key: cord-0911718-iiafud87
authors: Barnwal, P.; Yao, Y.; Wang, Y.; Juy, N. A.; Raihan, S.; Haque, M. A.; van Geen, A.
title: No excess mortality detected in rural Bangladesh in 2020 from repeated surveys of a population of 81,000
date: 2021-05-12
journal: nan
DOI: 10.1101/2021.05.07.21256865
sha: da8fac385113d521c039fec309041157c2bfb76f
doc_id: 911718
cord_uid: iiafud87

Background Excess mortality has demonstrated under-counting of COVID-19 deaths in many countries but cannot be measured in low-income countries where civil registration is incomplete. Methods Enumerators conducted an in-person census of all 16,054 households in a sample of 135 villages within a 350 km2 region of Bangladesh followed by a census conducted again in May and November 2020 over the phone. The date and cause of any changes in household composition, as well as changes in income and food availability, were recorded. For analysis, we stratify the mortality data by month, age, gender, and household education. Mortality rates were modeled by Bayesian multilevel regression and the strata aggregated to the population by poststratification. Results A total of 276 deaths were reported between February and the end of October 2020 for the subset of the population that could be contacted twice over the phone, slightly below the 289 deaths reported for the same population over the same period in 2019. After adjustment for survey non-response and poststratification, 2020 mortality changed by -8% (95% CI, -21% to 7%) relative to an annualized mortality of 6.1 per thousand in 2019. However, salaried breadwinners reported a 40% decline in income and businesses a 60% decline in profits in May 2020. Discussion All-cause mortality in the surveyed portion of rural Bangladesh was if anything lower in 2020 compared to 2019. Our findings suggest various restrictions imposed by the government limited the scale of the pandemic, although they need to be accompanied by expanded welfare programs.

population. As cases surge again, as they did in March-April 2021, policy makers may want to consider 56 limiting strict restrictions to urban areas while expanding a financial support throughout the country. 57 58 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 

Excess mortality has received increasing attention over the course of the pandemic as a robust measure of

Using an approach developed for conflict zones (15), we inferred mortality in our study area in 2020 by working in the field for two weeks; each of the phone surveys employed full-time another 60 enumerators 113 for an entire month. At the beginning of each phone call, each consenting respondent was asked to list 114 current household members without prompting. The electronic questionnaire was set up for the 115 enumerator to check off each household member recorded in January 2020 and identify discrepancies to 116 investigate subsequently. As a reference for identifying excess mortality in 2020, each household was 117 asked at the end of the first phone census about all deaths that occurred in 2019. Only for 2019 was a 118 respondent asked directly about a death that occurred in the household.

Repeated census surveys were conducted in 135 villages, or pre-defined portions ("paras") of larger 121 villages, located 30-100 km to the northeast of Dhaka (figure 1). The villages had been selected for a 122 randomized controlled trial, paused in February 2020, to reduce arsenic exposure from drinking well-123 water. The 135 study villages are distributed across 16 unions, an administrative unit of which there are a 124 total of 4,563 in Bangladesh. Government data show that the average age in these 16 unions is 3 years 125 higher than in the 2,188 other rural unions of the country and that the proportion of households engaged 126 in agricultural activity is somewhat lower ( Table 1) . Proxies for socioeconomic status such as education 127 and the number of rooms in the house are no different in the study unions compared to other rural unions.

From January 15 to February 3, 2020, enumerators sought to contact door-to-door all 17,538 households 130 identified in the study villages or paras. Among these households, 1,478 were absent and could not be 131 reached. Of the remaining 16,060 households, only 6 declined to participate in what was presented at the 132 time as a study of arsenic mitigation. Sharing a kitchen was the criterion used to define a household.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 Following consent, the name, age, gender, and relationship of each individual member of the household, as well as GPS coordinates of each house, were recorded electronically. Up to two mobile phone numbers 135 from all consenting households were also recorded. Most households were subsequently recontacted 136 using one of these numbers, or in some cases alternative numbers provided by their neighbors.

During the first phone survey conducted from May 8 to June 7, 2020, a total of 14,551 (91%) of the 138 households surveyed in person in January 2020 could be contacted and consented to respond. During the 139 second phone survey conducted between October 27 and December 14, 2020, 11,933 (74%) households 140 consented. The total population of consenting households surveyed in person amounts to 81,164. This 141 number is used as the reference population and includes 7,921 household members recorded during the 142 phone surveys who had reportedly been overlooked during the January 2020 in person census, as well as 143 484 deaths in the households that occurred reportedly in 2019. The reference population does not include 144 1,068 household members who joined the household from elsewhere in 2020.

Additional information related to 2019 and 2020 deaths was collected during the phone surveys.

Respondents were asked if a reported death was the result of injury, e.g. a road accident, and if treatment 148 from a doctor or at a hospital was sought. Respondents were asked if death was preceded by symptoms 149 related to COVID-19 such as fever, headache, cough, sore throat, breathing difficulty, loss of sense of 150 smell, muscle aches, and chills. Respondents were also asked to attribute reported deaths to a few broad 151 categories including stroke or heart disease (combined here because they are often confused in rural 152 Bangladesh), cancer, liver, or lung disease. Some deaths, often in the case of an elderly parent, were 153 reported more than once by different households. A total of 76 such cases were identified over 2019-2020 154 based on name and proximity and then confirmed with an additional phone call to avoid duplication.

In order to determine the economic impact of COVID-19 in Bangladesh, we requested additional 156 information from a randomly selected 20% of 16,054 households in the first phone survey and the same participated in the first survey and 3,151 households in the second.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 gender and education. To estimate aggregate mortality, we therefore first estimate mortality for various 165 age, gender, and education group and then calculate a weighted sum.

We stratify observed death counts into cells based on a four-way structure: 

In order to compare mortality before and during the pandemic, and to estimate the resulting excess consider February 2020 as the start month but also consider later months. We specify a comparisonstarting month , which splits time into the two periods (a) the baseline: Jan 2019, Feb 2019, …, 176 1, and (b) the potential excess period: , 1, …, Oct 2020. This approach allows us to 177 be agnostic about when COVID-19 started to impact mortality in the study villages.

Modeling baseline and excess mortality in each strata.

Both baseline and excess mortality may differ across strata. We set up a multilevel logistic regression 180 (17,18) to model baseline mortality in each stratum by a four-way interaction of month-of-year, age, sex 181 and education (Appendix A). To quantify the stratum-specific excess mortality (the additional risk in 182 2020 after month ), we set another layer of multilevel model, where the monthly excess mortality is 183 decomposed by a three-way interaction of age, sex and education.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) 

Raw data.

Considering first only the 11,256 households that could each be reached during both phone surveys, a 204 total of 639 deaths were reported between January 2019 and the end of October 2020 for a total 205 population of 58,806, excluding individuals who joined the household after the January 2020 census. This 206 corresponds to an average annualized mortality rate of 5.9 per 1000 over 22 months. For the same 207 households that could be reached twice over the phone, a total of 276 deaths were reported between 208 February and the end of October 2020, slightly below the 289 deaths reported for over the same months in 209 2019. Using all the available data, the monthly-sample-weighted annualized mortality rate between 210 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101/2021.05.07.21256865 doi: medRxiv preprint which mortality differed before and after the pandemic by taking into account all the available data, and 11 deaths per 1000 since January 2019, but without a clear seasonal pattern (figure 2b). A longer time 215 series from official statistics suggests that mortality in rural Bangladesh is on average 20% higher during 216 winter compared to summer (20). The highest mortality rate during our study period was recorded in May 

The model confirms the expected increase in mortality as a function of age, expressed as a log odds ratio 228 (figure 3a). There is no clear seasonal pattern in mortality in our data according to the model (figure 3b).

Excess mortality is the key output from the model. At first, we assume that the onset of the pandemic in CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021.

After poststratification across gender and education, the model shows no indication of an increase in 236 mortality rate during Feb-Oct for 2020 compared to 2019 in all age group (figure 4a). Comparing Feb-(95% C.I., -0.52 to 0.29) at ages 50 and older. The baseline monthly baseline mortality is estimated to be 240 0.51 (.41-.63) for all age average, or 2.4 (1.9-3.0)) deaths per thousand per month. Hence, using estimate 241 [eq_excess_m3] in Appendix A, our inferred mortality declines amount to -8% (-21% to +7%) or -5% (-242 21% to +12%) in percentage changes. This overall decline was largely due to the mortality decline in the 243 80+ group. Excess mortality does not vary much with household education level (a proxy for 244 socioeconomic level) and gender.

The excess mortality changes as the boundary between the two comparison window is shifted to later in 247 2020. Comparing Aug-Oct 2020 to the baseline, the model indicates that monthly mortality rate increased 248 by 0.10 (95% CI, -0.06 to 0.30), death per thousand people per month). For age 50+, the monthly 249 mortality increased by 0.58 (95% CI, -0.23 to 1.60). The inferred posterior mean corresponds to a 20% 250 (25%) increase in percentage in all (50+) population, with considerable uncertainty.

The circumstances and causes of a total of 795 deaths reported for 2019-20 during the two rounds of 253 phone calls did not vary much over time. In 2019 and 2020, respectively, 73 and 67% of deaths were 254 preceded by consultation with a doctor or nurse, 17 and 18% of deaths occurred at a hospital, and 6.9 and 255 7.6% were the result of injury. Heart disease and stroke combined reportedly caused 41 and 47% of 256 deaths in 2019 and 2020, respectively, and cancer reportedly caused 9.0 and 9.1% of deaths. The 257 proportion of deaths attributed to lung disease actually went down from 6.4 to 3.7%. Among the 30 258 deaths attributed to lung disease in 2019, COVID-19 related symptoms such as fever, headache, cough, 259 sore throat, breathing difficulty, loss of sense of smell, muscle aches, and chills were reported 64 times.

Among the 12 deaths attributed to lung disease in 2020, the same symptoms were reported 30 times.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101/2021.05.07.21256865 doi: medRxiv preprint Economic impacts.

The lockdown had a large economic impact on our study population. Survey responses show that 264 household income where the main bread earner had a salaried job declined by 40% on average in May 265 2020, although the decline was reduced to 30% by November 2020 (figure 6b). For self-employed bread 266 earners in business, the decline in profit was 60% and barely recovered later in the year. Over the same 267 period, 25% of households reported that they couldn't obtain an essential food item in May 2020 because 

On the basis of the census data collected on three occasions, once in person and twice over the phone, we 281 conclude that mortality did not increase in 2020 across our 135 study villages or paras. In fact, our best 282 estimate is that mortality declined by 8%. This decrease in mortality can be potentially explained by a 283 decline in mortality from other causes such as road accidents or the seasonal flu caused by reduced travel 284 and social interactions, respectively. On average, the net impact of COVID-19 therefore does not appear 285 to have come close to the levels of excess mortality of 20% and over in 2020 reported for over two dozen 286 countries including the US (2). The 95% confidence interval obtained from the model indicates a one in 287 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101/2021.05.07.21256865 doi: medRxiv preprint forty chance that mortality actually increased by more than 7% in our study villages. Applied to the months of February through October 2020 along with an annual mortality of 820,000 for the country 289 overall in 2019 (9), our upper bound corresponds to a one in forty chance of 43,000 additional deaths or 290 more. The official death count attributed to COVID-19 had reached 5,900 by October 31, 2020.

Combining these two figures indicates that there is only a one in forty chance that COVID-19 deaths were 292 under-reported by more than a factor of 7. This comparison assume that both COVID-19 mortality and 293 the reporting of COVID-19 mortality were similar in urban and rural areas, which probably was not the 294 case. Our data could suggest that all the reported COVID-19 mortality was limited to urban areas, which 295 account for about 1/3 of the Bangladesh population. Even under this scenario, reported COVID-19 296 mortality would correspond to only about of 3% or urban mortality in previous years.

We do not claim that our sample of 135 villages is representative of all of rural Bangladesh, although 299 Table 1 provide some reassurance. Another limitation of the study is that a quarter of the households 300 surveyed in January 2020 could not be reached over the phone by November 2020. We cannot exclude 301 that households with a member who recently died is less likely to pick up the phone or less willing to 302 participate in the survey. Our approach to calculate aggregate mortality using demographic group-level 303 mortality corrects for it, to the extent that mortality estimated for a particular age, sex and education 304 group is not biased by attrition. The repeated census approach may also not have entirely eliminated a 305 tendency not to report a recent death, especially if it was associated with COVID-19 symptoms because 306 of widely reported stigma, especially at the beginning of the pandemic (21). We have separate evidence 307 from our survey that COVID-19-like symptoms were under-reported by affected households on the basis 308 of responses from neighboring households integrated at the village and para level.

Various hypotheses have been proposed to explain the apparently lower impact of COVID-19 in some 311 low-income countries (22). The effect of a relatively young population cannot be a factor in our study 312 given that we are comparing the same population over two years. Spending more time outside or in well-

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101/2021.05.07.21256865 doi: medRxiv preprint ventilated houses has been invoked as an explanation, but another possibility is that previous infections in regions like rural Bangladesh could have dampened the symptoms of COVID-19 (23).

Mobility data and our economic impact data both indicate that the reach of the government's 317 interventions extended to rural areas of Bangladesh. The absence of excess mortality in our study 318 population suggests that limiting gatherings, encouraging masks, and maintaining social distance were 319 broadly successful in these areas in 2020 (24). At the same time, our data show that measures restricting 320 work, trade, and travel imposed an economic burden on rural households that extended over at least six 

A riskier interpretation of our results is that the de facto national lockdown imposed in March 2020 was 328 excessive given the high economic cost and that no excess mortality was observed in our study villages.

We do not endorse this view but the government of Bangladesh and other low-income countries might 330 want to consider in the future more regionally targeted lockdowns that distinguish urban and rural areas 331 among other factors. Such a targeted approach crucially depends, however, on monitoring across the 332 country and, therefore, on the widespread availability of COVID-19 testing. The rapid growth in the pace 333 testing at the beginning of the pandemic was arrested after the imposition of a charge for testing of 334 BDT200 (US$2.40) at government facilities, a charge halved since, and BDT500 (US$6.00) for samples 335 collected from home (27). This suggest free COVID-19 testing should be made available again 336 throughout the country and that increased testing capacity will be needed to handle the likely surge in 337 demand. By analogy to the impact free well-testing for arsenic had in rural Bangladesh over a decade ago 338 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. 

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101/2021.05.07.21256865 doi: medRxiv preprint during the COVID-19 pandemic: A geospatial and statistical analysis in Aden governorate, Yemen. . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 mortality rate is calculated relative to the size of the survey sample for that month, which is indicated by 473 the dot size. 474 475 476 . CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101 https://doi.org/10. /2021 CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101/2021.05.07.21256865 doi: medRxiv preprint CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101/2021.05.07.21256865 doi: medRxiv preprint Community Mobility Reports. The baseline mobility is the median of mobility index from January 3-. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) 

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (1) died from all causes, (2) died from "stroke/heart discese", (3) died from from "natural death", (4) died from "cancer", (5) died from "liver disease", (6) died from "lung disease". The comparison is between Feb-Oct 2019 and Feb-Oct 2020. The dot sizes are the accessible population of that age group in the survey. (headache, muscle aches, chill, cough, sore throat, lose-ofsense-of-smell, breathing difficulty, and fever) . The dot size indicates the total death cases during each month. The category (c) and (d) largely overlap, while none of them exhibited an increasing trend in 2020.

. CC-BY-NC-ND 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)

The copyright holder for this preprint this version posted May 12, 2021. ; https://doi.org/10.1101/2021.05.07.21256865 doi: medRxiv preprint

widely available COVID-19 testing could have an additional impact by encouraging health-REFERENCES 381

Excess mortality: The gold

We thank the large teams of enumerators involved in the field and phone surveys.

Baseline mortality rate. We use the following notation for the four discrete variables: month 1 = 1, . . . , 22; age during the evaluation month, 2 = 1, . . . , 9; gender 3 = 1, 2; and household education 4 = 1, 2. Any individual stratum can be written as ( 1 , 2 , 3 , 4 ). In each stratum , we count the number of surveyed individuals that were alive in the beginning of the month and accessible throughout the month, and the number of deceased individuals. Assuming independent sampling, the data model iswhere the parameter is what we want to estimate: the monthly mortality rate in month 1 for age group 2 , gender 3 , and household education level 4 .The baseline ( 1 < start ) mortality rate for stratum is modeled as a function of month and individual's age-sex-education attributes, (1) where the free parameters are• The month-of-year factor, or the seasonal trend, denoted by 1 , . . . , 12 for January toDecember each year. Due to periodicity, cell processes this seasonal factor [ 1 mod 12] ∈ R. For identification, we set January as baseline such that 1 = 0.• The age factor, denoted by 1 , . . . , 9 for 9 age categories. The age factor for cell is[ 2 ] ∈ R. • The male factor male ∈ R. We set female as reference, then the sex factor for cell is• The education factor edu ∈ R. We set non-education as reference, and the education factor for cell is 1 4 = 2 edu .Mortality in 2020. We take a flexible approach by specifying a comparison-starting month start , which splits time into the two periods (a) the baseline: Jan 2019, Feb 2019, . . . , start − 1, and (b) the potential excess period: start , start + 1, . . . , Oct 2020. During the second period, on top of the baseline model (1), we model the stratum-specific excess risk:= logit −1 [ 1 mod 12] + [ 2 ] + 1 3 = 2 male + 1 4 = 2 edu + excess [ 2 ] + 1 3 = 2 excess male + 1 4 = 2 excess edu , ∀ 1 ≥ start , (2) where additional parameters excess , excess male and excess edu represent the excess risk in 2020 associated in the age, sex, and education, on top of the baseline risk.