I was in a meeting the other week where we were wrestling with the issue of how to capture better labor supply in agricultural surveys. This is tough – the farms are often far from the house, tasks are often dispersed across time, with some of them being a small amount of hours – either in total or on a given day. Families can have more than one farm, weakening what household members know about how the others spend their time. One of the interesting papers that came up was a study  by Elena Bardasi, Kathleen Beegle, Andrew Dllon and Pieter Serneels. Before turning to their results its worth spending a bit more time discussing what could be going on.
Two things would seem to matter (among others). First, who you ask could shape the information you get. We’ve had multiple posts in the past about imperfections in within household information. These posts have talked about income and consumption and while labor would arguably be easier to observe, it may suffer from the same strategic motives for concealment and thus be underreported when the enumerator asks someone other than the actual worker to respond on this.
The second issue is how you ask the question. Take the case of unpaid work – think about helping out at the farm (for example doing weeding once in awhile) or helping out in the family kiosk. Or processing food for sale in the family kiosk. These could be underreported because norms dictate that these aren’t really “work” since they are not only non-salaried, they are somewhat more distant from the actual cash transaction. Moreover, who you ask can compound the problem. If you ask the husband, for example, he may be less likely to regard what his wife is doing as “work” because his wife is only doing what she is “supposed to be doing”.
These issues could add up to significant differences. Bardasi et. al. provide data from two national Tanzanian surveys from the same year. According to the Integrated Labour Force Survey, labor force participation for men is 90.6 for men, 89.5 percent for women. Meanwhile, the household budget survey s shows participation of 91.1 percent and 82.4 percent – a significant difference for women.
In addition to an informative literature review, Bardasi and co. provide some evidence on how different approaches can play out by doing a field experiment (of survey methods) in Tanzania. They randomly assigned respondents to a two by two set of possible labor questions – a longer module versus a short module and responses from the worker him or herself versus a response by a proxy. The proxy vs. self report is pretty straightforward – here they compare self reports in households where others (randomly selected adults) were also asked to cross-report on the respondent. In terms of the length of the household questionnaire, the analog for me was a survey in which labor is a main subject/outcome of interest versus another where I might think it interesting, but it’s not a top priority. In terms of labor force participation, in the long module Bardasi, et. al. separate out working outside the household, working on the household farm and working on the household enterprise. For the short module, the participation question is simply did the respondent work or not in the last 7 days.
So what do they find?
• Short vs long module. No significant difference in LFP for men. But a significant difference for women – about 7 percentage points lower. In terms of hours, there is no significant difference for women but there is a small one for men (2.3 hours less on the detailed module). And for earnings there is no significant difference for men, but there is a massive significant one for women – income is 187 shillings lower in the short module (and the mean is only 198).
• Proxy vs self report. This is useful for those domestic discussions about who is doing what in the household – proxy reports of LFP are 13 percentage points lower for men and 8 percentage points for women. This is repeated for hours – the proxies significantly underestimate male hours by 6.9 hours (self reported mean is 31.3 hours) and 4.2 hours for women (mean of self report is 24.2). Proxies however don’t report significant differences in income which is interesting because if you thought the difference in reports here was being driven by strategic factors, this variable should show a difference.
• Combining the two. The short+proxy module gets you significantly lower LFP rates than the other 3 options – about 10 percentage points lower for both men and women. It also clocks in significantly lower hours – 4.6 for men and 4 for women. For income, the proxy-short reports are only significantly lower for women.
• What’s work? My prior (as for Bardasi et. al.) was that the short module would be less likely to pick up domestic work. But this is not the case – the short module is actually much more likely to pick up a folks engaged in domestic duties than the long module. This makes sense when you recall that the long module actually is specific (off-farm work, farm, household enterprise) and implicitly rules out domestic work. What is perhaps more surprising is that the proxy report is also more likely to pick up domestic duties.
• Does it matter who the proxy is? Yes. Bardasi, et. al. look at age, gender and education. For subjects less than the median age in the sample, the age difference between the subject and the proxy matters a lot – the older the proxy, the lower the LFP and hours reported. There is no clear pattern for subjects above the median age. Gender also matters – female proxies report lower LFP and hours than male proxies do for female subjects. For male subjects on the other hand, females report significantly higher LFP and hours than male proxies do. Bardasi et. al. point out that these cross gender reports are likely to come from spouses and thus (contrary to my dismal view of within household information) these cross-gender, often spousal reports are going to be closer to the self reports than getting someone from the same gender within the household. Schooling (here either some or none) differences between the proxy and the subject doesn’t show any pronounced pattern.
So this is a pretty informative start on to how we might think about collecting this kind of data in developing countries – and it should give those of us who run “light” labor modules some pause when analyzing our data (although it is important to note that the caution might be different across the variable used). There is clearly more to do – for example, it is not clear how even these patterns might differ across seasons (their sample is too small to get insights into this). In addition, there are other areas that would be worth exploring – for example how the length of the recall period matters across different crops (e.g. asking annual labor for a one season crop with a discrete harvest period versus a more continuously harvested crop). We clearly have work to do – insights from other papers on this or ongoing work would be much appreciated.