Sexual Violence Trends before and after Rollout of COVID-19 Mitigation Measures, Kenya

COVID-19 mitigation measures such as curfews, lockdowns, and movement restrictions are effective in reducing the transmission of SARS-CoV-2; however, these measures can enable sexual violence. We used data from the Kenya Health Information System and different time-series approaches to model the unintended consequences of COVID-19 mitigation measures on sexual violence trends in Kenya. We found a model-dependent 73%–122% increase in reported sexual violence cases, mostly among persons 10–17 years of age, translating to 35,688 excess sexual violence cases above what would have been expected in the absence of COVID-19–related restrictions. In addition, during lockdown, the percentage of reported rape survivors receiving recommended HIV PEP decreased from 61% to 51% and STI treatment from 72% to 61%. Sexual violence mitigation measures might include establishing comprehensive national sexual violence surveillance systems, enhancing prevention efforts during school closures, and maintaining access to essential comprehensive services for all ages and sexes.


Statistical Analysis
We first conducted descriptive analyses. We assessed distributions and outliers for each indicator, decomposed the data to check for seasonality, secular trends, and random noise, and assessed the data for autocorrelation and partial autocorrelation. Next, we conducted unit root tests to estimate the number of lags required to make the data stationary.
Traditional quasi-experimental policy evaluation methods like difference-in-differences are unsuitable in the current scenario due to the lack of appropriate control groups. We, therefore, used time series approaches to compare sexual violence trends before and after the introduction of COVID-19 mitigation measures in Kenya on March 15, 2020.
Akin to difference-in-differences, time series methods assume that pre-policy trends, including seasonal variations and levels would remain unchanged in the post-policy period under a nonintervention counterfactual state. The estimated policy impact is therefore the difference between the counterfactual-state estimates and observed data. The validity of this approach hinges on accounting for any concurrent shocks that could affect these trends and levels, such as changes in concomitant policies, measurement processes or population composition (1).
Our models are based on the following assumptions. First, that there were no changes in data reporting during the pandemic. We check this assumption by examining data quality reports and through discussions with key public health program officials working on sexual violence in Kenya. Second, we assumed that there were no other concurrent events, other than the pandemic policy shock, that could drive the results. These competing events could include new legislation penalizing sexual violence or mass disruptive events like civil conflicts. We check this assumption using date falsification tests (changing the policy start dates several months before Page 2 of 8 and after March 2020), Supremum Wald tests for unknown structural breaks, and Wald tests for known structural breaks in the data (2)(3)(4).
Third, we assume that there were no anticipatory (i.e., Ashenfelter-type) pre-policy effects; that is, perpetrators could not adjust their behavior in anticipation of the lockdown policy because the shutdown date was driven by unanticipated global factors. Existence of prelockdown anticipatory effects would bias the estimation of counterfactual trends. We checked this assumption in part by tests outlined under the second assumption and by examining raw trend We used a forecasting approach for the SARIMA model. Having defined the number of appropriate lags to stabilize the data by using the steps described above, we selected the most appropriate SARIMA model by using Akaike information criterion and the Ljung-Box (Q) test (5). We then stabilized the model by using regular and seasonal differencing and rechecked for stationarity by using augmented Dickey Fuller (ADF) unit root tests.
We introduced an additional preprocessing step to confirm if the selected SARIMA model was appropriate. We split the data into training and testing datasets as follows: training was January 2015-July 2019, and testing was July 2019-February 2020. We then compared the SARIMA forecasts against the actual observed values in the testing dataset. We evaluated forecasting performance by using root mean square errors (RMSE) and mean absolute errors (MAE), and chose the appropriate autoregression orders, trend differences, and moving average orders. We then used the dataset through February 2020 to forecast for values through June 2021 and compared forecast estimates with the actual observed values. The difference between the forecasted and actual values represents the policy impact. The main ITSA model was specified as follows (1): Where Yt is an aggregated regressor (e.g., OPD visits) that is measured at equal monthly intervals t, Xt is a 0-1 indicator variable representing the COVID-19-related lockdown in March 2020, Tt is the time in months since January 2015 and are dummy variables for months to account for seasonality. Β0 is the intercept term, β1 is the pre-lockdown slope, β2 estimates the lockdown policy shock level change, and β3 estimates the long-term effect of the policy change (1,6). The error term, εt uses Newey-West standard errors to account for serial correlation (1,7).

Assumptions and Additional Robustness Checks
Kenya experienced a series of nationwide healthcare worker strikes in 2016 and 2017 that disrupted health services (8). There is a risk of obtaining spurious results if the effects of these strikes were significant and sustained. We, therefore, conducted additional robustness checks with an additional dummy variable (second interruption) for the onset of the strikes in the segmented regression models. We visually inspected the decomposed data for the strike period to determine if the trends were deterministic (recovered long-term trajectories after the strike ended) or stochastic (maintained a new trend after the strike ended).
The Wald and Supremum Wald tests identified 1 significant change (structural break) in trends in 2017 coinciding with the national health worker strikes. The inclusion of this break in the models did not change the results. Our results were also robust to date falsification tests with no impacts seen when the lockdown start date was varied by several months before and after March 2020.

Software
We performed the initial data manipulations using Python version 3.7 (Python Software