COVID-19 preparedness measures.

Two months before the first case of COVID-19 was reported in Kenya on 13^th March 2020, the government had increased preparedness to the pandemic. These included installation of surveillance systems to detect suspected COVID-19 cases at border points, additional medical staff at international airports and ports, in-country capacity to test and isolate COVID-19 cases, sensitization of healthcare workers on dealing with COVID-19 cases and establishment of a National Emergency Response Committee.

COVID-19 intervention measures.

The government started introduction of interventions to combat COVID-19 spread on 13^th March 2020. These included suspension of public gatherings and events, closing of schools, international travel restrictions, fumigation and disinfection of markets, closure of bars and restaurants, suspension of attendance to places of worship, limit of people attending weddings and funerals and national dust-to-dawn curfew. The month of April 2020 saw cessation of movement in and out of four counties with highest number of COVID-19 cases, restaurants were opened under strict guidelines of social distancing, handwashing and temperature checks. During the month of May 2020, cessation of movement into and out of Kenya through Tanzania and Somalia borders was affected while on 10^th June 2020 the government launched home-based care for patients with COVID-19 infection. In July 2020, certain measures were relaxed; cessation of movement into the four counties was lifted, phased re-opening of places of worship and resumption of local air travel. On 1^st August 2020, international air travel resumed and in September 2020, operation of bars resumed. This was followed by phased reopening of schools and lifting of suspension on political gatherings in October 2020 and November 2020 respectively. In December 2020, there was a nationwide health care workers strike precipitated by demands for better working conditions such as provision of adequate Personal Protective Equipment (PPE), enhanced risk allowances and a health insurance cover. The strike only affected public facilities. Public and private health facilities in Kenya have separate management processes and funding structures affecting service provision hence challenges experienced in the public sector do not mirror the situation in Private facilities. These Health worker strikes were as a result of pay disputes, contracts and COVID safety concerns, including availability of personal protective equipment [11]. By early 2021, the test positivity proportion was below 5% and all schools were re-opened on January 4^th. On 5^th March 2021, the government launched its COVID-19 vaccination campaign and on March 12^th, due to cases rising again during the third wave of the pandemic, funerals, weddings and other events were capped to including 100 persons and capacity of places worship was restricted. The timeline of events is presented in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. COVID-19 control measures timeline during the period of study.

https://doi.org/10.1371/journal.pgph.0000029.g001

Selection of hospitals.

Hospitals were selected based on service availability. Data from varied databases on service availability were consulted. These include; Service Availability and Readiness Assessment (SARAM) [12], Emergency Medicine Foundation of Kenya (EMFK) [13], Harmonised Health Facility Assessment (HHFA) [14], District Health Information System (DHIS2) [15] and the Master Facility List (MFL) [16]. A single database of facilities was then created using a combination of the data sources. Five common indicators used across different countries to delineate hospitals were used. These were availability of operating theatres, oxygen, blood transfusion, radiology and critical new-born care. Hospitals identified to have the five indicators were then matched with the number of inpatient beds in an iterative process to confirm the provision of inpatient services and to validate the use of inpatient bed numbers in the definition of hospitals. Different iterations were then conducted based on the number of admissions and service availability to identify facilities that were either likely to have been missed by all of the service availability assessment considered or may be classified as lower-level facilities but attract on average 5,408 admissions annually and therefore warrant upgrading to hospital status. A total of 201 hospitals fitted this criterion and were selected. The indicators used to select the hospitals is presented in Table 1 in S1 File. Three more hospitals, which are approved internship centres by the Kenya Medical Practitioners and Dentists Council [17] were added to the list totalling to 204 hospitals. The distribution of the 204 hospitals is represented in Fig 1 & Table 2 in S1 File and the definition of the indicators used in this study presented in Table 3 in S1 File.

Extraction of data.

Monthly data on inpatient admissions and deliveries in five patient groups (maternity, new-born units, paediatric, adult medical and surgical wards) were obtained from the District Health Information System (DHIS2), an open-source software platform for data reporting by most health facilities in the country. Data were extracted for the 204 hospitals, which are known to offer Comprehensive Emergency Obstetric and Newborn Care (CEmONC) services for the period January 2018 to March 2021, a total of 39 months. Adult medical data would be expected to include any admissions to general, high-dependency or intensive care wards although few hospitals will have these higher levels of care. We chose 2018 as a starting point for the data extraction because of prolonged health care workers’ strikes in 2017 which affected health services provision [18] and consequently reporting. For each hospital, we extracted data on administrative units, level of hospital (Level 3: comprehensive primary health care facility, level 4: primary referral hospitals, Level 5: secondary referral hospitals and Level 6: national teaching and referral hospitals), whether a hospital is privately or publicly owned (ownership) and whether the hospital belongs to a clinical information network (CIN). The CIN is a collaboration of selected hospitals, policy makers and health researchers aimed at improving information available on the quality of inpatient paediatric care in Kenya [19–21]. A few hospitals reported data into an alternative parallel database (the hospital administrative database) and these data were merged where appropriate.

Handling missing data.

To adjust for incompleteness in reporting during the period of study, multiple imputation (MI) was performed [22–24]. Missing monthly values were imputed using a mixed effects model in a joint modelling framework using the jomo package in R and obtained 100 imputations for each indicator [25, 26]. We assumed the data followed as missing at random (MAR) pattern. The imputed datasets were analysed individually, and the estimates pooled using Rubin’s Rules. Each inpatient indicator was used as an outcome in the imputation model. Ownership (public or private), whether hospital belongs to a clinical information network (CIN), time (month and year) and COVID-19 binary indicator (0 –months before pandemic and 1 –months post pandemic) were used as covariates with the hospital ID as a clustering variable. MI was performed for hospitals with more than 30% of months reported (at least 11/39 months reported) to reduce uncertainty in imputed values and promote generalizability of the estimates. This strategy was supported by results from a simulation study we carried out, indicating MI performance and efficiency was best when imputing data for hospitals that reached this threshold of months reported. The imputation model specification is outlined in Handling missing data in S4 File. The number of hospitals included in analyses based on this strategy varied for each ward as summarised in the results with further information presented in Table 4 in S1 File.

Exploratory analyses.

Admissions and deliveries were aggregated monthly across all reporting hospitals and stratified by hospital ownership. Temporal trends were plotted separately for private and public hospitals to identify significant changes in health services and Statistical Process Control (SPC) charts were produced using 2018–2019 monthly averages as baseline to detect significant shifts in monthly admissions in 2020–2021 [27].

We also carried out multiple change point analysis [28, 29] to identify inpatient indicators that were affected by the national healthcare workers strike between December 2020 –February 2021. The assumption was that the strike affected all the public hospitals since we couldn’t obtain a database that tracks strikes in hospitals nationally.

Segmented regression.

We conducted interrupted-time series analysis using monthly number of admissions by ward and deliveries as outcomes. Analysis was stratified by hospital type. For caesarean sections and admissions in NBU where we had an appropriate denominator (total deliveries and live births respectively), they were used as offset variable to convert them into a rate and adjust for any potential changes in utilization over time [30, 31]. We therefore report caesarean section rates and NBU admission rates. The period running from January 2018 through March 2020 was defined as pre-COVID-19 and April 2020 to March 2021 as post-COVID-19. For each indicator, a segmented regression model was fitted [30, 32]. The following equation specifies the model [32];

Where, is the number of admissions in month is a continuous indicator of time in months from January 2018; is an indicator of time occurring before ( or after ( the outbreak, which was implemented at April 2021 in the series, where COVID-19 measures were most stringent in Kenya [33]; and is a continuous variable counting the number of months after COVID-19 at time In the model, estimates the baseline level of admissions at time zero; estimates the change in the monthly number of deliveries before the intervention (the pre-existing trend); estimates the level change in monthly number of admissions immediately after COVID-19 outbreak, which is from the end of the preceding segment; estimates the change in the trend of monthly admissions after COVID-19, compared with the pre-existing trend. The random variability at time , which is not explained in the model is represented by the error term . A level change (immediate COVID-19 effect) and change in trend after COVID-19 were hypothesised [30].

A generalized linear model was applied, assuming a Poisson distribution. Since the count data was over dispersed, and after evaluating several models, we fitted the negative binomial model to account for overdispersion [30, 34, 35]. Model checking was conducted for autocorrelation using the Durbin-Watson statistic and autoregressive moving average (ARMA) models were fitted for indicators with serial autocorrelation [36–38]. For indicators with significant autocorrelation, various combinations of the ARMA parameters (order of the autoregressive and moving average polynomials) were tested and the model resulting to lower Akaike Information Criterion (AIC) selected. Seasonality was adjusted using Fourier terms [39]. Results were pooled across the multiple imputed datasets using Rubin’s rules [40]. The negative binomial segmented ITS model, which accounted for seasonality, autocorrelation and overdispersion was the best model when compared with negative binomial models that were not adjusted.

Sensitivity analyses were performed for indicators affected by the national health care workers strike by excluding months affected and comparing the estimates with those from data that included the strike period. In addition, we anticipated differences in reporting across hospitals. It is possible that the impact of the pandemic across hospitals was variable, for instance some hospitals were assigned as isolation centres or closed for some time. Also, the magnitude of the epidemic in some counties was higher. Therefore, in counties with higher case burden, some hospital staff were re-assigned or had to go for quarantine if exposed or isolation if they contracted COVID-19, leading to human resources shortages. The intraclass correlation coefficients across hospitals for each indicator is provided in Table 3 in S3 File. Therefore, a Poisson generalized estimating equations (GEE) was fitted for each indicator to test the effect of varying model assumptions, such as accounting for differences in hospitals, on the estimates [41].

Statistical significance was defined as p-values < 0.05. All analyses were performed using R (version 3.6.3).