Mortality in rural coastal Kenya measured using the Kilifi Health and Demographic Surveillance System: a 16-year descriptive analysis

Background: The Kilifi Health and Demographic Surveillance System (KHDSS) was established in 2000 to define the incidence and prevalence of local diseases and evaluate the impact of community-based interventions. KHDSS morbidity data have been reported comprehensively but mortality has not been described. This analysis describes mortality in the KHDSS over 16 years. Methods: We calculated mortality rates from 2003–2018 in four intervals of equal duration and assessed differences in mortality across these intervals by age and sex. We calculated the period survival function and median survival using the Kaplan–Meier method and mean life expectancies using abridged life tables. We estimated trend and seasonality by decomposing a time series of monthly mortality rates. We used choropleth maps and random-effects Poisson regression to investigate geographical heterogeneity. Results: Mortality declined by 36% overall between 2003–2018 and by 59% in children aged <5 years. Most of the decline occurred between 2003 and 2006. Among adults, the greatest decline (49%) was observed in those aged 15–54 years. Life expectancy at birth increased by 12 years. Females outlived males by 6 years. Seasonality was only evident in the 1–4 year age group in the first four years. Geographical variation in mortality was ±10% of the median value and did not change over time. Conclusions: Between 2003 and 2018, mortality among children and young adults has improved substantially. The steep decline in 2003–2006 followed by a much slower reduction thereafter suggests improvements in health and wellbeing have plateaued in the last 12 years. However, there is substantial inequality in mortality experience by geographical location.


Amendments from Version 1 Introduction
A majority of low-and middle-income countries (LMICs), especially in sub-Saharan Africa (sSA), lack comprehensive civil registration and vital statistics' systems (CRVS) necessary for monitoring mortality 1,2 . Tracking the progress in child and adult survival, therefore, relies on alternative data sources such as demographic and health surveys (DHS), national population censuses, and health and demographic surveillance systems (HDSSs). HDSSs are designed to monitor a small subpopulation of a nation in a defined geographical area and are a commonly used resource for health and demographic research in LMICs.
The Kilifi Health and Demographic Surveillance System (KHDSS) was established in 2000 by the KEMRI-Wellcome Trust Research Programme (KWTRP) to monitor mortality and morbidity caused by common diseases and to provide a sampling frame for epidemiological studies 3 . The surveillance area was selected to capture at least 80% of patients admitted to the Kilifi County Hospital (KCH) and over the subsequent two decades, the platform has been used to describe morbidity in children and adults. This includes incidence of malaria 4 , pneumonia 5 , lower respiratory tract infections 6 , rotavirus 7 , malnutrition 8,9 , sickle cell disease 10 , epilepsy 11 as well as the burden of mental health problems 12 , pregnancy-related disorders and chronic diseases that contribute substantially to overall mortality 13,14 . The KHDSS has also been used to evaluate the impact of new community-based interventions, such as vaccines and bed net use 5,15-19 .
Morbidity data have been reported comprehensively in Kilifi but mortality in this population has not been systematically described. In this paper, we describe mortality in children and adults over a 16-year period and analyse deaths by age, sex, season, geographical location, and temporal trend.

Data source and setting
The KHDSS surveillance area, which is located in Kilifi County, within the former Coast Province, is divided into 15 government administrative regions called locations, comprising a total of 891 km 2 ( Figure S1, Extended data 20 ). An initial census and mapping of the surveillance area was conducted in 2000 and was found to contain 189,148 residents in 20,978 households. Subsequently, the population has been continuously monitored for births, pregnancies, deaths, and migration events in re-enumeration rounds occurring approximately every 4 months, and the mapping was updated in 2017. At the end of 2018, there were 299,471 residents living in 41,536 households. From 2008, deaths registered in the KHDSS have also been investigated for the cause of death by verbal autopsy which has been reported separately 21 . KCH is located at the geographical centre of the KHDSS area and, during the study period, it was the only government facility offering inpatient care for the KHDSS population. A small number of private hospitals and lower-level facilities have a few inpatient beds.
The concept of the KHDSS is based on the INDEPTH (International Network for the Demographic Evaluation of Populations and Their Health) data model. Demographic and health data are collected at four points of contact: at re-enumeration when community interviewers make household visits to update the population register; at the inpatient wards of KCH where medical staff record patient history, clinical examination and outcome (death or discharge); at the maternity ward of KCH where staff record births and perinatal deaths; and in 34 vaccination clinics distributed across the surveillance area which collected data on childhood vaccination between 2008 and 2018. The eligibility for inclusion, the variables routinely measured, the structure of the KHDSS databases and the population structure have all been described previously 3 .
Initially, data collection during household visits was paper based but switched to electronic data collection using tablets in 2016. The tablets are loaded daily with the most recent copy of the residents' database and, after data collection, they are returned to the research unit where a two-way synchronization with the master database is performed. All other data collection points are linked in real time to the master database that has been specified using MySQL.
At re-enumeration rounds, information on all household members is sought from a single informant, usually a member of the household. If all household members are unavailable during the visit, information is obtained from neighbouring households. All field staff are debriefed on the quality of data collected after each enumeration cycle and re-trained where needed. The data collection applications are programmed with skip patterns and consistency checks to ensure mandatory information is collected. Additionally, within the database, there are built-in checks for missing or duplicated data.
To explore the accuracy of age data at the first census and among all new in-migrants, we calculated Whipple's Index 22 . Whipple's index measures the tendency for individuals to inaccurately report their age in rounded numbers, usually ending in 0 and 5, resulting in age heaping.

Statistical analysis
The analysis period, from 1 Jan 2003 to 31 Dec 2018, was stratified into four non-overlapping periods each lasting 4 years. We excluded 2000-2002 because of changes in the re-enumeration protocols designed to increase the ascertainment of deaths in neonates during these years . We used survival analysis and  routine demographic life table methods to calculate mortality  rates and life expectancy and examined seasonality, short-and  long-term trends over the 16 years. Age-sex mortality profile. The mortality rate was calculated as the number of deaths divided by person-years of observation (PYO). Entry to risk begins at the latest of birth, in-migration or study start date. Exit from risk is at the earliest of study end date, out-migration or death. If an out-migration is followed by an in-migration, the period between the out-migration and in-migration is excluded from the risk period to avoid survivor bias. The total PYO was computed for different age groups, sex, and locations.
For children aged less than five years, we have also calculated conventional mortality ratios where the number of deaths within a specific age group in a given time period is divided by the number of live births occurring during the same time period. Mortality ratios are commonly used in settings where risk time cannot be quantified. They can be confounded by varying birth rates as the deaths in the numerator are not always drawn from the denominator population.
Survival and life expectancy. We used two methods to estimate life expectancy: the period life table method which calculates the mean life expectancy at birth and the Kaplan-Meier (KM) survival method which calculates the median age at death. The main difference between the methods is in the age intervals used; the life table method computes survival probabilities within pre-define age intervals, e.g. 5-year intervals, whereas the KM method computes survival probabilities whenever there is a death in the cohort making the KM intervals smaller and of variable length 23 .
For purposes of comparison with other analyses, we also generated abridged life tables using data structured according to analytic methods developed by the Multi-centre Analysis of the Dynamics of Internal Migration and Health (MADIMAH) which was a working group within INDEPTH. In the MADIMAH method 24,25 , the definition of risk time considers the time between out-migration and a subsequent in-migration. If the difference is less than 180 days, this time is included in the risk period which increases the person-years of observation resulting in lower estimates of mortality rates.

Seasonality and trend.
We first assessed seasonality and long-term temporal trends for each age group by graphically reviewing a time series of monthly mortality rates. We then estimated trend and seasonality based on an STL (Seasonal and Trend using LOESS) decomposition and identified months with the highest and lowest mortality rates from the seasonal component. We investigated whether the difference in mortality rates between analysis periods could be attributed to random month-month fluctuations by fitting a negative binomial model with yearly counts as the outcome and period as a categorical explanatory variable. To test for seasonality, we fitted a model that included month, period and the interaction to monthly mortality counts.

Geographical heterogeneity in survival and mortality over time.
We produced choropleth maps for overall and age-specific mortality rates in the four 4-year periods to investigate the geographical variation of mortality in the administrative locations over time. For overall mortality, we accounted for temporal differences in the population age-sex structure by direct standardization against the 2011 KHDSS age-sex structure. All the maps were created at the administrative location level.
We used the quantile method to create five mortality rate classes for map reading. This method places equal numbers of data units (death rates) in each class resulting in classes centred on the median death rate. For each age group, the quintiles are derived from the entire mortality rate range between 2003 and 2018 and the resulting classification is applied across each of the 4-year-period maps for that age group. The quantile method, though simple, has been shown to be the optimal classification method for displaying geographically varying data in series in a map reading experiment 26 .
To assess geographical variation in mortality, for each period, we fitted a multi-level Poisson regression model adjusting for sex and age in which location was included as a random effect and used the variance of the random effect to quantify heterogeneity. We tested for between-location variation within each period using the likelihood ratio test and tested for temporal variation in mortality rates between the 2003-2006 period and each of the subsequent periods using a z-test. We also calculated the median age at death for each of the 15 administrative locations in the 4-year periods and assessed variation in life expectancy by location and time.

Age-sex profile
The cohort consisted of 699,841 individuals of whom 125,587 (18%) were followed from birth. In total, we observed 22,207 deaths in 3,897,529 person-years. More than 95% of the information on residence and vital status was collected from respondents living in the same household. Females contributed 53% of the total PYO and 48% of deaths (Table 1). There was no indication of age heaping or misspecification of sex (Table S1, Extended data 20 ).  (Table 2). Results from fitting a negative binomial regression model to yearly counts confirmed that mortality rates were significantly lower in the latter periods, for all age groups <55 years, compared to the first time period (Table S2, Extended data 20 ). Mortality ratios, per 1000 live births, for children aged <5 years, are presented in Table S3 (Extended data 20 ). Over the whole 16-year period mortality rates were lowest in children aged 5-14 years ( Figure S2b, Extended data 20 ). In this age group, mortality declined from 1.  (Table 2) and varied little thereafter. In adults, the steepest period-to-period decline (35%) was seen in the age group 15-54 years between the periods 2003-2006 and 2007-2010. Mortality changed very little over time for those aged ≥75 years. Overall, mortality was higher in males than females at all ages with differences being greater in adulthood than in childhood.  Seasonality and trend Figure 3a shows the age-specific mortality rates by calendar month for each of the analysis periods. A seasonal pattern appears only in children aged 1-4 years in the first period. Table S7 (Extended data 20 ) shows the estimated high and low mortality months. The months with the highest mortality for neonates, children aged 29-365 days and those aged 1-4 years were February, June, and July respectively. For adults aged 15-74 and ≥75 years, high mortality months were June and August, respectively. The interaction between period and seasonality was significant in children aged 1-4 years only (Table S2, Extended data 20 ).   Table S9 (Extended data 20 ).

Discussion
In the Kilifi Health and Demographic Surveillance System, overall mortality rates declined steeply in the first four years of the study period in all age-sex groups, except in older adults, and then declined much more slowly in the subsequent    12 years. Neonatal and under-5 years mortality rates declined by 48% and 59%, respectively, between the first period and the last. The mortality reduction observed between the first and subsequent 4-year time periods was greatest among those aged 1-4 years. Median survival was greater in women, by 6 years compared with men, and increased in both sexes by approximately 12 years during the study period. Seasonal effects on mortality were only evident in children aged 1-4 years and  only in the first 4-year period. Finally, location-specific mortality varied from the median value by ±10%, which represents an important inequality. There was no evidence that this variation has improved over time.
Adult mortality in LMICs has received little attention in recent decades but from available estimates, adult mortality in East Africa either stagnated or declined between 2003 and 2010 28-30 . The trends coincide with earlier patterns of HIV incidence but these data are not sufficient to provide a detailed description of mortality levels and age patterns at national and sub-national levels, which hinders comparison. Nonetheless, adult mortality has been characterized by higher rates in females between 15-34 years because of deaths related to childbirth and possibly an earlier age of infection by HIV 31 (The Gap Report 2014). We observed this phenomenon in the first 4 years of the study period, but it was later reversed with men being at a greater risk of dying ( Figure  Sub-national variation in child mortality, which is driven in part by inequitable distribution of health services and interventions, is a common observation across SSA 34-38 . In Kilifi, the magnitude of the overall variation can be understood as meaning that one location (5% of all locations) experienced a mortality rate that lies beyond ±10% of the average mortality rate. The magnitude of variation did not change significantly over time, which suggests sustained inequitable distribution in public health services and access to healthcare if we consider geographical variation in mortality as the measure of equity.
The decline in child mortality in Kilifi is consistent with independent observations over an extended period ( dependent on potentially unreliable recall provide higher estimates of the number of deaths. The differences between the two methods are more likely to be driven by different definitions of their target populations. DHS and census methods capture all residents observed at one point in time within the geographical locale; some of these may not meet the residence requirement of the HDSS cohort. These requirements include that they are, or intend to be, resident in this household for at least 3 months. Furthermore, the DHS covers the whole of Coast Province, which has four counties in addition to Kilifi and, even within Kilifi, the KHDSS is a sub-population (approximately 40%) of the county. The overall numbers of child mortality from the HDSS, DHS and census datasets may be different but the trends are consistent with each other and are also consistent with an analysis of multiple disparate datasets that shows a decline in child mortality beginning as far back as 1965 34 .
This descriptive analysis lays out the baseline trends in mortality in Kilifi over time. Several additional data sources may help to explore the underlying causes of these trends and geographic patterns. Within KHDSS there are data on the changing morbidity experience of residents from hospital records over the same period; for example, the incidence of admission to hospital with malaria declined sharply between  Underlying individual data include geo-located residence, date of birth and migration data and hence would be high risk for identifiability. This project contains the following intermediary data: • Master ASMR -aggregated data with age specific mortality rates

Introduction
The introduction is clear and well-written. 1.

Methods
The authors write "Subsequently, the population has been continuously monitored for births, pregnancies, deaths, and migration events in re-enumeration rounds occurring approximately every 4 months, and the mapping was updated in 2017." Is this first sentence saying that all data points (births, pregnancies, deaths) are collected from the reenumeration rounds, or just the migration data? 1.
In the next paragraph, the authors write that there are 4 points of contact, where reenumeration rounds refer to household visits. Births and pregnancies information is collected from the maternity wards. This sentence suggests that not all data is collected from the re-enumeration rounds (see comment #1). Please fix this to make the sentence and data collection methods clearer.

2.
What are the four-month re-enumeration rounds referring to specifically? The average time between household visits or the average time between data points (not just household data from a key informant but also from clinics and maternity wards). This is not clear.

3.
Do field workers visit only the households every for months, or do they also go to the wards, clinics and vaccination sites every four months. This is not clear.

4.
The authors note that all household members can be absent during a household visit. No data or information is given on participation rates in the surveillance program. Can more be said about this? Is the household participation rate low, moderate or good? Is there a paper from the Kilifi surveillance that documents and quantifies missed household visits and nonresponse rates in more detail? For example, at the Welcome Trust Surveillance site in northern Kwazulu-Natal, South Africa, there is the paper by Larmarange et al. 1

Ian Cook
University of Limpopo, Mankweng, Limpopo, South Africa

General comments:
Well written and constructed paper -achieves a balance between detail and overview. Also, caters for a specialised reader as well as a more general readership. Notable, the number of publications which have emanated from the KDHSS. A large database with almost 300 000 residents in 2018, 700 000 individuals and nearly 4 million p-y of observation, analysed over 16 years.

Specific comments:
It might be useful (if even in an appendix) to provide a map(s) (country and county), identifying the site and important landmarks (hospitals/towns, cities/major transport routes etc) in the site area.

Are sufficient details of methods and analysis provided to allow replication by others? Yes
If applicable, is the statistical analysis and its interpretation appropriate? I cannot comment. A qualified statistician is required.

Are all the source data underlying the results available to ensure full reproducibility? Yes
Are the conclusions drawn adequately supported by the results? Yes between indicators for different time periods.
I observe the interchangeable use of the terms "rates" and "ratios". Since mortality indicators are conventionally estimated as rates, I will suggest that the term "rates" be used consistently in the paper.

3.
Too many methods are used to characterise mortality for the first time in the Kilifi HDSS area. The purpose of the paper is mortality estimation and validation of the KHDSS data, and not comparison of methods using the same dataset. Therefore, it may be worthwhile to adopt the conventional methods of mortality estimation to effectively describe the characteristics of the region's mortality experience for the stated period; as well as facilitate easy comparison with estimates from other data sources, and for other regions in Kenya.
For instance, comparisons of life expectancy estimates from the Life Table method with similar estimates from census or DHS sources provides more analytical value in assessing the quality of the KHDSS data than comparing them with Kaplan-Meier-derived mean life expectancies from the same dataset. The Life Table method also provides the opportunity to cross-check and validate derived age-specific mortality rates.

SPECIFIC QUERIES
Page 4, left column, fifth para: Remove the word "survival" from the sub-title. It can be used interchangeably with "mortality" and the choropleth maps refer to mortality rates over time.

1.
Page 4, left column, fifth para: Provide a justification for applying the direct standardization of the populations against the 2011 KHDSS age structure. If the idea is to use the midperiod age structure for this purpose, then that of 1 st January 2013 should be used instead.

2.
Pages 5 & 6: Consider merging Tables 1 and 2 to provide comprehensive information on births, deaths, PYO and corresponding mortality rates (with 95% CI) by age-group and period. The childhood mortality estimates (i.e. neonatal, post-neonatal, infant [<1 yr]), child [1-4 yrs] and under-5) can be expressed in both per 1,000 PYO and per 1,000 live births where applicable. The authors are urged to reconsider the method used in estimating the childhood mortality rates presented in Table S2. These rates should be equivalent to the probabilities of dying within the respective age brackets, and can be estimated more accurately using the Lexis Diagram method, for instance. The under-5 mortality rate should not be the sum of the infant mortality (<1) and child mortality (1-4), but rather the complement of the product of the probability of surviving the first year of life and the probability of surviving to exact age 5 thereafter. The authors should therefore review their method of analysis to provide accurate measures of childhood mortality indicators. In fact, the mortality rate for children aged 1-4 years cannot be expressed in "per 1,000 live births", because it refers to a population defined by having already survived the first year of life.

3.
Page 5, left column, last para: The statement that a seasonal pattern appears only in children aged 1-4 years in the period 2003-2006 is speculative. Whether it is real or a data artefact can be ascertained by adopting more robust statistical analysis. An inclusion of 95% CI bars in Figure 3 would have enabled the reader to make an inference to confirm or dismiss such a seasonal mortality pattern claim.

4.
Page 10, right column, second para: The explanation linking oestrogen levels and female life expectancy is beyond the scope of the paper and should be removed.

5.
Whilst sufficient comparison is done between KHDSS childhood mortality indicators with estimates from other sources and regions, not much similar comparison is attempted for adult mortality.

6.
Is the work clearly and accurately presented and does it cite the current literature? Partly Is the study design appropriate and is the work technically sound? Partly

If applicable, is the statistical analysis and its interpretation appropriate? Partly
Are all the source data underlying the results available to ensure full reproducibility? Yes Are the conclusions drawn adequately supported by the results? Partly the area is necessary and relevant; and the attempt is therefore justified.
Whilst the paper is generally well-written, it can be enhanced further both in terms of content and presentation as indicated in the following queries.

GENERAL QUERIES
The frequent reference made to supplementary tables and figures in the paper makes reading it a rather arduous task. If tables and figures are important enough to be presented as part of the results, then they should appear in the paper accordingly. The authors should therefore decide what information to include in the paper and how they should be lucidly presented to enhance clarity and ease comprehension.
As mentioned in the review, Kilifi HDSS is the setting for a large number of studies of observational epidemiology and public health interventions and scientists reading these studies may wish to interrogate the background mortality experience of the population from a variety of different perspectives. This is what has motivated our analyses -this is expressed in the introduction. We divided the results between the main paper and a supplement to produce a main paper that was accessible to the general reader and simultaneously provide a compendium of tabulation sufficiently comprehensive to satisfy the more detailed analyst. We have decided what information we wish to emphasise and provided signposts to the supplementary data merely to inform the more curious reader.

1.
None of the mortality rates estimated and shown in the tables and graphs are presented with Confidence Intervals (95% CI). It is recommended that these be provided for all indicator estimates presented in tables, and as CI bars in levels/trends graphs. Not only will they add value to the results of the study but will also enable the reader to make inferences/deductions from the presented results beyond what the authors have reported, especially -for example -with respect to confirming statistically significant differences between indicators for different time periods.
In our analysis, the Kilifi HDSS is both the population of interest and the sample population. Since we are not inferring to a larger population, confidence intervals for our mortality rates estimates are unnecessary. However, we agree with the reviewer that it is important to rule out month-to-month fluctuations (e.g. due to conceptions, temperature, rainfall, crop yields, road access to hospital, vaccination coverage, etc) when comparing between time periods. To rule out random fluctuations (Table 2 and Figure 3a), we have fitted a negative binomial model with yearly rates as the outcome (yearly rates eliminate the issue of seasonality) and period as the explanatory variable. For all age groups under 55 years, we observed significantly lower mortality rates in the latter periods compared to the first period (2003)(2004)(2005)(2006). We have added a supplementary table (Table S2) showing these comparisons for all age groups.

2.
I observe the interchangeable use of the terms "rates" and "ratios". Since mortality indicators are conventionally estimated as rates, I will suggest that the term "rates" be used consistently in the paper.

3.
We have emphasised the calculation of rates because we have a sophisticated dataset with events linked to person years at risk. Much demographic data does not have this linkage and many of the estimates of mortality are calculated as the ratio of the number of deaths in any one age group in one place in one year to the number of live births in the same place in the same year. These ratios are used as an approximation of the cumulative mortality incidence but they have a number of shortcomings including the fact that where the birth rate fluctuates annually, they misrepresent the denominator for any group beyond infancy, and the fact that they calculate the mortality based on the starting population not the person years at risk. From this perspective they are risks (not rates) but the lack of correspondence between the numerator and denominator negates this description and so we, and others 1,2 , have referred to these as ratios. As much demographic data is captured with this method we have calculated the ratios in Kilifi for wider comparison. We have not used the terms interchangeably and in fact we defined their separate and distinct use in the statistical methods (Age-sex mortality profile); the first paragraph defines the calculation of rates, the second the calculation of (and justification for) ratios.
Too many methods are used to characterise mortality for the first time in the Kilifi HDSS area. The purpose of the paper is mortality estimation and validation of the KHDSS data, and not comparison of methods using the same dataset. Therefore, it may be worthwhile to adopt the conventional methods of mortality estimation to effectively describe the characteristics of the region's mortality experience for the stated period; as well as facilitate easy comparison with estimates from other data sources, and for other regions in Kenya. For instance, comparisons of life expectancy estimates from the Life

SPECIFIC QUERIES
Page 4, left column, fifth para: Remove the word "survival" from the sub-title. It can be used interchangeably with "mortality" and the choropleth maps refer to mortality rates over time.
Following directly from the response above -we have tried to signpost which analytic approach is being used at each point. The sub-header 'Survival' precedes a paragraph explaining the survival analysis of the mortality data and a reference to six Kaplan Meier survival curves. Later, under the heading Geographic heterogeneity in survival and mortality we signpost the variation in survival (from the survival analyses which we detail in the supplement) and then present here the location-specific mortality rates in the choropleth maps. Our judgment is that, whilst survival curves describe mortality by age efficiently, mortality rates provide a more efficient metric to understand the variation of mortality cartographically.

1.
Page 4, left column, fifth para: Provide a justification for applying the direct standardization of the populations against the 2011 KHDSS age structure. If the idea is to use the mid-period age structure for this purpose, then that of 1st January 2013 should be used instead.
We stand by our calculation -for a period beginning 1st January 2003 and ending 31st December 2018, a mid-point estimate of 1st January 2011 is perfectly reasonable (it could be argued that 31st December 2010 may be more accurate -in days…) The purpose is to find a reasonable reference population for a standardisation exercise -it is not necessary to fix it to the mid-study point, though this is what we did.

2.
Pages 5 & 6: Consider merging Tables 1 and 2 to provide comprehensive information on births, deaths, PYO and corresponding mortality rates (with 95% CI) by age-group and period. The childhood mortality estimates (i.e. neonatal, post-neonatal, infant [<1 yr]), child [1-4 yrs] and under-5) can be expressed in both per 1,000 PYO and per 1,000 live births where applicable. The authors are urged to reconsider the method used in estimating the childhood mortality rates presented in Table S2. These rates should be equivalent to the probabilities of dying within the respective age brackets, and can be estimated more accurately using the Lexis Diagram method, for instance. The under-5 mortality rate should not be the sum of the infant mortality (<1) and child mortality (1-4), but rather the complement of the product of the probability of surviving the first year of life and the probability of surviving to exact age 5 thereafter. The authors should therefore review their method of analysis to provide accurate measures of childhood mortality indicators. In fact, the mortality rate for children aged 1-4 years cannot be expressed in "per 1,000 live births", because it refers to a population defined by having already survived the first year of life.
We have extended Table 2 to include age and period specific numbers of deaths and pyo in addition to rates. In our selection of results we have emphasised mortality rates (over ratios) for the reasons outlined in general query #3 above and have only presented the infant mortality ratio, under 5 mortality ratio etc for the sake of comparison with other demographic sources, and only in the supplement. We agree with the reviewer that these 3.
ratios provide a metric which has no basis in reality ('the mortality rate in children aged 1-4 years cannot be expressed…') but this analytic approach is, nonetheless, in widespread use and is frequently taught (see for example the definition of 'Post-neonatal mortality rate' at https://www.cdc.gov/csels/dsepd/ss1978/lesson3/section3.html). For exact probabilities of dying from one age point to the next, we refer our readers to the Life  Tables in the supplement. Page 5, left column, last para: The statement that a seasonal pattern appears only in children aged 1-4 years in the period 2003-2006 is speculative. Whether it is real or a data artefact can be ascertained by adopting more robust statistical analysis. An inclusion of 95% CI bars in Figure 3 would have enabled the reader to make an inference to confirm or dismiss such a seasonal mortality pattern claim.
In addition to the visual interpretation of the data (a seasonal pattern appearing only in in children aged 1-4 years), results from a negative binomial regression model confirm a significant interaction between season and period in children aged 1-4 years only. We have reported the same in the in the methods, results and supplement.

4.
Page 10, right column, second para: The explanation linking oestrogen levels and female life expectancy is beyond the scope of the paper and should be removed.
We have removed it.

5.
Whilst sufficient comparison is done between KHDSS childhood mortality indicators with estimates from other sources and regions, not much similar comparison is attempted for adult mortality. Countries.
I confirm that I have read this submission and believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard.