Maternal and Child Health Care Service Disruptions and Recovery in Mozambique After Cyclone Idai: An Uncontrolled Interrupted Time Series Analysis

Timely and relevant information is vital to help identify and track areas of improvement after extreme weather events and during emergencies to prioritize limited resources. Routine data can provide useful evidence of health system performance during and after natural disasters, contributing to an effective and efficient response.


INTRODUCTION
I ncreasingly, extreme weather events such as floods, storms, and cyclones present a permanent threat to health systems and individual health outcomes, particularly in low-and middle-income countries. 1 Worldwide, roughly 7,300 natural disasters occurred between 2000 and 2019, resulting in 1.2 million deaths (yearly average of 35,000) and an economic loss of US$2.97 trillion. 1,2 Floods and storms are the most frequent events, accounting for 44% and 28% of natural disasters, respectively. 1 Exposure to extreme weather events leads to immediate service disruptions, increased burden of infectious and non-infectious diseases, and poor long-term health outcomes. [3][4][5][6][7][8] A systematic review from 2012 found a 47% to 50% increase in the population mortality at all ages and a 40% increase in mental health disorders among individuals older than age 14 years in the first year after a severe flood episode. 2,[9][10][11] After being exposed to an extreme weather event, older adults are 2.1 times more likely than younger people to experience post-traumatic stress and 1.7 times more likely to develop a subsequent adjustment disorder 12 ; pregnant women are at increased risk of pre-term delivery and low birth weight 13,14 ; and children may experience an 18% increase in diarrhea and a 15% increase in acute respiratory infections. 15 Regardless of the magnitude of the event, women and children are most vulnerable to disruptions in health care (e.g., immunization and maternal and child health services). Antenatal care (ANC) visits, institutional deliveries, and postpartum care visits for both mothers and newborns are significantly reduced in areas recurrently affected by floods, as suggested by a study conducted in Bangladesh. 16 The impact of extreme weather events on public infrastructure is expected, leading to reduced accessibility, availability, and quality of health care services, particularly with higher-magnitude events. 4,6 Power outages frequently occur and are consistently associated with poor patient outcomes, as they affect quality of care for both chronic and acute conditions. 17 Regardless of severity, resilient health systems are expected to maintain essential services while responding to initial shocks and recover quickly. 18 The speed of recovery is related to the shock's magnitude, its characteristics, the population's baseline vulnerability, and the health system's level of resilience. 1 Health system resilience is defined as the capacity to absorb external shocks and adequately and promptly adjust to respond effectively, while maintaining all essential functions, including recovering any observed losses. [18][19][20][21] Because of its geographic location, Mozambique is highly vulnerable to floods and cyclones. On March 14, 2019, category 4 Cyclone Idai hit Mozambique's central region (Zambezia, Tete, Sofala, and Manica Provinces), directly affecting 2.1 million people and causing 603 deaths, 1,641 injuries, and the displacement of 400,000 people. 22,23 Even though neighboring provinces (Tete and Zambézia) experienced high-speed winds and rains, the cyclone magnitude there was substantially lower and resulted in fewer fatalities, less destruction, and fewer health service disruptions were reported. No meaningful impacts from Idai were observed in any other provinces.
With the disruption of essential services (e.g., water, electricity, and communications) and significant damage to public infrastructure (including 90 health facilities and 3,145 health workers' homes, particularly in Manica and Sofala), the effects of Idai exacerbated ongoing challenges in sanitation, water supply, and food security in these 2 provinces. 24,25 One month after the cyclone, a cholera outbreak affected 4 districts in Sofala Province (Beira City, Dondo, Nhamatanda, and Buzi), with 6,768 reported cases and 8 deaths (0.12% case fatality). 22 Given the extent of the destruction, the budget needed to rebuild the health infrastructure was estimated at US$202 million over 5 years, with the first half needed in the first year (unpublished data). Domestic and international solidarity in the aftermath of Idai was impressive. The national government, bilateral organizations, and multilateral international institutions played a critical role in saving lives, bridging gaps in health service disruptions, and mobilizing resources for a comprehensive response plan to address immediate and long-term needs.
There is limited evidence on health service continuity and the speed of recovery during and after an extreme weather event, especially in low-and middle-income countries. Routine health information system (RHIS) data, which are frequently disparaged due to quality issues, might be the best source to describe health service impacts with high resolution, granularity, and availability; therefore, these systems provide a critical opportunity to understand how health systems adjust to external shocks and guide policy makers' decisions. However, they can also be affected during external shocks, which may impede efforts to distinguish whether disruptions reflect service discontinuity or simply a lack of data.
In the 3 years since Idai, Mozambique has made significant progress toward rebuilding its health infrastructure and ensuring the provision of primary health care services. Mozambique's experience with Idai has offered a unique opportunity to understand the nature of health service disruptions after an extreme weather event and the speed of the health system's recovery to predisaster levels. In this study, we aimed to assess Idai's impact on district-level maternal and child health care services in the 2 most affected provinces (Sofala and Manica), as well as evaluate the health system's recovery. We also intend to demonstrate the relevance of frequently overlooked RHIS data and propose a method to assess service disruptions and inform emergency response and preparedness plans. No other study has comprehensively investigated the effects of Idai on immediate health service utilization or Mozambique's recovery process. Furthermore, to the best of our knowledge, no other study has investigated health system recovery after an extreme weather event by applying the methods used in this study.

Study Design
Using a quasi-experimental design, we performed an uncontrolled interrupted time series analysis to assess monthly changes from November 2016 to March 2020 in 10 selected indicators. [26][27][28][29] These indicators covered the continuum of maternal and child health service delivery in 25 districts across 2 provinces (Manica and Sofala) before and after Cyclone Idai.

Setting
We assumed 10% as the cut-off value for the percentage of people affected by Idai. Manica and Sofala provinces were selected due to the percentage of people (41.2%) affected in those provinces, a proxy measure of Idai's impact. Zambezia and Tete provinces were excluded because less than 5% of the total province population was affected.
Sofala and Manica are neighboring provinces located in the central region of Mozambique ( Figure 1). Sofala is situated along the coast of the Indian Ocean, whereas Manica borders Zimbabwe to the west. Sofala is among the poorest of Mozambique's 11 provinces. In 2017, Sofala's population was 2.3 million, of which 60% were living in rural areas; Manica's population was 1.9 million, with 66% living in rural areas. In both provinces, the under-5 mortality rate was higher than 100 deaths per 1,000 live births in 2011. 24,30 For administrative purposes, the cities of Beira and Chimoio are considered districts.

Maternal and Child Health Service Delivery Outcomes
Ten indicators available through the RHIS were selected to reflect a range of maternal and child services at the primary health care level. Together they were used to assess Idai's impact on service utilization. The following indicators were assessed using monthly aggregated counts at the district level: (1) first antenatal care visits; (2) women completing at least 4 doses of intermittent preventive treatment for malaria during pregnancy (IPTp4); (3) institutional deliveries; (4) postpartum care visits within 3 to 7 days after delivery; (5) new users of modern contraceptives; (6) bacillus Calmette-Guerin (BCG) vaccinations; (7) diphtheria, pertussis, tetanus, and Haemophilus influenzae type b (DPT-Hib3) vaccinations; (8) measles vaccinations; (9) fully immunized children under age 1; and (10) first consultations for pediatric at-risk services.

Cyclone Idai
Cyclone Idai was characterized by high-speed winds reaching more than 118 miles (190 kilometers) per hour and heavy rains (200 mm per day). The cyclone led to destruction and flooding in districts and communities surrounding Buzi, a prominent regional river. On its trajectory, Idai inflicted the most damage to Sofala and Manica provinces. Within the provinces, 7 districts (4 in Sofala and 3 in Manica) experienced the highest impact. Beira City, where Idai made landfall in Sofala province, was mostly affected by highspeed winds. However, Buzi District, also in Sofala, suffered from both high-speed winds and dramatic floods that covered almost the entire district.
Given the heterogeneity in the levels of destruction, we characterized Idai's severity based on the proportion of people affected in each district (estimated as the number of people affected among the total district population). We defined this as the "resident population whose homes were affected by shelter damage and have not left Mozambique's experience with Idai has offered a unique opportunity to understand the nature of health service disruptions after an extreme weather event and the speed of the health system's recovery to pre-disaster levels.
the assessed locality." 31 We created 3 strata of districts: I) least affected, II) moderately affected, and III) highly affected. All districts with no affected people were included in strata I. We used the median (of the proportion of affected people) to separate strata II and strata III (Table 1).

Data Collection and Processing
Data on selected indicators were sourced from the health information system (Sistema de Informação de Saúde para Monitoria e Avaliação, or SISMA), based on the District Health Information Management System 2 (DHIS-2), from November 2016 to March 2020. Population data (including the total number of women of reproductive age and children under 5) were sourced from the 2007 Population and Housing Census, using district-level projections for 2017. After 2017, Manica District was divided into 2 districts (Manica and Vanduzi), and Gondola District was divided into Gondola and Macate districts. Data from the 2017 census were used to distribute the projected population of Manica and Gondola districts proportionately into these new districts.
We collected and analyzed data for 25 districts (13 in Sofala Province and 12 in Manica Province) Of the 25 districts, we classified 10 as least affected (strata I), 7 as medium affected (strata II), and 8 as highly affected (strata III) by Cyclone Idai, based on the number of people affected. 31

Statistical Analysis
We explored the data to assess completion and the presence of potential outliers to inform the analysis. Furthermore, we assessed district-specific time series plots to identify the parametrization of the models. For district comparisons, we used descriptive statistics (mean, standard deviation,

coefficient of variation [CV]
, minimum, and maximum). We modeled the monthly counts of service provision, accounting for yearly estimated population and potential overdispersion through a negative binomial regression with district-specific random effects (intercepts and slopes) using an equation (Box).
We computed the relative losses as the ratio between the observed counts for Idai and the expected counts for a district in a particular month since March 2019. The predicted counts are estimated from the model above with the on and off Idai scenarios set by the indicators I() terms as 1 or 0, respectively. The calculations are done at the district level. We assessed the relative loss by aggregating districts into their respective strata of cyclone damage.
The above equation is estimated as generalized linear mixed model (GLMM) 32 with negative binomial family and log link for each of the 10 service provision indicators. GLMMs are an extension of GLM to include random effects to address the nested and clustered nature of the data (e.g., 1 district has monthly counts making 42 observations). The GLMMs can be estimated through Maximum-Likelihood (ML), restricted ML (REML), and Bayesian approaches. 33 Due to numeric estimation challenges (multiple random effects and issues of estimation convergence) and less bias on the variance of the random effects, we chose to use the Bayesian approach. 34 All regression models were estimated using Stan programming language through the brms package in RStudio version 3.6.3 (RStudio). 35,36 As priors, we chose for fixed effects coefficients univariate normal distribution with 0 mean and 100 variance, for each standard deviation of the random effects a half Student-t with 3 degrees of freedom and scale parameter that is minimally 10, and for the negative binomial shape parameter a gamma distribution with 0.01 for shape and rate parameter. For correlations, the default priors were left unchanged. Four parallel chains were fit with 16,000 iterations, of which the first 1,000 were discarded as burn-in. We used Gelman-Rubin diagnostics (less than 1.02), trace plots, and autocorrelations to assess convergence, good mixing, and iteration autocorrelation. We applied a thin of 5, resulting in 12,000 iterations remaining for the posterior estimation. These were used to estimate the relative loss of service delivery in March, April, and May 2019 and to assess the immediate losses due to the cyclone and the subsequent overall recovery. We focused the analysis on 3 months to capture the first month of returning to pre-Idai levels. The exponentiated parameters b jump and b post BOX. Model Used to Calculate Monthly Counts for Service Provision After Cyclone Idai b 0 is the intercept (overall log-mean count on November 2016), b pre is the overall monthly change in log counts for all 25 districts before the cyclone, b post is the immediate overall change in log counts for all districts since March 2019 (affecting the overall post-Cyclone Idai period), b March2019 and b April2019 are March-and April 2019-specific deviations from the overall b post , b postslope is the overall monthly change in log counts during the post-cyclone period, and b* coefficients are district-specific deviations from the respective b coefficient. The above model works on counts per population; therefore, the exponentiated coefficients are to be interpreted as relative changes in the count services per population. So immediately after Idai, there was e b post , e b post þ b Mach2019 and e b post þ b April2019 for overall months, specific to March 2019, and specific to April 2019, respectively, associated level change in count services per population. We focus on these 2 months as they represent the period of greatest impact.
indicate the overall relative changes across 25 districts in the months after Cyclone Idai.

Ethics Approval
We used monthly district-level aggregated data, with approval from the Ministry of Health. We extracted data from the health information system devoid of individual-level identifiers.

RESULTS
Of the 4.44 million people living in Sofala and Manica provinces, 41.2% (1.83 million) were affected by the cyclone. Overall, the greatest impact was observed in Buzi, Dondo, and Nhamatanda, where almost all district inhabitants were affected. 31,37 Table 1 shows the district-specific population and the proportion of affected people. Compared to the number of institutional deliveries, the mean number of postpartum visits was lower than expected-less than one-sixth of the mean institutional deliveries for both provinces. Table 2 shows the monthly average counts for each indicator per district.

Regression Results
In November 2016, across all 25 districts, the average number of pregnant women per 100,000 women of reproductive age (WRA) who had completed a first ANC visit was 934 (95% CI=867, 1,007); of those, an average of 310 (95% CI=246,  Table 3 shows the regression coefficients for each model.    ) is the monthly multiplicative increase in the ratio of indicator count to 100,000 people before Cyclone Idai (e.g., there was a 0.2% relative decrease in first ANC visits in women of reproductive health age per month). d Post-slope (e b postslope ) change is the multiplicative change in the pre-slope coefficient (e.g., after Cyclone Idai, the monthly trend in first ANC visits was unchanged). e March 2019 is the specific change relative to the post-level (e b March2019 ) change. This is a multiplicative deviation from the overall post-Idai level in March 2019 (e.g., during the month when Cyclone Idai occurred, the ratio of first ANC visits to the population of women of reproductive health age was 23% lower than during the entire post-Idai period). f April 2019 is the specific change relative to the post-level (e b April2019 ) change. This is a multiplicative deviation from the overall post-Idai level in April 2019 (e.g., during the month following Cyclone Idai, the ratio of first ANC visits to the population of women of reproductive health age was 11% lower relative to the entire post-Idai period).
increase. Concerning the measles vaccine, Buzi had a relative loss of 38.0% (95% CI=0.36, 0.92) in March 2019. Table 4 presents model-based estimates for the relative losses by selected indicators.
The Supplement provides province-and districtspecific estimates for all relative losses for March, April, and May 2019. Two months after Idai, strata I districts were already returning to positive trends in all but 3 indicators (DPT-Hib3, IPTp4, and family planning); in contrast, highly affected strata III districts still had greater losses for all indicators except ANC visits and institutional deliveries.
Among strata III districts, the relative loss in institutional deliveries was 20.0% (95% CI=0.69, 0.94) in March 2019 and only 3.0% in May 2019. However, Buzi showed the most significant immediate relative loss in institutional deliveries, estimated at 55.0% (95% CI=0.36, 0.56) in March 2019. Despite its impressive recovery, Buzi still  Figure 3 shows the relative loss by strata.

DISCUSSION
To determine disruptions to maternal and child health service delivery indicators after Cyclone Idai, we conducted an analysis that simultaneously accounted for annual population growth at the district level, seasonality, and historical trends using RHIS data. Overall, the results showed significant relative losses in all 10 selected indicators in the 2 months following Idai (March and April 2019), and a quick recovery within 3 months back to or higher than levels seen before the cyclone. Not surprisingly, strata III districts that were most seriously affected showed higher disruptions. These results corroborate previous findings from similar studies 16,38,39 and are consistent with what could have been expected due to the massive destruction of the public health service infrastructure. 1,2,14 Even though returns to pre-Idai service delivery levels were observed by May 2019, this may not represent a full recovery, considering the need to recuperate losses from March and April. This is particularly true for indicators such as immunization, family planning, or antenatal care (for which care seeking could have been delayed), though not for other indicators that cannot be To determine disruptions to maternal and child health service delivery indicators after Cyclone Idai, we conducted an analysis that simultaneously accounted for annual population growth at the district level, seasonality, and historical trends using RHIS data. postponed (such as institutional deliveries). Therefore, studying the recovery process deserves special attention to the type of service disrupted; prospective studies may help researchers understand the full scope of recovery when service delivery trends return to levels before the shock. A study from Liberia quantified the cumulative losses not only to capture the return to expected trends but also to track whether losses from shock periods had been recovered. 26 While this approach could help better describe the recovery, limitations persist-particularly when using aggregated data, as in this article. With aggregated data, it is difficult to disentangle whether specific groups recovered, particularly for chronically underperforming services, as higher than expected performance may not represent a full recovery but could instead be a consequence of increased demand due to targeting new groups or community mobilization activities.
The number of first consultations for at-risk child services was least affected in March and April 2019, particularly in strata I and II districts; however, this indicator showed the highest relative increase in May 2019 across all districts (Manica: 36%, Sofala: 22%), regardless of their level of destruction. This pattern might be due to increased demand for services for acute conditions, particularly malnutrition, which is predictable following an extreme weather event. Most at-risk child consultation visits are for children exposed to HIV or tuberculosis, or children being treated for malnutrition. In fact, 6 months after the cyclone, Sofala Province reported 600 cases of pellagra-a chronic B3 vitamin deficiency-after decades with no episodes, which signals a critical nutrition issue following Idai. 40 Nutritional insecurity worsens after natural disasters, and vulnerable populations, particularly displaced people, face significant difficulties in securing daily meals due to possible increases in disruptions to travel, infrastructure destruction, and increased food prices, among others. Vulnerable groups often must rely on support from local and international organizations, which lacked resources or were delayed in some study areas. 41 Several factors may have contributed to a quick health system recovery. Because the catastrophic impact of Idai captured widespread international attention, we assume that the aid received might have contributed to minimizing the initial shock and speeding up the recovery process. Domestic and international support contributed to swiftly addressing essential health system needs, including drugs, supplies, and materials. Furthermore, significant support existed to restore water supplies, electricity, and infrastructure rehabilitation-all critical service readiness determinants-though most support did not continue for an extended period and was focused on highdemand areas. Mozambique's health sector has a well-established coordination mechanism with donors and implementing partners. Robust leadership and collaboration, together with massive individual solidarity, appear to have played an essential role in efficiently responding to Idai (at least during the initial shock period). 44,45 To the best of our knowledge, no other studies have focused on understanding how a health system recovers after the shock of an extreme weather event such as Cyclone Idai. Despite health system resilience being increasingly discussed in the field of global health, its definition, characteristics, and indicators of measurement are still subjects of debate. 42 Notwithstanding, definitions of health system resilience have reached consensus in their inclusion of a health system's ability to respond effectively to shocks and to maintain primary health system functions in the presence of external shocks or crises. 18 Resilience is contextspecific, adaptive, and builds on new learnings and knowledge translation into the health delivery system. 18,19,[42][43][44][45] The rapid and steady recovery seen across nearly all study indicators after the cyclone showcases critical signs of Mozambique's health system resilience, particularly in Sofala and Manica. The substantial gains recorded within 3 months after the cyclone (by May 2019) reinforce the evidence for resilience -as the immediate international aid, in some cases, did not appear to continue beyond this timeframe. 46  This study provides important lessons, particularly for low-and middle-income countries vulnerable to extreme weather events. First, the finding that highly affected areas had the most significant impact on services uptake reinforces the need to establish systems that quickly detect these areas during shock periods so that aid can be immediately channeled to support recovery plans. Second, even though we were able to observe a recovery to pre-Idai levels, this may not be enough since disruptions can be severe-as observed in Buzi-and a full recovery of the accumulated losses may take longer; therefore, robust monitoring mechanisms should be developed to continue tracking these losses. Third, external shocks can trigger acute conditions and create space for the emergence of related diseases, as seen in the cholera outbreak and resurgence of pellagra; in anticipation of this, RHIS data should be prioritized since it can facilitate rapid detection as conditions emerge and guide an informed response. Indeed, the level of granularity achieved with RHIS data is useful to understand district-level variability, as well as identify highly affected services. RHIS provides a unique opportunity to track the recovery process, particularly for indicators such as family planning or immunization for which a full recovery (not only returning to pre-shock levels but also recovering the losses accrued during periods of disruption) is a programmatic goal. Fourth, effective coordination mechanisms and strong leadership are critical in an emergency, mainly when massive solidarity exists and new stakeholders come in. We have learned that the existing tools between government agencies and partners were essential to avoid or minimize anarchy in the Robust leadership and collaboration, together with massive individual solidarity, appear to have played an essential role in efficiently responding to Idai.
The rapid and steady recovery seen across nearly all study indicators after the cyclone showcases critical signs of Mozambique's health system resilience, particularly in Sofala and Manica.
response and to direct aid to the most vulnerable areas during Idai.

Limitations
This study has some notable limitations. First, districts could have experienced disruptions to service delivery after the cyclone even without significant levels of destruction. Because districts were classified by severity level based on the number of people affected, some districts could have been misclassified when estimating relative losses. Second, we relied on routine data, which may have quality issues, including missing data. Indeed, our analysis included a small number of outliers; however, these were not influential. Third, we were not able to track aid directed to each district; therefore, we missed understanding whether the recovery had any association with the resources allocated. Fourth, given the study design (uncontrolled ITSA) and lack of covariates at the district level, causal inference should be avoided and result interpretation should be conservative. Fifth, although we did not statistically test for lead and lagged effects, our data exploration (by plotting individual district time series) did not suggest such patterns. Despite the limitations, the results presented are robust and are consistent with what could have been expected after an external shock of Idai's magnitude, with the added advantage of quantifying the effects across a set of essential service delivery indicators, using routine data that are frequently overlooked to track health system performance during and post shocks.

CONCLUSION
This study provides evidence of the negative impacts of extreme weather events on women's and children's ability to access essential evidencebased interventions. It also showcases how routine data is useful for tracking health system performance and resilience during and after shocks; therefore, it should be used and prioritized to guide decision making. Overall, Cyclone Idai led to massive disruptions in health service delivery, with all elements of maternal and child health services showing meaningful and statistically significant decreases immediately following the cyclone. Recovery to pre-Idai trends occurred quickly for most indicators, although highly affected districts took relatively much longer. However, describing the specific characteristics that most influenced the health system's recovery and accumulated losses should be investigated to more fully picture the recovery process. Finally, despite the focus on a single dimension of system resilience (ability to recover), this study contributes to evidence on features of health system resilience and alternative methods for assessing them.