A theoretical and econometric framework

Ž . Fig. 1. Labour force participation and maximum weekly benefits 1985:4–1995:4 . Aggregate, season- ally adjusted labour force participation rates are taken from Statistics New Zealand for all individuals 15 years old and over. Mean weekly guarantees are expressed in 1995:4 dollars, and come from the author’s calculations for those aged between 16 and 64. Inspection of the relationship between the aggregate series in Fig. 1 does not provide evidence that these benefit reforms had a positive impact on the LFPR. We are not holding constant the variety of other factors that independently influence labour supply, and we are not taking advantage of the available disaggregate data. One concern here is the cyclical effects of a deep recession in New Zealand prior to and during these benefit reforms, followed by a vigorous recovery in the post-reform period. For this reason, we use disaggregate data and regression analysis to isolate the effects of specific benefit changes that are relevant to the individuals in our sample, and to control for other factors including general macroeconomic conditions that may be independently influencing this behaviour.

4. A theoretical and econometric framework

Ž Suppose a representative individual maximises utility over income i.e., the . consumption of goods and services Y and leisure L. U s f Y , L 1 Ž . Ž . A fixed endowment of time T must be allocated between leisure and market work H. T s L q H 2 Ž . The individual faces a nonlinear budget constraint that captures the basic features of New Zealand’s three NIT programmes. Income comes from either labour Ž . market earnings hours of work multiplied by an exogenous wage W , or social welfare benefits B. Y s WH q B 3 Ž . The benefit formula is typical of NIT-type programmes. w x B s Max G y t Max WH y D,0 ,0 4 Ž . Ž . The maximum weekly benefit or guarantee G is received if the individual earns less than the earnings disregard D. When earnings exceed this threshold, the benefit received falls by the benefit reduction rate t , where 0 - t F 1. Once Ž . earnings exceed the ‘‘breakeven’’ level i.e., Grt q D , the benefit is zero and the person is off the programme. Fig. 2 depicts the overall budget constraint facing the individual. In the absence of this social welfare programme, the relevant budget constraint is the line segment AB. Under the benefit formula, the relevant ‘‘kinked’’ budget constraint is initially CDEB. The individual can choose not to participate in the labour force and receive the full benefit G at point C. He or she can work and receive benefits along either the ‘‘low’’ or ‘‘high’’ abatement segments of the budget constraint Fig. 2. Pre- and post-reform budget constraints. Ž . CD and DE, respectively . If earnings exceed Grt q D along line segment EB, then all income comes from the labour market and the individual is off the benefit. A decline in the real maximum weekly benefit from G to G X shifts the entire budget constraint downward. A decline in the real earnings disregard, due to either a rise in the price level or a cut in the nominal disregard, shifts the first kink in the budget constraint to the right from D to D X . The breakeven income level X X X Ž X . could decline to G rt q D from E to E with a fall in either the real maximum weekly benefit or real earnings disregard, or a rise in the benefit reduction rate. Although there are three abatement regions under the UB and DPB programmes Ž . 0, 30 and 70 , this complexity adds little to the theoretical discussion. Benefit reduction rates under these programmes were constant over our sample period. There are two abatement regions under Superannuation, with four changes Ž . in this benefit reduction rate or Tax Surcharge over our sample period of between 18 and 25. Nominal earnings disregards were constant under the UB and DPB programmes, while they changed five times under Superannuation between 1985 and 1992. The potential labour supply responses associated with all of these benefit reforms cannot be represented in a simple theoretical expression. However, we can characterise how changes in each of the three parameters within the benefit formula may have influenced labour supply behaviour. A decline in the real guarantee should increase labour supply as nonparticipants are drawn into the labour force and working beneficiaries increase their hours of work due to positive Ž income and substitution effects where the latter effect comes from a higher X . effective wage rate as the breakeven point falls E toward E . A decline in the real earnings disregard would have a theoretically ambiguous impact on labour supply. On the one hand, working beneficiaries might increase their hours of work due to positive income and substitution effects that are similar Ž to those associated with a falling guarantee particularly for those initially working . in the high abatement region . On the other hand, individuals initially working Ž . near their maximum hours in the low abatement region i.e., just below point D would see their effective marginal wage rates decline as this kink shifts rightward toward D X . A positive income effect could be offset by a negative substitution effect. In other words, these beneficiaries would have to reduce their hours of work if they want to remain in the low abatement region. The labour supply effect of an increase in the benefit reduction rate would also be uncertain. Offsetting income and substitution effects would occur as the high abatement region of the budget constraint rotates downward. While lower income levels would tend to encourage work, lower effective wage rates would discourage it. This labour supply response would be further complicated by a corresponding drop in the breakeven point. A positive substitution effect would occur if the individual earns enough to leave the benefit system. Two things can be gleaned from this simple theoretical framework. First, the very nature of the NIT system makes it difficult to characterise the likely labour supply effects of recent benefit reforms in New Zealand. Both the directions and magnitudes of these numerous behavioural responses are ultimately empirical Ž questions. Second, although all three parameters in the benefit formula benefit . guarantees, earnings disregards and benefit reduction rates have been altered across the three social welfare programmes in this country over this sample period, we propose collapsing these basic determinants into just two explanatory variables for our subsequent regression analysis. The first variable will capture changes in real maximum weekly benefits that varied across both time and subgroups within the population. The second variable will capture changes in both real earnings disregards and benefit reduction rates. With regard to this second regressor, note Ž . that breakeven income depends on all three programme parameters i.e., Grt q D . To remove the effects of changes in guarantees on this variable, we hold all real maximum weekly benefits constant at their initial values in the first quarter of our Ž . sample period 1986:1 . As a result, variation in our measure of breakeven income reflects the combined influence on labour supply of changes in real earnings disregards and benefit reduction rates across both programmes and time. Of course, a single variable could be used to capture changes in all three parameters in the benefit formula. We have already noted that the breakeven income level varies with guarantee levels, earnings disregards and benefit reduc- tion rates. The problem is that this single regressor might not adequately portray the complex set of labour supply incentives associated with social welfare programmes. For example, breakeven income could remain constant if both the guarantee and benefit reduction rate were cut. Yet, participation and hours of work could clearly be affected by such a policy. The two variables used in this study would capture at least some aspects of these programme changes. The drawback of our approach is that we are still collapsing changes in both earnings disregards and benefit reduction rates into a single variable. Again, this may not adequately capture the nature of labour supply disincentives associated with these pro- grammes. The alternative would be to include a series of either quantitative or qualitative variables to pick up changes in these separate programme character- istics. Unfortunately, with no changes in benefit reduction rates under UB and DPB and no legislated changes in earnings disregards under these same pro- grammes, we would find it difficult to measure these separate influences. In fact, the identification of the overall labour supply effect of earnings disregards and benefit reduction rates in our analysis depends heavily on the legislated changes in these programme parameters under Superannuation. A general econometric model of labour supply behaviour is used in this study. H s h q X X g q a ln G q b ln E q Q X d q u 5 Ž . i t t i t i t i t i t i t Ž . The generic dependent variable H is a summary measure of the labour supply it behaviour of individuals in group i in period t. Three different measures of labour supply are used to test the robustness of our findings. Descriptive statistics on the three variables are presented in Table 1. The first dependent variable is the LFPR. This is the proportion of individuals within a cell who are either employed or unemployed at the time of the survey. Thus, it takes on a continuous range of values within the 0–1 interval. 6 A second dependent variable is defined as the number of weekly hours of labour supplied to the market. The LFPR ignores any variation in the ‘‘amount’’ of labour participants are willing to supply. Benefit reforms could raise aggregate labour supply without affecting the LFPR. Since the HLFS does not provide information on the number of hours the unemployed are willing to work, we multiply the LFPR in each cell by the number of hours worked per week by those employed within the same cell. This assumes that the unemployed want to work the same hours per week as the employed within the same cell. A third dependent variable ‘‘broadens’’ this notions of labour supply. Current economic activity in the labour market could include both labour force participa- tion and human capital accumulation. 7 It would be easy to show in a simple model Ž . of human capital investment that reductions in current and future welfare benefits could increase optimal educational attainment. This concern is particularly relevant given that reforms like the rise in the age of eligibility for UB and DPB were directly targeted at youth. Students generally are not eligible for benefits through the UB programme, but are eligible for benefits through the DPB programme. This dependent variable is defined as the proportion of individuals within a cell who were either in the labour force, or reported ‘‘studying’’ as their main activity while out of the labour force. Table 1 shows that this broader measure of economic activity adds an average of 4.4 percentage points to the LFPR. Ž . Quarterly dummy variables h are included as regressors in Eq. 5 to capture t all time-specific, individual-invariant effects on labour supply. The use of time dummies represents a very general specification of this regression model. It implies that cross-sectional differences in benefit reforms will play a critical role in isolating the overall effects on labour supply. All relevant demographic characteristics like age, gender, ethnicity, educational attainment, marital status and number and ages of children in the family are included in the vector X . There are two reasons for including these regressors in i t the model. First, they may directly influence labour supply behaviour. In fact, we allow the effects of family circumstances like marital status, and the number and ages of children in the household to vary between men and women. Second, 6 This raises the issue of the ‘‘censoring’’ of the dependent variable at zero and one. Failure to incorporate this censoring into the estimation procedure could produce coefficient estimates that are both biased and inconsistent. However, this censoring problem is unlikely to be appreciable in this study. More than 97 of the individuals in our sample are located in cells with participation rates between the extremes of zero and one. 7 Ž Unfortunately, no information is available in the HLFS on the job training of individuals another . source of human capital accumulation . because we have no information on the potential market wages facing the individuals in the HLFS, at least some of these variables may capture systematic differences in earnings capacities. In this way, we ‘‘indirectly’’ control for the impact of wages on labour supply in this analysis. Ž . The natural logarithm of the real maximum weekly benefit ln G is included i t as an independent variable in this regression. Dividing this estimated coefficient by the mean of the dependent variable provides an estimate of the labour supply elasticity associated with this variable. 8 Our null hypothesis is that this effect will Ž . be negative i.e., lower guarantees, on average, increase labour supply . The Ž . natural logarithm of the breakeven income level ln E is also included as an i t explanatory variable. Since real maximum weekly benefits are held constant in the computation of this variable, it captures the effects of changes in both real earnings disregards and benefit reduction rates across the sample. Our simple theoretical model of labour supply behaviour does not provide a clear expectation as to the sign on the coefficient attached to this variable. Two variables are used to capture changes in the ages of eligibility for the three main social welfare programmes. These covariates are included in the vector Q . i t The first dummy variable captures the increase in the age of eligibility from 16 to 18 for UB and DPB. This variable takes on a value of one for unmarried 16- or Ž 17-year-olds with children after 1991:2 when the age of eligibility for DPB was . Ž raised , and a value of one for all other 16- or 17-year-olds after 1990:4 when the . age of eligibility for UB was raised . This variable is set equal to zero for all other age groups and all other time periods. Real maximum weekly benefits for those aged 16 or 17 are not recorded as zero after this increase in the age of eligibility, because of the creation of other welfare programmes for which 16- or 17-year-olds Ž are potentially eligible see the discussion on these alternative programmes in . Section 3 . There was little change in either measured guarantees or breakeven income levels as a result of this increase in the age of eligibility. This means that any labour supply effects associated with this tightening in eligibility criteria will be captured by this dummy variable. The second variable slowly rises from zero to one as the age of eligibility for Superannuation was gradually raised from 60 to 65. In 1992:2, the age of eligibility for Superannuation was raised from 60 to 61. Since this policy directly affected approximately 20 of the individuals in the 60- to 64-year-old cells, this variable jumped from zero to 0.2 for this age group. This minimum age of eligibility has since been raised in 3-month increments every 6 months until it will eventually reach age 65 in the year 2001. Our policy variable mimics this steady rise in the age of eligibility for Superannuation, reaching 0.55 in 1995:4. Our null 8 It would be impossible to take the natural logarithm of the dependent variable in this sample to give us a direct measure of this elasticity. Participation rates for some groups of individuals in this sample are zero. hypotheses are that increases in the ages of eligibility under all three programmes will increase the labour supply among the relevant age groups. This rise in the age of eligibility for Superannuation also reduced both real maximum weekly benefits and real breakeven income levels for this age group. Once individuals are no longer eligible for Superannuation, they become eligible for lower benefits under the UB programme with higher benefit reduction rates. It is important to note, however, that we expect the rise in the age of eligibility for Superannuation will have a larger impact on labour supply than the corresponding decline in maximum weekly benefits and breakeven points. The primary reason is that there is an expectation under the other social welfare programmes that individuals seek employment, while no such obligation is inherent under Superan- nuation. Ž . Weighted, Generalized Least-Squares GLS estimation will be used to estimate the parameters in this model. Observations will be weighted by ‘‘sample weights’’ constructed by Statistics New Zealand to extrapolate these random samples to the general population. Furthermore, the disturbance term u is assumed to contain a i t Ž . component Õ that is specific to a given cell of individuals: i u s n q ´ 6 Ž . i t i i t where E n s 0 Var n s s 2 Cov n ,´ s 0. 7 Ž . Ž . Ž . Ž . i i n i i t For a particular cell, disturbances in different periods are correlated due to this common component: s 2 n Corr u ,u s for t s 8 Ž . Ž . i t i s 2 2 s q s n ´ It would be inappropriate to use a ‘‘fixed-effects’’ estimation technique with these Ž . data i.e., where all variables are deviated from their sample means . The reason is that these data do not constitute a true panel data set. The same individuals are not followed over the entire sample period. Such a data set is commonly referred to in the literature as a ‘‘synthetic’’ panel. 9 Although rotation groups keep the same households in the HLFS for eight consecutive quarters, individuals within these households can ‘‘migrate’’ across cells due to changes in age, education, marital status, number and ages of children in the family. These groups of individuals are essentially random samples of all individuals in the population at a point in time who share the same characteristics. This means that the observed labour market 9 Ž . See Verbeek and Nijman 1992 for a discussion of estimation techniques using synthetic panel data. outcomes within the cells are sample statistics, and a random effects estimation procedure should be used.

5. Regression results