ESTIMATING THE LABOUR SUPPLY EFFECT OF THE WORKING TAX CREDIT FOR

ESTIMATING THE LABOUR SUPPLY EFFECT OF THE WORKING TAX CREDIT FOR CHILDLESS HOUSEHOLDS IN THE UK Ian Mulheirn and Mario Pisani * The Working Tax Credit (WTC) was introduced in April 2003. This was the first time a national programme of in-work financial support for childless people had been attempted in the UK. Childless claimants become eligible for support from the age of 25. This paper exploits the age requirement of the policy, using a difference-in-differences approach, to assess its impact on labour supply, in terms of both participation and hours worked. We find evidence of a positive 1.9 to 2.9 percentage point effect on labour market participation amongst eligible groups, which can be confidently described as robust and significant. At the intensive margin there is evidence that WTC causes only a small reduction in hours worked, although even this is dependent upon the hours measure used. Overall the results indicate a positive effect on labour supply. The paper also sets WTC in the context of other changes that occur at 25, including greater generosity of Housing Benefit and Income Support, and the end to eligibility for the New Deal for Young People. Regression Discontinuity analysis shows evidence of a net negative impact of these policies on participation around age 25, in the years prior to 2003. Repeating the analysis after policy introduction shows that the discontinuity has disappeared: this evidence suggests that WTC has helped to ameliorate this policy-driven deterioration of work incentives, which existed before its introduction. 1. Introduction Two well-known findings of labour economics are that long spells of unemployment worsen future labour market outcomes, while out-of-work benefits reduce incentives to work. To tackle both these effects simultaneously, in April 2003 the British Government introduced a national policy of in-work support for people without children for the first time. It was part of the Working Tax Credit (WTC), which also provides separate in-work support to families with children. Eligibility for this in-work support begins when childless claimants are 25 years old. However, two other policies that provide support, mainly to those out of work, also become significantly more generous once the claimant reaches 25. Consequently the policy-driven work incentives for this age-group pull in opposing directions. We would like to thank various member of staff at University College London, Her Majesty’s Treasury, and Her Majesty’s Revenue and Customs for their advice, comments and support. All errors remain the authors’ own. Views expressed in this research are made in a personal capacity. Both authors: HM Treasury, 1 Horse Guards Road, London, SW1A 2HQ. Corresponding author: Ian Mulheirn, ian.mulheirn@hm-treasury.gov.uk JEL classification: J38, J22, I38. * This paper aims to analyse the impact of WTC on the labour supply of childless households by exploiting the age-eligibility requirement, effectively using it as a natural experiment and exogenous discontinuity. Section 2 describes the policy, presents a brief theoretical framework, and reviews related empirical literature. Section 3 describes the estimation methods: difference-in-differences techniques, used to isolate the impact of WTC on labour supply decisions regarding participation and hours worked; and local linear regression to assess the net effect of the policies for which eligibility changes at the age of 25. Section 4 describes the data used, the Labour Force Survey (LFS). Section 5 reports the results and section 6 concludes. 2. Description of the policy environment, theoretical pointers, and existing evidence In-work support for families with children was introduced in Britain in 1971 with the Family Income Supplement, followed with the introduction of Family Credit in 1988 and Working Families’ Tax Credit (WFTC) in 19991. The Working Tax Credit (WTC) was introduced, alongside Child Tax Credit (CTC), in April 2003. These new tax credits replaced the previous patchwork of support available, the largest element of which was WFTC. WTC departs from the previous tax credits in a number of ways2, one of the most significant of which is the extension of in-work support to families without children (single persons or couples) working at least 30 hours per week. The aim of this aspect of the policy was to “tackle the problem of persistent low income among working people without children”. The UK Government estimated that, in 2002, there were around one million adults without children working at least 16 hours per week but in households with income below 60% of the UK household median. Eligibility among this group was restricted to those aged 25 or over. For a single person over 24 years-old and working 30-or-more hours per week WTC in 2004/05 was worth £2,210 per year. This is the tapered away at a rate of 37% beyond a household income of £5,060. This means that an individual could receive WTC on an income of up to £11,000 per year. For a couple the maximum support was £3,755 per A non-national precedent exists: in 1996 DWP launched the Earnings Top-Up pilots, which aimed to subsidise the wages of low-paid workers without children. The evaluation found no noticeable impact. 2 The Child and Working Tax Credits: The Modernisation of Britain’s Tax and Benefit System, Number Ten (April 2002) 1 2 year, enabling childless couples, with one person working 30 hours or more, to receive support on joint incomes of over £15,000. Only for those earning below the Minimum Wage is it possible to receive the maximum WTC entitlement (i.e. self employed people). The nature of the WTC 30-hour requirement creates a discontinuity in the budget constraint of low-paid workers. As the hourly wage rises, the amount of WTC remaining at 30 hours decreases. For sufficiently well-paid individuals (over £7 per hour), the credit provides no additional incentive to work, since all of the award is withdrawn by the time they have worked 30 hours. In this way the policy targets the biggest incentives at the group most likely to face poor work incentives: low productivity workers. Chart 1, below, shows what WTC does to the budget constraint of a 25-year-old earning the National Minimum Wage (£4.20 per hour at the time of WTC introduction): it creates a jump in disposable income for those moving from 29 to 30 hours. As can be seen in chart 1, there are other benefits for those working less than 16 hours per week. Income Support (IS) provides £44.50 per week for a single person aged 18-24. Once the individual is 25 they become eligible for the higher single person rate of £56.20 per week. The award includes an earnings disregard so that a person can earn a small amount without having any of their award withdrawn; after that the entitlement is reduced by £1 for every £1 earned until the entitlement is zero or the person is working 16 hours, at which point any remaining entitlement is withdrawn. Another major source of support is housing benefit (HB). Unemployed under-25s are entitled to claim for the entirety of their housing rental costs up to the market rental value of a single room in a shared house. This is typically about £60 per week. Once 25, the same individual can claim the full value of their housing cost up to some level deemed reasonable by the local authority. As a result, an individual living alone will be able to get far greater financial support through HB once 25. Claimants are entitled to maximum HB (the full value of their rent) provided they continue to receive IS. Once IS has been withdrawn HB starts to be taken away at a rate of 65%. The increased generosity of each of these policies for claimants aged 25-and-over significantly lowers the relative return from entering employment. 3 Chart 1: budget constraint for a 25-year-old minimum wage earner 200 180 160 Disposable Income (£/week) 140 120 100 CTB 80 60 40 20 0 0 12 24 Hours 36 48 HB IS WTC Take-home A final element of the policy environment is the end to eligibility for the New Deal for Young People (NDYP) at age 25. NDYP provides unemployed young people with up to 12 months’ subsidized on-the-job training or education. The impact of NDYP at 25 is mitigated by the existence of a (albeit less intensive) New Deal 25+ programme, and by the fact that many continue to participate in NDYP if they qualified for the programme when 24. The theoretical rational for such policies are clear: governments recognise the benefits of providing some minimal financial safety net for all members of society including those of working age. However such a safety net involves the concomitant dilemma of how to withdraw the financial support as an individual progresses up the income distribution: high rates of withdrawal damage work incentives, while low rates cost more money as they involve making payments to people further up the income distribution who were not the initially intended recipients of the support. Long tapers also have the side-effect of putting more people on high marginal deduction rates and therefore reducing their work incentives. 4 The WTC is based upon household income which means that, strictly, it differs from a wage subsidy. However, for individuals (and childless couples if we are happy to treat these as one unit) WTC can be considered to be a wage subsidy, which can be analysed in the same way as the negative income tax first proposed by Milton Friedman in 1962. The basic idea of wage subsidy programmes such as negative taxation and WTC is to reduce the withdrawal rate of financial support to those on low incomes, thus reducing the effective tax rate on each extra pound earned, and improving the incentives to work. The lower effective tax rate means an individual is able to reach a higher indifference curve: the increase in the return to work generates a positive substitution effect (of consumption for leisure) that outweighs any negative income effect that might arise due to recipients finding that they can achieve the same income as before the policy, while spending less time at work. The cost of improving work incentives by this means is, however, dependent upon the level of minimum income guaranteed to individuals who do not work. With higher levels, for any given withdrawal rate, the negative income tax will involve people further up the income distribution. It is often argued that the immediate cost of such a scheme is outweighed by the impact it would have on labour supply. This would be more likely to be true if the labour supply effects of wage subsidies were unambiguously positive, but this is not the case. The reduction of the 100% effective tax rate for benefit recipients facing a pound-forpound reduction in their entitlement unambiguously improves the work incentives faced by this group. They may now decide to work for some positive number of hours per week, where this was not in their interests under the pound-for-pound withdrawal regime. The change for this individual is represented by the blue indifference curves in chart 2. The incentive is to increase labour market participation as shown by arrow C. This occurs because the financial support is unavailable to those not working, so there is no negative income effect for those working zero hours, while the substitution effect is positive. The net effect on participation should therefore be unambiguously positive. 5 However, for those not previously on any form of negative tax or tax credit taper, it is now possible to increase their earnings and reduce their labour market participation. This unambiguously negative effect on labour supply at the intensive margin occurs because both the income and substitution effects are working in the same, negative direction. Arrow D in chart 2 illustrates the move to the higher red indifference curve that may be undertaken by someone on this part of the income distribution. Finally, for those above the newly-extended taper of a negative income tax policy, there may be an incentive to reduce their income and hours of work, and move onto the taper - a move represented by arrow E. Whether this occurs depends entirely on the tastes of the individuals and their willingness to substitute income for leisure at the margin. Chart 2: The ambiguous impact of lower withdrawal rates. The net effect on labour supply of lowering the withdrawal rate by extending the taper up the income scale is ambiguous, while it is unambiguously positive on participation. Consequently the justification for such a policy aimed at increasing labour supply must rest upon either empirical observations that demonstrate the net effect to be positive, or a societal welfare function that weights the benefits of incentivising those at the bottom of the distribution more than the costs of disincentivising those further up. However such a policy would be theoretically justifiable if the priority was to increase participation. 6 Under the Working Tax Credit for people without children, the picture is complicated slightly by the existence of a 30-hour work requirement before the individual or couple qualify for the payable credit. This has a number of interesting effects. Firstly, the fact that (in 2003/04) the credit is withdrawn at a rate of 37 pence for each pound earned above £5,060 per year, means that the system does not put higherproductivity (and therefore higher hourly-paid) workers onto any incentive-damaging taper since their eligibility will be zero by the time they reach the 30-hour qualification. Denying these high productivity workers support is justified by the reasoning that if their higher earning power is insufficient to incentivise them to work, then the costs of providing them with a WTC incentive would probably outweigh the benefits, as laid out above. Secondly, it provides positive work incentives to a large group of people, namely those on low hourly pay rates who work less than 30 hours. The size of the subsidy when the claimant work 30 hours provides a significant cliff-edge in the claimant’s favour for those on hourly wages of up to around £7. Chart 1 shows the effective budget constraint for an individual on the 2003 minimum wage of £4.20 per hour. Compared to a strictly income-based credit, in which the effect on intensive labour supply is predicted to be unambiguously negative, the 30-hour requirement makes the net effect on average hours among those in work ambiguous. In the presence of WTC those previously working less than 30 hours are incentivised to move above that threshold, while those working over 30 can reduce their hours towards that point without being worse-off than they were before the introduction of the policy. This element of the design of the system makes it possible that the effect on intensive labour supply will be positive. Similar theoretical predictions, for other child-dependent policies, have been empirically tested in the past. Most are attempts to evaluate the impact of WFTC on the labour supply of people with children, focussing on the impact on lone parents. Brewer et al (2003, 2005) use a structural model of individuals’ preferences to analyse the effect of WFTC on labour supply and take-up: they estimate that WFTC increased the labour supply of lone mothers by around 5.11 percentage points, reduced the labour supply of mothers in couples by 0.57 percentage points, and raised that of fathers in couples by about 0.75 7 percentage points. They also conclude that the net effect of all other tax and benefit reforms of the period ran counter to the generally positive labour supply impact of WFTC. Francesconi and Van der Klaauw (2004) used longitudinal data from the British Household Panel Survey to analyse the labour supply impact of WFTC. The authors use a difference-in-differences (DiD) approach to isolate the policy effect: the analysis indicates that WFTC increased the proportion of lone mothers working 16 hours or more by about 7 percentage points, with most of the increase being for full-time work. Gregg and Harkness (2003) examined the package of reforms introduced in 1998/99 to assess their impact on lone parent employment and hours worked. Again the authors use a difference-in-differences approach, but in order to control for differences in observed characteristics between lone parents and the control group, the authors use propensity score matching. They find a positive 4.6 percentage point effect on lone parent employment rates. Given the Brewer et al finding of a net negative effect on lone parent employment of reforms other than WFTC in this period, this finding supports that of Francesconi and Van der Klaauw in suggesting WFTC had a large positive impact on this group. Research into in-work support in the United States, such as Eissa and Liebman (1996), finds that it caused lone mothers’ participation increased by 2.8 percentage points. Contrary to what theory would predict for this sort of in-work credit (without an hoursrequirement), they find no evidence of a fall in hours worked by those in employment. This is in line with much past labour supply literature, which finds that hours worked are less responsive than participation to changes in the net wage. 3. Evaluation techniques Our analysis employs two estimation techniques: Difference-in-Differences (DiD) and Regression Discontinuity Design (RD). These are discussed in turn. 8 Difference-in-differences estimation3 This technique exploits the start of the policy to compare the outcomes of the target group before and after the policy launch. Straight-forward ‘before-and-after’ estimation assumes the counterfactual post-policy outcomes to be the same as for those observed before the start of the policy. This will obviously not be the case if macroeconomic changes affect the outcome of all those observed. Consequently DiD uses a comparator group, unaffected by the policy, to strip out any macroeconomic effects. Effectively this technique compares the before-and-after effects on the treated group with those for the untreated group. Assuming that the comparator group is a sufficiently close match, the remaining effect can be considered to be the impact of the policy alone on the treated individuals. If the comparator group is not a close match it may be the case that macroeconomic changes affect the treated and untreated groups differently. This would bias the estimate of the policy effect. In the quasi-experimental setting of the national launch of WTC it is impossible to find a group that matches the treated in all aspects but the fact of their treatment (i.e. an experimental control group). However, if one is prepared to assume that macroeconomic fluctuations have the same impact on the labour market prospects of 22 to 24-year-olds as on those of 25 to 27-year-olds, then the younger group can be used as a control group in the estimation. This assumption means that, pre-policy, changes in employment probability for those in the control group should not be significantly different from changes for those in the treatment group. The other identifying assumption needed in DiD is that the policy to be evaluated only affects those in the treatment group. The validity of these assumptions will be tested in section 5. The raw or basic DiD estimator is given by equation 1, where y is the outcome (probability of being in employment or number of hours worked each week). Subscript T represents those individuals and couples over 25, the treated group, while C represents Numerous recent papers use DiD techniques to analyse UK labour market outcomes, two good examples are Connolly and Gregory (2002) and Stewart and Swaffield (2004, 2005). 3 9 those under 25, the control group. Subscript A represents those observations taken after policy implementation (April 2003), and B those taken before that time. θ = ( yTA − yTB ) − ( yCA − yCB ) ∧ (1) This estimator can be generated using the basic DiD regression model shown in equation 2. Here γ is a coefficient on the time dummy Tt , which is positive after the introduction of the policy and zero before. Coefficient δ is for the treatment dummy Ai , which takes the value one if the respondent is 25-or-over and zero otherwise. Finally, θ is the DiD estimator, and ε it is an i.i.d. variable with a mean value of zero. yit = α + γTt + δAi + θ (T ⋅ A) it + ε it (2) We can make two improvements to this specification aimed at generating a closer fit. First, by using a vector of individual characteristics Zit it is possible to create a ‘regression adjusted’ DiD estimator (3). This takes into account any differences between the treatment and comparison group not already accounted for by the time and treatment dummies. The vector of characteristics also includes a cubic polynomial function of the age of the respondent. yit = α + βZ it + γTt + δAi + θ (T ⋅ A) it + ε it (3) Secondly, to more accurately control for time variation we can use data on interview dates to generate a deterministic trend variable that changes daily. This allows us to replace the simplistic pre/post time dummy with any given function of the trend variable. In equation (4) this variable is named DTt , and its cubic function is used to replace Tt . The advantages of such a functional form are intuitively attractive: it measures time more closely, allowing for changes in macroeconomic conditions, and it allows for the rate of change of time-variation to differ across time. yit = α + βZ it + γ 1 DTt + γ 2 DTt 2 + γ 3 DTt 3 + δAi + θ (T ⋅ A) it + ε it (4) 10 To overcome the usual problems encountered when estimating binary dependent-variable models by linear methods, this analysis uses a probit model to analyse participation. In the analysis of intensive labour supply, OLS is used. In both cases, the DiD estimator θ can be interpreted as measuring the average treatment effect on the treated group. Here the treated group only covers 25 to 27 year-olds, so one may conclude that the estimator is not measuring the average effect on all possible beneficiaries, as some are older than this. However, in the absence of close control groups for all other ages, the quasi-experimental design created by the age-eligibility criteria is the only way to measure the average effect on recipients. This effect should, on average, apply to all age groups – although the validity of this statement cannot be evaluated, as the counterfactual can neither be observed nor derived. Other techniques, beyond the scope of this paper, may be able to address this question. Furthermore, DiD cannot measure the net effect of all the eligibility changes that occur at age 25. The local linear regression approach, by demonstrating discontinuities in a trending line, can illustrate the net impact on labour supply. Regression Discontinuity and Local Linear Regression Since eligibility for WTC occurs for all childless individuals or couples from the age of 25, there is a sudden change in the budget constraint of low wage people once they cross this threshold. The Regression Discontinuity (RD) technique exploits this fact. It can be effectively applied in a situation where an individual’s treatment depends upon an observable characteristic, in this case age, where there is a known point at which the probability of treatment changes discontinuously. The idea is that in the neighbourhood of the discontinuity, treatment is effectively randomly assigned. It is for this reason that RD is sometimes referred to as a quasi-experimental design. The RD approach can be used to estimate the treatment effect of the various policy changes at 25 in two ways. One method is to use a parametric approach employing a treatment dummy taking a value of one for those respondents who are 25 or over. However, the design of the DiD model discussed above effectively incorporates such an approach within it. Furthermore, assuming a functional form for the data in this way will 11 lead to underestimation of the impact of the changes at 25 if there are lags and/or anticipation effects at work. Consequently a non-parametric method is employed here. Non-parametric techniques avoid the need for imposing a pre-determined structure on the data and allow the data flexibility to characterise its own shape. Such a technique presents a trade-off regarding the bandwidth to be used for smoothing: with an insufficient sample size a small bandwidth causes the local regression to suffer from noise that obscures the underlying function. Too large a bandwidth, on the other hand, can result in estimates of the underlying function being biased if observations at the extremes of the local regression are systematically different from those at its centre. A key point to note here is that analysing a regression discontinuity at age 25 means taking into account all policies for which eligibility changes at that age. This technique therefore measures the net impact of all policies that affect labour supply and for which eligibility changes at age 25. Consequently the results will measure the net impact of the increased Housing Benefit and Income Support, and the end of NDYP, as well as WTC eligibility at 25. For this application, since treatment is a deterministic function of age, the policy changes represent a ‘sharp’ discontinuity: the probability of being treated jumps from zero to one as an individual turns 25. Moving up the age groups from 20 to 33 (the age range of the sample used) there may be other effects at play. For example it may be the case that the sample is not randomly selected due to the fact that living arrangements and fertility decisions may be endogenous and the group of single childless people at 33 is systematically different to that at age 20. However, as long as these effects occur as a smooth function of age, the existence of a discontinuity at 25 can be put down to the local impact of the policy. RD can be invalidated if the observable characteristic on which eligibility is determined is able to be manipulated through people misreporting their age. With social security numbers being used to determine age, this does not present a threat to the validity of the technique in this application. As with DiD, a possible further threat is posed by general 12 equilibrium effects, but for the reasons given below this is not thought to be a significant effect. By performing a local linear regression (as suggested by Fan 1992) on each side of the age threshold it is possible to estimate the function g in regression equation (5) for the group under 25 and for those 25-and-over. Given the changes in the policy environment at 25, it might be anticipated that the result will show a discontinuity between the two regression functions. yi = g ( xi )+ ε i (5) The function, g, is estimated for the whole sample by finding the local linear estimate of employment probability given age: E [ yi | x ] = g ( x) ∧ (6) where ∧  n 2  g ( x)  = arg min ∧ ∑ ( yi − g ( zi − z0 ) ) K  zi − z0 1( zi − z0 ) .  h  g ( x ) i =1       (7) In expression (7), h represents the bandwidth used to smooth the regression. z0 represents the value of the regressor at which estimation is performed and zi other points away from z0 that influence the estimation of the outcome at z0 . The bandwidth determines the maximum distance of observations away from z0 that contribute to the estimation. This procedure is repeated for a number of points along the age axis, yielding an estimate of the true function underlying the data. The kernel function used is the Epanechnikov kernel4. This kernel function weights observations close to the point being estimated higher than those further towards the extremes of the bandwidth, unlike the uniform kernel. However none of the commonly used kernels obscure the salient points of the regression plots produced in this analysis. 3 (1 − u 2 ) I (| u |≤ 1) , where u = zi − z0 4 h 4 The Epanechnikov kernel is K ( u ) = 13 Here, too, we can make improvements to the model by controlling for characteristics other than age. An important limitation on the non-parametric approach is that it is difficult to include many variables in the regression. One way around this is to use a model of the following form yi = β Zi + g ( xi )+ ε i (8) where Zi is a vector of individual characteristics. This kind of semi-parametric model is estimated by first generating parametric coefficients for β , then subtracting from each value of yi the fitted values given by the parametric regression (the sum of the estimates for β multiplied by the value of Zi ) for the given observation (i.e. yi − β Zi = g ( xi )+ ε i ). The resulting dependent variable has now been stripped of the effects contained within the control vector, allowing a local linear regression to be performed to generate an estimate of the function, g ( xi ) , in the same way as before. ∧ ∧ ∧ For spring quarters, the LFS also contains a question about the employment status of individuals one year before. This suggests a further way of determining what happens to participation at the age of 25. The local linear regression models discussed above rely on the assumption that people close to each other in age are not systematically different from each other. While this seems to be a plausible assumption it is difficult, if not impossible, to test. The first difference estimator, based on comparing employment at the time of interview with employment one year before, has the advantage of controlling for such individual heterogeneity by exploiting the longitudinal nature of the data on employment status, and comparing the same person, say, at age 24 and age 25. The first-difference model provides a technique to demonstrate what happened to the rate of change in employment by age. The model we estimate is of the form: yi = g ( xi1 − xi 0 )+ ε i . (9) 14 Controlling for eligibility In both the DiD and RD frameworks, the analysis has to take into account one of the key features of the policy: the WTC is targeted at the lowest-income households. As described in section 2, coverage is determined by household circumstances, namely whether the household has an adult aged 25 or over, working at least 30 hours per week and with no children. The level of support is determined by earnings: the WTC award is tapered away at 37 pence for each pound earned above the £5,060 threshold. This means that individuals with high earning capacities will not be encouraged to enter employment by policies such as the WTC. Consequently, analysis is conducted for groups estimated to be eligible for the policy, as there is little use in including people who would always earn too much to be incentivised to work through WTC. Potential eligibility has been determined in two ways. Firstly, the highest educational levels of respondents are used to screen out those with high earning capacity. Results are shown for respondents with qualifications up to (but not including) A-Level; this group represents 33% of the sample. Secondly, because there is no information about earnings for out-of-work respondents, it is necessary to run an auxiliary wage regression. After this regression a wage can be predicted based on other characteristics (see appendix, table 5). This allows us to choose a wage level above which an individual’s potential earnings would be too high for WTC eligibility. While the DiD models are run for both approaches to determining eligibility, the local linear regressions use only the low qualification eligibility criteria. This is due to the difficulty of producing a meaningful estimate of the underlying function on age when using the fitted-wage eligibility. The local linear regression technique relies on having large amounts of data, and the numbers estimated to be eligible by the fitted wage decline rapidly from 22 upwards. Another consequence of considering eligibility is that couples have to be excluded from the analysis. This is because these households can potentially earn two separate incomes, and as the following section explains, the dataset used only contains individual-level income information. 15 4. Data The analysis makes use of the Labour Force Survey (LFS). This is an extensively-used and nationally-representative UK panel survey, in which around 60,000 participants are interviewed for five consecutive quarters. The questionnaire covers a range of issues, including labour market behaviour and household characteristics. For the DiD analysis it is necessary to have observations for those aged 22 to 27 without children, before and after the policy’s implementation in April 2003. Overall the sample used in the DiD analysis involves around 125,000 observations from 25,000 people interviewed between Q1 2001 and Q4 2005. For the local linear regression, observations are split into those occurring before and after the implementation of the policy. Under the first difference local linear regression model, observations from Q2 2004 and Q2 2005 are compared to observations from Q2 2001 and Q2 2002, since the second quarter is the only time when the respondents are asked about their employment status one year earlier. Given the need for a sufficiently large dataset for the local linear regression approach, it was necessary to use data from the very first days of the policy. The analysis of the impact of WTC presented here is therefore qualified by the fact that the policy had no time to bed-down over the period of the observations for DiD in particular and, to some extent, for RD. Given that the numbers benefiting from childless-WTC increased in the months and years following the policy’s launch, results can be considered a lower bound of its impact. It has been noted that the income data on the LFS is not particularly reliable5, particularly for the hourly pay-rate variable ‘hourpay’, not least because that variable is imputed from weekly earnings and hours data. This makes determining eligibility through hourly earnings rates problematic. Consequently, when determining eligibility for the policy using the potential wage rate (as described above), we place extra caution on these results. 5 Dickens and Draca (2005) 16 The analysis considers the impact of WTC on reported actual hours worked and on usual working hours. How respondents define their usual as opposed to their actual hours is unclear, but it is likely that the major difference between the two answers is caused by overtime undertaken by the respondent. Usual hours represent the important variable determining eligibility for WTC, since it is an annualised support system that requires the claimant to report the number of hours worked in a typical week. It is also likely that usual hours are more rigid than actual hours in the face of changing incentives, since they may be set by contract for some employees. This measure of hours worked is more widely reported than the measure of actual hours, so cell sizes increase when using usual hours. Table 1 includes some descriptive statistics for the key variables used. Table 1: descriptive statistics for key sample variables Variable description Mean Currently in employment (binary) 0.69 Usual hours worked per week 39.52 Actual hours worked per week 41.68 Hourly pay 7.87 Age, in years 24.20 Ethnicity: white British (binary) 0.71 (and a full set of dummies on ethnicity) Female (binary) 0.41 Highest qualification degree (binary) 0.266 Highest qualification other higher (binary) 0.068 Highest qualification A-level (binary) 0.230 Highest qualification other low vocational (binary) 0.089 Highest qualification GCSE (binary) 0.159 No qualifications (binary) 0.054 Current student (binary) 0.11 London (binary) 0.091 (and a full set of regional dummies) Disabled (binary) 0.096 Rented accommodation (binary) 0.384 (and full set of dummies on accommodation status) Std dev 0.462 10.58 14.78 3.85 1.69 0.454 0.491 0.442 0.251 0.421 0.245 0.367 0.226 0.313 0.287 0.486 0.489 5. Results The results for the DiD and RD estimation methods are presented in turn, along with tests of robustness. 17 Difference in differences For completeness, table 6 in the appendix reports average treatment effects following estimation of the simple DiD model without controls. The introduction of WTC is estimated to have had a small positive effect on employment probabilities, of between 1.1 and 1.8 percentage points. At the intensive margin, WTC is estimated to have resulted in a small reduction in hours worked, of around 0.6 hours. The results using the simple DiD model, however, are largely non-significant. Using the regression-adjusted DiD model, with a cubic function of time and a full vector of control variables, should allow for greater distillation of the policy’s effect. This is done in table 2, which reports average treatment effects on the probability of employment and hours worked from estimating equation 4. Table 2: DiD treatment effects for eligible groups Employment Hours worked Eligibility criteria probability (usual hours) 0.029* -0.74* Based on low qualifications (0.01) (0.006) 0.019 -0.70* Based on fitted wages (0.16) (0.047) Where probit model: marginal effects reported p-values in parentheses * denotes significance at the 95% level Hours worked (actual hours) 0.63 (0.358) 0.08 (0.922) Looking at the full DiD specification, all participation results are positive and there is strong evidence of an effect, with the coefficients on policy treatment ranging between 1.9 and 2.9 percentage points. Furthermore, once the vector controlling for characteristics is included most of the results become statistically significant at confidence levels ranging between 99-84%. The result using low qualifications as the eligibility criteria shows the most significant increase in employment probability. Table 2 also shows the impact of the policy on the labour supply decision at the intensive margin. As discussed above, theory suggests that the participation effect should be unambiguously positive. On the other hand, the change in labour supply per worker could 18 go in either direction depending upon whether a person worked more or less than 30 hours per week before becoming eligible for WTC. As expected from the theoretical discussion earlier, the effect of WTC at the intensive margin changes according to the variable used to measure hours worked. However, only results using the usual hours measurement are statistically significant (as discussed in section 4, sample sizes are greater for this hours measurement). Therefore, it is likely that WTC introduction resulted in a very small reduction in the number of hours worked, with an average reduction for those in employment of less than one hour per worker per week. Given the relatively sizes of the coefficients on participation and on hours worked, it is reasonable to conclude that the introduction of WTC resulted in an overall increase in labour supply. Robustness analysis Because of the nature of DiD estimation, which relies on a very reduced-form specification and few (but strong) assumptions, robustness tests under this technique tend to be minimal. The most frequently used method is to conduct tests on whether the DiD identifying assumptions hold for the sample. This is effectively a test on whether the analysis at hand is indeed a quasi-experiment. We complement this approach with some sensitivity analysis and some further tests on inference efficiency. The central DiD assumption states that, in the period before policy introduction, changes in the outcome variable should not be significantly different for respondents in the control and treatment groups. This assumption can be tested by running the various DiD regressions over the pre-policy period, including interaction terms between the treatment and time dummies6. A significant coefficient on an interaction term would mean that there is some time-varying effect which only affects one of the two groups, and which cannot be accounted for by the treatment, time or control vectors. We run equation (4) with these interaction terms; we find that for all three outcome variables (employment probability, usual hours and actual hours) the interaction terms are consistently non6 Stewart (2004) 19 significant, with p-value terms ranging between 0.40 and 0.95. This validates the DiD assumption of no interaction effects. The other identifying assumption needed in DiD is that the policy to be evaluated only affects those in the treatment group. In the context of WTC this means one has to rely on the absence of any general equilibrium effects generated by the policy. In this case the main possible source for such effects might be through employers’ substitution of wagesubsidised 25+ workers for unsubsidised, younger workers. This is unlikely to be a problematic effect however, not least because many of the target group are already on the minimum wage. Previous evaluative work on the Working Families Tax Credit found no evidence of such substitution effects7. Another issue concerning the DiD analysis in the sensitivity of the results to different definitions of policy eligibility, particularly when using fitted wages as the determinant. The theoretical lower and upper thresholds for eligibility are given by the minimum wage, which during this period averaged around £4.50, and the wage of a 30-hour worker who has their award completely tapered away, around £7.40. The results above use a wage cut-off of £6.60 as the central case, because people earning above this limit would have to be working more than 30 but fewer than 35 hours in order to be eligible for the policy. Consideration of usual hours responses in the sample reveals that very few people are in this situation. The choice of wage cut-off involves some discretion from the analyst, but sensitivity analysis can help determine the plausibility of any assumption. Table 3 reports the result of a sensitivity analysis regarding the wage cut-off. Four inferences can be made from this exercise. Firstly, it can be seen that our key result, concerning the effect of WTC on participation, is not overwhelming sensitive to the wage cut-off: the effect of the policy remains consistently positive. 7 Blundell, Brewer and Shepherd (2005) 20 Table 3: DiD treatment effects using different wage cut-offs for eligibility Employment probability p-value Wage cut-off (marginal effect) 8 0.26 0.78 7.8 0.48 0.62 7.6 0.5 0.62 7.4 1.11 0.3 7.2 1.39 0.21 7 1.83 0.14 6.8 1.77 0.17 6.6 1.91 0.16 6.4 1.73 0.25 6.2 2.57 0.12 6 2.59 0.15 5.8 1.96 0.31 5.6 2.29 0.29 5.4 1.73 0.47 5.2 4.81 0.07 5 4.72 0.13 4.8 5.72 0.1 Marginal effects for probit regressions of employment probability using equation (4) Grey denotes non-significance at an illustrative 80% confidence level Secondly, the size of the coefficients are largely within the expected range. For wages ranging between £5 and £7.40, the coefficients range between 1.1 and 4.8 percentage points, with an average value of 2.3 percentage points. This is in line with results using the alternative definition based on qualifications. Thirdly, the variation of the coefficients is in line with expectations from economic theory and the design of the policy. As chart 3 shows, the marginal effect of the policy increases as the DiD regression focuses on those respondents which could potentially benefit most from WTC. Finally, most of these results are significant at an arbitrary 20% significance level. This significance level is chosen to take into account both the small size of the parameter being estimated and the large amount of variation resulting from using fitted wages as a measure of eligibility. Other robustness checks were conducted on these results. These included testing whether inference efficiency was significantly reduced when taking into account the clustering of the data by respondent. It was found that using much smaller samples, with only one response per individual or with explicit personal identifiers to allow for clustering, did 21 not significantly improve the measurement of variance. This was also the case when forcing the estimation to produce robust standard errors. Chart 3: plot of sensitivity analysis regarding the wage cut-off 7 6 5 Policy coefficient 4 3 2 1 0 4.5 5 5.5 6 6.5 Im puted hourly w age 7 7.5 8 8.5 20% significance n/s Log. (20% significance) Regression Discontinuity We use a bandwidth of 1 year for the non-parametric local linear regression of employment probability upon age. Local linear smooths are run either side of the threshold age of 25. The resulting plots for the data from before the introduction of WTC show a clear discontinuity in the regression, with employment as a function of age falling by around 3.9 percentage points at the age of 25 prior to the introduction of WTC in 2003, (see chart 6, appendix) on an otherwise steadily upward-trending line. After the introduction of the new policy, employment levels among the eligible groups show no evidence of a discontinuity at the age the policy environment changes. More interesting is the result of the semi-parametric local linear regression, which is represented in chart 4 and, using a bandwidth of 1.2, clearly demonstrates the net impact of the policies that change at the age of 25. These regressions, which control for other 22 characteristics of respondents, give an indication of the effect on participation of the increased generosity of HB and IS (theoretically negative) before 2003, represented by the blue line. The impact of the end of New Deal for Young People (NDYP) may also influence the regressions. The red line represents the relationship after 2003, when all of the above policies were in operation alongside the new WTC. Chart 4: Semi-parametric local linear regression of employment on age, bandwidth=1.2 0.1 0.08 Employment probability 0.06 0.04 0.02 0 20.0 -0.02 Age 21.0 22.0 23.1 24.1 25.0 26.6 28.3 29.9 31.5 2001-2003 2003-2005 For data from the two years before April 2003, the regressions show evidence of a large discontinuity at 25, attributable to the policy environment and causing the age-contingent probability of employment to fall by around 3.2 percentage points. Secondly, there is evidence that since the introduction of WTC the detrimental impact of age on participation around 25 years appears to have been effectively countered, with no evidence of a discontinuity at 25 in the data since 2003. Again, this change in the age function can be attributed to the introduction of WTC. Some robustness checks were conducted for the discontinuity estimates. One obvious way of allowing inference is to use bootstrapping. This has to be done manually when 23 using non- and semi-parametric methods, as in the case of these local linear regression smooths. Using the bootstrapped standard errors (after 30 replications8) we find that the semi-parametric regression discontinuity estimate for the pre-WTC data is statistically significant (see table 4), while there is no statistically or economically significant discontinuity in the 2003-05 data. Another way of testing for robustness is to run the estimation procedure across an age threshold at which no change in the policy environment occurs, to check that that the procedure does not suggest the existence of any discontinuity. This is done for hypothetical discontinuities at age 23 and 28: neither resulting regression discontinuity coefficients is statistically different from zero. We conclude that the regression discontinuity estimates are robust. Table 4: regression discontinuity estimates Local linear regressions Non-parametric Semi-parametric -0.0386 -0.0319* Pre-WTC (2001-03) (n/a) (0.014) 0.0045 -0.0003 Post-WTC (2003-05) (n/a) (0.012) Bandwidth 1 year 1.2 years Smoothed over actual or predicted employment probability Eligibility defined by low qualifications Bootstrapped standard errors in parentheses (30 replications) * denotes significance at the 95% level That participation starts to fall before the age of 25 may well indicate an anticipation effect. It is reasonable to think that, a few months before they become eligible for the more generous IS, some individuals who have become unemployed decide to remain jobless in order to be able to claim it. This is not unlikely since would-be claimants are not eligible for IS if they have voluntarily left unemployment in order to claim. Non-parametric estimates of the age function on the first-difference model show a steady downward trend in the first difference of the employment rate before 2003 (see appendix, chart 7). After the introduction of WTC the regression plot shows an increase in the firstdifference of the employment rate after the introduction of WTC. The regressions do not show evidence of a discontinuity at 25. This might be expected since the full impact of 8 While a higher number of replications would be desirable, computational time is high. Improvements in replication size are planned. 24 the policy on the first difference (particularly on exit rates from employment) would be seen over the two years after eligibility begins. The post-April 2003 plot shows evidence of the first difference becoming greater between 25 and 27, where before it declined. While it would be unwise to put too much weight on this result given the smaller sample size available under the first-difference model, it indicates that WTC has contributed to a rise in the rate of change of employment as a function of age. 6. Conclusion This paper set out to estimate the impact of WTC on young childless individuals’ participation and labour supply, and to put this in the context of the net impact of all policies that affect labour supply around the age of 25. Isolating the impact of WTC is achieved by exploiting the start of the programme and the fact that people become eligible for it at 25, using a difference-in-differences approach. The increased generosity of Income Support and Housing Benefit at 25, as well as the end of NDYP eligibility, creates incentives that run in the opposite direction to the intended effect of WTC. The net effect of all changes at 25 is captured by the regression discontinuity technique. The DiD analysis uses proxies based on education and an auxiliary wage regression to identify a group for whom pay rates are likely to be low. It is unlikely that higher paid workers would see any WTC entitlement if working full-time. As theory predicts, results on the impact of WTC strongly suggest an increase in participation of between 1.9 and 2.9 percentage points due to the policy, among the eligible groups. The results suggest that there is only a small reduction in the number of hours worked at the intensive margin although even this is dependent upon the hours measure used. Overall, the picture is one of a positive impact of the policy on labour supply among low-pay workers and a stronger positive impact on participation. Since WTC for childless people was part of the much bigger package of tax credits launched in April 2003, it did not receive as much publicity as it might have had it been launched separately. Consequently take-up over the first two years has been improving rapidly as people become more aware of the support available. This lag in take-up may have depressed estimates of the policy impact. 25 The regression discontinuity analysis uses a non-parametric local linear regression to demonstrate the net impact of all the policies with an age-contingent aspect that change at the age of 25. Before the introduction of WTC, the impact of the policy environment around the age of 25 appears to reduce the employment rate of those with lower qualifications by over 3 percentage points on an otherwise steady upward trend in employment rates by age. Use of non-parametric bootstrapping to determine standard errors suggests that this discontinuity is significant. It seems likely that more generous Income Support and Housing Benefit available from the age of 25 were the primary causes of the discontinuity, although other policies may contribute. Out-of-work IS increases by £11.70 – 26% higher than the 18-24 year-old level. Single Housing Benefit recipients can claim support for, typically, twice as much rent as they could at the age of 24, if out of work. Consequently the out-of-work income of someone aged 25 could be very much higher than that of someone aged 24, with concomitant effects on work incentives. The local linear regression models provide no evidence of any statistically or economically significant discontinuity after the introduction of WTC, indicating that the policy appears to have done a lot to reverse the deterioration in incentives at 25 by boosting the available gain to work for low-wage workers. Overall, this paper finds strong evidence that WTC is encouraging eligible people to enter work. Furthermore, the analysis supports the hypothesis that WTC helps to ameliorate the policy-driven deterioration of work incentives at 25, which existed before its introduction. References Adam, C (2004) Seven Habits of Effective Reviewers: A checklist for effective evaluation of econometric evidence. Angrist, J.D. and Krueger, A.B. (1999) Empirical Strategies in Labour Economics in Ashenfelter and Card ed. Handbook of Labour Economics Volume 3. 26 Battistin, E. and Rettore, E. (2003) Another Look at the Regression Discontinuity Design. Cemmap Working Paper CWP01/03 Blundell, Brewer and Shepherd (2005) Evaluating the impact of Working Families’ Tax Credit using Difference-in-differences. HM Revenue & Customs Blundell, Brewer and Shepherd (2004) The Impact of Tax and Benefit changes between April 2000 and April2003 on Parents’ Labour Supply. IFS Briefing Note No.52 Blundell and Costa Dias (2000) Evaluation Methods for Non-Experimental Data. Fiscal Studies col. 21, no.4, pp.427-468 Blundell and Reed (2000) The Employment Effects of the Working Families Tax Credit. IFS Briefing Note No. 6 Brewer, Duncan, Shepherd and Suarez (2003) Did Working Families’ Tax Credit Work? Analysing the impact of in-work support on labour supply and programme participation. Inland Revenue Brewer, Duncan, Shepherd and Suarez (2005) Did Working Families’ Tax Credit Work? The final evaluation of the impact of in-work support on parents’ labour supply and takeup behaviour in the UK. Inland Revenue Connolly and Gregory (2002) The National Minimum Wage and Hours of Work: Implications for Low Paid Women. Oxford bulletin of economics and Statistics, 64 De Giorgi (2005) Long Term Effects of a Mandatory multistage Programme: The New Deal for Young People in the UK. IFS Working Paper W05/08 Dickens and Draca (2005) The Employment Effects of the October 2003 Increase in the National Minimum Wage. Queen Mary, University of London Working Paper No. 532 27 DiNardo and Lee (2002) The Impact of Unionization on Establishment Closure: A Regression Discontinuity Analysis of Representation Elections. NBER Working Paper 8993 DiNardo and Tobias (2001) Non-parametric Density and Regression Estimation. The Journal of Economic Perspectives, Vol. 15, No. 4, pp. 11-28. Eissa and Liebman (1996) Labour Supply Response to the Earned Income Tax Credit. The Quarterly Journal of Economics, Vol. 111, No. 2, pp 605-637 Francesconi and Van der Klauuw (2004) The Consequences of ‘in-work’ Benefit Reform in Britain: New Evidence from Panel Data. ISER Working Paper No 2004-13 Greene, W Econometric Analysis. New York University, Prentice Hall Gregg and Harkness (2003) Welfare Reform and Lone Parent Employment in the UK. Department of Economics and Leverhulme Centre for Market and Public Organisation, University of Bristol Hanh, Todd and Van der Klauuw (2001) Identification and Estimation of Treatment Effects with a Regression-Discontinuity Design. Econometrica, Vol. 69, No. 1, pp 201209 HM Treasury (2002) The Child and Working Tax Credits: The Modernisation of Britain’s Tax and Benefit System. Number 10. Lemieux and Milligan (2004) Incentive Effects of Social Assistance: A Regression Discontinuity Approach. NBER Working Paper 10541 Low, H. (2004) Presentation: ‘How Effective are in-work Tax Credits? New Perspectives’. LSE Taxation Seminar, March 28 Lydon and Walker (2004) Welfare-to-Work, Wages and Wage Growth. Inland Revenue Marsh, A. Stephenson, A. Dorsett, R. and Elias, P. (2001) Earnings Top-up Evaluation: Effects on Low Paid Workers. Policy Studies Institute on behalf of The Department of Social Security Moffitt, A (2003) The Negative Income Tax and the Evolution of U.S. Welfare Policy. Journal of Economic Perspectives, Vol. 17, No. 3, pp 119-140 Myck and Reed (2000) Labour Supply Estimation Project: Report 1. IFS for HM Treasury Smith, J (2004) Evaluating Local Economic Development Policies: Theory and Practice. University of Maryland Stewart, M.B. (2004) The impact of the introduction of the UK Minimum Wage on the employment probabilities of low-wage workers, Journal of the European Economic Association, 2(1), p.67-97 Stewart and Swaffield (2004) The Other Margin: do Minimum Wages cause Working Hours Adjustments for Low-wage Workers? Low Pay Commission. Stewart and Swaffield (2005) Conference paper, RES conference 2005. Stine, R. (1989) An Introduction to Bootstrapping Methods. Socialogical Methods and Research, Vol. 18, No. 2 & 3 Wooldridge, J (2001) Econometric Analysis of Cross Section and Panel Data. MIT Press, Cambridge, Massachusetts 29 APPENDIX Table 5: auxiliary wage regression Explanatory variable Coefficient Female* -0.71 Couple* 0.57 Currently student* -1.0 Highest qualification: degree* 1.99 Highest qualification: higher* 0.53 Highest qualification: GCSE* -0.72 Highest qualification: other low* -0.96 Highest qualification: no qualifications* -1.63 Employment duration: under three months or -0.78 unemployed* Employment duration: Between 3 and 6 months* -0.48 Employment duration: Between 6 and 12 months* -0.43 Employment duration: Between 2 and 5 years* 0.45 Employment duration: Between 5 and 10 years* 0.49 Employment duration: Between 10 and 20 years -0.15 Disabled* -0.36 Owner-occupier* 0.27 Rent free* -0.73 Social housing* -0.57 Asian or Asian British -0.32 African or Africa-born -0.14 Black or black British* -0.46 Other ethnic group* -0.46 White non-British* 0.38 Outer London* 1.02 London* 2.29 North East* -1.24 North West* -1.17 Merseyside* -1.19 Yorkshire and the Humber* -1.26 East Midlands* -0.97 West Midlands* -0.88 East 0.01 South West* -1.02 Wales* -1.26 Scotland* -1.16 Northern Ireland* -1.74 Age function, levels parameter 0.004 Age function, squared parameter -4.5E-7 Age function, cubic parameter 2.1E-11 P-value 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.38 0.00 0.00 0.00 0.00 0.07 0.39 0.04 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.89 0.00 0.00 0.00 0.00 0.89 0.88 0.85 30 Time function, levels parameter* Time function, squared parameter Time function, cubic parameter* Dependent variable hourly pay Ordinary least squares estimation, robust standard errors R-squared: 0.25 Number of observations: 23,479 * denotes significance at the 95% level Chart 5: Fitted wage distribution .2 0.004 -2.2E-6 4.5E-10 0.03 0.06 0.05 0 0 .05 Density .1 .15 5 10 wagea 15 20 Table 6: simple DiD treatment effects for eligible groups Employment Hours worked Eligibility criteria probability (usual hours) 0.018 -0.81* Based on low qualifications (0.071) (0.005) 0.011 -0.69 Based on fitted wages (0.386) (0.090) Where probit model: marginal effects reported p-values in parentheses * denotes significance at the 95% level 31 Chart 6: Non-parametric local linear regression of employment on age, bandwidth = 1 0.8 0.75 Employment probability 0.7 0.65 0.6 0.55 0.5 20.0 20.6 21.2 21.8 22.4 23.1 23.7 24.3 24.9 25.7 26.6 27.6 28.6 29.6 30.6 31.5 32.5 Age 2001-2003 2003-2005 Chart 7: Non-parametric local linear regression of first-difference of employment on age, bandwidth = 1.5 0.12 First difference of employment 0.1 0.08 0.06 0.04 0.02 0 20.0 21.0 22.0 23.1 24.1 2001-2002 25.0 Age 26.6 2004-2005 28.3 29.9 31.5 32