Secondary Logo

Share this article on:

The Effect of Electrical Load Shedding on Pediatric Hospital Admissions in South Africa

Gehringer, Christiana,b; Rode, Heinzb; Schomaker, Michaelc

doi: 10.1097/EDE.0000000000000905
Pediatric Epidemiology

Background: South Africa faced repeated episodes of temporary power shutdowns, or load shedding, in 2014/2015. The effect of load shedding on children’s health is unknown.

Methods: We determined periods of load shedding using Twitter, Facebook, and data from the City of Cape Town. We obtained the number of unscheduled hospital admissions between June 2014 and May 2015 from Red Cross Children’s Hospital, Cape Town, and weather data from the South African Weather Service. We used quasi-Poisson regression models to explore the relationship between number of hospital admissions and load shedding, adjusted for season, weather, and past admissions. Based on assumptions about the causal process leading to hospital admissions, we estimated the average treatment effect, that is, the difference in expected number of admissions per day had there been load shedding each day or on any of the preceding 2 days compared with if there had not been any load shedding.

Results: We found a 10% increase (95% confidence interval: 4%, 15%) in hospital admissions for days where load shedding was experienced on the same day, or no more than 2 days prior, compared with when there was no load shedding in the past 2 days. The increase was more pronounced during weekdays (12% [7%, 18%] vs. 1% [−9%, 11%]), and for specific diagnoses (e.g., respiratory system: 14% [2%, 26%]). The average treatment effect was estimated as 6.50 (5.12, 7.87) highlighting that about 6 additional admissions a day could be attributed to load shedding.

Conclusions: The association we measured is consistent with our hypothesis that failures of the power infrastructure increase risk to children’s health. See video abstract at, http://links.lww.com/EDE/B409.

From the aDivision of Internal Medicine, University Hospital Basel, University of Basel, Basel, Switzerland

bRed Cross War Memorial Children’s Hospital, Cape Town, South Africa

cCentre for Infectious Disease Epidemiology and Research, University of Cape Town, Cape Town, South Africa.

Submitted December 6, 2017; accepted July 26, 2018.

Supported by the US National Institute of Allergy and Infectious Diseases through the International epidemiological Databases to Evaluate AIDS, Southern Africa (IeDEA-SA), grant 5U01AI069924-05.

The authors report no conflicts of interest.

Availability of code and data: The R-code for this analysis is part of the supplementary material; http://links.lww.com/EDE/B389. The data are available from the authors upon request to verify the results.

Supplemental digital content is available through direct URL citations in the HTML and PDF versions of this article (www.epidem.com).

Correspondence: Michael Schomaker, Centre for Infectious Disease Epidemiology and Research, University of Cape Town, Anzio Road, Observatory 7925, Cape Town, South Africa. E-mail: michael.schomaker@uct.ac.za.

The Republic of South Africa faced repeated episodes of temporary power shutdowns in 2014–2015. Owing to its inability to satisfy the power demand (because of loss of power generation) and to prevent uncontrolled blackouts, the monopoly power supplier ESKOM implemented this practice, which is also known as rotational load shedding, for several hours a day in most of the country. Load shedding is an intervention of last resort when power demand exceeds supply: times and areas affected by load shedding have been communicated by ESKOM to the public on short notice, for example, via schedules published on Twitter and different dedicated homepages.

Even though electricity is the main power source for heating, cooking, and lighting throughout South Africa,1 the consequences of load shedding are predominately discussed with regard to its economic implications, probably because of South Africa’s challenging economic situation.2 Unfortunately, surprisingly little information can be found about health-related implications and costs. This is worrying, as case reports from hospitals suggest a direct link between blackouts and health outcomes, such as an increased burden on already overworked staff, for example, during surgeries.3

Failures of the electrical infrastructures are known to have increased hospital admissions, health-related complications, and mortality during both the “Northeast blackout” of 2003 in the United States and Canada and a power blackout in Italy, Europe, the same year.4–9 Reasons for increased admissions included carbon-monoxide intoxications because of the use of portable generators,7 more emergencies owing to failure of electrical medical devices,6 more domestic accidents,5 and a higher rate of food poisoning.7 Other studies investigated natural disasters and extreme weather conditions, which were accompanied by power failures and affected the health of the respective population by an increased number of emergency presentations, carbon monoxide poisoning, among other reasons.10–13 Although spontaneous power shutdowns as described above, and repeated power shutdowns during load shedding in South Africa have different causes, the implications, that is, lack of electricity are very much the same. One may therefore assume similar underlying etiological mechanisms. Nevertheless, to the best of our knowledge, no study has yet investigated the health effects of load shedding, which differs from unexpected power failures in the sense that people can partly adapt their lives around load shedding schedules and face shorter durations of no electricity supply. The effect of load shedding on health outcomes is of particular interest in a developing country like South Africa, where resources are scarce, electricity is the main source of heating and cooking, and health care facilities are often overburdened.14

This study analyzes the effect of load shedding on admissions to the Red Cross War Memorial Children’s Hospital (RCCH), a 300-bed tertiary pediatric hospital in Cape Town.

Back to Top | Article Outline

METHODS

Study Design

This is a retrospective, single-centre observational study analyzing the relationship between unscheduled hospital admissions to the RCCH and load shedding in the period between 1 June 2014 and 31 May 2015.

Back to Top | Article Outline

Study Population and Number of Admissions

The RCCH catchment area is the city of Cape Town except for specialty consultations, for example, burns, from the whole Western Cape and beyond.

The number of direct admissions of children up to 13 years of age, excluding internal transfers, was used as primary outcome in this analysis. Planned admissions, that is, patients with appointments, were excluded from the analysis. We thus considered unplanned admissions—for example, owing to emergencies, external transfers for specialized care (e.g., burns, surgeries, and intensive care), and presentations because of proximity or personal experience—as the main quantity of interest. We used International Statistical Classification of Diseases and Related Health Problems (ICD) codes to group patients according to their leading diagnosis, and the responsible speciality, that is, medicine or surgery.

Back to Top | Article Outline

Load Shedding as Documented on Twitter and Facebook

Our primary exposure variable is binary and indicates whether load shedding was implemented on the respective day (or preceding days, see below). In secondary analyses, which are descriptive and spatial in nature, the exposure relates to the event that load shedding was implemented in a specific area.

The authors identified days of load shedding from the available Twitter (San Francisco, CA) tweets from the ESKOM account @Eskom_SA (accessed on 12/05/2015 and 14/06/2015) and cross-checked them with Facebook (Menlo Park, CA) entries (https://www.facebook.com/EskomSouthAfrica/), documenting the respective day of an event but neglecting the exact time span (usually between 6 AM and 10 PM) and the outage severity because those details were only inconsistently reported and information quality differed by area and time. This information was then validated, and also updated for 9 days, by using data from the electricity generation and distribution department of the City of Cape Town, which also provided information on the length and area of load shedding (for those areas which were under direct control of the city).

Back to Top | Article Outline

Weather and Other Potential Confounders

Weather data, identified as a possible confounder (see below), was obtained from the South African Weather Service. Relative humidity (in %), pressure (in hectopascal), precipitation (rainfall in mm), temperature (in degrees Celsius), and wind speeds (in meter/second) were obtained for the five weather stations in Cape Town: Cape Town airport, South African Astronomical Observatory, Royal Yacht Club, Molteno Reservoir, and Kirstenbosch. Sunshine (hours per day) was only measured at Cape Town airport. We defined the arithmetic mean of the measurements of Cape Town Airport, the Observatory, and Molteno Reservoir as our weather indicators, based on the proximity to the catchment area of the RCCH. Kirstenbosch was excluded because it lies on the slopes of a mountain and has weather conditions that are not representative for the rest of Cape Town (see eFigure 1; http://links.lww.com/EDE/B390 for smoothed weather data for different stations). The Yacht Club was excluded as well because humidity and temperature measurements were missing for 119 consecutive days and the sea climate may not perfectly resemble the weather conditions of the study population.

We further identified seasonal trends as another potential measured confounder.

Back to Top | Article Outline

Statistical Analysis

We used kernel density plots to look at the distribution of hospital admissions depending on whether load shedding occurred and if it was a weekday. We used quasi-Poisson regression models to explore the relationship between the number of hospital admissions and load shedding. We considered month, weather, last week of the month (pay week), a weekly trend modeled with sine and cosine terms, past weather indicators up to a lag of 2 days, and past admissions up to a lag of 28 days to be potential confounders or relevant to model the admission process. The quasi-Poisson model can be interpreted as any other Poisson model, but allows the variance of the model to be different from the mean, and can therefore deal with overdispersion, that is, greater variability in the data than expected under the specified model. All continuous variables were included nonlinearly in the model using p-splines.15 In the main model, load shedding is a binary variable, which indicates if there was a load shedding event on the same day or up to 2 days prior of the day of interest, as hospital referral or admission may not occur immediately after load shedding.

Models used for sensitivity analyses used different definitions and different model classes (negative binomial model, linear model, INGARCH model, see eText1). In secondary analyses, we looked at alternative exposure variables, that is, (1) the event of load shedding, on the same day or any of the two preceding days, in one of Cape Town’s 16 load shedding areas; and (2) the exposure in interaction with length of load shedding, modeled nonlinearly with p-splines, and defined as the number of minutes of load shedding per day averaged over the respective areas. For both secondary exposures data on areas that were not under direct control of the city, which include both populated and unpopulated (mountainous) areas (in total 59% of Cape Town’s official size), were not available and were thus not part of the calculations.

We used frequentist model averaging16 , 17 to estimate the importance of the inclusion of different lags, that is, to what degree load shedding on the same day versus previous days is important to describe hospital admissions in the above Quasi-Poisson models. Briefly, frequentist model averaging means calculating Akaike’s Information Criterion for all possible models. Then, a higher weight is given to models that are more plausible according to Akaike’s Information Criterion. The sum of the weights of those models that include the variable of interest are used as a variable importance (VI) measure (0 ≤ VI ≤ 1). We used VI > 0.5 as a rule to include a lag variable.

We also used frequentist model averaging to determine the inclusion of past admissions, weather indicators, and complexity of the weekly trend. More details on the final model, and more methodological background, are given in eText1.

The directed acyclic graph (DAG) in Figure 1 represents our assumptions about the causal process leading to LS and hospital admissions.

FIGURE 1

FIGURE 1

Because local weather conditions in Cape Town may affect hospital admissions, such as viral infections and weather-related accidents, and electricity demand and therefore the probability of experiencing load shedding, local weather may be a confounder.18 National weather conditions may affect the implementation of load shedding but is likely unrelated to admissions at RCCH. Under the assumptions represented in the DAG, the causal effect of load shedding on hospital admissions can be estimated by adjusting for local weather and seasonal patterns using appropriate methodology, for instance targeted maximum likelihood estimation (TMLE).19 , 20 We estimated the average treatment effect, that is, the difference in expected number of admissions per day had there been a load shedding event each day or on any of the preceding 2 days, during the whole year, compared with if there had not been any load shedding, using TMLE with super learning.21 We refer the reader to eText1, and the references therein, for a more technical background. Briefly, TMLE first standardizes the data with respect to the confounders presented in Figure 1. In a second targeted step, estimation of the average treatment effect as defined above is potentially improved by utilizing information from the treatment assignment mechanism, which is the probability of load shedding conditional on the potential confounders.

All analyses were conducted in R,22 using packages “SuperLearner” and “tmle23 for the causal inference analysis, package “MuMIn”for model averaging, and packages “MASS” and “tscount” to fit the negative binomial and INGARCH model, respectively. We obtained ethical approval from University of Cape Town’s Human Research Ethics Committee for this study (Ref#: 901/2016).

Back to Top | Article Outline

RESULTS

During the study period between June 2014 and May 2015, Cape Town experienced 72 days of load shedding, 48 during the week and 24 on the weekend. Load shedding started as soon as 11 June 2014, but many events (38) occurred in April/May 2015. Figure 2 shows that there are more hospital admissions during days of load shedding, but typically during weekdays. The mean number of unscheduled admissions during the study period was about 57. On days of load shedding there were on average 61.3 admissions a day, and on days without there were about 56.7 admissions.

FIGURE 2

FIGURE 2

As speculated in the DAG, weather conditions, such as precipitation, were associated with both the probability of load shedding, and the rate of hospital admissions, supporting our initial assumption that weather may be a confounder (eFigure 2; http://links.lww.com/EDE/B390).

The Table shows that load shedding leads to a 10% increase (95% confidence interval [CI]: 4%, 15%) in hospital admissions, after adjustment for weather indicators, month, week of payment, seasonal trends, and past admissions. Similar results are obtained under a standard Poisson model or a linear regression model, but these models violated certain assumptions, including overdispersion and normality of residuals.

TABLE 1

TABLE 1

Using a negative binomial regression model led to an estimate of 10% (5%, 15%), an INGARCH model to 6% (2%, 10%), though not all assumptions were met for the latter model (eFigure 6; http://links.lww.com/EDE/B390). Using another definition of a LS event (same day, only 1 day prior, only 2 days prior, only 3 days prior) led to incidence rate ratios (IRRs) which suggest an increase in hospital admissions between 5% and 9% (Table). The VI measure obtained from frequentist model averaging suggests that inclusion of a 3- or 4-day lag period does not add much (VI ≤ 0.4) information. This demonstrates the usefulness of evaluating the effect of LS for events that occurred up to 2 days prior of admission.

The IRRs, estimated for occurrence of load shedding in each of the city’s official 16 load shedding areas, are visualized in Figure 3.

FIGURE 3

FIGURE 3

LS events in the southern peninsula of Cape Town and in the residential areas of Durbanville produced the lowest rate ratios. High IRRs were found for the township of Philippi, the satellite town of Atlantis, and areas close to Red Cross hospital (Hanover Park, Lansdown, Observatory, Rondebosch, Newlands) and areas of mixed population and income groups (Hout Bay, southern suburbs, Parow, Goodwood). These areas were sometimes, but not always, located close to areas of lower median household income (eFigure 4; http://links.lww.com/EDE/B390). There are associations of varying strength between the implementation of load shedding in different areas, as this followed a schedule, highlighting the complex spatial dependence structure (eFigure 5; http://links.lww.com/EDE/B390).

Hospital admission rates did not differ substantially when comparing overall surgical with medical specialties, but results differed with respect to the different diagnoses based on ICD-10 code chapters (Table). The highest IRR was observed for diseases of the eye and ear (12% [−16%, 48%]), the digestive system (11% [−8%, 33%]) and the respiratory system (14% [2%, 26%]). No relevant changes in admission were observed for intoxications or infections not defined in other ICD-10 chapters.

In exploratory analyses, we found that the increase in admissions occurred primarily during the week as shown by inclusion of an interaction with weekend/weekday in the model (12% [7%,18%] vs. 1% [−9%, 11%]). Moreover, we could not find evidence that length of load shedding affected the rate of admissions (eFigure 3; http://links.lww.com/EDE/B390).

Using TMLE, we estimated the average treatment effect as 6.50 (95% CI: 5.12, 7.87), or in other words that about six additional admissions a day could be attributed to load shedding.

Average treatment effects and IRRs for individual diagnoses are listed in eTable 1; http://links.lww.com/EDE/B390. Most of these estimates are not precise enough to conclude that specific diagnoses would occur more often in days following load shedding events.

Back to Top | Article Outline

DISCUSSION

Our analyses demonstrate that load shedding as implemented in South Africa is associated with a substantial increase in hospital admissions of children, on the same day and up to 2 days following the power interruption. Under the assumptions that we have identified and included all confounding variables in our analysis, that is, that the DAG from Figure 1 is correct, and that the modeling approach discussed in eText1 is appropriate, this effect is causally interpretable. Note that this applies to the average treatment effect estimated by TMLE, as the IRRs estimated with quasi-Poisson regression require stronger assumptions to be causally interpretable, for example, a constant treatment effect across all covariate strata.

A strength of our study is our rigorous approach of data collection using Twitter, Facebook, and the City of Cape Town after local authorities, including ESKOM, did not support our request for data sharing. Moreover, contacted radio stations only kept short-term records of less than 2 weeks. This approach of data collection may serve as a future model for surveys where data that normatively should be publicly available are withheld. Moreover, we have clearly communicated our assumptions under which our effect estimates are causally interpretable and used state-of-the-art methodology to facilitate this analysis.

Our study has some limitations. First, we did not have access to individual patient folders and higher numbers of specific diagnoses to further expand our hypotheses on what are the biggest risks of load shedding. Our results on individual diagnoses are imprecise. Moreover, while we had been able to identify the areas of load shedding events, these areas are large and often cover populations of different household income groups and ethnicities. Our results may indicate that poorer areas could be affected more heavily by load shedding than wealthier areas. However, wherever load shedding was not directly controlled by the city but ESKOM, data were unavailable. This includes townships such as Khayelitsha and Nyanga, and small residential areas in Cape Town’s north. There is also a complex spatial-temporal relationship with respect to the implementation of load shedding in different areas. We may not have been able to model this in all detail and it may therefore be advisable to interpret our spatial analysis with care. In addition, there is the possibility of migration between different areas, though this may be negligible as the study period is very short.

In general, the methods we use (quasi-Poisson regression, TMLE with super learning) require that observations be independent conditional on the covariates; this assumption is needed for the validity of the likelihood functions used and for the application of super learning. Although we have tried to include seasonal trends, past admissions and past weather indicators to meet this assumption, we cannot exclude the possibility that there remains dependence which we have not been able to model. In this case, inference may be affected, and CIs may not be correct. Apart from missed seasonal trends, there could be other confounders we have not measured: however, these could only be variables that cause an unusually high energy demand in the country, such as international sporting events, although we are unaware of any such events in the relevant time period. Last, it is important to note that our results may not be generalizable to high-income countries or settings where living conditions differ greatly from those in South Africa.

The effect of load shedding on health may be best explained with case reports from RCCH. For example, there was a patient with a skin burn because of handling candles during load shedding and the father admitted the patient after the skin got infected 2 days later. Another admission was related to injuries caused by a pan containing hot fat, which had been placed near an outdoor fire since the electric kitchen stove could not be used. These two cases happened at night and at home, where accidents may happen most often. Since Capetonians have long daily commutes,24 often more than 2 hours one-way, it may well be that accidents (related to limited lighting and heating options) happen after they return home from work. This could explain why, by our estimates, load shedding affected admissions primarily during weekdays, where people arrive home late. Besides such obvious relationships, there might be less obvious ones that may not be immediately clear to treating doctors: for example, inhalation of fumes from an improvised stove, or the ingestion of food from an interrupted cold chain. Increased admissions owing to eye and ear, lung, and digestive system suggest external noxious influences, for example, owing to combustion, as a possible trigger.

We did not find ICD-10 codes of the mixed chapter “injuries, poisoning, and other external causes” to be contributing to the increased number of admissions as described for other blackouts. Since we deal with children only, it may well be that access to toxins such as gasoline is limited. Furthermore, the diverse diagnoses covered in this chapter might make an observable effect less likely.

Moreover, infections are partly covered by the respective organ specific ICD-10 chapter, for example, for the respiratory system. This explains the missing relationship of the ICD-10 chapter “Certain infectious and parasitic diseases” with load shedding and more generally highlights the challenges of the interpretation of grouped disease categories. In-depth analyses implied trends of higher incidences of burns, traumatic fractures, meningitis, and other individual diagnoses, but the number of cases per diagnosis were too small for reliable interpretation. Bigger studies are needed to enhance our understanding of (indirect) causal relationships and, most importantly, to prevent casualties in situations when power failures occur.

An increased number of hospital admissions during load shedding leads to an increased burden of already overwhelmed health care facilities. Additional resources are not necessarily available, and it remains unclear what the consequences of the additional costs are. This consideration is relevant in the current and very lively debate on South Africa’s future energy mix. While costs of generating energy, political considerations, and CO2 emissions are certainly relevant aspects of this discussion, security of an uninterrupted power supply should remain a priority not only from an economic perspective, but also from a public health point of view. As we have shown, the above measured association is consistent with our hypothesis that failures of the power infrastructure increase risk to children’s health.

Back to Top | Article Outline

ACKNOWLEDGMENTS

We greatly acknowledge the help from the staff of the electricity generation and distribution department of the City of Cape Town. We would like to particularly thank Peter Jaeger in his help of administering and interpreting load shedding data. We are also immensely grateful to Mary-Ann Davies and Kenneth Sinclair-Smith for their committed support of this project and their active engagement in the logistics of it. We further thank Elsa DeJager from the South African Weather Service, and Mark Seiderman from the National Centers for Environmental Information, for their help in acquiring our weather data. We also thank Annibale Cois for his insights on this topic; and Daniela Bandic, Thomas Derungs and Sabine Bélard for constructive support throughout this project.

Back to Top | Article Outline

REFERENCES

1. General household survey 2015.Accessed August 2018.Statistics South Africa (STATS SA). Available at: http://www.statssa.gov.za.
2. Inglesi R, Pouris A. Forecasting electricity demand in South Africa: a critique of Eskom’s projections. South African J Sci. 2010;106:50–53.
3. Bateman C. Eskom debacle: health care risks, frustrations climb. S Afr Med J. 2008;98:171–173.
4. Anderson GB, Bell ML. Lights out: impact of the August 2003 power outage on mortality in New York, NY. Epidemiology. 2012;23:189–193.
5. Eachempati SR, Mick S, Barie PS. The impact of the 2003 blackout on a level 1 trauma center: lessons learned and implications for injury prevention. J Trauma. 2004;57:1127–1131.
6. Greenwald PW, Rutherford AF, Green RA, et al. Emergency department visits for home medical device failure during the 2003 North America blackout. Acad Emerg Med. 2004;11:786–789.
7. Klein KR, Herzog P, Smolinske S, et al. Demand for poison control center services “surged” during the 2003 blackout. Clin Toxicol (Phila). 2007;45:248–254.
8. Lin S, Fletcher BA, Luo M, Chinery R, et al. Health impact in New York City during the Northeastern blackout of 2003. Public Health Rep. 2011;126:384–393.
9. Pironi L, Spinucci G, Paganelli F. Effects of the September 28 2003 blackout in Italy in patients on home parenteral nutrition (HPN). Clin Nutr. 2004;23:133.
10. Abir M, Jan S, Jubelt L, et al. The impact of a large-scale power outage on hemodialysis center operations. Prehosp Disaster Med. 2013;28:543–546.
11. Johnson-Arbor KK, Quental AS, Li D. A comparison of carbon monoxide exposures after snowstorms and power outages. Am J Prev Med. 2014;46:481–486.
12. Kelman J, Finne K, Bogdanov A, et al. Dialysis care and death following Hurricane Sandy. Am J Kidney Dis. 2015;65:109–115.
13. Nakayama T, Tanaka S, Uematsu M, et al. Effect of a blackout in pediatric patients with home medical devices during the 2011 eastern Japan earthquake. Brain Dev. 2014;36:143–147.
14. Luque-Fernandez MA, Van Cutsem G, Goemaere E, et al. Effectiveness of patient adherence groups as a model of care for stable patients on antiretroviral therapy in Khayelitsha, Cape Town, South Africa. PLoS One. 2013;8:e56088.
15. Wood SN. Generalized Additive Models: An Introduction with R. 2006.Chapman and Hall/CRC;
16. Burnham K, Anderson D. Model Selection and Multimodel Inference. A Practical Information-theoretic Approach. 2002.New York: Springer;
17. Schomaker M, Heumann C. Model selection and model averaging after multiple imputation. Computational Stat Data Anal. 2014;71:758–770.
18. Sigauke C, Chikobvu D. Daily peak electricity load forecasting in South Africa using a multivariate non-parametric regression approach. ORiON (Operations Research Society of South Africa (ORSSA)). 2010;26:16.
19. Pearl J. Causality: Models, Reasoning, and Inference. 2009.Cambridge University Press;
20. Van der Laan M, Rose S. Targeted Learning. 2011.Springer;
21. Van der Laan M, Polley E, Hubbard A. Super learner. Stat Appl Genet Mol Biol. 2008;6:Article 25.
22. R-Core-Team. R: A Language and Environment for Statistical Computing. 2017.Vienna, Austria: R Foundation for Statistical Computing. Available at: http://www.R-project.org/.
23. Gruber S, van der Laan MJ. tmle: An R Package for targeted maximum likelihood estimation. J Stat Software. 2012;51:1–35.
24. Sinclair-Smith K, Turok I. The changing spatial economy of cities: an exploratory analysis of Cape Town. Dev Southern Africa. 2012;29:391–417.
Keywords:

Load shedding; Power failure; Pediatrics; Causal inference; TMLE

Supplemental Digital Content

Back to Top | Article Outline
Copyright © 2018 Wolters Kluwer Health, Inc. All rights reserved.