Study setting and data source
This research aims to examine the relationship between the size and composition of the health workforce (input) and the volume of service utilization (output) of Health Centers (HCs) from three regions of Ethiopia: Addis Ababa, Oromiya, and Southern Nations, Nationalities and People’s Region (SNNPR). We used secondary data of Ethiopian Fiscal Year 2007 (Gregorian Calendar 2016) from three sources: the HMIS for outputs (17), Human Resource Information System (HRIS) for inputs and facility attributes (18,19), and subnational-level (woreda or district) population projections of Ethiopia from the US Census Bureau’s International Data Base (20).
HCs in Ethiopia provide both preventive and curative services – including family planning, perinatal care, facility-based delivery, vaccination services, and outpatient consultations. Some HCs also provide limited inpatient service with five beds. To provide this wide range of PHC services, an average of 20 staff are sanctioned in each HC (21), which includes emergency surgical officers, health officers, nurses, midwives, pharmacy professionals, laboratory technicians, and administrative staff (22). While there are few HCs where doctors are posted, in most HCs, clinical service is provided by non-physician health workers such as emergency surgical officers and health officers (HO) (23).
Measurements
Output measures
We selected five indicators representing the annual volume of curative and reproductive, maternal, and child health services provided by the HCs – number of outpatient visits (OPD), number of acceptors of modern family planning methods (FP), total first antenatal care visits (ANC1), annual number of facility-based deliveries by skilled birth attendance (SBA), and number of children with three Pentavalent vaccines received within their first year (PENTA3). HCs report monthly the volume of these services through the HMIS system (21). We acquired the annual service utilization of the five indicators from 2,163 HCs, along with their name and geographical locations from the HMIS repository.
Input measures
The HRIS reports the number of all types of healthcare workers posted in each HC at the beginning and the end of the fiscal year (18). We identified 2,005 HCs from the three regions and retrieved the health workforce information at the beginning of the fiscal year, along with the name and geographical locations of the HCs. We organized the healthcare workers into three categories – clinical, para-clinical, and administrative staff. To adjust the variability of the skills of the different clinical staffs, we constructed a HO equivalent clinical staffs considering the four years of training of HO as reference (Weights: HO = 1, doctor = 1.5, emergency surgical officer = 1.5, nurse = 1, midwife = 0.75). The sum of weighted-clinical, para-clinical, and admin staff represents the total workforce of an HC.
Contextual covariates
Many contextual factors can also confound the estimation of productivity of health workers (24). These factors can be either facility’s intrinsic characteristics (25) – for example, infrastructure, provider-mix, financing, management, etc. – or extrinsic factors such as geography, demography, and the healthcare market structure (26,27). As intrinsic factors, we included the number of beds of the HCs as a proxy for facility size and the provider-mix of clinical, para-clinical, and admin staff. As extrinsic contextual covariates, we included the geographical location of the HCs and the woreda population where the HC is situated, estimated by the US Census Bureau (20).
Analytical approach
To develop the productivity measure of the health workforce, we followed these analytical steps: (1) development of the analytical dataset, (2) estimating a summary measure of the five outputs, and exploring its distribution (3) constructing a productivity ratio by using the summary measure and total staff, and exploring its characteristics, and (4) providing two examples of practical applications of the productivity ratio that could be a part of routine health service monitoring and provide the basis for interventions to improve TE.
We exported the input and output measures of the HCs, and the woreda-level population estimates from Microsoft Excel spreadsheets to Stata 15.1 (28) for data management. We performed data cleaning by checking the frequency and missingness and found that some HCs reported a high volume of utilization and staff numbers. We identified outliers from the input and output measures using the Interquartile range (IQR) method (29). After performing listwise deletion of any missing and outliers, 1,582 HCs with all five output measures and 1,483 HCs with the input information remained. Combining the output and input measures by matching the name of the facility and location (woreda and zone), we developed a unified dataset of 1,128 HCs and merged the woreda level population estimates with the dataset.
Developing a summary measure of output
To estimate health worker productivity as a measure of TE for multi-function HCs, we need to solve the complexity of these facilities producing multiple outputs. We used two different statistical methods – principal component analysis (PCA) and factor analysis (FA) – to estimate the summary measure of outputs (SMO) from the five output measures. Both PCA and FA are data reduction techniques that allow us to build a single measure from multiple variables capturing the most variability in the data, with some fundamental differences in the underlying theory and assumptions (Figure 2).
As indicated in Figure 2, using the PCA, we can develop a single index measure – also called a component (C) – which is the weighted average of indicators Y1 to Y5 (30). From a causal perspective, it signifies that the five outputs are cumulatively producing the index measure that reflects the overall output of an HC. In contrast, FA considers there is a latent variable (F) – in this case, the overall or system-level outputs produced by an HC – which we cannot directly measure (31). This latent construct represents itself through the common variance shared by the individual outputs, which we can measure. If Y1, Y2,…Y5 are highly correlated – indicating the same latent construct – we will see strong associations (λ1, λ2,…λ5) between the outputs and the latent variable (Figure 2). The unique variance not explained by F is considered as the measurement error (ε1, ε2,…ε5). Parameterizing these equations, we can statistically estimate the factor score representing the latent construct’s value.
After estimating the PCA and factor score for each HC, we explored their consistency using the Pearson correlation coefficient and visualizing their distribution. As FA is theoretically suited for this analysis and produces a more precise measure, we used factor score as the SMO of each HC. We rescaled factor scores between 0 to 100 because the standardized factor scores generated from the FA presents a mean of 0 and a standard deviation of 1.
Calculating the health worker productivity measure
We calculated the productivity of each HC by dividing the SMO by the total number of health staff, which is the cumulative number of HO equivalent clinical staff, paraclinical staff, and admin staff.
The crude productivity score represents the average unit of the SMO per staff of the facility. We have also examined the relationship of the productivity score with the SMO and the staffing level of facilities.
Developing examples of practical application of productivity score
We provided examples of how this kind of analysis could be used in practice by health system managers: (1) investigating the determinants of productivity as an explanatory tool for policymaking, and (2) ranking of the HCs and higher administrative levels using the adjusted productivity score. To develop these examples, we implemented multilevel linear mixed-effects regression models accounting for the confounding effect of the contextual factors. Contextual factors affect HCs’ capacity to produce outputs by influencing the service utilization volume, and subsequently, its productivity (24). HCs may yield higher outputs when situated in an urban area due to higher demand. Likewise, a cluster of HCs located in a geographical area (woreda or zone) may have more health workers because of policy measures. Ranking of the HCs based on the crude productivity can be misleading due to the confounding effect of the intrinsic and extrinsic contextual factors.
We accounted for the contextual factors and the clustering effects in the regression model to explore the determinants of productivity. We performed a log-log transformation of the dependent variable and the provider-mix covariates (number of HO equivalent clinical, para-clinical, and admin staffs) as they are highly skewed to the right (32). The regression was used to estimate the predicted productivity of HCs, which is a more precise measure of productivity adjusted for the contextual factors. The descriptive analysis, PCA, and regression models were performed using Stata 15.1 (28), FA was performed using Mplus 8.3 (33), and visualizations of the results were developed using the R package ggplot2 3.3.3 (34).