Labor Data and Quantity/Quality Tradeoffs

|

This page in:

In their new paper on fast internet and employment in Africa, Jonas Hjort and Jonas Poulsen examine the impact of submarine internet cables on employment outcomes in several African countries. They report differences-in-differences estimates, exploiting the gradual arrival of the cables along the coast of Africa. (David noted this work in his Weekly links March 29.) In addition to the compelling topic of the paper, what drew my eye to this work was the use of the Demographic and Health Surveys and Afrobarometer data to look at labor outcomes.

Neither survey is known for its employment content – the former focuses on covering in-depth issues on reproductive and child health and the latter are public attitude surveys on democracy, governance, and society. In the case of the Barometer surveys, to identify employment status, there is one question, “Do you have a job that pays cash income?” The DHS questionnaire identifies working women by asking them “Aside from your own household work, have you done any work in the last seven days?” with a follow-up if the response is no, which asks “As you know, some women take up jobs for which they are paid in cash or kind. Others sell things, have a small business or work on the family farm or in the family business. In the last seven days, have you done any of these things or any other work?” Men are asked one question: “Have you done any work in the last seven day?”

Based on my experience (and not based on a recent systematic review of multi-topic household), in multi-topic household surveys (like LSMSs) these screening questions are in a designated labor module of the questionnaire. They are typically broken out by employment type (such as wage/salary work, work on the household farm, being an own-account working of a household enterprise, working as a contributing family worker, and sometimes a question on being an apprentice). My inclination (and that of colleagues I have discussed this with over the years – yes, I have had conversations about this…) is that surveys like the DHS and Afrobarometer are not well-suited to study employment in Africa. This is because of the concern that they under-report non-wage/salary work, though perhaps less so for the DHS with its extra question for women. Notwithstanding this concern, it is easy to understand why one would use the DHS or Afrobarometer surveys to explore cross-country employment in Africa. These data are harmonized and packaged very nicely on their respective web sites, and geo-identifiers can be obtained. Compiling similar data from multi-topic national surveys and labor force surveys (LFSs) in the region entails considerably more effort.

The point of this blog is not to try replicate Hjort and Poulsen (H&P) and explore if using data on employment from multi-topic household surveys or labor force surveys would yield different results. This is partly for practical reasons. H&P have made their analysis files available (programs and data). But their study uses the GPS data for DHS and Afrobarometer, and this requires requesting access and then linking to the geo-referenced internet cable data… extra work I opted to skip for this blog. Rather, I use their paper as motivation to explore, albeit lightly, if I have been too stubborn or orthodox in my approach to what constitutes appropriate labor data and too quick to dismiss these much-easier-to-access data.

But first some minor housekeeping. Two things caught my attention in H&P’s paper. First, the mean level of employment between DHS and Afrobarometer surveys was similar: 68% and 58% respectively (Table 1). Such a comparison, however, is not straightforward because the DHS sample is heavily skewed towards women and is much younger (both by sample design) than the Afrobarometer. This would tend to drive employment rates in DHS down (also, the set of countries are not the same). Offsetting this is the specific Afrobarometer question (with its reference to cash) which struck me as likely to under-report work in household enterprises or on a household farm. Upon closer look at their replication files, the employment indicator constructed from the Afrobarometer excluded people who reported not working and not looking for work (a large share of those not working) – perhaps an error. Going to the data and constructing the same employment-to-population ratio as reported for the DHS, for men and women 18+ with sample weights I find stark difference in employment rates between the surveys: 87% and 39% for men in the DHS and Afrobarometer respectively. For women, the rates are 73% and 27%. Clearly, these surveys are not measuring the same concept of employment.

The second thing that caught my attention, also in Table 1, was that 85% of employed people in the DHS were classified as skilled, what H&P also label as high-productivity occupations. This seem extremely high given that 56% of those working in the DHS sample have less than secondary schooling and 21% with more than primary have less than 3 years of any secondary. The details in footnote 44 explain their definition of skilled as those in occupational categories of professional, sales, services, skilled manual, clerical or employee in agricultural sector (presumably agriculture wage workers). The unskilled occupations are self-employed farmers, domestic workers, unskilled manual. The categories of sales and services likely largely captures informal microenterprises in urban settings. This strikes me as a very liberal concept of skilled employment. And it left me wondering about how much one can say about skills-based technology change in Africa using these data and this definition, as H&P set out to do.

Circling back, how closely do the employment rates in the DHS surveys match the rates from multi-topic surveys or LFSs? Comparing employment rates between 12 DHS and 12 multitopic surveys from 2009-2018 for the 8 countries in the H&P paper, controlling for age and country, rates in the DHSs are 3 percentage points higher (not statistically significant) for women. Rates for men are 9 percentage points higher. The higher employment rate for men in the DHS was a surprise to me. One explanation is that some domestic activities are being included by men in the DHS as employment, given that the single question is vague. For women, these are explicitly excluded in the question wording.

There is variation across countries in these differences for both men and women. The level difference is not necessarily a big concern; but the magnitude of the differences also varies by traits. To illustrate, the table below shows the comparisons of employment rate for Ghana, Kenya, and Tanzania for all women (men) and by education levels. No single pattern emerges, if anything more questions are raised. And it leaves me wondering if the findings in H&P on education (where only those with no education do not benefit from fact internet) would hold if using more ‘traditional’ sources of employment data.
   

From this first cut of the data, I am left feeling very leery about using Afrobarometer data to study employment in Africa. As for DHS, I was pleasantly surprised at how close the employment rates are, but still concerned by the variation across countries (which I did not describe here) and across education levels.  So probably I am back to where I started, inclined to recommend using the labor data from LFSs or multitopic household surveys to study labor – despite the extra work to compile which this entails.
 
 

Authors

Kathleen Beegle

Lead Economist with the World Bank Gender Group

Join the Conversation

Jonas Hjort
June 24, 2019

We are grateful to Kathleen Beegle for flagging her concerns with some of the publicly available datasets used in our paper. We here briefly respond to what we understand to be her primary concerns.

1. When analyzing whether individuals in the Afrobarometer dataset are employed or not, we defined employed individuals as those who say that they have a job, and unemployed individuals as those who say that they do not have a job and are looking for one. We exclude those who do not have a job and say that they are not looking for one, and who are therefore out of the labor force, from the analysis. The more conventional alternative is to not exclude out-of-the-labor-force individuals and simply consider the unemployed to be everyone who does not have a job.

With the definition of (un)employment used in the paper, the mean of the outcome in our sample is 0.58, the estimated percentage point increase in the probability of employment with access to fast Internet is 0.077, the standard error of this estimate 0.037, and the percent increase in the probability of employment (evaluated at the mean) therefore 13.28% (see Table 3 of the paper). If we instead use the more conventional definition of (un)employment, the corresponding numbers are 0.41, 0.064, 0.055, and 15.61%.

We agree that we should have been clearer about this definition. Nevertheless, the results are not sensitive to this choice.

2. Two of the datasets we used to analyze how fast Internet affects employment in Africa (DHS and Afrobarometer) are household surveys rather than labor force surveys.

Our third dataset, however, is a labor force survey, from South Africa (QLFS). Our results using this dataset are generally quite similar to those from the multi-country DHS and Afrobarometer surveys (though the estimated magnitudes differ across the three samples/datasets, see e.g. Table 3).

We also find an increase in reported hours worked and a decrease in ``wants to work more'' in the QLFS sample (see Panel B of Table 3), suggesting that our estimated impact on employment itself captures a real employment response.

We agree that the self-reported employment rates in DHS and Afrobarometer are high (and this is true also for QLFS). We do not have reason to believe that the way respondents answer employment questions change when submarine cables arrive on the coast, and differentially so for connected vs unconnected locations, as would need to be the case for employment-measurement issues to explain the differential employment response we estimate.

It is possible that we underestimated the extent to which labor force surveys comparable to the QLFS from other African countries exist and can be accessed by researchers back when we started the project that led to this paper. We agree that future work analyzing how employment outcomes in Africa respond to availability of fast Internet using labor force surveys and other forms of data will be very valuable.

3. We use our three primary datasets to inform debates over the skill-bias of fast Internet in Africa.

Our coding of job types simply follows ILO's categories. We show results not only for Skilled/Unskilled jobs, but also for Highly/Somewhat/Moderately/Unskilled jobs.

In our view, how we classify a given technology should depend not only on its impact on whether different types of individuals (e.g. the more versus less educated) hold skilled jobs, but also on whether they hold any job.

Hopefully it will be possible in the future to systematically record data on each job’s task-content in Africa.

Our paper clearly opens many new questions. We are currently working on several of these.

Jonas Hjort and Jonas Poulsen