This study utilized an interrupted time series design used in epidemiology, public health, and other fields to assess the impact of an intervention or policy change over time. For our study, the researchers collected data at multiple time points, both before and during the COVID-19 lockdown. This design allowed us to examine the trends in the outcome of teenage pregnancy over time and evaluate whether the lockdown had a significant and sustained effect on the teenage pregnancy trend.

The researcher considered one year before the institution of the lockdown and one year after the institution of the lockdown. In total, the researcher collected data for 25 months, with March 2020 as the starting point when our intervention was implemented.

The study area was Pakwach District, located in the West Nile Region of northern Uganda. According to Uganda Bureau of Statistics figures, the district had a total population of 158,037 in 2014 and was projected to increase to 181,400 by 2018, and 51.7% of these are females. Antenatal care first-visit attendance is very high in our study setting, above 95%, as reported in the DHIS2. The District Health Information System 2, which collects aggregate data on ANC attendance for women, was then accessed to retrieve our dataset.

The target population in this study was a record of teenage girls aged 10–19 who got pregnant and attended their first ANC visits to health facilities in Pakwach district from March 2019 to March 2021.

In our analysis, the researcher obtained teenage pregnancy data directly from the District Health Information System 2 (DHIS2) database, which is accessible at the DHO office. We specifically searched and retrieved data related to the first antenatal care (ANC) service for adolescent girls aged 10–19 during the period from March 2019 to March 2021. The indicators required for this analysis (period name, ANC first visit for women less than 15 years, and ANC first visit for women aged 15–19 years) were correctly specified by the researcher to the district biostatistician at the DHO office. Using the district Health Information System 2 dataset, which is one of the most reliable databases for keeping aggregate health records in Uganda, the researcher logged into the system and retrieved data as per our specified indicators into an Excel sheet. The extracted records the researcher collected over a span of 25 months, encompassing both pre-lockdown and lockdown periods. During that period, the district recorded a total of 4422 pregnancies among adolescents during these two selected periods.

Before conducting our data analysis using the interrupted time series (ITS) design, the researcher performed several important tests to assess key characteristics. These tests include visual examination by plotting the time series data, evaluating descriptive statistics, testing model assumptions to validate the chosen ITSA model, and checking for autocorrelation, which assesses the correlation and proximity of collected data. The researcher also conducted tests for secular trends, examining whether the dataset exhibited consistent increases or decreases over time regardless of any intervention, as well as for seasonality. Therefore, in our analysis, the researcher diligently examined all these essential attributes in our dataset and generated the necessary graphs, particularly autocorrelation graphs. Furthermore, the researcher examined the overdispersion in our data to ensure it adhered to the assumption of the Poisson distribution, where the variance is equal to the mean, and confirmed that our data were suitable for the chosen model. The autocorrelation checks ultimately guided our decision to employ an OLS model, specifically the Newey model with a specified lag of 1.

The researcher conducted two-level analyses to compare teenage pregnancy trends before and during COVID-19 in Pakwach. In the first analysis, the researcher loaded the data, declared the dataset as a panel in Stata, and specified a single group ITS analysis. For this analysis, the intervention start was set for March 20th, 2020, when the COVID-19 lockdown was instituted in Uganda. The researcher then conducted an ITS analysis using OLS and plotted the results. See Table 1.

Firstly, to compare the trend in teenage pregnancy before and during COVID-19, the researcher utilized the Stata itsa (interrupted time series analysis) command. After importing the data to Stata, the researcher started the analysis by ensuring that our time variable was in the proper format, e.g., 722 became 2020 m3 (March). After that, the researcher then declared the new time variable created to become a time set.

At this stage, the researcher performed some descriptive analysis to just check our trends by summarizing our variable Preg. This was followed by a time-series analysis conducted in Stata using the itsa command. The result was presented in a table and a graph. (See Table 1 and Fig. 1). More importantly, we made sure that our model fit by checking for autocorrelation, using p < 0.05 at any lag to show autocorrelation, finding no autocorrelation in the data, indicating that there was no need to adjust for it, and plotting its graphs.

Showing results of OLS plots

In the second analysis, the researcher used the poison regression command to estimate the effect of the COVID-19 lockdown on teenage pregnancy trends. In this poison regression analysis, our outcome became the number of pregnancies, intervention (COVID-19 lockdowns), and time (_n of time periods chronologically arranged). After specifying all conditions of the family of Poisson regression and plotting a log (link) eform for it, the researcher generated the predicted values based on this model and produced a plot of the model along with a scatter graph. The researcher then generated the counterfactual by removing the effect of the intervention (_b [smokban]) for the post-intervention period and adding the counterfactual to our plot.

To ensure the soundness of the model, the researcher further checked for autocorrelation by examining the autocorrelation and partial autocorrelation functions. Additionally, the researcher also adjusted for seasonality as one of the important attributes in the analysis. The results were presented in the form of a table and graphs (See Table 2 and Fig. 1).