Association Between COVID-19 and Self-Harm: Nationwide Retrospective Ecological Spatiotemporal Study in Metropolitan France

Abstract Background The COVID-19 pandemic has not been associated with increases in suicidal behavior at the national, regional, or county level. However, previous studies were not conducted on a finer scale or adjusted for ecological factors. Objective Our objective was to assess the fine-scale spatiotemporal association between self-harm and COVID-19 hospitalizations, while considering ecological factors. Methods Using the French national hospital discharge database, we extracted data on hospitalizations for self-harm of patients older than 10 years (from 2019 to 2021) or for COVID-19 (from 2020 to 2021) in metropolitan France. We first calculated monthly standardized incidence ratios (SIRs) for COVID-19 between March 2020 and December 2021, using a Besag, York, and Mollié spatiotemporal model. Next, we entered the SIRs into an ecological regression in order to test the association between hospital admissions for self-harm and those for COVID-19. Lastly, we adjusted for ecological variables with time lags of 0 to 6 months. Results Compared with a smoothed SIR of ≤1, smoothed SIRs from 1 to 3, from 3 to 4, and greater than 4 for COVID-19 hospital admissions were associated with a subsequent increase in hospital admissions for self-harm, with a time lag of 2 to 4 months, 4 months, and 6 months, respectively. Conclusions A high SIR for hospital admissions for COVID-19 was a risk factor for hospital admission for self-harm some months after the epidemic peaks. This finding emphasizes the importance of monitoring and seeking to prevent suicide attempts outside the epidemic peak periods.

1 Spatio-temporal models and ecological regressions

Spatio-temporal models
Let denote by y it and E it the observed number and age-and -gender standardized expected number of either hospital admissions for self-harm or COVID-19 incident cases for the ith spatial unit and t-th month t, respectively.The expected number of hospital admissions for COVID-19 was calculated as follows: where n kit is the population at risk in spatial unit i in month t and age and sex stratum k, and p k corresponds to the incidence rate of COVID-19 in stratum k.
A more novel approach was used to calculate the expected number of hospital admissions for self-harm.In fact, the incidence of hospital admissions for self-harm before the pandemic was relevant for inclusion in the model because the association between hospital admissions for self-harm and COVID-19 might change over time.This information was included in the calculation of the expected number of cases, as follows: where p 2019 kt is the incidence rate of hospital admissions for self-harm in month t in 2019 in stratum k in the study area.This approach differs from the classical approach but has the advantage of using the 2019 monthly incidence rates as a reference.
The general formulation of the model is the following: where θ it is the relative risk of either hospital admissions for self-harm or COVID-19 incidence in the i-th spatial unit and t-th time, and β 0 is an intercept (overall risk).u i is a spatially structured random effect modeled by an intrinsic conditional autoregressive model (ICAR) in which the weights w ij of adjacent spatial units are equal to 1 and all others are equal to 0, of variance σ 2 u .v i is an unstructured spatially random effect of variance σ 2 v .ξ t is unstructured temporal random effet (i.i.d.) of variance σ 2 ξ and γ t is temporally structured random effect defined as a random walk of order n (RWn), n = 1 or 2, of variance σ 2 γ .δ it corresponds to the type I interaction of both temporal and spatial components, referring to unstructured overdispersion in time and space, and modeled as an i.i.d.random effect of variance σ 2 δ [Besag et al., 1991, Knorr-Held, 2000].
As a first step, we considered the following different spatio-temporal models whose fit was assessed by the Watanabe-Akaike information criterion (WAIC)(Table 1).For self-harm, based on the criterion values, we selected the model that presented the best fit (i.e.minimizing the WAIC) which had an unstructured temporal random effect, a structured temporal random effect (RW1) and type I interaction: For COVID-19, the model that presented the best fit had an unstructured temporal random effect, a structured temporal random effect (RW2) and type I interaction: To describe spatio-temporal patterns of both self-harm and COVID-19, we reported the mean posterior of the exponentiated θ it as smoothed spatio-temporal Standardized incidence ratio (SIR).
Table 1: Different spatio-temporal models and their Watanabe-Akaike information criterion (WAIC).The minimum value of the WAIC criterion has been highlighted in bold and corresponds to the spatio-temporal model selected for self-harm and COVID-19.267,159.19 359,967.83

Ecological regression models
The association between between hospital admissions for self-harm and for COVID-19 was investigated using the previously selected Bayesian model, introducing each ecological variable X it as a fixed-effect covariate: where β 1 is the log-increase of hospital admissions for self-harm incidence for 1 unit in ecological covariate.For each covariate, we reported this on natural scale (exp(β)) as the relative risk together with its 95% Bayesian credible interval (BCI).
Initially, the covariates considered were the smoothed SIR for COVID-19 obtained in the selected spatiotemporal model in subsection 1.1.Smoothed SIRs were transformed into a qualitative variable with spatial units with SIR ≤ 1, corresponding to space-time units for which the incidence rate of hospital admissions for COVID-19 was lower than the average over the study period and area.SIR ≤ 1 served as the reference class.The other SIR classes were ]1-2], ]2-3], ]3-4], and >4.For example, the class ]2-3] corresponds to excess incidence by a factor of 2 to 3, compared with the average incidence over the whole study area and period.Given that COVID-19 might have induced a time lag in hospital admissions for self-harm, time intervals ranging from 0 to 6 months were tested.
After selection of the best model for the time lag between self-harm and COVID-19 hospital admissions, ecological covariates were added to the models through forward selection (based on a decrease in the WAIC).All the covariates were centered and reduced.