Less load shedding? Fewer deaths

Image credit: Matthew Henry on Unsplash.

On 25 May 2023, three-year-old Neyamiah James died after extended electricity outages in  Johannesburg caused her home oxygen machine to run out of backup power. In June 2015, a woman died after a load shedding-induced power failure at Letaba Hospital in Limpopo caused her ventilator to stop functioning. On 4 March 2023, Yengiwe Mthimunye died after being turned away from Waterval Clinic in Mpumalanga, because of load shedding. Bhekisisa relayed an anonymised account of a load shedding-induced death at Charlotte Maxeke Hospital in Johannesburg in July 2022. On 25 November 2023, a man died after being refused entry to a clinic in Limpopo, reportedly because of load shedding. In January 2023, load shedding was alleged to have contributed towards five farmworkers’ deaths from heatstroke in Kakamas in the Northern Cape.

These deaths are likely the tip of iceberg when it comes to the health effects of load shedding. Refrigeration failure increases the risk of foodborne illnesses in the home, while bulk water treatment and distribution failures impede effective sanitation (Laher 2019). Load shedding degrades healthcare services, potentially severely (Apenteng et al, 2018). Communications and emergency response is hampered, while non-functioning street and traffic lights likely increase road accidents, and lighting and alarm system failures may increase crime. Load shedding has been found to increase residential fires and paediatric hospital admissions (Gehringer 2018, Lawson 2022). An international literature has shown that electricity rationing has large effects on elderly mortality, due to inability to regulate ambient temperatures (Chirakijja et al. 2024; He and Tanaka 2023, Neidell et al. 2021).

Supported by SA-TIED, I investigated load shedding’s mortality effects in South Africa, focusing on Cape Town. Below, I outline the key features of that research, which was published as a UNU-WIDER working paper.

 

Determining the causal effects of load shedding is not straightforward

The chief difficulty when it comes to identifying the causal effects of load shedding is that it is not implemented randomly. Many existing analyses implicitly assume that it is random, by drawing causal conclusions from regressions of national load shedding incidence on various national outcomes. The problem is that load shedding has seasonal patterns (it tends to be worse in summer) and it has a long-run trend (it has mostly become worse over time). This means that load shedding will, to some extent, be spuriously correlated with anything that has a similar seasonal component or long-run trend, even if that thing is unrelated to load shedding. Does load shedding increase ice cream consumption? Or the success of the Springbok rugby team? Probably not, but you are likely to find statistically significant regression coefficients.

In this study, I try to estimate the causal effect of load shedding by comparing outcomes in Cape Town to the rest of the country. Cape Town differs from the rest of South Africa because, since June 2015, it has used the Steenbras pumped-storage hydroelectric plant to mitigate load shedding. I combine three different episodes—one in each of 2015, 2018 and 2019—where Cape Town introduced mitigation after at least 18 weeks of having the same load shedding as the rest of the country.

I use weekly mortality data from the Medical Research Council, and load shedding data from the City of Cape Town and EskomSePush. Consistent with the existing literature, I focus on deaths among those 65-and-older.

 

Constructing a counterfactual “synthetic Cape Town”

The difficulty with this exercise is that Cape Town differs from the rest of the country in many ways, not just in its load shedding policy. Thus, we cannot simply attribute differences in outcomes to the difference in load shedding intensity.

The way I get around this is by using a method called “synthetic control”, to construct a “synthetic Cape Town” as a weighted average of other parts of the country. The weights are constructed so that synthetic Cape Town is as similar as possible to actual Cape Town (along dimensions specified by the researcher), in periods when there is no load shedding mitigation.

If the method works, the differences in outcomes between actual Cape Town and synthetic Cape Town when load shedding mitigation starts can then be attributed to the causal effects of the policy. But it does mean changing the question slightly: I investigate the mortality effects of Cape Town’s load shedding mitigation scheme, not load shedding itself.

 

Results

In my main specification, the estimated weights result in a “synthetic Cape Town” which is comprised of about 50% Western Cape district municipalities, 40% metropolitan municipalities, and 10% other district municipalities. Is this a plausible “synthetic Cape Town”? It doesn’t seem obviously wrong, but also not obviously correct.

One way to judge the plausibility of the method is to look at the results. Figure 1 shows the difference, or “gap”, between 65+ mortality rates of actual and synthetic Cape Town, for the 18 weeks before (periods -17 to 0) and 9 weeks after (periods 1 to 9) load shedding mitigation starts.

Figure 1: Difference in weekly deaths per 10,000 population (age ≥ 65), Cape Town versus synthetic Cape Town

It is encouraging that there is no systematic gap between actual and synthetic Cape Town before mitigation starts. This could be due to spurious “over-fitting” (I use this same period to estimate the synthetic control weights), but I show in the paper that this is unlikely. This suggests that synthetic Cape Town is a good counterfactual for actual Cape Town.

Then, immediately after mitigation starts, there is a (statistically significant) decrease in 65+ mortality in Cape Town relative to synthetic Cape Town. This effect is loaded in the first few periods after mitigation starts, which I show in the paper is consistent with the time-pattern of mitigation intensity.

 

How large is the effect?

What do the results imply in terms of actual numbers of deaths averted? Some adjustments need to be made to the raw numbers for two reasons. First, only part of Cape Town is supplied electricity by the City and gets the mitigation policy (about 73% of the City’s total 65+ population), and second, the intensity of mitigation varies over time.

I end up with a large, but very imprecisely estimated effect. Specifically, the results suggest that Cape Town’s mitigation policy averted 547 premature deaths among the city-supplied 65+ population between 2015 and 2019. But because the 95% confidence interval is so wide, I can’t rule out that the true effect over this period was between 94 and 1002 premature deaths averted.

 

Limitations

One can extend these results to get an idea of how many premature deaths were caused by load shedding, rather than averted by the City’s mitigation policy. The same numbers suggest that load shedding caused 2623 (95% confidence interval 452 to 4805) premature deaths in the city-supplied 65+ population between 2014 and 2019 (while Cape Town started mitigating load shedding in 2015, there was load shedding in 2014 too). These numbers are large – perhaps implausibly so at the extremes – and I think that it is more likely that the true number is on the lower end of the confidence interval than the higher end.

What about extending the results to after 2019, when load shedding was at its most intense? Unfortunately, the synthetic control method I use above is not straightforwardly applicable to the post-2019 period. I also argue in the paper that it would be ill-advised to try to extrapolate the 2015-2019 results to post-2019, because the dramatic increase in load shedding made it a qualitatively different phenomenon with different societal responses.

Or what about estimating the number of 65+ deaths caused by load shedding across Cape Town as a whole, rather than just the City-supplied region? Or across the country? In theory, one could just extend the results of this analysis to these broader populations. But the problem is that the City-supplied population is quite unique, and so we probably shouldn’t assume that other populations have the same mortality response to load shedding. Figure 2 shows City-supplied versus Eskom-supplied areas in Cape Town. Looking at this map, people who are familiar with Cape Town’s geography would be unsurprised to learn that City-supplied regions are much older and richer than the rest of the City and the rest of the country.

Figure 2: City of Cape Town (CoCT) versus Eskom electricity supply regions in Cape Town (Source: “Project 90 by 2030”, 2021)

 

Inequality

The scale of the inequality might be surprising though. While about 81% of whites in Cape Town live in a City-supplied region, only about 27% of black Africans in Cape Town live in these load shedding mitigation areas. This extreme disparity seems to have both historical and present-day causes, which I discuss in the paper.

It is to the City’s credit that it took advantage of its pre-existing pumped storage hydroelectric plant and electricity distribution infrastructure to reduce load shedding for some of its residents. This research suggests that that has actually averted premature deaths among these residents. But it is also unsurprising that the use of this historical infrastructure has, to some extent, entrenched historical inequalities, given who this infrastructure was likely originally designed to serve.

 

References

Apenteng, B.A., S.T. Opoku, D. Ansong, E.A. Akowuah, and E. Afriyie-Gyawu (2018). “The Effect of Power Outages on In-Facility Mortality in Healthcare Facilities: Evidence from Ghana”. Global Public Health, 13(5):545–55. https://doi.org/10.1080/17441692.2016.1217031

Chirakijja, J., S. Jayachandran, and P. Ong (2024). “The Mortality Effects of Winter Heating Prices”. The Economic Journal, 134(657): 402–17. https://doi.org/10.1093/ej/uead072

Gehringer, C., H. Rode, and M. Schomaker (2018). “The Effect of Electrical Load Shedding on Pediatric Hospital Admissions in South Africa”. Epidemiology, 29(6): 841. https://doi.org/10.1097/EDE.0000000000000905

He, G., and T. Tanaka (2023). “Energy Saving May Kill: Evidence from the Fukushima Nuclear Accident”. American Economic Journal: Applied Economics, 15(2): 377–414. https://doi.org/10.1257/app.20200505

Laher, A., B. Van Aardt, A. Craythorne, M. Van Welie, D. Malinga, and S. Madi (2019). “‘Getting Out of the Dark’: Implications of Load Shedding on Healthcare in South Africa and Strategies to Enhance Preparedness”. South African Medical Journal, 109(12): 899–901. https://doi.org/10.7196/SAMJ.2019.v109i12.14322

Lawson, K. (2022). “Electricity Outages and Residential Fires: Evidence from Cape Town, South Africa”. South African Journal of Economics, 90(4): 469–85. https://doi.org/10.1111/saje.12329

Neidell, M., S. Uchida, and M. Veronesi (2021). “The Unintended Effects from Halting Nuclear Power Production: Evidence from Fukushima Daiichi Accident”. Journal of Health Economics, 79: 102507. https://doi.org/10.1016/j.jhealeco.2021.102507

Project 90 by 2030 (2021). “Understanding residential electricity tariffs in the cape town metropolitan area”. Information Sheet, Friedrich-Ebert-Stiftung. Project_90_Electricity_Tariff_Updated-June-2023-24.pdf (90by2030.org.za)