(This is a joint post with Donal O’Neill (NUIM) and Frank Walsh (UCD))
Because of the media storm last week about the Tol et al. paper on working costs, myself and a few colleagues (separately) decided to read the paper to see what all the fuss was about. We are all labour economists, used to using data on individuals, households and firms to address questions relevant to public policy issues. Our assessment of the paper is below.
The basic approach of the paper is as follows: it uses Household Budget Survey (HBS) data to examine the consumption patterns of different types of households. HBS data is collected at the household, rather than the individual level so the analysis distinguishes between households whose chief earner is employed and those whose chief earner is not employed. In particular, it examines the consumption patterns of these two household types under four headings: transport, childcare, heat & light, and takeaway food. To the extent that the amount spent by these households differ, the difference is designated a cost of working.
Without going any further, the problems with this approach are clear. Households headed by earners would have higher expenditure than households headed by unemployed people even if the earners incurred zero costs of working, simply because they have higher incomes. If you have a higher income, you are more likely to go for a spin in the car at the weekend, to buy a takeaway on Friday night, to employ a nanny rather than use the local childminder, and to leave the heating on if it’s chilly. It is therefore very important to take this income effect – which is going to be substantial – into account before labelling the difference in expenditures as working costs.
This is actually very tricky to do at all, and even trickier to do well. However the approach adopted by Tol et al. paper is not statistically valid – instead of modelling selection into employment and taking that into account in predicting expenditure for the two groups, they model positive expenditures; the econometric analysis in the first half of the paper is therefore flawed in that it fails to address the non-random selection of individuals into unemployment.
The use of these estimates in the second part of the paper is even more confusing: having obtained predicted probabilities of having positive expenditures (for each of the four categories) separately for households headed by workers and those headed by unemployed people, they multiply the conditional mean expenditures by these probabilities. For example, they multiply the mean transport expenditure for worker households (which as far as we can see includes those with zero expenditure, since the minimum value is reported as zero) by the probability of observing positive transport expenditure in worker households; then they multiply the mean transport expenditure for non-worker households by the probability of observing positive transport expenditure in non-worker households. The difference is then called costs of transport to work. This makes no sense to us; the measure of cost obtained in this way does not correspond to any meaningful summary statistic that we are aware of and should not be used as the costs of transport to work.
Furthermore, it does not appear to be possible to distinguish between one- and two-earner households in the data. This means that where costs of working do arise (assuming they were properly estimated in the first place), it is not possible to tell whether they should be spread over one or two individuals. In the case of childcare (and to a lesser extent, heat & light), this is crucial: if the chief earner is employed but his wife is not, he will not incur work-related childcare costs, whereas if his wife is working, he will. The Tol et al. paper does not address this problem because it seems that the data don’t allow it. If this is true then the HBS is just not appropriate for addressing this question.
In addition, we have several, more detailed, technical criticisms of the econometric methodologies employed at various points in the paper, including concerns about the estimation of incomes using a Tobit model when the zeroes appear to be due to missing data rather than any labour market behaviour. The inappropriate use of the Tobit model in this context will introduce biases in the estimates of incomes in addition to the biases in the costs data outlined earlier.
We have sent these comments to Richard, and he has not rebutted them. We agree with him that this is an important topic: costs of working do arise, and they may indeed be substantial, thereby creating a disincentive to work. But while Richard thinks that his estimates, though not perfect, are 90% there, we disagree; in our view, the problems with the data and methodology are so severe that these estimates are not informative about the disincentives associated with the cost of working.