Matching hours worked

[jdebacker]

I'd like to calibrate the $\chi_n^s$ parameters to match the lifecycle labor supply profile in South Africa.

To do so, I'm trying to match hours worked by age. My data source is the Quarterly Labour Force Survey. From these data, I'm using the Q418HRSWRK variable, which is the respond to how many hours the respondent typically works in a week. The lowest reported value is 1, so I interpret a missing value as a zero.

When averaging these reported hours by age (using sampling weights), I find the following picture:

If I turn these into the fraction of total "potential" labor supply (16 hours per day), and compare to the a similar survey in the US, I find the following (with the US data in red above the South African data):

So it looks like adults in South Africa work less than 1/2 the time of those in the U.S. My question is, does this seem correct or is there something I'm missing when making these averages from the QLFS data.

Both countries seem to have somewhat similar labor force participation rates (~60%), while the unemployment rates are very different: around 33% in South Africa vs about 4% in the U.S.

[NadiaMatulich]

Hi Jason. Let me ask questions about the method- I'm not sure I fully follow it?

Are we averaging including unemployed or excluding unemployed. I'd imagine that you have similar average hours worked between US and SA (i.e. most people who are working are working a 40 hour week), but that the difference is driven by a lot of people not working. i.e., there are a lot of people of working age working zero hours pulling the average down.

If it is just amongst the employed then I'm not sure and would have to look deeper. I'd imagine the general survey weights don't quite apply the same to employed as they do to general working age population as our unemployment crisis is heavily skewed with youth unemployment being much higher than general unemployment.

If we're aiming to include average hours worked for the whole working age population this sounds about right to me, given the high unemployment.

(We are very jealous of the US and its 4% unemployment rate.)

[jdebacker]

@NadiaMatulich Yes, good question. I am averaging over all adults (so this includes those out of the labor force and those in the labor force, but who are unemployed). This corresponds to the model where we don't have an extensive margin (employed or not), so we want each skill/cohort to match average hours for all those in that group (which will include those employed and those not).

I do think that we might need a modification to OG-Core to handle the SA situation well. As it stands, the main parameter that will help us match the labor supply patterns shown above is the disutility of labor, $\chi_n^s$. This parameter represents costs to going to work and so can include opportunity costs (e.g., for the young, this could be the opportunity cost of not going to additional schooling) or frictions in the labor market (e.g., skill mismatch that causes unemployment). But I think a limitation, especially in the context of very high unemployment that is skewed towards lower skill groups, is that we don't have variation in this parameter across skill groups, $j$. So while we can adjust $\chi_n^s$ to match overall patterns, my guess is the rate of unemployment is much different for workers with high levels of education than those with no formal education and it would be hard to fit that pattern without $j$-specific parameters.

[NadiaMatulich]

@jdebacker yes I think this is a worthwhile extension. The literature does show that youth unemployment is high, but graduate unemployment is not (well, relative to youth unemployment). https://ideas.repec.org/p/sza/wpaper/wpapers175.html (an older working paper but the patterns persist, albeit slightly higher graduate unemployment). Unemployment is largely structural and doesn't exhibit a lot of churn. People keep jobs, and if they lose them they struggle to find new jobs. (I'm thinking of a specific masters thesis written in 2022, I will get my hands on a link for anyone else that wants to read it).

As is true of many things in SA there is a dichotomy- unemployment is lower amongst older people than younger people, and lower for skilled than unskilled, and many potential policy reforms working on unemployment will specifically target the skills mismatch.

[DaanSteenkamp]

Have a look at this comparison: https://codera.co.za/which-countries-work-the-hardest/ aggregate

[aidanhorn]

@jdebacker how are you defining skill groups? This could be wage categories, occupation categories (ranking by mean wage by occupation), or a join with O*NET skill data.

[jdebacker]

@DaanSteenkamp very neat plot. Those are average hours per worker -- so we are seeing high hours intensity among those working, but with large parts of the working-age population not working in SA.

@aidanhorn "skill groups" in the model are defined by "potential lifetime income". These are identified by using panel data to observe hourly earnings and then extrapolating those to full time work for everyone to determine potential lifetime income. But they would probably be highly correlated with wage categories (especially if conditional on age).

The model is making the assumption that a person's "skill type" is fixed, but their labor productivity within any time does vary by age. Here's what this looks like:

[aidanhorn]

I suggest the following QLFS variables to investigate disutility of labour:

• Q44YEARSTART
• Q44MONTHSTART
• Q314RSNSTOPWRK
• Q416ASATISFIED
• Q422MOREHRS
• Q424WRKXHRS
• Q425STARTXWRK
• InactReason