacement rates for lower qualified workers (a labour supply effect), less ineligibility on
means-testing grounds e.g. related to partner income (an institutional effect), and the multiple
disadvantage of having a disability and few qualifications in labour markets at below full
employment (a labour demand effect).
Second, disability is not homogenous (Jones, 2011), and we find that the benefit impact of
onset of a mental health disability is much bigger than for the onset of a physical disability,
albeit our analysis of mental health disability onsets is based on a relatively small sample.18
That those with mental health impairments account for an increasing share of disability
benefit recipients in both the UK and elsewhere is well known (e.g. see Banks et al., 2015;
Burkhauser et al., 2014). In part this reflects increased prevalence or at least diagnosis and/or
18
That mental health impairment is more strongly associated with labour market outcomes than physical
impairments has also been reported in the wider cross-section literature on disability and labour markets (e.g.
see Kidd et al., 2000).
23
reporting of such conditions (e.g. Moncrieff and Pomerleau, 2000). Indeed, coupling the
increasing prevalence of such impairments with their larger benefit onset effects has
significant implications for disability benefit roll growth. We also find bigger onset effects on
benefit recipiency for those reporting multiple conditions than for those reporting a single
condition, which we interpret as a proxy for disability severity. There is also a very clear
contrast in disability onset effects by the pre-onset existence of a long-term health condition,
with much bigger impacts for those not reporting a long-term health condition in wave 1 than
for those reporting such a condition. This suggests that gradual disability onset has a lesser
impact on benefit outcomes than sudden onset, perhaps because it is easier to accommodate
the gradual onset of a disabling condition in work than a sudden onset condition.
In contrast, we find only small differences in onset effects by gender, age, local
unemployment, wave 1 employment status, and the presence of a long term health condition
prior to wave 1. Although it does appear that disability onset has a bigger effect on receipt of
non-sickness benefits in high unemployment areas than in low unemployment areas, the
mostly ‘null’ finding with regards to local unemployment rates is perhaps surprising given
evidence presented elsewhere of the impact of labour market context on disability benefit
claiming (e.g. Black et al., 2002; Autor and Duggan, 2003; McVicar, 2006). Similarly, the
mostly ‘null’ finding with regards to wave 1 employment status is also perhaps surprising
given evidence of big differences in employment effects of disability onset by prior
employment status found by Polidano and Vu (2015).
Our final sample split is for onset pre and post-2009, i.e. onset under the old IB regime in
place up to the end of October 2008 compared to onset under the new ESA regime in place
subsequently. In doing so we provide a within-country parallel to the cross-country study of
Garcia-Gomez (2011) who finds bigger employment impacts of negative health shocks in
countries where disability benefits are more generous and conditioned on not working. There
are also several examples in the wider disability benefits literature where benefit reforms
reducing payments, increasing the stringency of medical screening, and/or conditioning on
work-related activity – all of which are aspects of the shift from IB to ESA – have impacted
on program growth in the desired direction (e.g. Gruber, 2000; Adam et al., 2010; Staubli,
2011; de Jong et al., 2011). Having said that there are also contrasting findings where such
reforms appear to have had little impact (e.g. Campolieti, 2004; Karlstrom et al., 2008). The
estimates presented in Table 7, however, are consistent with Garcia-Gomez (2011) and the
former set of studies, i.e. we find bigger disability onset impacts on benefit recipiency under
24
the pre-reform regime than under the post-reform regime, with the biggest differences
unsurprisingly for the sickness benefits and IB/ESA measures. There is no apparent
countervailing positive impact of the reforms on receipt of non-sickness benefits which,
consistent with Banks et al., (2015), suggests only limited displacement onto non-sickness
benefits of those who might otherwise have claimed sickness and disability benefits. Of
course the introduction of ESA approximately coincided with the Great Recession – from
2007Q4 to 2009Q1 in the UK – and the post-recession labour market through to 2012 was
slacker than the pre-recession labour market from 2004. If anything, however, this would lead
us to underestimate the impact of the reform on the benefit receipt effects of disability onset
because we would expect higher recipiency rates across all income-replacement and means-
tested benefits in the 2009-2012 period than in the 2004-2008 period.19
6. The Benefit Recipiency Impacts of Disability Exit
Post-matching estimates of DDA disability exit treatment effects are reported in Table 8. As
in the case of onset, for each of the benefit measures we first present PSM estimates for the
difference between treatment and control group recipiency rates in each wave. Under the
standard CIA, and assuming negligible within-year reverse causality, the wave 3 estimates
give the ATTs in the year of exit and the wave 4 estimates give the ATTs one year on. For
three of the four outcome measures the wave 3 estimates suggest large negative disability exit
effects in the year of exit, all of which are statistically significant at the 99% level. Contrast
this with Jones et al. (2013) which finds no impact of disability exit on employment. For
IB/ESA and our more general sickness benefits measures, those experiencing disability exit
have wave 3 benefit recipiency rates of 3.5% and 6.1% respectively, against control group
recipiency rates of 21.7% and 31.2% respectively. For the any benefit measure those
experiencing disability exit have a recipiency rate half that for the controls (19.7% versus
37.9%). For these three outcomes, the gap in recipiency between the treatment and control
groups also grows between waves 3 and 4, in each case because recipiency increases for the
control group (likely reflecting further deterioration in health and falling income from other
sources) while it stays essentially flat for the treatment group. The suggestion is that receipt
of disability benefits may not be an absorbing state in the UK for all recipients: many people
19
On the other hand, the discontinuity in disability measurement between 2009 and 2010 referred to in footnote
3 may lead us to over-estimate the difference between pre-2008 and post-2008 if the latter period includes those
with less severe disabilities under the DDA definition than was the case prior to the discontinuity.
25
recover from temporary disabilities and flow off disability and sickness benefits; although
others do remain on IB/ESA one to two years after disability exit. Note there is no apparent
impact of disability exit on the recipiency rate for non-sickness benefits, either in wave 3 or
wave 4. If disability exit leads to moves from sickness to non-sickness benefits, then this
appears to be balanced out by moves off (or fewer moves onto) non-sickness benefits.
For three of the four outcomes there is a statistically significant difference between the
treatment and control groups even in wave 1, which again makes us question the standard
CIA. Table 8 therefore also presents the range of DID-PSM estimates discussed in Section 4.
First consider the difference between wave 1 and 2 recipiency rates. For all four outcome
measures the gap between the treatments and controls grows between waves 1 and 2, and in
contrast to the onset case the magnitude of this effect is large, and in all four cases
statistically significant at the 99% level. If we are prepared to interpret this pre-exit
divergence as part of the treatment, e.g. as a result of gradually improving health which leads
to reported disability exit once some subjective threshold is met, then DID-PSM estimates
differencing waves 3 and 1 and waves 4 and 1 can be interpreted as ATTs, subject to the
weaker CIA and assuming negligible within-year reverse causality. All are negative, large
and statistically significant at 99%.
Given the extent of pre-exit divergence, however, we cannot ignore the possibility that at
least some of this divergence is being driven by time-varying confounders. This points us
away from the initial DID-PSM estimates (and the PSM estimates discussed above) towards
the estimates differencing waves 3 and 2 (if we assume any divergence between waves 1 and
2 is the result of time-varying confounders that do not drive ongoing divergence post-onset)
or the estimates differencing waves ((3-2)-(2-1)) (if we assume confounding effects would
continue linearly post-onset). Whichever of these two assumptions we make, the evidence for
disability exit effects on benefit recipiency appears fragile. The estimates differencing waves
3 and 2 are all negative – suggesting disability exit is associated with lower benefit recipiency
rates in the year of exit – but none are large in magnitude and only one (sickness benefits) is
statistically significant at 95%. The second set of estimates are all positive, suggesting a fall
in the rate of divergence in benefit recipiency rates between the treatment and control groups
following disability exit, although again only one (any benefit) is statistically significant at
95%. These results point to the potential existence of long-lasting effects of temporary
disability, e.g. through impacts on human capital accumulation and through state dependence,
in line with those argued for labour market outcomes by Charles (2003), Mok et al. (2008),
26
Oguzoglu (2012a) and Meyer and Mok (2013). They are also more in line with the Jones et
al. (2013) zero impact of disability exit on employment. Note, however, that this conclusion
stems from the particular interpretation of the pre-treatment effects as being driven by time-
varying confounders and not by gradual improvements in health.
The evidence for exit effects continues to look fragile if we compare the exit treatment group
to the alternative control group, for which we can reject diverging prior trends for all four
outcome measures. Here both the DID-PSM estimates differencing waves 3 and 1 and
differencing waves 3 and 2 suggest large, negative treatment effects, which are statistically
significant at 95%, for the any benefits and non-sickness benefits measures (see Table 9). In
contrast, estimated exit impacts on sickness benefit recipiency and on the narrower IB/ESA
measure are small, of inconsistent sign, and statistically insignificant. Even if these are under-
estimates of the impact of disability exit (because they wash out the effect of gradual
improvements in health for exiters), again there is little here to support a conclusion that
disability exit per se has any impact on disability benefit receipt. Although it is not robust, the
non-sickness benefit result suggests that the impact of disability exit on benefit receipt is not
for those on disability benefits but for those who either did not apply for such benefits, who
applied but were denied such benefits, or who had moved off such benefits in advance of
reporting disability exit. We have no way of untangling the underlying reason for this
contrast here, but it is consistent with differences in exit effects by severity of disability (we
return to this below) and/or differences in exit effects by distance from the labour market.
As for disability onset, estimates of exit effects are robust to matching method (see Table
A15). The WL definition of disability, which is closer to eligibility criteria for IB/ESA but
may also be more susceptible to justification bias, similarly suggests diverging prior trends
and zero estimates according to the DID-PSM estimates which differences waves ((3-2)-(2-
1)). DID-PSM estimates differencing waves 3 and 2, however, suggest large, negative exit
effects for all but non-sickness benefits (see Table A10) as do estimates comparing WL
exiters to the alternative control group (Table A11). Again it appears difficult to draw any
firm conclusions regarding the benefit receipt impacts of disability exit from these data.
Finally, interpreting any of the estimates above as an ATT rests on the assumption that
reverse causality, specifically within-year reverse causality, is negligible. As for onset, we
can get some indication of the reasonableness of this assumption by following the approach
of Garcia-Gomez (2011) and exact matching additionally on outcomes in the exit year,
27
thereby ensuring that disability exit precedes any change in benefit status. Table A16 presents
the resulting wave 4 PSM estimate for our four outcome measures, for both DDA and WL
definitions of disability. For all four benefit receipt measures the estimated exit effect remains
positive, large and statistically significant at 95%. Reverse causality does not drive our
estimates.
6.1. Heterogeneous impacts of disability exit
As for disability onset, we split the sample along a number of dimensions to examine
heterogeneity in the benefit recipiency impacts of disability exit. The DID-PSM estimates
differencing waves 3 and 1 and waves 3 and 2 are reported in Table 10. The equivalent
estimates using the WL definition of disability are reported in Table A12. In contrast to the
estimates of heterogeneous disability onset effects, however, many of our estimates of
disability exit effects by subgroup appear sensitive to disability definition. Even just focusing
on the DDA definition, smaller sample sizes for the exit analysis means fewer differences are
statistically significant.
For onset we found large differences in benefit receipt effects by qualification level (bigger
effects for less qualified), type of disability (bigger effects for mental health) and severity of
disability (effects increasing with severity). For exit the picture is less clear. We might expect
higher benefit replacement rates and lower chances of finding work for lower qualified
workers to make the lower qualified less responsive to disability exit in terms of benefit
receipt. For three out of four outcomes (not IB or ESA), however, we find equal or bigger
exit effects (in absolute terms) on benefit claiming for those with low qualifications
compared to those with higher qualifications, which appear robust to disability definition at
least in direction. As for onset we also find bigger (absolute) exit effects on sickness benefit
and IB/ESA benefit receipt for those whose disability is related to mental health but we find
no robust differences in exit effects by disability severity.
In partial contrast to onset effects which appeared similar by age, we do find bigger disability
exit effects for younger compared to older exiters, although the differences are mostly not
large in magnitude. There are also no clear, robust differences in exit effects by local
unemployment rate quartile.
28
7. Conclusions
There is consensus in the dynamics of disability and labour market outcomes literature that
disability onset leads to a decline in employment and earnings and an increase in applications
for and receipt of disability insurance. But important questions remain. How does disability
onset impact on welfare benefit receipt beyond disability insurance? How does disability exit
impact on benefit receipt? To what extent do these relationships vary for individuals with
different socio-economic characteristics, in different labour-market contexts, and under
different disability benefit regimes? This paper addresses these questions exploiting rarely-
used longitudinal data constructed from the UK Local Labour Force Survey, combining PSM
with DID methods to draw plausibly causal inferences under alternative but explicit
assumptions.
We first show that disability does not imply benefit receipt. Many individuals with disability
work and few among them claim welfare benefits. Many of those reporting disability who do
not work also do not claim benefits, and fewer than half receive IB/ESA (the DI equivalent).
Having said that we show that disability onset (robustly) increases receipt of sickness and
disability benefits within one year, although the impact on non-sickness benefits is small and
non-robust. Onset effects are larger for individuals with lower qualification levels, for those
experiencing onset of disability linked to mental health conditions or more severe disability
onset, and for those under a disability benefit regime with less rigorous screening and
conditionality. This suggests a number of policy implications. First, disability benefit
characteristics – in this case the differences between IB and ESA – matter. Second,
interventions aimed at keeping more of those who experience disability onset (or whose
health gradually declines in advance of disability onset) in employment would help to reduce
inflows not only to disability benefits but also to other benefits, given that many individuals
take an indirect route to disability benefits via other benefits, or remain on such benefits after
disability onset. Third, assistance to help overcome disability–related barriers to employment
is likely needed across multiple benefit payments. Fourth, the estimates of heterogeneous
disability onset effects suggest targeting interventions at the low-skilled and those
experiencing onset of disability linked to mental health conditions or onset of multiple
conditions.
The evidence on disability exit effects presented here is more mixed: some estimates suggest
negative impacts on benefit receipt for some benefit measures; others suggest zero impacts.
29
Arguably the most convincing exit estimates presented here – comparing those who report
disability exit with those who report disability exit one year later – suggest exit itself does not
impact on receipt of sickness and disability benefits up to two years on. The existence of a
sizeable pre-exit decline in sickness benefit receipt among those who exit is, however, a
possible reflection of gradual recovery from disability and is consistent with the possibility
that labour market adjustment occurs pre-exit (Jones et al., 2013). There are also few clear
differences in estimated disability exit effects by individual characteristics or labour market
context.
Whilst we acknowledge that exiters are unlikely to representative of all those who experience
disability onset, the apparent asymmetry between the benefit receipt impacts of disability
onset and disability exit provides tentative evidence that disability benefits do have some
‘absorbing state’ characteristics in Britain, and further (indirect) evidence that even
temporary disability can have long lasting economic effects. Two explanations for such long
term effects of temporary disability put forward in the literature on disability and labour
market outcomes are declining human capital and state dependence. There are several
potential state-dependence type explanations for what we observe in this paper, albeit that we
cannot untangle here. First, some of those reporting disability exit may still qualify for
disability benefits even if rescreened immediately following the change in their self-reported
disability status. Second, some of those reporting disability exit in survey data may not report
disability exit to benefit administrators, or may delay doing so, and in the absence of regular
systematic rescreening, may therefore remain in receipt of such benefits. Third, even those
reporting changes in disability status may not be rescreened, or may not be rescreened
quickly.20
Human capital depreciation alone might suggest disability exit leading to
wholesale switching from disability benefits to other benefits, which we do not find here.
There is, however, an alternative interpretation, if one is willing to consider pre-exit effects as
part of the treatment itself, that is, the results reflect gradual withdrawal from welfare support
prior to rather than at disability exit.
Despite the insights afforded by these data we are unable to explore the alternative
interpretation of our findings in relation to disability exit without panel data spanning a
longer period. There are also limitations as to the representativeness of our sample which
limit the extent to which these results are fully generalisable even within the UK. These latter
20
Of course screening is costly and in recent years the UK has a rather poor track record in terms of screening
recommendations overturned on appeal.
30
points also apply to our conclusions regarding disability onset. While longitudinal
administrative data on benefit records exist and would overcome at least some of these issues,
such analysis would require that these data are linked to survey or medical data on health
which is not currently permitted in the UK.
31
References
Abadie, A., and Imbens, GW. (2006). Large sample properties of matching estimators for
average treatment effects. Econometrica 74, 235–267.
Adam, S., Bozio, A. and Emmerson, C. (2010). Reforming Disability Insurance in the UK:
evaluation of the Pathways to Work programme. Institute for Fiscal Studies, London.
Autor, DH. and Duggan, MG. (2003). The rise in disability recipiency and the decline in
unemployment. Quarterly Journal of Economics 118, 1, 157–205.
Autor, DH. and Duggan, MG. (2006). The growth in the Social Security disability rolls: A
fiscal crisis unfolding. Journal of Economic Perspectives 20, 71–96.
Banks, J., Blundell, R. and Emmerson, C. (2015). Disability benefit receipt and reform:
reconciling trends in the United Kingdom. Journal of Economic Perspectives 29, 2, 173–190.
Beatty, C. and Fothergill, S. (2015). Disability Benefits in an Age of Austerity. Social Policy
& Administration 49, 2, 161-181.
Beatty, C., Fothergill, S. and Macmillan, R. (2000). A theory of employment, unemployment
and sickness. Regional Studies 34, 7, 617–30.
Benitez-Silva H., Buschinski, M., Chan, HM., Sheidvasser, S. and Rust, J. (2004). How large
is the bias in self-reported disability? Journal of Applied Econometrics 19, 6, 649–670.
Black D., Daniel K. and Sanders, S. (2002). The impact of economic conditions on
participation in disability programs: Evidence from the coal boom and bust. American
Economic Review 92, 27–50.
Bound, J. (1989). The health and earnings of rejected Disability Insurance applicants.
American Economic Review 79, 3, 482–503.
Bound, J. (1991) Self-reported versus objective measures of health in retirement models,
Journal of Human Resources, 26, 1, 106-138.
Browne, J. and Hood, A. (2012). A survey of the UK benefit system. IFS Briefing Note
BN13.
Burchardt, T. (2000). The dynamics of being disabled. CASE Paper 36, Centre for Analysis
of Social Exclusion, London School of Economics.
Burkhauser, RV., Daly, MC., McVicar, D. and Wilkins, R. (2014). Disability benefit growth
and disability reform in the US: lessons from other OECD nations. IZA Journal of Labor
Policy 3:4.
Campolieti, M. (2004). Disability Insurance benefits and labor supply: some additional
evidence. Journal of Labor Economics 22, 4, 863-887.
32
Charles, KK. (2003). The longitudinal structure of earnings losses among work-limited
disabled workers. Journal of Human Resources 38, 3, 618–646.
Colman, G. and Dave, D. (2014). Unemployment and health behaviors over the business
cycle: a longitudinal view. Working Paper No. 20748, National Bureau of Economic
Research, Cambridge MA.
Deb, P., Gallo, WT., Ayyagari, P., Fletcher, JM. and Sindelar, JL. (2011). The Effect of Job
Loss on Overweight and Drinking. Journal of Health Economics 30, 2, 317-327.
De Jong, P., Lindeboom, M. and van der Klaauw, B. (2011). Screening Disability Insurance
applications. Journal of the European Economic Association 9, 1, 106-129.
Disney, R., Emmerson, C. and Wakefield, M. (2006) ‘Ill health and retirement in Britain: a
panel data based analysis’, Journal of Health Economics, 25(4), 621-649.
Garcia-Gomez, P. (2011). Institutions, health shocks and labour market outcomes across
Europe, Journal of Health Economics 30, 200-213.
García-Gómez, P., van Kippersluis, H., O’Donnell, O. and van Doorslaer, E. (2013). Long-
Term and Spillover Effects of Health Shocks on Employment and Income. Journal of Human
Resources 48, 4, 873-909.
Gruber, J. (2000). Disability Insurance benefits and labor supply. Journal of Political
Economy 108, 6, 1162-1183.
Haveman, R.H. and Wolfe, B.L. (1984). Disability transfers and early retirement: a causal
relationship? Journal of Public Economics 24, 1, 47–66.
Heckman, J., Ichimura H. and Todd, P. (1997). Matching as an econometric evaluation
estimator: evidence from evaluation. Review of Economic Studies 64, 605-54.
Jenkins S. and Rigg J. (2004). Disability and disadvantage: Selection, onset and duration
effects. Journal of Social Policy 33, 3, 479-501.
Jones, MK. (2011). Disability, employment and earnings: an examination of heterogeneity,
Applied Economics 43, 8, 1001–1017.
Jones, MK., Davies, R. and Drinkwater, S. (2013). The dynamics of disability and work.
University of Swansea, Swansea.
Jones, M. K., Latreille, P. L. and Sloane, P. J. (2006). Disability, gender and the British
labour market. Oxford Economic Papers 58, 3, 407-459.
Karlström, A., Palmeb, M. and Svensson, I. (2008). The employment effect of stricter rules
for eligibility for DI: Evidence from a natural experiment in Sweden. Journal of Public
Economics 92, 2071-2082.
33
Kidd, MP., Sloane, PJ. and Ferko, I. (2000). Disability and the labour market; an analysis of
British males. Journal of Health Economics 19, 961–81.
Leuven, E and Sianesi, B (2003), PSMATCH2: Stata module to perform full Mahalanobis
and propensity score matching, common support graphing, and covariate imbalance testing.
http://ideas.repec.org/c/boc/bocode/s432001.html.
Lindeboom, M. and Kerkhofs, M. (2009). Health and work of the elderly: subjective health
measures, reporting errors and endogeneity in the relationship between health and work.
Journal of Applied Econometrics 24, 1024–1046.
Lindeboom, M., Llena-Nozal, A. and van der Klaauw, B. (2006). Disability and work: the
role of health shocks and childhood circumstances. IZA DP No. 2096.
Maestas, N., Mullen, K. J. and Strand, A. (2013). Does Disability Insurance receipt
discourage work? Using examiner assignment to estimate causal effects of SSDI receipt.
American Economics Review 103, 5, 1797-1829.
McVicar, D. (2006). Why do disability benefit rolls vary between regions? A review of
evidence from the US and UK. Regional Studies 40, 5, 519–33.
McVicar, D. (2008). Why have UK disability rolls grown so much? Journal of Economic
Surveys 22, 1, 114–39.
Meyer B. D. and Mok W. K. C., (2013). Disability, earnings, income and consumption.
NBER Working Paper No. 18869.
Mok, WK., Meyer, BD., Charles, KK. and Achen, AC. (2008). A note on ‘The longitudinal
structure of earnings losses among work-limited disabled workers’. Journal of Human
Resources 43, 3, 721–728.
Moncrieff, J. and Pomerleau, J. (2000). Trends in sickness benefits in Great Britain and the
contribution of mental disorders.’ Journal of Public Health Medicine 22, 59-67.
Mora, R. and Reggio, I. (2012). Treatment effect identification using alternative parallel
assumptions. Discussion paper 12-33, University Carlos III Madrid.
National Audit Office (2010), Support to Incapacity Benefits Claimants through Pathways to
Work, London: Stationery Office.
Norman, P., Boyle, P. and Rees, P. (2005) ‘Selective migration, health and deprivation: a
longitudinal analysis’, Social Science and Medicine, 60(12), 2755-2771.
Oguzoglu, U. (2012a). Dynamics of disability and work in Canada. IZA Discussion Paper
6603. IZA, Bonn.
Oguzoglu, U. (2012b) Is there a better measure of self-assessed disability? Applied
Economics Letters 19, 14, 1335-1338.
34
Organization for Economic Co-operation and Development (2010). Sickness, Disability and
Work: Breaking the Barriers: A Synthesis of Findings across OECD Countries. OECD, Paris.
Parsons, DO. (1980). The decline in male labor force participation. Journal of Political
Economy 88, 1, 117–34.
Polidano, C. and Vu, H. (2015). Differential labour market impacts from disability onset.
Health Economics 24, 3, 302-317.
Rosenbaum, P.R. and Rubin, D.B. (1983). The central role of the propensity score in
observational studies for causal effects, Biometrika 70, 41-55.
Singleton, P. (2014). The dynamic relationship between disability onset, earnings, and
disability insurance application and receipt. Economics Letters 124, 374-377.
Sissons, P., Barnes. H. and Stevens, H. (2011). Routes onto Employment and Support
Allowance. Department for Work and Pensions Research Report No 774, London.
Staubli, S. (2011). The impact of stricter criteria for Disability Insurance on labor force
participation. Journal of Public Economics 95, 1223-1235.
Trevisan, E., Zantomio, F. (2016). The impact of acute health shocks on the labour supply of
older workers: evidence from sixteen European countries. Labour Economics
http://dx.doi.org/10.1016/j.labeco.2016.04.002.
35
Figure 1: Proportion reporting receipt of welfare benefits, 2004-2012
Notes: LLFS working-age population (2004-2012).
0
.05
.1.1
5.2
Pro
po
rtio
n
2004 2005 2006 2007 2008 2009 2010 2011 2012
Year
Any benefit Non-sickness benefit
Sickness benefit IB or ESA
IB ESA
36
Table 1: Disability Onset and Exit Treatment and Control Groups
DDA WL
N % N %
Onset Control (0000) 33,216 67.74 35,486 72.37
Control (0001) 1,730 3.53 1,346 2.75
Treatment (0011) 905 1.85 596 1.22
Exit Treatment (1100) 420 0.86 431 0.88
Control (1111) 4,762 9.71 4,703 9.59
Control (1110) 579 1.18 486 0.99
Other 7,427 15.13 5,991 12.20 Notes: Sample is restricted to a balanced panel sample with a minimum of 4 waves within the LLFS. Onset and
exit groups are defined using a two wave definition of consistent disability onset or exit.
37
Table 2: Proportions receiving welfare benefit by wave and treatment status, DDA
disability
Any Benefit Wave 1 Wave 2 Wave 3 Wave 4
Onset Control (0000) 0.058 0.064 0.064 0.063
Control (0001) 0.089 0.105 0.123 0.146
Treatment 0.135 0.158 0.211 0.232
Exit Treatment 0.267 0.242 0.196 0.200
Control (1110) 0.380 0.389 0.384 0.346
Control (1111) 0.700 0.708 0.724 0.724
Non-sickness Benefit Wave 1 Wave 2 Wave 3 Wave 4
Onset Control (0000) 0.050 0.055 0.054 0.053
Control (0001) 0.075 0.084 0.095 0.098
Treatment 0.110 0.128 0.138 0.144
Exit Treatment 0.192 0.171 0.162 0.154
Control (1110) 0.209 0.218 0.237 0.223
Control (1111) 0.351 0.372 0.384 0.383
Sickness Benefit Wave 1 Wave 2 Wave 3 Wave 4
Onset Control (0000) 0.007 0.008 0.007 0.008
Control (0001) 0.017 0.023 0.030 0.066
Treatment 0.023 0.039 0.099 0.132
Exit Treatment 0.113 0.093 0.051 0.068
Control (1110) 0.274 0.263 0.255 0.204
Control (1111) 0.589 0.608 0.625 0.633
IB or ESA Wave 1 Wave 2 Wave 3 Wave 4
Onset Control (0000) 0.001 0.001 0.001 0.002
Control (0001) 0.003 0.007 0.016 0.036
Treatment 0.015 0.021 0.064 0.083
Exit Treatment 0.068 0.064 0.029 0.034
Control (1110) 0.165 0.182 0.169 0.123
Control (1111) 0.390 0.414 0.417 0.420 Notes: See notes to Table 1.
38
Table 3: Benefit Receipt by DDA Disability and Employment Status
Disabled Non-disabled
All Employed Non-
employed
All Employed Non-
employed
Any Benefits 50.6 11.9 79.7 9.5 3.1 33.4
Non-sickness Benefit 30.3 5.1 49.3 8.4 2.5 30.4
Sickness Benefit 37.3 7.9 59.5 1.2 0.5 3.8
IB or ESA 24.1 2.3 40.4 0.4 0.1 2.0
N 137,508 58,966 78,542 647,219 510,056 137,163 Notes: LLFS working-age population (2004-2012).
39
Table 4: Descriptive Statistics for Explanatory Variables by Control and Treatment
Groups
Onset Exit
Treatment
(0011)
Control
(0000)
(pre-
matching)
Control
(0000)
(post-
matching)
Treatment
(1100)
Control
(1111)
(pre-
matching)
Control
(1111)
(post-
matching)
Age 45.791 41.158*** 45.576 45.509 48.085*** 45.402
Gender
Male 0.495 0.476 0.474 0.497 0.515 0.506
Highest
qualification
Degree 0.192 0.236*** 0.182 0.116 0.082** 0.127
Other Higher
Education 0.123 0.113 0.098 0.116 0.081** 0.107
A level 0.218 0.231 0.231 0.223 0.185* 0.220
O level 0.193 0.227** 0.201 0.243 0.181*** 0.254
Other 0.112 0.086** 0.102 0.095 0.135** 0.121
None 0.163 0.107*** 0.185 0.208 0.335*** 0.171
Students
Full-time student 0.026 0.063*** 0.018 0.038 0.021** 0.020
Marital Status
Single 0.235 0.280*** 0.215 0.217 0.239 0.228
Married 0.618 0.613 0.640 0.613 0.539*** 0.621
Widowed/divorced 0.147 0.107*** 0.144 0.171 0.222** 0.150
Children
Dependent child in
household 0.316 0.442*** 0.329 0.309 0.224*** 0.286
Housing Tenure
Owned outright 0.238 0.213* 0.230 0.257 0.234 0.266
Mortgaged 0.549 0.651*** 0.522 0.494 0.352*** 0.503
Rented 0.214 0.137*** 0.248 0.249 0.414*** 0.231
Region
Tyne and Wear 0.021 0.030 0.029 0.032 0.037 0.035
Rest of North East 0.052 0.045 0.054 0.066 0.056 0.066
Greater Manchester 0.072 0.066 0.076 0.052 0.072 0.078
Merseyside 0.049 0.033** 0.047 0.029 0.043 0.029
Rest of North West 0.033 0.033 0.020 0.026 0.036 0.017
South Yorkshire 0.022 0.017 0.017 0.026 0.026 0.029
West Yorkshire 0.013 0.013 0.016 0.012 0.016 0.006
Rest of Yorkshire &
Humberside 0.028 0.033 0.047** 0.020 0.029 0.014
East Midlands 0.024 0.019 0.033 0.026 0.018 0.020
West Midlands
Metropolitan county 0.035 0.040 0.039 0.032 0.035 0.014
Rest of West
Midlands 0.030 0.031 0.033 0.017 0.028 0.023
East of England 0.041 0.030* 0.033 0.035 0.024 0.066*
Inner London 0.021 0.022 0.017 0.026 0.020 0.012
Outer London 0.029 0.030 0.030 0.020 0.021 0.020
South East 0.067 0.089** 0.080 0.069 0.052 0.052
South West 0.076 0.067 0.066 0.038 0.055 0.032
Wales 0.171 0.168 0.156 0.194 0.220 0.208
40
Strathclyde 0.094 0.101 0.077 0.113 0.096 0.116
Rest of Scotland 0.122 0.134 0.130 0.168 0.116*** 0.162
Economic
Conditions
Employed 0.751 0.833*** 0.752 0.668 0.269*** 0.671
Local area
unemployment rate 0.065 0.063*** 0.066 0.065 0.067* 0.064
Year of observation
2004 0.142 0.181*** 0.142 0.153 0.166 0.153
2005 0.133 0.172*** 0.133 0.199 0.158** 0.199
2006 0.199 0.186 0.199 0.171 0.186 0.171
2007 0.158 0.174 0.158 0.165 0.169 0.165
2008 0.199 0.146*** 0.199 0.159 0.166 0.159
2009 0.169 0.141** 0.169 0.153 0.156 0.153
Any benefit 0.134 0.058*** 0.134 0.269 0.706*** 0.269
N 762 27,762 737 346 4,173 303 Notes: All characteristics are measured at Wave 1. Treatment is defined by consistent onset/exit of DDA
disability and is estimated using a NN(1) matching algorithm. *,**,*** denote significance from the treatment
group at the 10%, 5% and 1% level respectively.
41
Table 5: DID-PSM, DDA Disability Onset Treatment Effects: Proportions Receiving
Benefits
Any Benefit
Onset Control (0000) Difference T stat
Wave 1 0.134 0.134 0.000 0.00
Wave 2 0.151 0.125 0.026 1.91
Wave 3 0.209 0.115 0.093 5.69
Wave 4 0.223 0.127 0.096 5.53
Difference (2-1) 0.017 -0.009 0.026 1.91
Difference (3-1) 0.075 -0.018 0.093 5.69
Difference (4-1) 0.089 -0.007 0.096 5.53
Difference (3-2) 0.058 -0.009 0.067 4.11
Difference (3-2)-(2-1) 0.041 0.000 0.041 1.61
Non-sickness Benefit
Onset Control (0000) Difference T stat
Wave 1 0.110 0.112 -0.001 -0.21
Wave 2 0.119 0.115 0.004 0.30
Wave 3 0.139 0.102 0.037 2.59
Wave 4 0.142 0.108 0.034 2.32
Difference (2-1) 0.009 0.004 0.005 0.39
Difference (3-1) 0.029 -0.009 0.038 2.61
Difference (4-1) 0.031 -0.004 0.035 2.38
Difference (3-2) 0.020 -0.013 0.033 2.25
Difference (3-2)-(2-1) 0.011 -0.171 0.028 1.16
Sickness Benefit
Onset Control (0000) Difference T stat
Wave 1 0.021 0.020 0.001 0.19
Wave 2 0.038 0.012 0.026 3.31
Wave 3 0.097 0.016 0.081 7.21
Wave 4 0.126 0.017 0.109 9.22
Difference (2-1) 0.017 -0.008 0.025 2.69
Difference (3-1) 0.076 -0.004 0.080 6.60
Difference (4-1) 0.105 -0.003 0.108 8.22
Difference (3-2) 0.059 0.004 0.055 4.87
Difference (3-2)-(2-1) 0.042 0.012 0.030 1.80
IB or ESA
Onset Control (0000) Difference T stat
Wave 1 0.014 0.004 0.010 2.41
Wave 2 0.021 0.003 0.018 3.45
Wave 3 0.064 0.004 0.060 6.86
Wave 4 0.083 0.003 0.080 8.34
Difference (2-1) 0.007 -0.001 0.008 1.36
Difference (3-1) 0.050 0.000 0.050 5.32
Difference (4-1) 0.068 -0.001 0.070 6.87
Difference (3-2) 0.043 0.001 0.042 4.85
Difference (3-2)-(2-1) 0.037 0.003 0.034 3.00 Notes: See notes to Table 1. ATT are based on a NN(1) matching algorithm and are estimated over the region of
common support. T statistics are based on Abadie and Imbens (2006) standard errors.
42
Table 6: DID-PSM, DDA Disability Onset Treatment Effects: Proportions Receiving
Benefits, Alternative Control Group
Any Benefit
Onset Control (0001) Difference T stat
Wave 1 0.134 0.134 0.000 0.00
Wave 2 0.151 0.147 0.004 0.23
Wave 3 0.209 0.170 0.039 1.98
Difference (2-1) 0.017 0.013 0.004 0.23
Difference (3-1) 0.075 0.035 0.039 1.98
Difference (3-2) 0.058 0.022 0.035 1.74
Non-sickness Benefit
Onset Control (0001) Difference T stat
Wave 1 0.110 0.118 -0.008 -0.93
Wave 2 0.119 0.122 -0.003 -0.16
Wave 3 0.139 0.139 0.000 0.00
Difference (2-1) 0.009 0.004 0.005 0.32
Difference (3-1) 0.029 0.021 0.008 0.46
Difference (3-2) 0.020 0.017 0.003 0.14
Sickness Benefit
Onset Control (0001) Difference T stat
Wave 1 0.021 0.029 -0.008 -0.87
Wave 2 0.038 0.024 0.014 1.34
Wave 3 0.097 0.035 0.062 4.70
Difference (2-1) 0.017 -0.005 0.022 1.84
Difference (3-1) 0.076 0.007 0.070 4.80
Difference (3-2) 0.059 0.012 0.047 3.35
IB or ESA
Onset Control (0001) Difference T stat
Wave 1 0.014 0.009 0.005 0.92
Wave 2 0.021 0.009 0.012 1.77
Wave 3 0.064 0.022 0.042 3.97
Difference (2-1) 0.007 0.000 0.007 0.90
Difference (3-1) 0.050 0.013 0.037 3.35
Difference (3-2) 0.043 0.013 0.030 2.97 Notes: See notes to Table 5.
43
Table 7: Heterogeneity in the DID-PSM (wave 3-1, wave 3-2), DDA Disability Onset
Treatment Effects: Proportions Receiving Welfare Benefits
DID
Any Benefit Non-sickness Benefit Sickness Benefit IB/ESA
DID_PSM 3-1 3-2 3-1 3-2 3-1 3-2 3-1 3-2
Male 0.077 0.074 0.024 0.023 0.077 0.066 0.056 0.048 Female 0.091 0.055 0.049 0.023 0.068 0.055 0.044 0.010
Low Qual 0.124 0.067 0.084 0.036 0.084 0.073 0.062 0.051
High Qual 0.067 0.054 0.020 0.017 0.054 0.035 0.039 0.032
Older 0.082 0.067 0.037 0.039 0.087 0.056 0.050 0.045
Younger 0.088 0.071 0.057 0.037 0.054 0.051 0.047 0.037
Mental (treatment n=73) 0.260 0.137 0.151 0.027 0.192 0.151 0.151 0.110
Physical 0.074 0.049 0.025 0.010 0.069 0.059 0.045 0.039
Single 0.070 0.035 0.044 0.006 0.047 0.041 0.021 0.018
Multiple 0.115 0.077 0.053 0.036 0.110 0.077 0.077 0.060
Pre-2009 0.069 0.064 0.030 0.039 0.080 0.055 0.055 0.053
Post-2009 0.062 0.060 0.017 0.013 0.072 0.057 0.042 0.032
Unemploy Q1 0.065 0.046 0.014 0.018 0.074 0.023 0.037 0.023
Unemploy Q2 0.074 0.051 0.045 0.040 0.080 0.057 0.045 0.040
Unemploy Q3 0.090 0.042 0.079 0.042 0.074 0.042 0.053 0.042
Unemploy Q4 0.072 0.072 0.006 0.006 0.100 0.094 0.072 0.067
Employed 0.080 0.049 0.030 0.009 0.073 0.051 0.049 0.040
Not employed 0.074 0.084 0.079 0.090 0.063 0.058 0.037 0.032
Any benefits 0.110 0.119 0.085 0.186 0.161 0.025 0.110 0.051
No benefits 0.071 0.071 0.030 0.022 0.059 0.062 0.033 0.037
Long-term health 0.043 0.030 0.030 0.023 0.020 0.023 0.015 0.028
No long-term health 0.121 0.101 0.058 0.055 0.112 0.074 0.082 0.060
Past health 0.167 0.106 0.091 0.045 0.152 0.091 0.045 0.045
No past health 0.106 0.075 0.065 0.032 0.063 0.059 0.055 0.051 Notes: See notes to Table 5. Samples are defined on the basis of information in wave 1 or at onset as
appropriate. Bold indicates statistically significant from zero at the 95% confidence level.
44
Table 8: DID-PSM, DDA Disability Exit Treatment Effects: Proportions Receiving
Welfare Benefits
Any Benefit
Exit Control (1111) Difference T stat
Wave 1 0.269 0.269 0.000 0.00
Wave 2 0.246 0.384 -0.139 -5.09
Wave 3 0.197 0.379 -0.182 -6.53
Wave 4 0.199 0.431 -0.231 -7.83
Difference (2-1) -0.023 0.116 -0.139 -5.09
Difference (3-1) -0.072 0.110 -0.182 -6.53
Difference (4-1) -0.069 0.162 -0.231 -7.83
Difference (3-2) -0.049 -0.006 -0.043 -1.60
Difference (3-2)-(2-1) -0.026 -0.121 0.095 2.04
Non-sickness Benefit
Exit Control (1111) Difference T stat
Wave 1 0.191 0.142 0.049 2.43
Wave 2 0.173 0.188 -0.014 -0.59
Wave 3 0.156 0.179 -0.023 -0.91
Wave 4 0.145 0.182 -0.038 -1.44
Difference (2-1) -0.017 0.046 -0.064 -2.72
Difference (3-1) -0.035 0.038 -0.072 -2.73
Difference (4-1) -0.046 0.040 -0.087 -3.35
Difference (3-2) -0.017 -0.009 -0.009 -0.33
Difference (3-2)-(2-1) 0.000 -0.055 0.042 1.32
Sickness Benefit
Exit Control (1111) Difference T stat
Wave 1 0.118 0.223 -0.104 -6.47
Wave 2 0.101 0.301 -0.199 -7.86
Wave 3 0.061 0.312 -0.251 -9.95
Wave 4 0.081 0.367 -0.286 -10.68
Difference (2-1) -0.017 0.078 -0.095 -3.76
Difference (3-1) -0.058 0.090 -0.147 -5.39
Difference (4-1) -0.038 0.145 -0.182 -6.43
Difference (3-2) -0.041 0.012 -0.052 -2.10
Difference (3-2)-(2-1) -0.023 -0.067 0.043 1.03
IB or ESA
Exit Control (1111) Difference T stat
Wave 1 0.072 0.145 -0.072 -3.86
Wave 2 0.066 0.217 -0.150 -6.33
Wave 3 0.035 0.217 -0.182 -8.38
Wave 4 0.040 0.257 -0.217 -9.64
Difference (2-1) -0.006 0.072 -0.078 -3.17
Difference (3-1) -0.038 0.072 -0.110 -4.43
Difference (4-1) -0.032 0.113 -0.145 -5.68
Difference (3-2) -0.032 0.000 -0.032 -1.47
Difference (3-2)-(2-1) -0.026 -0.072 0.046 1.18 Notes: See notes to Table 5.
45
Table 9: DID-PSM, DDA Disability Exit Treatment Effects: Proportions Receiving
Welfare Benefits, Alternative Control Group
Any Benefit
Exit Control (1110) Difference T stat
Wave 1 0.269 0.269 0.000 0.00
Wave 2 0.246 0.309 -0.064 -1.89
Wave 3 0.197 0.335 -0.139 -3.75
Difference (2-1) -0.023 0.041 -0.064 -1.89
Difference (3-1) -0.072 0.067 -0.139 -3.75
Difference (3-2) -0.049 0.026 -0.075 -2.55
Non-sickness Benefit
Exit Control (1110) Difference T stat
Wave 1 0.191 0.127 0.064 2.77
Wave 2 0.173 0.124 0.049 1.54
Wave 3 0.156 0.194 -0.038 -1.10
Difference (2-1) -0.017 -0.003 -0.015 -0.50
Difference (3-1) -0.035 0.066 -0.101 -3.06
Difference (3-2) -0.017 0.069 -0.087 -2.70
Sickness Benefit
Exit Control (1110) Difference T stat
Wave 1 0.118 0.214 -0.095 -4.62
Wave 2 0.101 0.231 -0.130 -3.88
Wave 3 0.061 0.208 -0.147 -4.63
Difference (2-1) -0.017 0.017 -0.035 -1.04
Difference (3-1) -0.058 -0.006 -0.052 -1.61
Difference (3-2) -0.041 -0.023 0.017 -0.82
IB or ESA
Exit Control (1110) Difference T stat
Wave 1 0.072 0.156 -0.084 -3.29
Wave 2 0.066 0.159 -0.093 -3.58
Wave 3 0.035 0.142 -0.107 -4.32
Difference (2-1) -0.006 0.003 -0.009 -0.40
Difference (3-1) -0.038 -0.014 -0.023 -0.98
Difference (3-2) -0.032 -0.017 -0.014 -0.78 Notes: See notes to Table 5.
46
Table 10: Heterogeneity in the DID-PSM (wave 3-1, wave 3-2), DDA Disability Exit
Treatment Effects: Proportions Receiving Welfare Benefits
DID
Any Benefit Non-sickness
Benefit
Sickness Benefit IB/ESA
DID_PSM 3-1 3-2 3-1 3-2 3-1 3-2 3-1 3-2
Male -0.174 -0.035 -0.064 0.017 -0.122 -0.047 -0.081 -0.029
Female -0.190 -0.098 -0.138 -0.098 -0.109 -0.012 -0.057 0.006
Low Qual -0.191 -0.085 -0.111 -0.053 -0.153 -0.069 -0.063 -0.021
High Qual -0.153 -0.064 -0.051 -0.013 -0.140 -0.038 -0.076 -0.032
Older -0.164 -0.056 -0.117 -0.037 -0.098 -0.014 -0.070 -0.009
Younger -0.182 -0.045 -0.061 0.000 -0.182 -0.076 -0.091 -0.030
Mental (treatment
n=35) -0.286 -0.114 0.000 0.057 -0.314 -0.171 -0.257 -0.143
Physical -0.256 -0.116 -0.097 -0.047 -0.217 -0.112 -0.173 -0.087
Single -0.228 -0.074 -0.142 -0.049 -0.142 -0.043 -0.136 -0.068
Multiple -0.170 -0.050 -0.050 -0.044 -0.187 -0.060 -0.121 0.000
Unemploy Q1 -0.226 -0.072 -0.060 0.012 -0.226 -0.131 -0.155 -0.107
Unemploy Q2 -0.217 -0.141 -0.141 -0.087 -0.130 -0.087 -0.130 -0.087
Unemploy Q3 -0.152 -0.098 -0.054 -0.054 -0.152 -0.043 -0.043 0.011
Unemploy Q4 -0.308 -0.115 -0.231 -0.077 -0.192 -0.115 -0.103 -0.090 Notes: See notes to Table 5. Samples are defined on the basis of information in wave 1 or at exit as appropriate. Bold indicates statistically significant from zero at the 95% confidence level.
47
Supplementary Appendix: Further Data Details and Additional Results
Figure A1: LLFS and WPLS Proportion reporting receipt of IB and ESA, 2004-2012
Notes: Data from the Work and Pensions Longitudinal Study (WPLS) and is access via NOMIS. Claimant rates
relate to Great Britain and are created using claimant numbers as of November each year and mid-year working-
age population estimates.
0
.02
.04
.06
.08
Pro
port
ion
2004 2005 2006 2007 2008 2009 2010 2011 2012
Year
IB ESA
IB WPLS ESA WPLS
48
Table A1: Representativeness of LLFS Balanced Sample
APS (2004-2012) LLFS (2004-2012)
All waves All waves 4 waves
WL disabled 15.13 15.49 17.48
DDA disabled 15.57 16.12 18.11
Long-term health problem 28.05 28.15 32.70
Past long-term health problem 7.76 7.59 8.39
Employment 72.44 71.54 76.34
Any Benefit 15.49 16.76 14.74
Non-sickness Benefit 11.68 12.70 9.58
Sickness Benefit 6.53 7.03 8.00
IB or ESA 3.89 4.27 4.92
Gender
Female 49.54 49.70 49.70
Male 50.46 50.30 50.30
Age 38.45 38.11 42.83
Highest qualification
Degree 19.84 18.74 20.00
Other Higher Education 8.62 8.97 10.99
A level 22.69 22.93 22.76
O level 22.73 22.62 21.83
Other 11.60 11.72 9.82
None 14.51 15.02 14.60
Students
Full-time student 8.70 8.74 4.83
Not full-time student 91.30 91.26 95.17
Marital Status
Single 39.79 40.88 25.49
Married 47.58 46.16 61.53
Widowed/divorced 12.63 12.95 12.99
Children
Dependent child in household 39.20 39.24 41.74
No dependent child in household 60.80 60.76 58.26
Housing Tenure
Owned outright 17.29 16.38 21.57
Mortgaged 50.18 49.38 59.87
Rented 32.53 34.24 18.56
Region
Tyne and Wear 2.66 3.67 3.19
Rest of North East 3.47 4.89 4.72
Greater Manchester 5.21 6.65 6.77
Merseyside 2.59 3.27 3.42
Rest of North West 4.36 3.54 3.43
South Yorkshire 2.16 1.97 1.91
West Yorkshire 2.95 1.42 1.36
Rest of Yorkshire & Humberside 2.99 3.31 3.23
East Midlands 5.46 2.62 1.93
West Midlands Metropolitan county 3.97 3.85 3.79
Rest of West Midlands 3.76 2.65 2.95
East of England 6.79 3.28 2.96
Inner London 4.28 4.24 2.08
Outer London 5.57 3.56 2.77
49
South East 11.69 9.18 7.96
South West 7.40 6.46 6.48
Wales 9.57 16.80 17.98
Strathclyde 5.47 8.15 9.83
Rest of Scotland 7.21 10.49 13.24
Northern Ireland 2.43 - -
Local Area Unemployment Rate 0.063 0.066 0.064
Year of observation
2004 19.72 22.45 20.07
2005 12.09 13.37 19.43
2006 10.18 9.17 16.80
2007 10.11 8.84 15.96
2008 9.86 8.79 14.03
2009 9.51 8.75 13.72
2010 9.47 9.07 -
2011 9.53 9.83 -
2012 9.52 9.72 -
Interview type
Face-to-face 78.82 74.42 85.59
Telephone 21.18 25.58 14.41
Sample
QLFS 59.89 - -
LLFS 40.11 100 100
N 1,099,439 440,947 49,071 Notes: All characteristics are measured at Wave 1. The APS sample excludes the APS boost.
50
Table A2: Proportions receiving welfare benefit by wave and treatment status, WL
disability
Any Benefit Wave 1 Wave 2 Wave 3 Wave 4
Onset Control (0000) 0.054 0.060 0.060 0.059
Control (0001) 0.100 0.131 0.141 0.188
Treatment 0.137 0.196 0.308 0.318
Exit Treatment 0.251 0.260 0.199 0.184
Control (1111) 0.749 0.755 0.764 0.761
Control (1110) 0.477 0.480 0.470 0.432
Non-sickness Benefit Wave 1 Wave 2 Wave 3 Wave 4
Onset Control (0000) 0.047 0.051 0.050 0.049
Control (0001) 0.085 0.098 0.105 0.120
Treatment 0.104 0.151 0.194 0.183
Exit Treatment 0.161 0.175 0.156 0.126
Control (1111) 0.379 0.402 0.410 0.412
Control (1110) 0.285 0.274 0.273 0.271
Sickness Benefit Wave 1 Wave 2 Wave 3 Wave 4
Onset Control (0000) 0.006 0.007 0.007 0.008
Control (0001) 0.014 0.030 0.040 0.090
Treatment 0.030 0.055 0.142 0.182
Exit Treatment 0.124 0.106 0.053 0.068
Control (1111) 0.628 0.643 0.659 0.662
Control (1110) 0.303 0.316 0.312 0.259
IB or ESA Wave 1 Wave 2 Wave 3 Wave 4
Onset Control (0000) 0.001 0.001 0.001 0.002
Control (0001) 0.004 0.012 0.018 0.051
Treatment 0.009 0.028 0.093 0.113
Exit Treatment 0.078 0.070 0.034 0.034
Control (1111) 0.415 0.435 0.441 0.446
Control (1110) 0.180 0.237 0.202 0.147 Notes: See notes to Table 1.
51
Table A3: Benefit Receipt by WL Disability and Employment Status
Disabled Non-disabled
All Employed Non-
employed
All Employed Non-
employed
Any Benefits 57.0 15.4 80.5 9.0 3.0 32.4
Non-sickness Benefit 34.3 6.4 50.1 8.1 2.4 29.7
Sickness Benefit 41.8 10.4 59.6 1.1 0.5 3.2
IB/ESA 27.0 3.1 40.5 0.4 0.1 1.6
N 125,115 45,061 80,054 659,612 523,961 135,651
Notes: LLFS working-age population (2004-2012).
52
Table A4: Descriptive Statistics for Explanatory Variables by Control and Treatment
Groups, DDA, Alternative Control Groups
Onset Exit
Treatment
(0011)
Control
(0001) (pre-
matching)
Control
(0001)
(post-
matching)
Treatment
(1100)
Control
(1110) (pre-
matching)
Control
(1110)
(post-
matching)
Age 45.791 45.919 45.887 45.509 46.626 45.324
Gender
Male 0.495 0.472 0.501 0.497 0.520 0.541
Highest
qualification
Degree 0.192 0.176 0.170 0.116 0.150 0.176**
Other Higher
Education 0.123 0.131 0.129 0.116 0.092 0.095
A level 0.218 0.226 0.205 0.223 0.218 0.218
O level 0.193 0.212 0.212 0.243 0.146*** 0.231
Other 0.112 0.114 0.121 0.095 0.138* 0.098
None 0.163 0.142 0.164 0.208 0.257 0.182
Students
Full-time student 0.026 0.029 0.022 0.038 0.016* 0.026
Marital Status
Single 0.235 0.202* 0.231 0.217 0.273 0.1999
Married 0.618 0.654* 0.632 0.613 0.559 0.627
Widowed/divorced 0.147 0.144 0.137 0.171 0.168 0.173
Children
Dependent child in
household 0.316 0.340 0.309 0.309 0.263** 0.286
Housing Tenure
Owned outright 0.238 0.246 0.259 0.257 0.287 0.234
Mortgaged 0.549 0.563 0.523 0.494 0.421** 0.515
Rented 0.214 0.191 0.218 0.249 0.292 0.252
Region
Tyne and Wear 0.021 0.029 0.026 0.032 0.027 0.038
Rest of North East 0.052 0.052 0.043 0.066 0.068 0.075
Greater
Manchester 0.072 0.078 0.068 0.052 0.068 0.035
Merseyside 0.049 0.040 0.329 0.029 0.027 0.026
Rest of North
West 0.033 0.043 0.049 0.026 0.039 0.035
South Yorkshire 0.022 0.017 0.012 0.026 0.025 0.020
West Yorkshire 0.013 0.014 0.009 0.012 0.014 0.012
Rest of Yorkshire
& Humberside 0.028 0.027 0.024 0.020 0.043* 0.009
East Midlands 0.024 0.015 0.029 0.026 0.021 0.017
West Midlands
Metropolitan
county 0.035 0.0390 0.039 0.032 0.045 0.011*
Rest of West
Midlands 0.030 0.019 0.028 0.017 0.021 0.023
East of England 0.041 0.042 0.036 0.035 0.021 0.032
Inner London 0.021 0.016 0.014 0.026 0.021 0.026
Outer London 0.029 0.027 0.029 0.020 0.031 0.026
53
South East 0.067 0.075 0.087 0.069 0.084 0.121*
South West 0.076 0.065 0.087 0.038 0.066* 0.035
Wales 0.171 0.169 0.175 0.194 0.150* 0.202
Strathclyde 0.094 0.093 0.114 0.113 0.123 0.116
Rest of Scotland 0.122 0.141 0.108 0.168 0.109** 0.142
Economic
Conditions
Employed 0.751 0.824*** 0.735 0.668 0.532*** 0.685
Local area
unemployment
rate 0.065 0.063 0.065 0.065 0.066 0.064
Year of
observation
2004 0.142 0.142 0.142 0.153 0.142 0.153
2005 0.133 0.147 0.133 0.199 0.185 0.199
2006 0.199 0.168* 0.200 0.171 0.195 0.171
2007 0.158 0.228*** 0.158 0.165 0.154 0.165
2008 0.199 0.174 0.200 0.159 0.148 0.159
2009 0.169 0.141* 0.168 0.153 0.177 0.153
Any benefit 0.134 0.093*** 0.134 0.269 0.359*** 0.269
N 761 1,469 544 346 487 197 Notes: All characteristics are measured at Wave 1. Treatment is defined by consistent onset/exit of DDA
disability and is estimated using a NN(1) matching algorithm. *,**,*** denote significance from the treatment
group at the 10%, 5% and 1% level respectively.
54
Table A5: Descriptive Statistics for Explanatory Variables by Control and Treatment
Groups, WL
Onset Exit
Treatment Control
(pre-
matching)
Control
(post-
matching)
Treatment Control (pre-
matching)
Control
(post-
matching)
Age 45.233 41.698*** 45.528 44.218 47.892*** 46.008
Gender
Male 0.502 0.474 0.480 0.549 0.526 0.500
Highest
qualification
Degree 0.147 0.237*** 0.157 0.143 0.075*** 0.155
Other Higher
Education
0.092 0.116 0.064 0.098 0.077 0.130
A level 0.243 0.232 0.255
0.232 0.182** 0.235
O level 0.229 0.226 0.237 0.199 0.177 0.197
Other 0.108 0.085* 0.127 0.134 0.134 0.113
None 0.181 0.105* 0.161 0.193 0.356*** 0.169
Students
Full-time student 0.052 0.059 0.028* 0.036 0.019** 0.040
Marital Status
Single 0.267 0.266 0.233 0.249 0.260 0.218
Married 0.588 0.625* 0.618 0.599 0.507*** 0.649
Widowed/divorced 0.145 0.108*** 0.149 0.151 0.232*** 0.132
Children
Dependent child in
household
0.311 0.433*** 0.317 0.325 0.216*** 0.342
Housing Tenure
Owned outright 0.265 0.215*** 0.247 0.261 0.237 0.342**
Mortgaged 0.486 0.653*** 0.534 0.493 0.321*** 0.486
Rented 0.249 0.131*** 0.219 0.247 0.442*** 0.172**
Region
Tyne and Wear 0.024 0.031 0.010* 0.025 0.038 0.0283
Rest of North East 0.060 0.046 0.036* 0.064 0.057 0.040
Greater
Manchester
0.076 0.067 0.070 0.050 0.069 0.065
Merseyside 0.040 0.034 0.052 0.034 0.044 0.031
Rest of North
West
0.042 0.032 0.046 0.028 0.037 0.034
South Yorkshire 0.020 0.017 0.024 0.028 0.027 0.028
West Yorkshire 0.024 0.013** 0.018 0.014 0.015 0.025
Rest of Yorkshire
& Humberside
0.020 0.032 0.030 0.045 0.029* 0.071
East Midlands 0.016 0.018 0.014 0.025 0.019 0.014
West Midlands
Metropolitan
county
0.044 0.040 0.052 0.036 0.035 0.037
Rest of West
Midlands
0.026 0.030 0.018 0.028 0.029 0.028
East of England 0.034 0.031 0.030 0.022 0.026 0.045*
Inner London 0.016 0.021 0.006 0.017 0.022 0.023
Outer London 0.040 0.029 0.046 0.039 0.020** 0.017*
55
South East 0.082 0.087 0.108 0.010 0.051*** 0.096
South West 0.048 0.068* 0.054 0.048 0.053 0.037
Wales 0.195 0.168 0.219
0.171 0.219** 0.141
Strathclyde 0.090 0.100 0.076 0.101 0.097 0.097
Rest of Scotland 0.100 0.136** 0.088 0.123 0.114 0.144
Economic
Conditions
Employed 0.739 0.841*** 0.719 0.647 0.213*** 0.678
Local area
unemployment
rate
0.066 0.063*** 0.064 0.065 0.068*** 0.066
Year of
observation
2004 0.157 0.178 0.157 0.190 0.171 0.192
2005 0.171 0.171 0.171 0.176 0.157 0.178
2006 0.171 0.185 0.171 0.179 0.180 0.181
2007 0.159 0.175 0.159 0.149 0.175 0.150
2008 0.175 0.149 0.175 0.157 0.161 0.158
2009 0.169 0.142 0.169 0.149 0.157 0.141
Any benefit 0.139 0.054*** 0.139 0.241 0.755*** 0.243
N 498 29,778 483 354 4,117 294 Notes: All characteristics are measured at Wave 1. Treatment is defined by consistent onset/exit of WL
disability and is estimated using a NN(1) matching algorithm. *,**,*** denote significance from the treatment
group at the 10%, 5% and 1% level respectively.
56
Table A6: Descriptive Statistics for Explanatory Variables by Control and Treatment
Groups, WL, Alternative Control Groups
Onset Exit
Treatment Control
(0001)
(pre-
matching)
Control
(0001)
(post-
matching)
Treatment Control
(1110) (pre-
matching)
Control
(1110) (post-
matching)
Age 45.205 45.354 46.183 44.218 45.500 44.96
Gender
Male 0.501 0.473 0.497 0.549 0.512 0.546
Highest
qualification
Degree 0.147 0.158 0.165 0.143 0.117 0.120
Other Higher
Education
0.093 0.115 0.099 0.098 0.090 0.137
A level 0.242 0.208 0.2173
0.232 0.204 0.283
O level 0.229 0.2305 0.217 0.199 0.157 0.157
Other 0.109 0.116 0.092 0.134 0.147 0.132
None 0.181 0.172 0.179 0.193 0.286*** 0.171
Students
Full-time student 0.052 0.046 0.038 0.036 0.032 0.022
Marital Status
Single 0.268 0.232 0.243 0.249 0.318** 0.235
Married 0.588 0.617 0.591 0.599 0.525** 0.633
Widowed/divorced 0.145 0.151 0.165 0.151 0.157 0.132
Children
Dependent child in
household
0.311 0.332 0.264* 0.325 0.216*** 0.342
Housing Tenure
Owned outright 0.265 0.243 0.262 0.261 0.271 0.252
Mortgaged 0.487 0.547** 0.503 0.493 0.391*** 0.549
Rented 0.249 0.210* 0.235 0.247 0.338*** 0.199
Region
Tyne and Wear 0.024 0.029 0.026 0.025 0.042 0.028
Rest of North East 0.060 0.046 0.068 0.064 0.050 0.070
Greater
Manchester
0.076 0.074 0.062 0.050 0.057 0.081
Merseyside 0.040 0.028 0.032 0.034 0.037 0.050
Rest of North
West
0.042 0.044 0.046 0.028 0.045 0.025
South Yorkshire 0.020 0.013 0.012 0.028 0.027 0.025
West Yorkshire 0.024 0.014 0.014 0.014 0.017 0.014
Rest of Yorkshire
& Humberside
0.020 0.023 0.026 0.045 0.039 0.028
East Midlands 0.016 0.013 0.012 0.025 0.012 0.014
West Midlands
Metropolitan
county
0.044 0.042 0.030 0.036 0.044 0.022
Rest of West
Midlands
0.026 0.028 0.032 0.028 0.025 0.014
East of England 0.034 0.028 0.032 0.022 0.020 0.042
Inner London 0.016 0.020 0.006 0.017 0.015 0.020
57
Outer London 0.040 0.035 0.024 0.039 0.042 0.059
South East 0.082 0.082 0.087 0.010 0.082 0.112
South West 0.048 0.070* 0.052 0.048 0.057 0.056
Wales 0.195 0.184 0.233
0.171 0.201 0.129
Strathclyde 0.090 0.090 0.091 0.101 0.080 0.092
Rest of Scotland 0.100 0.135* 0.099 0.123 0.104 0.118
Economic
Conditions
Employed 0.739 0.772 0.744 0.647 0.458*** 0.689
Local area
unemployment
rate
0.066 0.063** 0.065 0.065 0.066 0.066
Year of
observation
2004 0.157 0.144 0.157 0.190 0.164 0.190
2005 0.171 0.138* 0.171 0.176 0.132* 0.176
2006 0.171 0.181 0.171 0.179 0.152 0.179
2007 0.159 0.213** 0.159 0.149 0.139 0.148
2008 0.175 0.168 0.175 0.157 0.211* 0.157
2009 0.169 0.156 0.167 0.149 0.202* 0.148
Any benefit 0.139 0.102** 0.139 0.241 0.473*** 0.241
N 497 1,128 365 357 402 194 Notes: All characteristics are measured at Wave 1. Treatment is defined by consistent onset/exit of WL
disability and is estimated using a NN(1) matching algorithm. *,**,*** denote significance from the treatment
group at the 10%, 5% and 1% level respectively.
58
Table A7: DID-PSM, WL Disability Onset Treatment Effects: Proportions Receiving
Welfare Benefits
Any Benefit
Onset Control (0000) Difference T stat
Wave 1 0.139 0.139 0.000 0.00
Wave 2 0.195 0.127 0.068 3.57
Wave 3 0.311 0.102 0.209 9.00
Wave 4 0.309 0.118 0.191 7.94
Difference (2-1) 0.056 -0.012 0.068 3.57
Difference (3-1) 0.173 -0.036 0.209 9.00
Difference (4-1) 0.171 -0.020 0.191 7.94
Difference (3-2) 0.117 -0.024 0.141 5.98
Difference (3-2)-(2-1) 0.060 -0.012 0.072 2.01
Non-sickness Benefit
Onset Control (0000) Difference T stat
Wave 1 0.104 0.116 -0.012 -1.46
Wave 2 0.151 0.104 0.046 2.62
Wave 3 0.197 0.082 0.114 5.88
Wave 4 0.183 0.104 0.078 3.74
Difference (2-1) 0.046 -0.012 0.058 3.35
Difference (3-1) 0.092 -0.034 0.127 6.36
Difference (4-1) 0.078 -0.012 0.090 4.16
Difference (3-2) 0.046 -0.022 0.068 3.47
Difference (3-2)-(2-1) 0.000 -0.010 0.010 0.32
Sickness Benefit
Onset Control (0000) Difference T stat
Wave 1 0.032 0.020 0.012 1.58
Wave 2 0.054 0.012 0.042 3.97
Wave 3 0.145 0.006 0.139 8.34
Wave 4 0.179 0.014 0.165 9.03
Difference (2-1) 0.022 -0.008 0.030 2.81
Difference (3-1) 0.112 -0.014 0.127 7.23
Difference (4-1) 0.147 -0.006 0.153 8.10
Difference (3-2) 0.090 -0.006 0.096 5.76
Difference (3-2)-(2-1) 0.068 0.002 0.066 3.01
IB or ESA
Onset Control (0000) Difference T stat
Wave 1 0.010 0.002 0.008 1.63
Wave 2 0.028 0.002 0.026 3.47
Wave 3 0.094 0.000 0.094 7.01
Wave 4 0.114 0.006 0.108 7.22
Difference (2-1) 0.018 0.000 0.018 2.41
Difference (3-1) 0.084 -0.002 0.086 6.48
Difference (4-1) 0.104 0.004 0.100 6.80
Difference (3-2) 0.066 -0.002 0.068 5.18
Difference (3-2)-(2-1) 0.048 -0.002 0.050 2.99 Notes: See notes to Table 1. ATT are based on a NN(1) matching algorithm and are estimated over the region of
common support. T statistics are based on Abadie and Imbens (2006) standard errors.
59
Table A8: DID-PSM, WL Disability Onset Treatment Effects: Proportions Receiving
Welfare Benefits, Alternative Control Group
Any Benefit
WL Onset Control (0001) Difference T stat
Wave 1 0.139 0.139 0.000 0.00
Wave 2 0.195 0.163 0.032 1.49
Wave 3 0.312 0.191 0.121 4.48
Difference (2-1) 0.056 0.024 0.022 1.49
Difference (3-1) 0.173 0.052 0.121 4.48
Difference (3-2) 0.117 0.028 0.089 3.21
Non-sickness Benefit
WL Onset Control (0001) Difference T stat
Wave 1 0.104 0.113 -0.008 -0.74
Wave 2 0.151 0.135 0.016 0.76
Wave 3 0.197 0.141 0.056 2.41
Difference (2-1) 0.046 0.022 0.024 0.92
Difference (3-1) 0.093 0.028 0.064 2.55
Difference (3-2) 0.046 0.006 0.040 1.68
Sickness Benefit
WL Onset Control (0001) Difference T stat
Wave 1 0.032 0.012 0.012 1.01
Wave 2 0.054 0.024 0.024 1.56
Wave 3 0.145 0.052 0.093 4.30
Difference (2-1) 0.022 0.010 0.012 0.91
Difference (3-1) 0.113 0.032 0.081 4.05
Difference (3-2) 0.091 0.022 0.068 3.26
IB or ESA
WL Onset Control (0001) Difference T stat
Wave 1 0.010 0.006 0.004 0.65
Wave 2 0.028 0.010 0.018 1.99
Wave 3 0.095 0.034 0.060 3.69
Difference (2-1) 0.018 0.004 0.014 1.75
Difference (3-1) 0.085 0.028 0.056 3.57
Difference (3-2) 0.066 0.024 0.042 2.70 Notes: See notes to Table 1. ATT are based on a NN(1) matching algorithm and are estimated over the region of
common support. T statistics are based on Abadie and Imbens (2006) standard errors.
60
Table A9: Heterogeneity in the DID-PSM (wave 3-1, wave 3-2), WL Disability Onset
Treatment Effects: Proportions Receiving Welfare Benefits
DID
Any Benefit Non-sickness Benefit Sickness Benefit IB/ESA
DID_PSM 3-1 3-2 3-1 3-2 3-1 3-2 3-1 3-2
Male 0.188 0.120 0.096 0.036 0.120 0.104 0.088 0.068
Female 0.214 0.113 0.145 0.060 0.113 0.085 0.081 0.060
Low Qual 0.240 0.155 0.143 0.062 0.136 0.120 0.093 0.074
High Qual 0.138 0.104 0.067 0.042 0.096 0.071 0.075 0.058
Older 0.211 0.147 0.107 0.060 0.120 0.090 0.077 0.054
Younger 0.186 0.121 0.121 0.055 0.126 0.095 0.106 0.085
Mental (treatment
n=56) 0.304 0.143 0.214 0.071 0.196 0.107 0.196 0.089
Physical 0.158 0.109 0.087 0.041 0.092 0.087 0.061 0.061
Single 0.123 0.057 0.090 0.024 0.052 0.043 0.038 0.024
Multiple 0.243 0.148 0.109 0.046 0.176 0.134 0.120 0.099
Pre-2009 0.178 0.101 0.073 0.020 0.162 0.024 0.134 0.117
Post-2009 0.156 0.104 0.108 0.032 0.056 0.044 0.044 0.024
Unemploy Q1 0.183 0.122 0.061 0.017 0.139 0.078 0.078 0.052
Unemploy Q2 0.187 0.151 0.086 0.072 0.101 0.115 0.094 0.086
Unemploy Q3 0.262 0.146 0.208 0.092 0.123 0.077 0.092 0.054
Unemploy Q4 0.140 0.105 0.105 0.070 0.105 0.088 0.079 0.070
Employed 0.201 0.133 0.109 0.054 0.133 0.106 0.092 0.079
Not employed 0.208 0.131 0.100 0.085 0.123 0.054 0.092 0.046
Any benefits 0.088 -0.059 0.103 -0.074 0.132 0.103 0.103 0.118
No benefits 0.214 0.131 0.114 0.056 0.121 0.089 0.084 0.061
Long-term health 0.166 0.103 0.108 0.040 0.072 0.067 0.040 0.036
No long-term health 0.257 0.173 0.151 0.088 0.147 0.107 0.121 0.096
Past health 0.163 0.184 0.102 0.122 0.122 0.061 0.102 0.061
No past health 0.212 0.111 0.127 0.029 0.117 0.101 0.091 0.081 Notes: See notes to Table 5. Samples are defined on the basis of information in wave 1 or at onset as
appropriate. Bold indicates statistically significant from zero at the 95% confidence level.
61
Table A10: DID-PSM, WL Disability Exit Treatment Effects: Proportions Receiving
Welfare Benefits
Any Benefit
WL Exit Control (1111) Difference T stat
Wave 1 0.243 0.243 0.000 0.00
Wave 2 0.254 0.328 -0.074 -2.39
Wave 3 0.201 0.410 -0.209 -6.45
Wave 4 0.172 0.415 -0.243 -7.47
Difference (2-1) 0.011 0.085 -0.073 -2.39
Difference (3-1) -0.042 0.166 -0.209 -6.45
Difference (4-1) -0.071 0.172 -0.243 -7.47
Difference (3-2) -0.054 0.082 -0.136 -4.36
Difference (3-2)-(2-1) -0.065 -0.003 -0.062 -1.18
Non-sickness Benefit
WL Exit Control (0000) Difference T stat
Wave 1 0.153 0.088 0.065 3.61
Wave 2 0.172 0.133 0.040 1.50
Wave 3 0.153 0.150 0.003 0.11
Wave 4 0.113 0.141 -0.028 -1.05
Difference (2-1) 0.020 0.045 -0.025 -1.04
Difference (3-1) 0.000 0.062 -0.062 -2.56
Difference (4-1) -0.040 0.054 -0.093 -3.46
Difference (3-2) -0.020 0.017 -0.037 -1.33
Difference (3-2)-(2-1) -0.040 -0.028 -0.011 -0.24
Sickness Benefit
WL Exit Control (0000) Difference T stat
Wave 1 0.119 0.203 -0.085 -5.30
Wave 2 0.107 0.266 -0.158 -5.84
Wave 3 0.059 0.333 -0.274 -9.15
Wave 4 0.071 0.359 -0.288 -9.55
Difference (2-1) -0.011 0.062 -0.073 -2.79
Difference (3-1) -0.059 0.130 -0.189 -6.10
Difference (4-1) -0.048 0.155 -0.203 -6.66
Difference (3-2) -0.048 0.068 -0.116 -4.24
Difference (3-2)-(2-1) -0.037 0.006 -0.042 -0.97
IB or ESA
WL Exit Control (0000) Difference T stat
Wave 1 0.073 0.127 -0.053 -3.12
Wave 2 0.067 0.166 -0.099 -4.52
Wave 3 0.037 0.234 -0.198 -8.08
Wave 4 0.034 0.232 -0.198 -7.95
Difference (2-1) -0.006 0.040 -0.045 -2.24
Difference (3-1) -0.036 0.107 -0.144 -5.81
Difference (4-1) -0.040 0.105 -0.144 -5.52
Difference (3-2) -0.031 0.068 -0.099 -4.41
Difference (3-2)-(2-1) -0.025 0.028 -0.054 -1.55 Notes: See notes to Table 1. ATT are based on a NN(1) matching algorithm and are estimated over the region of
common support. T statistics are based on Abadie and Imbens (2006) standard errors.
62
Table A11: DID-PSM, WL Disability Exit Treatment Effects: Proportions Receiving
Welfare Benefits, Alternative Control Group
Any Benefit
WL Exit Control (1110) Difference T stat
Wave 1 0.241 0.241 0.000 0.00
Wave 2 0.252 0.277 -0.025 -0.75
Wave 3 0.199 0.289 -0.090 -2.69
Difference (2-1) 0.011 0.036 -0.025 -0.75
Difference (3-1) -0.042 0.048 -0.090 -2.69
Difference (3-2) -0.053 0.011 -0.064 -1.81
Non-sickness Benefit
WL Exit Control (1110) Difference T stat
Wave 1 0.151 0.129 0.022 2.38
Wave 2 0.171 0.137 0.037 1.32
Wave 3 0.151 0.154 -0.003 0.09
Difference (2-1) 0.020 0.008 0.011 -1.02
Difference (3-1) 0.000 0.025 -0.025 -2.32
Difference (3-2) -0.020 0.017 -0.036 -1.13
Sickness Benefit
WL Exit Control (1110) Difference T stat
Wave 1 0.118 0.137 -0.020 -0.83
Wave 2 0.106 0.188 -0.081 -3.00
Wave 3 0.059 0.185 -0.126 -4.65
Difference (2-1) -0.011 0.050 -0.062 -2.09
Difference (3-1) -0.059 0.048 -0.106 -3.52
Difference (3-2) -0.048 -0.003 -0.045 -2.03
IB or ESA
WL Exit Control (1110) Difference T stat
Wave 1 0.073 0.087 -0.014 -0.63
Wave 2 0.067 0.118 -0.050 -2.08
Wave 3 0.036 0.120 -0.084 -3.54
Difference (2-1) -0.006 0.031 -0.036 -1.49
Difference (3-1) -0.036 0.034 -0.070 -2.67
Difference (3-2) -0.031 0.003 -0.034 -1.54 Notes: See notes to Table 1. ATT are based on a NN(1) matching algorithm and are estimated over the region of
common support. T statistics are based on Abadie and Imbens (2006) standard errors.
63
Table A12: Heterogeneity in the DID-PSM (wave 3-1, wave 3-2), WL Disability Exit
Treatment Effects: Proportions Receiving Welfare Benefits
DID
Any Benefit Non-sickness
Benefit
Sickness Benefit IB/ESA
DID-PSM 3-1 3-2 3-1 3-2 3-1 3-2 3-1 3-2
Male -0.201 -0.103 -0.072 -0.026 -0.180 -0.108 -0.082 -0.031
Female -0.119 -0.088 -0.006 0.006 -0.150 -0.100 -0.056 -0.044
Low Qual -0.155 -0.059 -0.032 -0.011 -0.193 -0.075 -0.128 -0.075
High Qual -0.226 -0.125 -0.048 -0.018 -0.202 -0.113 -0.131 -0.083
Older -0.213 -0.099 -0.079 -0.040 -0.178 -0.084 -0.094 -0.035
Younger -0.234 -0.058 -0.091 -0.019 -0.221 -0.045 -0.169 -0.052
Mental (treatment
n=40) -0.100 -0.050 -0.025 -0.125 -0.125 -0.050 -0.075 0.000
Physical -0.189 -0.096 -0.075 -0.039 -0.167 -0.089 -0.100 -0.050
Single -0.115 -0.028 -0.011 0.050 -0.126 -0.060 -0.093 -0.060
Multiple -0.231 -0.121 -0.064 -0.058 -0.179 -0.081 -0.121 -0.075
Unemploy Q1 -0.209 -0.174 -0.047 -0.058 -0.209 -0.174 -0.093 -0.070
Unemploy Q2 -0.228 -0.119 -0.020 0.010 -0.257 -0.168 -0.188 -0.109
Unemploy Q3 -0.186 -0.070 -0.023 -0.023 -0.186 -0.047 -0.151 0.012
Unemploy Q4 -0.203 -0.089 -0.203 -0.152 -0.215 -0.089 -0.101 0.000 Notes: See notes to Table 5. Samples are defined on the basis of information in wave 1 or at exit as appropriate. Bold indicates statistically significant from zero at the 95% confidence level.
64
Table A13: DID-PSM, DDA disability onset, sensitivity to matching method
Any Benefit
NN(5) LL Kernel
Difference T stat Difference T stat Difference T stat
Difference (2-1) 0.027 2.20 0.026 1.72 0.035 3.03 Difference (3-1) 0.088 6.00 0.091 5.14 0.100 7.22 Difference (4-1) 0.102 6.43 0.105 5.55 0.116 7.74
Non-sickness Benefit
NN(5) LL Kernel
Difference T stat Difference T stat Difference T stat
Difference (2-1) 0.018 1.57 0.019 1.34 0.026 2.54 Difference (3-1) 0.042 3.29 0.046 2.98 0.054 4.54 Difference (4-1) 0.041 3.05 0.046 2.81 0.056 4.51
Sickness Benefit
NN(5) LL Kernel
Difference T stat Difference T stat Difference T stat
Difference (2-1) 0.019 2.55 0.018 2.04 0.019 2.63 Difference (3-1) 0.077 7.10 0.077 6.53 0.079 7.49 Difference (4-1) 0.109 8.86 0.106 7.75 0.108 8.94
IB or ESA
NN(5) LL Kernel
Difference T stat Difference T stat Difference T stat
Difference (2-1) 0.007 1.25 0.007 1.16 0.007 1.38 Difference (3-1) 0.051 5.77 0.051 5.43 0.051 5.86 Difference (4-1) 0.070 6.89 0.069 6.53 0.069 6.88 Notes: Alternative matching estimators include using the 5 nearest neighbours NN(5), local linear regression
(LL) and kernel density (Kernel).
65
Table A14: DID-PSM, DDA disability onset, sensitivity to controls for health
Any Benefit Wave 1 Health Past Health
Control Exact Match Control Exact Match
Difference T stat Difference T
stat
Difference T
stat
Difference T
stat
Difference (2-1) 0.034 2.24 0.011 0.68 0.030 1.81 0.030 1.68
Difference (3-1) 0.099 5.65 0.074 4.15 0.084 4.25 0.102 5.00
Difference (4-1) 0.114 6.00 0.072 3.77 0.086 4.07 0.073 3.40
Non-sickness Benefit
Wave 1 Health Past Health
Control Exact Match Control Exact Match
Difference T stat Difference T
stat
Difference T stat Difference T
stat
Difference (2-1) 0.021 1.47
-0.005
-
0.38
0.020 1.24
0.025 1.47
Difference (3-1) 0.053 3.39 0.032 2.06 0.039 2.22 0.063 3.29
Difference (4-1) 0.053 3.10 0.007 0.41 0.029 1.52 0.029 1.49
Sickness Benefit
Wave 1 Health Past Health
Control Exact Match Control Exact Match
Difference T stat Difference T
stat
Difference T stat Difference T
stat
Difference (2-1) 0.018 2.09 0.026 2.79 0.018 1.76 0.023 2.12
Difference (3-1) 0.072 6.05 0.074 5.88 0.075 5.63 0.082 5.86
Difference (4-1) 0.108 8.25 0.111 7.82 0.093 6.21 0.097 6.08
IB or ESA
Wave 1 Health Past Health
Control Exact Match Control Exact Match
Difference T stat Difference T
stat
Difference T
stat
Difference T
stat
Difference (2-1) 0.005 0.92 0.009 1.57 0.007 0.99 0.013 1.62
Difference (3-1) 0.047 5.27 0.049 5.22 0.057 5.13 0.057 5.04
Difference (4-1) 0.070 6.86 0.070 6.59 0.063 5.19 0.061 4.90 Notes: See notes to Table 1. ATT are based on a NN(1) matching algorithm and are estimated over the region of
common support. T statistics are based on Abadie and Imbens (2006) standard errors.
66
Table A15: DID-PSM, DDA disability exit, sensitivity to matching method
Any Benefit
NN(5) LL Kernel
Difference T stat Difference T stat Difference T stat
Difference (2-1) -0.127 -5.92 -0.120 -4.21 -0.116 -6.41
Difference (3-1) -0.191 -7.35 -0.199 -5.95 -0.194 -8.44
Difference (4-1) -0.206 -8.27 -0.204 -6.16 -0.201 -9.35
Non-sickness Benefit
NN(5) LL Kernel
Difference T stat Difference T stat Difference T stat
Difference (2-1) -0.055 -2.77 -0.065 -2.62 -0.060 -3.22
Difference (3-1) -0.075 -3.16 -0.090 -3.12 -0.083 -3.74
Difference (4-1) -0.091 -3.94 -0.107 -3.70 -0.102 -4.78
Sickness Benefit
NN(5) LL Kernel
Difference T stat Difference T stat Difference T stat
Difference (2-1) -0.092 -4.65 -0.090 -3.26 -0.089 -4.96
Difference (3-1) -0.158 -7.21 -0.162 -5.33 -0.161 -8.15
Difference (4-1) -0.162 -7.49 -0.154 -5.05 -0.155 -8.17
IB or ESA
NN(5) LL Kernel
Difference T stat Difference T stat Difference T stat
Difference (2-1) -0.068 -3.89 -0.063 -2.51 -0.064 -3.95
Difference (3-1) -0.099 -5.03 -0.102 -3.84 -0.105 -5.64
Difference (4-1) -0.115 -5.97 -0.101 -3.64 -0.105 -5.84 Notes: Alternative matching estimators include using the 5 nearest neighbours NN(5), local linear regression
(LL) and kernel density (Kernel).
Top Related