The effect of a guaranteed income on household labour supply
decisions: the case of the Canadian Mincome experiment
BSc Thesis
Abstract:
This paper analyses the effect of the implementation of a negative income tax on household labour supply decisions. A difference-in-difference approach on experimental data is used to isolate the treatment effect. The main results of the data show that on average households tend to decrease their hours worked for most tax rates and guarantee levels. Attrition and non-participation pose a threat to the external validity. They misrepresent the sample treatment effect negatively, suggesting there are combinations of the policies parameters to be found that are able to minimize, or even completely remove a potential drop in labour supply on a population level.
Name: Bram Bossink Student number: 10335609
Email address: brambossink@gmail.com Date: 26-06-2018
Name of supervisor: Simon ter Meulen Second reader:
Statement of Originality
This document is written by Student Bram Bossink who declares to take full responsibility for the contents of this document.
I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.
The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Contents
1. Introduction……….4
2. Literature review………..5
2.1 Guaranteed income……….5
2.2 General labour supply theory………...8
3. The Mincome experiment……...………..11
3.1 Experimental setup……… .11
3.2 First labour supply responses………14
4. Statistical analysis………..16
4.1 The difference-in-difference methodology………16
4.2 Results………..19 4.3 Discussion………..23 5. Conclusion………..26 References………...28 Appendix A………..30 Appendix B………..31 Appendix C………..33 Appendix D……….37 Appendix E………..39
1. Introduction
Technological progress has brought positive welfare effects on today’s economy by creating large efficiency gains. However, it is not straightforward how these gains should be distributed (Sachs, Benzell & Lagarda, 2015). The current technological progress, together with changes in the tax system, increase inequality in developed countries. First of all, Autor (2015) suggests that technological progress primarily crowds out middle-income jobs resulting in increasing income inequality. Sachs et al. (2015) argue that technological progress can increase inequality over generations through a reduction in relative wages. Second, the average effective tax rate on labour has risen faster over the past years than on capital (Carey & Tchilinguirian, 2000). Since capital owners are often more concentrated and wealthier, this contributes to the increasing inequality gap (Wolff, 2010). Current welfare programs partly address issues of increasing income inequality but are criticized. They are inefficient, have high costs, and create a disincentive to work (Murray, 2008).
Two popular suggestions to reform the current tax and benefit systems are a universal basic income (UBI) and a negative income tax (NIT). These systems provide a guaranteed unconditional income for every member of society ensuring that nobody will fall under a certain income boundary. Potentially this could be paid for by a tax increase on capital, or on the people who benefit from technological progress. As a result, such a program might be able to reduce the increasing income inequality. Moreover, these measures could replace a number of current welfare programs and reduce the necessity of control through its unconditional character. They are therefore expected to be an improvement over current systems based on increasing efficiency and decreasing bureaucracy costs. However, if people decide to reduce their working hours significantly the affordability of such a measure could become a serious issue. This paper will therefore focus on trying to measure the effect of a guaranteed income on household labour supply.
In Canada and the United States five prominent experiments to assess the responses to an implementation of a guaranteed income have been carried out. Difficulties of these experiments consisted of the prevalence of non-participation, attrition, and non-random assignment. The assignment problems resulted from using an optimal allocation model to reduce the costs and increase the precision of the experiments (Hum & Simpson, 1993). Existing literature on these experiments using structural labour supply models include Keeley, Robins, Spiegelman and West (1978), and Hum and Simpson (1991). These papers adopt a fixed effects (FE) regression model, or the within-estimator, relying on specific utility functions. This paper will focus on a difference-in-difference (DD) approach, making the effects of the treatment easier to interpret by reducing the necessity of such a utility function. Furthermore, in contrast to the earlier research on the Mincome experiment of Hum and Simpson (1991), this paper will try to correct for an observed lagged response in the first period after treatment. The model will also allow for non-linearity in the wage rate, and different treatment plans will be merged to reduce variability.
The data that will be used is provided by the Canadian ´Mincome´ experiment. This experiment was conducted from 1974 to 1977 in the Canadian city Winnipeg of the province Manitoba. The dataset provides an extensive list of individual characteristics and labour supply
responses before and after the implementation of a negative income tax. The treatments consisted of different combinations of guarantee levels and tax rates. The DD approach will be used to isolate the effects of the different treatment plans on three different household types: single headed male households, single headed female households, and double headed households.
The first results suggest that people reduce their labour supply when receiving a negative income tax. Double headed households tend to decrease theirs more than single headed households. The results also show that the responses differ over treatment groups. The group with the highest tax rate and the medium income level shows no or a low negative effect. Additionally, an analysis of attrition and non-participation suggests that the sample results influence the true values of the population mostly negatively. This leads to conclude that it is possible to construct a negative income tax that will have only a marginal, or no effect on labour supply decisions.
This paper will continue by providing an overview on the existing literature on labour supply in general, and the interaction with benefit systems in particular. Secondly, a summary of the experimental design and a first look at the labour supply responses will be provided. The third part explains the DD approach and applies it to the data, followed by an extensive discussion of the shortcomings of the experiment in the fourth part. At last this paper will produce a conclusion together with an advice on enquiries to pursue.
2. Literature review 2.1 Guaranteed income
According to Moffit (2003), most currently implemented welfare programs do not consider incentives of working. He argues that most welfare programs use a 100 percent tax rate. This means that for every euro earned, benefits decrease with one euro. A small enough increase in income, by accepting work, will be totally offset through a deduction of the benefits in place. Consequently, such a welfare program causes a disincentive to work. Gamel, Balsan and Vero (2005) call this effect an inactivity trap. Figure 1 shows a graph of this inactivity gap. The initial choice of hours of work, without the implementation of a welfare program, is given by H. The
initial budget constraint is tangent to the indifference curve of workers. The implementation of a welfare program with a 100 percent tax rate would change the budget constraint from the line BC to BC’. It becomes horizontal at the point where income reaches the level of the benefit, meaning that every number of hours worked on the horizontal part of the budget constraint result in the same level of income. Therefore, workers are able to shift from indifference curve U to U’, creating higher utility by reducing their working hours from H to H’, where they do not work at all.
An example of the inactivity trap is to be found in France. They implemented the minimum income guarantee (GMI) , using such a tax rate. It did indeed induce people to decline work under a certain income level. France’s reaction was the implementation of an employment premium existing together with the GMI. Gamel, Balsan and Vero argue that the simplest and most effective solution would have been the abolishing of GMI and replacing it with a stronger incentive such as a NIT. Figure 2 shows the potential effect of a NIT. The new budget constraint is still sloping downward instead of horizontally, causing the optimum point of hours worked for households to be closer to the original decision of hours worked. In this situation people are less likely to give up working all together. Appendix A shows a graph of the potential effect of replacing current welfare systems with a NIT.
The inefficiency of welfare programs is also addressed by Amétépé, through the non-take-up rate (2012). This rate measures the number of eligible individuals, or households, that do not receive the amount of benefit that they are legally entitled to. In other words, this rate can measure the effectiveness of welfare policies. A high non-take-up rate is considered a failure of such policies. In the case of a negative income tax or a basic income this problem is non-existent. In the case of a basic income the benefits are unconditional so there is no uncertainty of eligibility. Moreover, in the case of a negative income tax the ‘application’ for the benefit is embedded in the tax returns. These returns are required by law making it impossible for eligible people not to receive the benefit.
A UBI and a NIT are both examples of a guaranteed income. They ensure that, without exclusions, every economic agent has a minimum level of income. As defined by Tondani (2009) a UBI is a universal and unconditional lump-sum transfer. Whereas, NIT uses a baseline and everyone earning below this baseline receives a negative tax. He also shows that these two policies can achieve the exact same distributive outcomes with the same marginal and average tax rates.
To illustrate this, the disposable income equations for the different measures are derived. The determination equation of the net benefit for a UBI is straightforward and given by:
𝐵 = 𝑔 − 𝑡𝑌
Where B is the net benefit (when positive), or the tax paid (when negative), Y is gross income, g is the fixed and universal level of benefit and t is the tax rate. For the NIT the benefit determination consists of two equation, conditional on the level of pre-tax income.
𝐵 = 𝐺 − 𝑡 𝑌 if 0 ≤ 𝑌 < 𝑘 𝐵 = 𝑡 (𝑘 − 𝑌) if 𝑌 ≥ 𝑘
Where G is the maximum amount paid, t’ is the tax rate for the NIT and k is the threshold level where the tax shifts from positive to negative, or vice versa. So, for the case if Y = k the level of the benefit equals zero. The schemes for disposable income for UBI are given by:
𝑌 = 𝑌 + 𝐵 = 𝑔 + (1 − 𝑡)𝑌 And for NIT:
𝑌 = 𝑌 + 𝐵 = 𝐺 + (1 − 𝑡′)𝑌 if 0 ≤ 𝑌 < 𝑘 𝑌 = 𝑌 + 𝐵 = (1 − 𝑡 )𝑌 if 𝑌 ≥ 𝑘
Where Yd is disposable income after receiving benefit or paying taxes. From these equations,
when imposing 𝑡 = 𝑡 , the equilibrium outcome of the two programs will be 𝑔 = 𝑡 𝑘 if 𝑌 ≥ 𝑘. Using this equilibrium, it is possible to construct UBI and NIT schemes that produce the same level of disposable income given any level of gross income. Moreover, they can produce the same marginal and average tax rates. Tondani extends this to more realistic progressive tax schemes. The difference between the two measures is that a UBI uses a system with two transfers, namely a positive transfer for the lump-sum payment and a negative transfer for the taxes to be paid. Consequently, its costs are considerably higher, whereas NIT uses only one transfer.
Although equivalent in distributive outcomes, a UBI and NIT might have a different effect on the propensity to work through framing effects (Avram, 2015). He states that people react stronger to losses than gains. In other words, loss aversion has a greater effect on the utility function than gains in income. As UBI is constructed with a transfer of a lump sum first and a (higher) tax rate, people might be less incentivized to work more because they feel their losing a part of their benefit. A NIT does not cause such a disincentive, every extra hour worked increases their income. Considering these findings a NIT seems to be the first best welfare system from an economic point of view. Therefore, this paper will continue focusing on a NIT rather than on a UBI.
In short, a guaranteed income in the form of a NIT might be able to address the main criticisms on current welfare policies. However, there remains a great deal of disfavour on the unconditional character of these measures. Foregoing moral and ethical implications of this character, which are addressed in other fields of science, one main criticism remains. A guaranteed income possibly encourages idleness or, it creates a windfall for the people already working tempting them to reduce the hours of work. (Gamal, Balsan and Vero, 2006).
2.2 General labour supply theory
Classic microeconomic labour theory suggests there is a trade-off between hours of work creating income, and hours of leisure time (Pindick, 2013). If the wage rate increases, the opportunity cost of leisure time increases as well considering every hour of leisure time could have been spent on earning the wage rate by working. Figure 3 shows this relationship and how a change in wage affects the choice of labour supply. This effect can be decomposed in a substitution and an income effect. In the case of a wage increase from W to W’’ the substitution effect creates an increase in working hours as the opportunity costs of leisure time rise. This change is graphically represented by the switch from H to H’, which is tangent with the hypothetical budget constraint BC’. The income effect is negatively related to the hours of work. An increase in wage provides workers with the possibility to work less and consume the same amount of leisure, while increasing their utility from U to U’’. Whether the total effect is negative or positive ultimately depends on the shape of the indifference curves.
According to Gamel, Balsan and Vero (2006), the implementation of a basic income does not create a substitution effect as a basic income does not affect the wage rate. They therefore argue that the effect on labour supply ultimately depends on leisure being a normal or an inferior ‘good’.
Only if leisure is considered inferior, a basic income will undoubtedly create an incentive to work. This argument seems rather thin as it assumes that the net wage would not be affected. It seems more plausible that a basic income would be accompanied by a rise in the tax rate to be able to afford such a measure. As shown before this would result in a similar scheme as the NIT, thereby creating a substitution effect as shown in figure 4. The hourly wage rate is negatively affected through the implementation of a NIT. The budget constraint becomes less steep and a substitution effect is observed. However, when compared to figure 1, the decrease in hours worked is less severe than in the case of a w elfare program with a 100 percent tax rate. Nonetheless, the argument of Gamel et al (2006), that the effect on labour supply depends on leisure being an inferior or normal ‘good’, stays valid. A rise in the wage, or disposable income, would decrease the amount of leisure time when it is considered an inferior good, whereas they would increase it when leisure is normal.
Spencer (2004) refers to the famous example of the backward bending labour supply curve to provide an alternative explanation as to why in some cases the substitution effect, and in other cases the income effect might be stronger. Figure 5 shows how the decision of working hours depends on the wage rate. If lower income families, displayed at W, experience a wage increase to W’, they will increase their hours worked from the level H to the threshold point H’. When the wage rate increases even further to the point W’’, workers tend to reduce their hours to the point H’’. A possible explanation for this change might be that not only the quantity of leisure time is of importance, but the quality as well. At first people want to increase their quality of leisure time by acquiring a higher level of consumption and consuming higher valued goods. When this quality is satisfactory an increase in wage will provide them with the possibility to maintain this level with fewer hours of work. In other words, at the threshold W’ the quantity of leisure is inferior to consumption capacity. After the threshold it switches to being normal. So, in the area beneath W’ in figure 5 the substitution effect is larger because the marginal utility of income is higher than leisure. Above W’ the income effect is stronger as the marginal utility of leisure has surpassed the one of income. Figure 6 shows how this change in marginal utilities affects the labour supply. For the lower income wage categories an increase from W to W’ increases their hours of work from H to H’. They enjoy a relatively higher utility from work than from leisure. When the wage shifts up
even more to W’’ workers st art to work less up until the point H’’ indicating that their relative utility has changed in favour of leisure, so the turning point where the substitution effect changes from being larger to smaller than the income effect, is somewhere between W’ and W’’. If all changes in wage and worked hours are combined the backward bending supply curve from figure 5 can be derived.
Another view on labour supply is described in a paper by MaCurdy (1981). He states that workers determine their current labour supply in a life-cycle environment. Rather than determining it solely on the current wage level, workers determine their labour supply considering expectations about future wages, trying to smooth consumption over one’s lifetime. Conventional research considers labour supply in a one-period context. Consequently, omitting the expectation on disposable income in different periods will misrepresent total lifetime labour supply responses. For example, it might be the case that one’s wage increases in the current period and that this worker reacts by increasing his hours worked, only to be able to retire at a lower age. In this case, when only looking at the current period wage rate and hours worked, a positive relationship will be found. But looking over the lifetime of the worker might show he has lowered them, or at least increased them less. So, considering this case, the coefficient on wage will be upwardly biased in the short run looking solely in a one period context. This argument can be extended to the implementation of a NIT, as it will increase disposable income in the current period, its effect might be biased in the same way.
Another way that the influence of NIT might be biased in a one period context comes from job search models as the one from Mortensen (1970). These theories suggest that duration of job search depends on the wage a worker is willing to accept, also called the ‘asking wage’. As the asking wage of an individual increases, the probability of finding a job decreases. In this case an individual needs, on average, more time to find a job. An important determinant of the asking wage is the level of wealth when being unemployed. As Moffit states in his article on unemployment benefits, people who do not have, or have few assets or savings, cannot maintain a level of consumption when unemployed (2014). Their asking wage is therefore lower, decreasing the duration of unemployment. With the implementation of a NIT an individual can maintain a higher level of consumption when unemployed, increasing the asking wage and the
duration of unemployment. This argument can be extended to employed individuals. They might, because of the NIT, decrease their working hours to look for another job that suits them better. So, in the short run the NIT should have a negative effect on hours worked, but in the long run they might be induced to increase their worked hours because the new job possibly provides better working conditions and a higher wage.
In short, there are different arguments why and how people would adjust their level of hours worked after an implementation of a NIT. All these factors combined make it hard to anticipate what the overall effect will be. This paper will consider experimental data from the Mincome experiment to provide a better-informed expectation. The next part will provide a description of this experiment.
3. Mincome design 3.1 Experimental setup
Mincome was an experimental study that collected data on labour supply characteristics from 1974 until 1979. The main goal of the experiment was to test the feasibility of a NIT. The longitudinal dataset consists of observations on low-income, double and single headed households of the city of Winnipeg. People who were ineligible, on grounds of not being able to provide relevant labour supply responses, were excluded from participation. The main reason for ineligibility was having a higher average yearly income than $13 000 in the years 1972 and 1973. This number was indexed according to family size with a family of four being the basis. A list with the other ineligibility criteria is provided in appendix B. The data only includes intact households, that is, households that had no head split (divorce) or head join (marriage or common law).
The experiment consisted of 11 periods. It should be noted that the surveys were not taken on the same day for every participant. In every period the same order of interviewing participants was used to make sure that the period between the surveys was approximately 4 months for all. The first survey consisted of baseline interviews before the experimental treatment began. It used a longer period to increase the reliability of the baseline dataset. In the remaining 10 periods households received treatment or were employed as a control group. The treatments consisted of receiving a NIT at different support levels and tax rates. The payment P of benefits depended on the guarantee level G and the normal reduction rate NRR t, determined within the experiment, and the family net worth W and its net worth tax rate r, which were independent of the experiment. The NRR will be addressed as the tax rate in the continuation of the paper. The payment can be algebraically represented as:
𝑃 = 𝐺 − 𝑡 ∗ 𝑌 − 𝑟 ∗ 𝑊
Three different guarantee levels and tax rates were used, omitting the most generous treatment with the highest guarantee level combined with the lowest tax rate. These eight treatment groups and the control group are shown in table 1. Not all households in a given plan enjoyed the exact same level of guarantee. The guarantee levels shown in the table refer to a reference family that is double-headed, with a single earner and a family size of four. The actual
level was indexed over the years according to family size and adjusting for inflation. The participants were assigned to the different treatments or the control group using an optimal allocation model to get the least expensive allocation of families. Families were divided over the treatment groups based on family composition and ‘normal income’. Bigger households and households with relatively low income were less likely to be assigned to the most generous treatment groups. A more extensive explanation on the use of the optimal allocation model is provided in appendix B. The main message is that differences existed between the experimental groups prior to the experiment, so the analysis of the data should control for these pre-experimental differences to avoid encountering a sample selection bias.
Another problem prevalent in the Mincome experiment arises due to attrition. As can be seen in table 1 this attrition resulted in the collapse of plan 6 into plan 7 due to the fact that this plan wasn’t making enough datapoints to make inferences about. Special care must be taken when evaluating the effects of this treatment group. In table 2 frequencies of the different family types and treatment groups are given divided into completions and non-completions. The table clearly shows that attrition was prevalent, especially under male participants. The observations per treatment group are also quite low for single males and females, this poses challenges on achieving clear results. In table 3 the correlations of participant’s characteristics with attrition are presented. The table shows no to very low correlation. These numbers suggest that on average the people who did not complete the experiment do not share the same characteristics. In this sense it is unlikely that including non-completion to the analysis will result in an attrition bias. However, it must be noted that the attrition might be correlated with unobserved variables. The DD approach will be able to control for these omitted variables if these variables are individual specific and time-fixed.
Non-participation could possibly cause bigger problems. Although the Mincome experiment used a multi-phase random selection up until the enrolment stage, table 4 suggests that the actually enrolled do not provide a random selection. In the table some characteristics of households are presented by treatment group and enrolment status. The table suggests that refusals to enrol may cause a difficulty to make inferences about the population, because the characteristics between the enrolled and unenrolled show some big differences. People who were not enrolled had on average a higher age, a lower family income, fewer hours worked, a lower wage, and almost twice the level of welfare benefits compared to the people who did enrol. A lot of these characteristics seem to suggest that relatively more non-participants were expected to increase their working hours following from treatment. For example, theory suggests that a NIT increases the incentive to work relative to existing welfare benefits. Besides, people who work more hours have fewer hours left in the week for them to increase their hours worked. Also, the backward bending labour supply curve predicts people in a lower income category to increase their hours worked as their av erage disposable income increases. Together these findings seem to suggest that the experiment primarily consists of participants who are less likely to increase their working hours. The estimated effect of treatment on hours worked in the sample might therefore not be completely representative for the population of low-income households. It is expected that the people who refused would have had a relatively more positive effect on hours worked than the actual sample. This would imply that the estimates of this experiment will be negatively biased and should be dealt with accordingly. Appendix B provides a more extensive table including c haracteristics that did not seem to differ between participants and non-participants.
Another difficulty that could present itself in the context of the Mincome design are spill over effects on the control group. Take for example the situation that arises if participants will lower their working hours after receiving treatment. As a result, there are more hours of work available for other individuals in the economy. If these hours are primarily filled by people in the control group of the experiment, who might not have been able to expand their hours otherwise, the estimator of the treatment effect on the treated will be biased, it would not only measure their choice on decreasing labour supply. The effect on the control group will also contribute to the differences between the treatment and control, causing the estimator to overstate the effect of the treatment. However, since the sample is small relative to the population of Winnipeg at that time, this effect is assumed to be nihil. It is very unlikely that the spill overs will primarily affect the control group when they might just as well be appropriated by a large quantity of non-participants.
3.2 First labour supply responses
Graph 1 shows the trend in average weekly hours worked for all household types together and all treatment groups combined before and after treatment began. Two trends in the graph stand out. Firstly, the treatment and the control group seem to show the same trend in their hours worked prior to the start of the experiment. This suggests that the negative deviations from this trend after treatment are primarily caused by the treatment. The second trend that stands out is the similar response they display in the first period after treatment had started, that is, period three. This is probably an indication that the effect of the treatment suffers from a lagged response. It might be the case that participants had to get used to the new benefit system before they adjusted their hours accordingly. In consequence, the observations in this period might mitigate the real average effect of the treatment, causing the estimator to be inconsistent.
Graph 2 shows the responses divided per treatment group. The same two trends about the first three periods are observed here. However, not all treatment groups behave in the same way after treatment has started. It also seems that there are differences between the treatment groups
prior to treatment had started. A lso, it looks like treatment plans 1, 3 and 6, with the lowest guarantee level, show a reduction in their hours worked. Whereas plans 2, 4, and 7, with the middle guarantee level, seem to be stable, or even increase them. Furthermore, in contrast to wat was expected, given a guarantee level, the higher tax rates seem to return the highest supply of hours worked. This might be explained through the decreasing marginal returns to wage. With the higher tax rate, an increase in wage will return a lower dispensable income than under a lower one. People might reach the turning point of the backward bending supply curve later and are therefore less likely to decrease their hours facing a higher tax rate.
Graph 3 shows the average hours worked for male single headed households. As the previous graphs, this one also seems to behave fairly consistent in the periods preceding the treatment. However, overall the graph shows a very high variance which might be due to the fact of the smaller sample size per treatment group observed before in table 2. This high variance decreases the chance of finding significant results substantially. As the sample size decreases, the labour response of one participant has a very big effect on the average hours worked in the
treatment group. This way it will be very hard to make inferences on a population basis. A graph for female single-headed households, which is provided in appendix C, although less volatile, shows more or less the same problem. The data on double-headed households behaves as the graph of the total sample and seems to cause less problems.
Different ways are used to merge observations in order to create a less volatile group of observations. One of these methods included merging male and female headed households. Females where more represented in the experiment than males. And, as can be seen in the graphs in appendix C, they showed different labour supply responses than males. Merging them would therefore misrepresent the expected effect in the population. Graph 4 shows the labour supply responses for male headed households combined on the tax rate. It looks like the volatility has substantially decreased, achieving the main goal of the merge. Combining the treatment plans on their guarantee level shows similar results in reducing volatility. This graph, together with the graphs for all household types are presented in Appendix C. However, special care must be taken when interpreting the results of these newly formed groups. It might well be the case that, given a certain income level, the tax levels show very different responses, making it harder to provide inferences on the combination of different guarantee levels and tax rates. Only in combination with other results these estimates provide meaningful information. This paper will continue by identifying a suited way to estimate the data.
4. Statistical analysis
4.1 The difference-in-difference methodology
As noted in the explanation of the Mincome experimental design, participants were allocated to different treatment groups based on characteristics of income and family size. As can be seen from the work response graphs, this resulted in differing hours of labour supply across treatment groups before treatment started. A simple linear estimation on the effect of the treatment would in this case include the pre-treatment differences and consequently bias the estimator when comparing the outcome in the control and treatment groups. The difference-in-difference (DD)
estimator might be able to control for these differences. It compares the differences over time between the treatment groups. This way the DD estimator controls for unobservable individual fixed effects and common macro shocks (Blundell & Dias, 2009). Formally the estimator for the simple model consisting of only two periods and two groups can be written as:
𝑌 ̇ = 𝛼 + 𝛾𝑇 + 𝜆𝐷 + 𝛽𝐷 ∗ 𝑇 + 𝑢
Where Yist is the average hours worked for the ith household of treatment group s in period t . Tt is
the indicator variable taking on the value of one for the period after treatment has started. Ds is
the indicator variable that takes on the value of one when assigned to the treatment group. Ds * Tt
is the interaction term to capture the isolated effect of the treatment on hours worked, and uist is
the residual term. In this model α captures the expected hours worked by the participants in the control group before treatment. γ covers the expected differences over time independent of the treatment assignment. Otherwise said, it captures the effect of the trend in the economy. λ estimates the expected difference between the treatment groups and the control group before treatment starts. The coefficient β is the parameters of interest, it captures the average treatment effect on the treated. uit is the residual term, which is assumed to equal 0 (Angrist & Pischke,
2009).
The equation above, models the regression for one treatment group in two periods. Since the Mincome data set deals with eleven periods and eight different treatment plans, this equation can be extended by including indicator variables Pt, for every time period excluding the first, and
expanding Ds and Ds * Tt for all treatment groups excluding control (Stock & Watson, 2015). The
term Xnistrepresents the possibility to extend the model with n different control variables. These
variables might be included to capture non-constant individual specific effects, or different macro-effects that differ between groups. This model is given by the following scheme:
𝑌 = 𝛼 + 𝛾 𝑃 + 𝜆 𝐷 + 𝛽 𝐷 ∗ 𝑇 + 𝜃 𝑋 + 𝑢
Where Ptare the indicator variables for the t different time periods in which the surveys where
undertaken. These variables replace the variable Tt, that only captured the effect when two time
periods were used. The term for the treatment effect and the interaction term are extended to include all different treatment plans. The coefficients γt and λs can be interpreted like the ones in
the simple regression model, only extended for different time periods and different treatment groups respectively. When there are no additional regressors in the model to control for the group and individual time-variant effects, α can also be interpreted as in the simple model. However, when including controls its meaning will change. Part of the pre-treatment differences between the groups will then be explained by those variables. The initial differences would then become 𝛼 + 𝜃 𝑋 .
The DD estimator relies on two critical assumptions. The time effects, or macro-shocks, across the different treatment groups should be common. The DD estimator captures the effect of a deviation from the trend across treatment groups. If part of this deviation is due to differing time effects across treatment groups, the DD estimator will be biased. Since the different treatment groups and the control group are all situated in the same city it is safe to assume that in this
experiment the macro-effects are the same, and that this assumption should automatically hold. The second assumption is that average individual specific effects are fixed. This also implies that there may not be any systematic composition changes within a group which could influence the average. If the effects are not constant over time, differencing will not control for this change and the individual specific effects will bias the DD estimator (Blundell & Dias, 2009). If non-constant individual effects or non-constant macro-effects are observed, it is possible to control for them by including additional regressors that are represented in the model by Xnist.
In the Mincome experiment attrition has been a big problem. Although the table of correlations suggest there not to be a relation to observed individual specific characteristics, there might be unobserved effects that do influence the average, causing the DD estimator to be inconsistent. One way of controlling for this problem is excluding the non-completions per period from the regressions, resulting in a big loss of information. This way the panel data is balanced, making the estimator consistent. However, making inferences about the population will be more difficult if the characteristics of the non-completed participants are unknown. Lechner, Rodriguez-Planas and Fernández Kranz (2016) argue that if the coefficients on the balanced and unbalanced panel differ, attrition causes serious problems for the estimator. In this case the balanced data should be used to keep it consistent. Table 5 displays the average labour response per period of participants who completed the experiment versus those that did not. This table shows that the average response in hours differed significantly between the two groups. The people who did not finish the experiment, are completely represented in the pre-treatment period. After this period, they are only partly represented, biasing the group average to the value of the participants that did complete the experiment. This way the difference over time might show a different trend than it would have if the drop-outs had completed the experiment. For example, especially in plans 1 and 5, the participants who eventually dropped out show much lower average supply of hours before treatment than the group of completions. Together with the
fact that the average hours after treatment has risen faster, it might indicate that households with a lower initial supply of labour dropped out more. This means that the lower supply of labour will be underrepresented in the treatment group after treatment has started and will not represent the real common time trend in these groups. In the case of plans 1 and 5, this would result in a lower estimate of the average treatment effect, as a higher part would be attributed to the common trend of the treatment group. Appendix D provides a formal derivation of the statement, as to why the difference-in-difference estimator will show causal effects. This paper continues by analysing the results of the regressions.
4.2 Results
Table 6 shows the estimated coefficients1 of the model for all households where the treatment
groups are merged to form one group2. For readability the coefficients on the period dummies are
omitted. Appendix E provides a table including these coefficients. The model is modified in different ways, creating 6 models in total. In all six models the coefficient on the treatment dummy is not significantly different from zero, which indicates that, on average, the labour response in the treatment groups and the control group did not differ significantly prior to the start of the treatment. Furthermore, the first 4 models all provide roughly the same estimate for the constant and are all statistically significant. Concluding from the definition of the model this estimate should be the value of the pre-treatment average hours worked for all households in the control group. As can be seen from the lines in figure 1 this value comes very close to the real value. This is a first indication that the model seems to fit the data pretty well.
The first model, unmodified, includes all observations and does not show a significant result on the DD estimator. When excluding non-completion in model 2, it becomes apparent that the
1Robust standard errors are used to control for heteroskedasticity. Figure 8 in appendix A shows a scatterplot of the residuals against the fitted values and clearly rejects the assumption of constant variance of the residuals.
2The households are differently represented in the sample and show different results. Estimations for these house-holds will therefore not accurately depict the population.
attrition problem does indeed cause a bias in the estimator. The DD estimates for the unbalanced and the balanced model deviate considerably, implying that the individual specific characteristics were different for the participants who did not finish the experiment relative to those that did. Also, the DD estimator becomes significant at the 5 percent level, increasing the accuracy of the model. This leads to the decision to only include the balanced panel-data into the model.
Model 3 tries to deal with the problem of the lagged treatment effect mentioned before. This model treats period 3 as-if being pre-treatment, allowing all observations to stay within the model. As opposed to what was expected this method decreases the magnitude of the DD coefficient. The fourth model excludes period 3 altogether to control for the lagged response. It seems to do better than model 3, since the coefficient increases in magnitude as expected. The r-squared of this model is also marginally higher than the one of model 3, suggesting that this model performs better in explaining the variance of the labour supply decisions. It is from here on assumed that this model predicts reality better and this paper will therefore exclude this period in the following models.
Models 5 and 6 include wage and wage squared respectively, the most important determinant of labour supply decisions. These variables are included to control for the individual specific non-fixed effect of the wage rate. Wage for double headed households is calculated by taking a weighted average, using the relative hours worked between the male and female as weights. The variable wage squared is added because, as mentioned before using the backward bending supply curve, wage is not expected to be linearly related to the supply of worked hours. Instead, hours are expected to increase less, or even decrease, in the wage rate. This variable is therefore supposed to control for the expected non-linearity in wage. As can be seen, the coefficients on wage in these two models are significant and have, as expected for low-income households, a positive effect on hours worked. As the coefficient of the squared wage is also significant, it can be assumed that the model is in fact not linear. The negative value also indicates that the marginal utility of wage does in fact decrease. Following regressions will control for this non-linearity by including the wage and wage squared variables. One curious thing about model 6 is that the DD estimator is no longer significant at the 5 percent level, although the explained variability of the model has increased considerably. This possibly means that the effect of the treatment is less strong than observed before, or perhaps even non-existent.
Table 7 shows the regression estimates including the regressors for all different treatment plans. Again, the coefficients of the periods are omitted for readability. The first thing that stands out is that the standard errors for the single headed households, especially males, are much higher than those of the double headed households. This is probably caused by the lower sample size per treatment group for single headed households. The same is observed for male headed households in relation to females. It is therefore too soon to infer, on account of the non-significant results, that the treatment did not have an effect on single-headed households. Moreover, the regression for males show DD estimates that are significant at the ten percent level for plans 2 and 4. These plans both have the middle guarantee level of $ 4800 dollars and differ on tax rate, where plan 2 faces 35 percent and 4 faces 50 percent. The coefficient for plan 2 is strongly positive, whereas the one for plan 4 is strongly negative. Double headed households show predominantly negative coefficients. The DD coefficients of plans 2 and 3 are negative and significant at the 10 percent
Table 7. Regression estimations for all plans and households using model 6
VARIABLES Male head Female head Double head All households
Plan 1 1.510 1.285 3.249 2.204 (3.214) (2.333) (2.628) (1.748) DD Plan 1 -1.680 -0.627 -6.836** -5.124*** (3.908) (2.699) (2.881) (1.950) Plan 2 -6.765** 2.812 0.331 0.492 (2.906) (2.984) (2.125) (1.717) DD Plan 2 7.050* 0.201 -4.540* -1.586 (3.746) (3.428) (2.453) (1.965) Plan 3 -2.318 -0.187 -4.136* -2.461 (3.001) (2.247) (2.276) (1.584) DD Plan 3 1.759 1.864 -4.904* -2.052 (3.377) (2.591) (2.650) (1.802) Plan 4 2.539 1.576 -0.235 0.379 (3.562) (2.601) (2.234) (1.732) DD Plan 4 -7.205* 0.425 -2.830 -1.500 (4.075) (3.215) (2.460) (1.944) Plan 5 3.189 2.705 3.118 2.328 (4.802) (3.869) (3.207) (2.390) DD Plan 5 -3.151 -2.990 -7.989** -5.124** (5.259) (4.073) (3.404) (2.538) Plan 6 1.833 -2.729 3.140 1.140 (4.504) (2.269) (2.462) (1.794) DD Plan 6 -1.164 2.421 -4.123 -1.905 (5.105) (2.657) (2.753) (2.016) Plan 7 -8.637** 1.779 -0.606 -2.236 (3.777) (2.685) (2.911) (2.226) DD Plan 7 4.061 -1.758 2.188 1.096 (4.407) (3.779) (3.274) (2.582) Plan 8 -2.724 -0.610 1.503 0.720 (8.049) (2.839) (2.894) (2.145) DD Plan 8 3.206 1.965 -1.651 0.111 (8.425) (3.344) (3.206) (2.392) Wage 0.0833*** 0.129*** 0.127*** 0.126*** (0.00420) (0.00562) (0.00543) (0.00327)
Wage squared -5.43e-05*** -0.000115*** -0.000128*** -0.000109***
(6.37e-06) (1.32e-05) (7.58e-06) (5.33e-06)
Constant 5.330** 6.071*** 16.72*** 10.56***
(2.458) (1.417) (1.764) (1.114)
Observations 567 1,888 3,399 5,854
R-squared 0.557 0.511 0.403 0.478
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
level and the DD coefficient of plans 1 and 5 are negative and significant at the 5 percent level. The coefficients on plans 1 and 5 are enforced by the results for all households combined, which also show a significant negative effect. All the coefficients combined show that the plans with the lower tax rates respond by decreasing their hours, while plans 7 and 8, with the highest tax rate, show a coefficient which is not significantly different from zero and the one on plan 7 even has a positive coefficient. This result is corroborated when looking at the other household types and suggests that people respond the least when facing the highest tax rate. This can be explained by noting that the dispensable income increases the least for these people implying that they suffer less from the decreasing marginal utility of wage. They therefore do not decrease their hours as much as the other groups. This paper will continue by investigating the effects for the merged treatment groups.
Table 8 shows the new composition of the treatment plans after having been merged on their tax rate and on their guarantee level respectively. For the plans merged on the level of guarantee the original plan 6 is omitted because this plan was collapsed in plan 7 during the course of the experiment. The participants in plan 6 have thus experienced two different guarantee levels. Merging them together on basis of one of these values will therefore misrepresent the true average effect of the level of guarantee. Merging on the tax rate does not
cause any problems as plans 6 and 7 had the same rate for the complete duration of the experiment.
Table 9 displays the regression estimates of the newly combined treatment plans. Increasing the sample size per treatment group seems to provide some more reliable results. The standard errors for the single headed households come closer to those of the double headed households. On average it looks like single headed households react less strong to the policy change than double headed households. This seems to be coherent considering single headed households do not have a partner with an income to fall back on. Also, corresponding to the previous regressions, the highest tax rate gives the least negative estimate for double headed households and the coefficient on single headed households even turns positive. The results for the plans merged on level show a larger effect for the highest and the lowest guarantee level. For the highest guarantee level this can be explained through the income effect. As disposable income increases people need to work less hours to maintain a certain consumption level. In case of the lowest guarantee level, the turning point, where the negative income tax turns positive, is reached at a lower level of hours. People need to decrease their hours more to be able to reap some of the benefits of the treatment.
Overall the data shows that, on average, the implementation of a negative income tax primarily has a negative effect on the supply of labour. However, the effects across treatment groups and households show different results. Double headed households seem to reduce their hours more than single headed households do. Also, the negative labour response is less severe, and often not significantly different from zero for the highest tax rate. The effect of the level of the guarantee is subtler. When implementing a guarantee level that is too low, the labour supply tends to be largely negative and when it is too high the same effect is observed. The level of guarantee should therefore be somewhere in the middle to minimize a negative labour supply effect. The next section provides a discussion on whether it is possible to make inferences on a population based on these results.
4.3 Discussion
The results found in the previous section imply a predominantly negative effect on labour supply. However, there are treatment groups that show no to a very little effect. To apply these results on a population basis, the sample used should be a good representation of it. Mincome experimented on low-income households, meaning that households with a higher income might show different responses. However, the implementation of a negative income tax primarily affects the people with a lower income directly. They will receive a payment if their income is lower than the
Table 10. Treshold values k for various combinations of guarantee level and tax rate. Tax Rate
Guarantee level 35% 50% 75%
$ 3800 $ 10 857 $ 7 600 $ 5 067
$ 4800 $ 13 714 $ 9 600 $ 6 400
threshold value k. People earning closely above this threshold value might be induced to reduce their supply of labour to be able to reap some of the benefits of the policy. If they are too far above, the negative effect of the decrease in income will almost certainly be stronger than the perceived positive effect of the payment. Table 10 shows the threshold values for the combinations of the tax rate and guarantee level used in the treatment plans. Since Mincome only included low-income households with an income up until $13 000, the two most generous treatment plans with a tax rate of 35 percent and a guarantee level of $4 800 and $5 800 seem to be misrepresenting the population. Since the threshold value of these two plans lies above the maximum income to be enrolled, it will not represent the higher income households that are also expected to be affected by treatment. The treatments with a guarantee of $3 800 and a tax rate of 35 percent, and with a level of $5 800 and a tax rate of 50 percent have a threshold value that is below, but close to, the maximum level of income. For these treatments there also might be people not included in the experiment that would have decreased their hours. The treatments discussed above are therefore likely to show a larger negative effect in the population. For the other treatment plans the difference between the maximum income and the threshold value seems large enough to be a good representation. The treatment plans 7 and 8, that showed the most promising results in favour of a NIT, are therefore expected to be unaffected by the exclusion of higher income households and show estimates that are a good representation on an economy wide scale.
Another reason why the treatment effects might not be universal is that economies are different across countries and over time. Since the experiment was conducted in Canada during the seventies, it might be hard to make inferences about economies now, not situated in Canada. Firstly, it is only possible to make inferences about populations that show the same economic characteristics as the Canadian one. If characteristics differ a lot, it might well be the case that the labour supply decisions of economic agents also differ. Canada´s economy might be best described as a highly developed one. It is assumed that most developed countries share, for a large part, the same economic characteristics and will provide similar labour supply responses. Moreover, from the introduction of this paper it follows that the case of a guaranteed income is mostly relevant for higher developed countries. These countries show increasing income inequality and have a higher share of capital relative to labour. The experiment is therefore highly representative for countries that face the necessity of a tax and benefit reform. The second difference, over time, might cause bigger problems. Although the economy of Canada belonged to the developed ones as it does now, all developed countries have evolved over the years. For example, in the seventies the family has been the centre of civilized society, whereas nowadays it has become a lot more individualistic and women are participating in the labour force more than ever. Consequently, households might respond very differently than they did in the seventies. However, it is expected that this development only favours the case of a guaranteed income. The results of this paper suggest that single headed households showed a lower negative labour supply than double headed households. Since single headed households share more individualistic characteristics it is expected that a NIT would provide less negative results than it did in the seventies. Nevertheless, only a new experiment conducted in today’s economy can provide conclusive statements about the change in labour supply responses of economic agents.
A more intrinsic flaw of the Mincome design comes from the duration of the experiment. As discussed before, it is expected from the lifecycle hypothesis that economic agents adjust their supply of labour according to expected future wages. Since the duration of Mincome was approximately three years and the experiment did not go on indefinitely, people only saw a change in the current net wage level. If the net wage change would have been expected to be the same for all coming periods, this would affect labour supply decisions less. Because people knew the experiment was going to end, they might have changed their current hours upwards to reap the benefits of the temporary increase in net wage maximally. However, this experiment has not shown any convincing evidence suggesting that households showed a positive response to the policy change. This possibly suggests that people make their labour supply decision more based on the current period than in the context of the lifecycle hypothesis. Also, as argued before, the job search models predict people who are looking for a job will take more time, resulting in a lower labour supply on the short term. However, the increased time of the job search also increases the possibility of a better fit between the occupation and the individual. On the long run they might even increase their hours because of the increased job satisfaction. In this context the estimators will show a downward bias in the short run. A more extensive study should be provided to be conclusive about the different effects between the short and the long run.
The effect of non-participation earlier discussed, might cause the sample estimate to deviate from the true value in a more apparent way. Almost all characteristics that were more prevalent under non-participators, suggest that in the final sample households that were expected to increase their labour supply following a NIT were underrepresented. For example, the exclusion of people that had higher unemployment benefits are expected to increase their hours following treatment. Additionally, on account of the inactivity trap, that would still prevail in the control group, these participants were expected not to increase their hours. So the effect of the exclusion is twofold. The treated participants in the sample show a lower increase in hours than expected for the population, whereas the control group shows a higher increase. The sample therefore shows a larger negative effect of the treatment than probable on a population basis. On ground of the characteristics of the non-participators, it is therefore expected that people decrease their hours less, or maybe even increase them, resulting from an implementation of a negative income tax.
The last problem within the Mincome experiment is also the hardest to analyse. This problem is the one of attrition. Although table 3 showed no correlation between attrition and observed individual specific characteristic, there might be unobserved specific effects that differed between the people that did and did not complete the experiment. The values of the DD estimator are expected to be misrepresented, on account of the unbalanced panel supplying very different DD estimates than the balanced one. However, it remains unclear if the estimator misrepresents the population value in negative or positive way. A closer look at the labour supply responses of non-completions might clarify this.
Table 5 showed the average hours worked per period for the completed and non-completed participants. The hours of the latter behave rather ambiguously over the different periods. Looking at the average hours in the control group before and after treatment seems to show a trend. The non-completed participants showed higher average hours before treatment started,
indicating that high supply of labour is underrepresented in the group of completions for the control group. The average hours after treatment has started for the non-completed participants is impossible to interpret as it is unknown how these participants would have responded in the periods after they dropped out. For the control group it might be the case, seeing that the non-completed increased their hours less, that this group consisted of people who would also have given a lower labour supply response in the periods after they dropped out. On the other hand, this average might be caused by the drop out of participants that would have increased their hours relatively more than the completed group. In the first case, the estimator of the labour supply in the control group shows a higher value than for the real population, resulting in a stronger negative measured effect for the treatment groups. If the second case is true, the group of non-completed participants would have shown a higher increase in their labour supply resulting in the average of this group coming closer, or even passing the increase in labour supply of the group that did complete the experiment. For the control group it seems more likely that the attrition bias causes a negative effect on the estimates of the treatment effect. Only if the second case is true and the non-completed participants would have surpassed the increase in hours of the completed group, the bias in the estimator would have been positive. Since the non-completed group started out at a higher initial value of average hours worked, it seems unlikely that they would have increased their hours more, since there is less leisure time left in the week for them to substitute for work. Additionally, people face an increase in the marginal utility of leisure time when increasing their hours of work, implying they would be less likely to it for labour. So, based solely on the difference between the average hours worked before and after treatment has started for the completed and non-completed group, it is possible to form an expectation on which way the estimators might misrepresent the population.
In case of the treatment groups, plans 1 and 5 show a lot lower average supply of hours for the non-completed groups. Following the argumentation above this means it is expected that with the inclusion of these participants the average increase in hours would have been higher. The estimators of these groups are therefore expected to be lower than for the population. Groups 3, 4 and 6 showed a lot higher average before treatment started. It is therefore expected that they show a labour supply that is too high. The remaining groups show smaller differences and are harder to make inferences about. However, the negative effect resulting from the attrition in the control group affects all treatment plans. This way the positive results will be mitigated, and the negative ones will be amplified. These findings suggest that the exclusion of the non-completion mostly causes the sample estimators for the treatment groups to misrepresent the population values in a negative way. It is therefore expected that the actual coefficients for the population will show, on average, a less negative, and in some cases maybe even a positive value.
5. Conclusion
The main analysis of this paper shows that a NIT has a negative treatment effect on the participants. The external validity of this experiment is low due to the design, timing and placement. This makes it difficult to provide hard evidence on the effect of the implementation of a NIT on labour supply. Moreover, it is likely that for a sub-population a particular combination of
the level of guarantee and the tax rate could produce low, negligible or even positive effects on the supply of labour. This opposes the view that a negative income tax will surely induce idleness. It seems therefore likely that combinations of the tax rate and guarantee level minimizing negative, or maximizing positive labour supply responses, could be implemented without causing a substantial increase in costs by a significant decrease in labour supply. What the values of these levels are and whether this measure is even affordable without a significant decrease in labour supply, should be the focus of future research. Another major point is that this paper merely examines a quantitative effect. Whether the observed effects are socially desirable depends on the reasons why people change their labour supply and what effects it has on other aspects of society. Hence, using results of qualitative approaches on the desirability is deemed necessary to ultimately plead in favour or disfavour of implementing a negative income tax.
References
Angrist, J. D., & Pischke, J. S. (2009). Mostly harmless econometrics: An empiricist's companion. Princeton university press.
Avram, S. (2015). Benefit losses loom larger than taxes: The effects of framing and loss aversion on behavioural responses to taxes and benefits (No. 2015-17). ISER Working Paper Series.
Amétépé, F. (2012). The effectiveness of Luxembourg's minimum guaranteed income. International Social Security Review, 65(1), 99-116.
Autor, D.H. (2015). Why are there still so many jobs? The history and future of workplace automation. Journal of economic perspectives, 29(3), 3-30
Blundell, R., & Dias, M. C. (2009). Alternative approaches to evaluation in empirical microeconomics. Journal of Human Resources, 44(3), 565-640.
Carey, D., & Tchilinguirian, H. (2000). Average effective tax rates on capital, labour and consumption. OECD Economics Department Working Papers. No. 258. OECD Publishing. Paris.
Gamel, C., Balsan, D. & Vero, J. (2006). The impact of basic income on the propensity to work Theoretical issues and micro-econometric results. The Journal of Socio-Economics. 35. 476–497
Hum, D., & Simpson, W. (1991). Income Maintenance, Work Effort, and the Canadian Mincome Experiment. Economic Council of Canada.
Hum, D., & Simpson, W. (1993). Economic response to a guaranteed annual income: experience from Canada and the United States. Journal of Labor Economics, 11(1, Part 2), S263-S296. Keeley, M. C., Robins, P. K., Spiegelman, R. G., & West, R. W. (1978). The estimation of labor supply
models using experimental data. The American Economic Review, 68(5), 873-887. Lechner, M., Rodriguez-Planas, N., & Fernández Kranz, D. (2016). Difference-in-difference
estimation by FE and OLS when there is panel non-response. Journal of Applied Statistics, 43(11), 2044-2052.
MaCurdy, T. E. (1981). An empirical model of labor supply in a life-cycle setting. Journal of political Economy, 89(6), 1059-1085.
Moffitt, Robert, A. (2003). "The Negative Income Tax and the Evolution of U.S. Welfare Policy." Journal of Economic Perspectives, 17(3), 119-140.
Moffitt, R. (2014). Unemployment benefits and unemployment. IZA World of Labor.
Mortensen, D. T. (1970). Job search, the duration of unemployment, and the Phillips curve. The American Economic Review, 60(5), 847-862.
Murray, C. (2008). Guaranteed income as a replacement for the welfare state. Basic Income Studies, 3(2).
Pindyck, S. & Rubinfeld, D. (2013). Microeconomics. New Jersey, US: Pearson. pp 540-544 Sachs, J. D., Benzell, S. G., & LaGarda, G. (2015). Robots: Curse or blessing? A basic
framework (No.w21091). National Bureau of Economic Research.
Stock, J. H., & Watson, M. W. (2015). Introduction to econometrics: Global edition. Boston, MA: Pearson Education.
Watts, H., & Conlisk, J. (1969). A model for optimizing experimental designs for estimating response surfaces. American Statistical Association Proceedings, Social Statistics Section (pp. 150-156).
Wolff, E. N. (2010). Recent trends in household wealth in the United States: Rising debt and the middle-class squeeze-an update to 2007. Levy Economics Institute Working Papers Series, No. 159.
Data
Mason, G. (2017). Mincome Baseline Data (Minc1). Doi/10.5203/FK2/NV200L. University of Manitoba.V1.0. Retrieved from http://dx.doi.org/10.5203/FK2/NV200L
Mason, G. (2017). Mincome Longitudinal Labour Market Data (Minc4). Doi/10.5203/FK2/PO1F6R. University of Manitoba. V1. Retrieved from http://dx.doi.org/10.5203/FK2/PO1F6R
Appendix B Mincome characteristics List of additional ineligibility criteria:
1. Households with either head over 57 years of age as of September 1, 1974 2. Mentally incompetent households
3. Households with a language barrier to answering in English 4. Households with one or more heads in the armed forces 5. Households with disabled adult members
6. Members of religious order 7. Institutionalized households 8. Employees of Mincome Manitoba
9. Households with more than 5 roommates living in the same dwelling Mincome optimal allocation model:
The allocation model used in mincome was based on the model developed by Conlisk and Watts (1969). It analyses the benefits versus the costs of various allocations subjected to a budget constraint. Reductions in variance are the benefits and the costs are the expected costs of payments. The model can be formally stated using the notation in Hum and Simpson (1991) as Minimize F(n1,…,nm) = tr [W var(Pb)],
Subject to
ci ni < C (L[n1,….,nm] = 0),
ni > 0 (i = 1,…..,m),
where
m = the number of design points
ci = the cost of one observation at the ithdesign point
ni = the number of observations at the ith design point
C = the total available budget
F (n) represents the criteria for the optimal design and is specified in terms of estimation error var(Pb). Pbare obtained using a regression model estimating a response function. The optimal
allocation is the one that minimizes the weighted sum of the estimation error. The weights (W) are chosen on the basis of the policy relevance of the elements of Pb .
Table 11. Household characteristic by treatment plan and enrollment
Enrollment
Treatment Plan 1 2 3 4 5 6 7 8 Control Yes No
Age male 30.1 32.2 32.2 33.6 32.3 33.5 34.6 33.1 34.2 32.9 39.0
Age female 29.4 30.5 31.2 32.2 31.5 28.9 33.1 33.4 33.9 31.6 36.1
Homeowner 0.2 0.3 0.2 0.4 0.3 0.3 0.3 0.4 0.3 0.3 0.4
Total fam. income 1974 8461 8141 7785 8479 8399 7377 8164 8676 7229 8079 6941
Male years of school 11.4 10.3 9.6 10.2 10.1 9.9 10.5 11.1 9.6 10.3 7.6
Female years of school 10.2 10.2 9.6 9.6 10.0 10.5 9.3 10.0 9.5 9.9 8.8
Debt 1568 1529 1306 1903 1774 1294 1241 1620 1135 1485 1260
Male job satisfaction * 2.1 2.1 2.1 2.2 2.0 1.9 2.0 2.5 2.0 2.1 2.0
Female job satisfaction * 2.4 2.0 2.1 2.1 2.3 1.9 1.8 1.9 2.2 2.1 2.0
Number of children 0.5 0.6 0.5 0.6 0.5 0.9 0.6 0.6 0.6 0.6 0.5
Index family size 75.0 81.4 81.7 92.4 83.3 87.3 97.9 90.0 88.7 86.4 87.7
Welfare benefit 1974 186 171 277 237 191 189 141 258 294 216 419
Male number of jobs 1974 1.1 1.1 1.1 1.2 1.2 1.1 1.1 1.1 1.1 1.1 0.9
Female number of jobs 1974 0.5 0.8 0.7 0.6 0.7 0.8 0.7 0.7 0.6 0.7 0.5
Male worked hours last week 391 358 364 379 381 381 382 397 349 376 327
Female worked hours last week 348 360 344 320 327 331 322 351 336 338 318
Male wage last week 269 225 176 293 250 254 264 243 222 244 181
Female wage last week 231 222 236 212 216 232 240 245 225 229 205
