Globally, the COVID-19 pandemic has caused a surge in anxiety [8, 28]. With a survey covering 58 countries and over 100,000 respondents in March and April 2020, Paudel et al. demonstrated the link between the number of reported COVID-19 cases and anxiety levels [29]. Moreover, containment policies including self-isolation or social distancing orders were shown to be associated with increased anxiety and emotional instability for those who stayed at home [29]. A meta-analysis covering 204 countries pointed to a particularly strong psychological burden of the pandemic for women and younger populations [30]. Studies from the SSA region also indicate a heightened incidence of anxiety during the COVID-19 pandemic [8, 28].

A recent systematic review of the relationship between FI and mental health in Africa revealed a dose-response relationship [31]. Emerging evidence on the underlying mechanisms suggests that FI directly deteriorates mental health by depriving from fulfillment of a basic need and by creating uncertainty over the ability to meet these needs in the future [32] other mechanisms such as nutrient deficiencies affecting brain function or one’s perceived relative economic status manifested in FI were shown to be of secondary importance [31–33]. The inherent link between FI and mental health is also supported by the study of Jones and colleagues who found a robust association regardless of culture or a countries’ wealth [34].

Our study provides new evidence on the prevalence and associated factors of GAD in the African context. To this end, we draw on novel phone-based survey data providing estimates of prevalence rates of GAD according to the DSM-5 alongside with direct COVID-19 exposure and FI in the general adult population in four African low- and middle-income countries–Mozambique, Sierra Leone, Tanzania and Uganda. We then analyze whether the pandemic adversely affected mental health directly and indirectly by testing whether FI is a mediator of the relation between COVID-19 exposure and GAD.

As part of the Life with Corona (LwC) project [35], the LwC-Africa study collected phone survey data from adults (>17 years) in Mozambique, Sierra Leone, Tanzania and Uganda throughout the year 2021 [36]. During the COVID-19 pandemic, conducting face-to-face interviews could have posed a risk to both the interviewer and the respondent. Moreover, utilizing online surveys in low- and middle-income countries would have introduced a potential bias by selectively including individuals with internet access. To mitigate these challenges, telephone interviews were employed, as a substantial portion of the population possesses a cell phone, offering a more inclusive and accessible mode of data collection. The countries were selected for their variance regarding geography, demographics, and response to the COVID-19 pandemic (e.g., [25, 37]).

The survey was implemented in 12 rounds of data collection from January to December 2021 by a non-governmental organization (BRAC International in Sierra Leone, Tanzania and Uganda) and a local company (Intercampus in Mozambique) using large in-house databases of mobile phone contacts from active and previous research; in each round, participants were randomly selected, but the composition of the population in each country in terms of gender, age, location (urban/rural) was accounted for aiming at 500 new randomly selected participants per month and country. Our sample size was informed by and consistent with the literature [36]. Before the interview, informed oral consent was obtained from all individual participants included in the study by reading out the consent statement at the start of the interviews. The statement contains information regarding the purpose of the interview and the confidentiality of their personal information. The statement also contains information regarding the respondent’s rights to refuse to respond to any individual question or the entire interview. The calls to the respondents were made during the day from the call centers of the survey companies. Usually, two to three calls were made to secure an interview. When repeated attempts to reach a selected respondent were not successful, a replacement satisfying the stratification criteria was used. In line with similar studies [38], the response rate was about 50% whereby the main reasons for non-response were inactive phone contacts and non-response to the phone calls. The average interview duration was 16.5 min. The study was approved by the ethical commission of UNU-WIDER.

A total of N = 23,943 participants were interviewed (n_Mozambique = 6000; n_{Sierra Leone} = 6217; n_Tanzania = 6033; n_Uganda = 6033). To obtain a representative sample and more precise estimation of prevalence rates per country, we applied post-stratification weights for age, gender and location [39]. Resulting sociodemographic estimates and country comparisons are presented in Table 2.

Computer-assisted telephone interviews were conducted by trained enumerators; the 4 days training included relevant concepts, mock interviews, administration of the computer-assisted telephone interviews (CATI) protocol, piloting of the survey, and debriefing sessions. The original English version of the questionnaire was translated into Luganda, Runyankole, Acholi, Lusoga, Kiswahili, Krio/Creole and Portuguese. Validity of the translations for use in the respective countries was ensured based on repeated mock interviews among the enumerators, fluent speakers of the local languages, during the training as well as the pre-survey piloting of the questionnaire.

GAD according to DSM-5 was measured with the 7-item Generalized Anxiety Disorder scale (GAD-7) [16]. Each item (e.g., Feeling nervous, anxious, or on edge; Not being able to stop or control worrying; Worrying too much about different things) was rated on a scale from 0 (not at all) to 3 (nearly every day) according to the presence of symptoms during the last 2 weeks. A sumscore ranging 0–21 was used to indicate symptom severity. Following a systematic review/meta-analysis conducted by Plummer and colleagues, who presented a sensitivity of 0.83 and specificity 0.84 [40] we used a cut-off value of 8 to estimate the diagnosis of GAD. We present prevalence rates at alternative thresholds in the Supplementary Table S7. Although validation studies are not available for the specific countries included in our study, the GAD-7 is a standardized instrument validated for various (African) countries [41]. Recognizing the importance of context-specific validation, we strongly advocate for further validation studies in African countries to enhance the reliability and applicability of mental health assessment tools in diverse cultural settings.

COVID-19 exposure was measured with four dichotomous (yes [1]/no [0]) questions asking whether the respondent was infected by the Coronavirus, met anyone who acutely suffered from COVID-19 symptoms, felt to live in an area with high incidence or knew someone who died from the consequences of Coronavirus infection. We had to rely on subjective self-assessment of perceived exposure as there has been a wide gap in testing in most African countries [42]. Similar to other sensitive topics in surveys, data quality regarding questions about COVID-19 exposure may be constrained by survey biases such as reporting bias due to social desirability, fear of stigma and discrimination. For the prevalence rates, we analyzed calculated a score summing up all yes [1] answers; COVID-19 exposure was counted as positive when at least the answer to one of the four questions was yes.

FI was measured using the Food Insecurity Experience Scale (FIES) [43]; a globally used and validated measure. The eight questions focus on experiences and behaviors related to difficulties in accessing food (e.g., worrying about food scarcity, eating less nutritious food, skipping meals or running out of food) in a 4 weeks recall period and can be answered with yes [1]/no [0]. Using a probabilistic approach, the respondents’ status with regard to a policy relevant level of FI (moderate-or-severe FI) was determined as suggested by FAO [43, 44].

Furthermore, we collected information about sociodemographic characteristics (gender, age, years of education, and marital status), living in a rural/urban area, household composition and wealth indicating the socio-economic status of the respondent (hereafter: SES index); the latter was calculated via Principal Component Analysis to reduce the dimensionality of numerous variables (see Supplementary Material SⅡ). This approach captures household living standards and therefore avoids common measurement biases related to income or expenditure measures [45]; higher SES index indicates more assets, better infrastructure, and housing.

To estimate prevalence rates of GAD diagnosis and exposure to COVID-19 FI, we use frequencies and percentages on our sample with post-stratification weights applied for age, gender and location (urban/rural). Respondents were excluded when information in any main variable was missing (missing values: n_GAD-7 = 1; n_COVID-19 = 3,576; n_FI = 230). To examine the association of the pandemic and FI and the risk of developing GAD, we conducted logistic regression analyses. Furthermore, we performed a mediation analysis using structural equation modeling [46] hypothesizing that FI mediates the association of COVID-19 and GAD. To allow for full convergence of the mediation model, we omit country-fixed effects and only include region and survey round dummies. The reliability of mediation analysis relies on certain vital assumptions, including the lack of multicollinearity, minimal measurement errors (especially in the mediator), control of omitted variable bias by incorporating relevant confounders, and linearity. We have found that these necessary assumptions are met, thus confirming the validity of the use of mediation analysis in our study. The statistical analyses were performed in Stata version 17.

Acknowledging the limitations inherent in our study, we stress that our analysis can only estimate associations and not infer causal effects due to the cross-sectional nature of the observational data. Moreover, FI prevalence might be underestimated due to the sampling strategy via mobile phones. For the 2019–2021 timeframe, an estimated 79% of the adult population (<17 years) in Uganda own a cell phone, 75% in Tanzania; 76% in Sierra Leone; and 62% in Mozambique [50]; individuals without a mobile phone potentially show higher FI values. Consequently, GAD prevalence rates might also be higher. However, since we control for household composition and wealth, we guard against possible bias in our analyses. Another limitation is that self-administered instruments such as the GAD-7 may not allow for differential diagnosis, but rather provide complementary information. Nevertheless, sensitivity and specificity tests in validation studies show that GAD-7 has acceptable properties for identifying GAD compared to the established gold standard clinical diagnosis [40].

Our analysis suggests two main insights: Firstly, we learn from our data that about 2–4 in ten individuals –a considerable proportion– in Mozambique, Sierra Leone, Tanzania and Uganda presented with excessive and uncontrollable worries during the COVID-19 pandemic. Secondly, persistent stressors such as COVID-19 and moderate/severe FI but particularly FI contribute significantly to the risk of GAD. What does this tell us about the African experience of the pandemic, its legacy for development and avenues for policy action?

Anxiety induces the action disposition of avoidance or defense. While fueling anxiety may carry the advantage of stronger compliance to contagion containment measures, e.g., [48, 51], it may also cause higher GAD rates (amongst other consequences): For instance, a meta-analysis with 72,585 participants [52] revealed that participants with GAD presented with a higher incidence of lifetime perpetration of intimate partner violence. Assessing the transgenerational impact of GAD, Woodruff et al. found mothers with high anxiety levels less engaged with their children [53] and Moore et al. found less warmth, and higher levels of control and criticism [54]. Other transgenerational studies consistently indicate high GAD incidence rates in the offspring. Moreover, experimental laboratory studies show that patients with GAD are more likely to accept and react less strongly to unfair decision [55]. Thus, community-level social corrective reactions may decline while actual and subjective safety and stability in the families may decrease. Higher levels of day-to-day violence in communities and families should be considered as immediate consequence alongside with additional waves of migration. This indicates priority areas for peri- and post-pandemic development programs which include the reduction of violence at the family and community level, as well as an increase in social support.

Moreover, our findings underscore the detrimental effects of concurring persistent stressors on GAD risk. In contrast to other risk factors for anxiety disorders highlighted in previous research, such as immediate threats to survival (psychological trauma), genetic predisposition, and/or health status, we broaden the focus to include persistent threats to survival. In this context, the array of possible policy responses expands: Ensuring that all people can meet their basic needs would not only address these long-term stressors but would also have an important multiplier function. Thus, our findings have immediate policy implications for supporting food programs, as food security is a fundamental determinant of resilience and particularly important when other threats to human life arise. Moreover, addressing food insecurity lends itself more readily to population-wide interventions [56]. Thus, even during global health crises such as the COVID-19 pandemic, addressing FI as a key driver of GAD should be a top priority for policymakers.