Time Series Analysis and Short Term Forecasting of Female Contraceptives Consumption in the Public Health Sector

Abstract

Background: Contraceptives security is crucial in ensuring access and delivery of family planning services and improving the contraceptive prevalence rate. Commodity security ensures that clients have access to commodities when and where they need them. It involves the integration of distribution systems, finances, health services, and policy guidelines. Proper forecasting, quantification, and procurement are critical in ensuring an adequate supply of contraceptives. It is important to study consumption patterns and apply forecasting techniques so as to adjust for any changes in the choice of contraceptives. Objective: The overall objective was to evaluate trends in consumption and develop forecasting models for contraceptives consumption (injectables, pills, implants and intrauterine contraceptive devices) from 2014 to 2018 health facility reports. Methods: The study was a time series analysis of family planning data. Data on consumption of implants, injectables, pills, and intrauterine contraceptive devices was extracted from the District Health Information System (DHIS2) which is an online platform for reporting health-related data in Kenya. Data cleaning and transformation was done to improve accuracy and data quality. The first part of the analysis was exploratory where data for each contraceptive was evaluated for trend, seasonality, autocorrelation, and stationarity. This involved visual inspection of time series, correlogram and partial correlogram graphs and also carrying out statistical tests such as Ljung-Box test for autocorrelation and Augmented Dickey-Fuller test for stationarity. The data was also decomposed to evaluate the trends and seasonal components of the family planning commodity data. Comparison of consumption data on contraceptives and service point data was done using the time series correlation plots and correlation coefficients. The second part of the analysis involved short-term forecasting (six months) using the Autoregressive Integrated Moving Average (ARIMA) models and the exponential smoothing with underlying state space models. Model diagnostics were done on the residuals of optimal models. Measures of the accuracy such as mean absolute percentage errors and root mean square errors were used to determine the optimal model. Validation of the models was done to estimate the prediction error of the models and this was done by comparing the forecasted consumption from January to June 2018 with the actual consumption. Results: The consumption of pills, injectables and intrauterine contraceptive devices declined while that of implants increased from 2014 to 2017. There was seasonality in the consumption

patterns for each of the contraceptives. The lowest consumption was in December except for injectables and intrauterine contraceptive devices. There were differences in the data reported for consumption and service point data for injectables, implants and intrauterine contraceptive devices. The exponential smoothing models (ETS) were the best for forecasting consumption of all the contraceptives except for one-rod implants in which the Autoregressive Integrated Moving Average (ARIMA) model was more accurate. The ETS (M, N, N) was the best model for predicting consumption of progestin-only pills and intrauterine contraceptive devices. It tended to give underestimates with a mean error of -0.109 and -0.054 respectively. The ETS (A, N, N) was optimal for predicting consumption of combined oral contraceptives and injectables. For combined oral contraceptives the forecasts tended to overestimate with a mean error of 0.136 while for injectables it underestimated with a mean error -0.117. The ETS (A, A, N) was the optimal model for two-rod implants and it gave overestimates with a mean error of 0.052. The only contraceptive for which the ARIMA model was superior to the ETS models was for the one-rod implant. For this model, ARIMA (1, 1, 3) gave the lowest mean error for all methods considered with a mean error of 0.048. Conclusion: There was a general shift towards the use of long term reversible methods especially implants in Kenya. The difference in the reporting of consumption and service point data for injectables, implants and intrauterine contraceptive devices showed that there was a gap in the documentation and the reporting of the consumption and service point data. The ETS models were generally superior to the ARIMA models for predicting consumption of contraceptives.