The COVID-19 pandemic could worsen the psychological well-being of people with disabilities in Cambodia

Abstract

How the COVID-19 pandemic affected people with disabilities in the Global South is little known. This paper examines how disability was associated with people’s psychological well-being during the pandemic in Cambodia, and explores pandemic-related shocks as potential mechanisms. We collaborated with the national statistical office of Cambodia to design and implement a module on psychological well-being and COVID-19-related shocks in the 2021 Cambodia Socio-Economic Survey (official statistics). Our analysis of this unique, nationally representative data employs careful sample trimming (excluding disabilities due to aging), matching (of 276 disabled working-age adults with those without disabilities, within villages and districts), and regression analysis. The results show that disability was associated with lower subjective well-being (happiness and life satisfaction) and higher depression. The mediation analysis suggests distinct pathways for these adverse associations: pandemic-related economic shock (nonemployment) for physical disabilities and health shock (perceived risk of infection) for non-physical disabilities. Our findings provide novel representative evidence for the potentially intense vulnerability to the pandemic and underlying mechanisms among disabled people in the Global South, suggesting disability-inclusive policies addressing economic and health shocks to sustain their psychological well-being for future pandemics.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-43087-6.

Introduction

The estimated 1.3 billion people with disabilities (about 16% of the global population) are one of the most vulnerable and excluded groups in our society¹. Although the disability-inclusive response to COVID-19 was called for^2–5, available evidence suggests that disabled people were disproportionately affected by the pandemic in terms of access to healthcare^6,7, the risk of COVID-19-related mortality⁸, and mental health^7,9,10. In low- and middle-income countries, where approximately 80% of disabled people live², they faced multiple exclusions across the sectors of health, education, economy, community, and pandemic management^11,12. Due to a lack of data, however, how the pandemic affected disabled people in the Global South is little known^13,14. Although nationally representative evidence is crucial for effective policymaking, such evidence is notably lacking.

To fill this lacuna, we examined how disability was associated with people’s psychological well-being during the pandemic in Cambodia. Suppose that the association is shown to have been linked to pandemic-related shocks, which existed only during the pandemic. In that case, it provides evidence that the pandemic could exacerbate the psychological well-being of disabled people. We collaborated with the national statistical office of Cambodia to design and implement the module on psychological well-being and COVID-19-related shocks in the 2021 Cambodia Socio-Economic Survey (CSES) (official statistics based on a cross-sectional survey; ‘Data’ in Methods). These pieces of information are scarce in national surveys in the Global South. Our analysis is based on this unique, nationally representative data.

We employed a quasi-experimental framework to obtain more credible estimates of the association of people’s own disability with their psychological well-being (‘Empirical design’ in Methods). Specifically, because our data contain a small number of disabled working-age adults relative to those without disabilities, we employed careful sample trimming (excluding disabilities due to aging) and matching procedures to improve comparability in observed characteristics and outcome distributions between them within villages and districts. We then conducted a corresponding mediation analysis to assess pandemic-related shocks as potential mechanisms (pathways) underlying the association.

The results show that working-age adults’ disability (not due to aging) was associated with lower subjective well-being (happiness and life satisfaction) and higher depression. The patterns tend to have been stronger for physical disability than non-physical disability (sensory and mental disabilities) (these two types of disabilities are not mutually exclusive). These adverse associations of physical and non-physical disabilities were linked to pandemic-related economic shock (nonemployment) and health shock (perceived risk of infection), respectively. For non-physical disabilities, less common preventive health behaviors during the pandemic were associated with a higher updated perceived risk of infection, implying that disability-related constraints on preventive behaviors may have underlain the pathway through health shock. Comparing different types of non-physical disabilities revealed strong associations between underrepresented non-physical disabilities, such as mental disabilities, and lower psychological well-being. These novel representative findings have important policy implications to sustain the psychological well-being of disabled people in the Global South for future pandemics.

This paper contributes to the literature on vulnerable populations during the COVID-19 pandemic^15–17 in four important ways. First, we study subjective well-being and depression among people with disabilities in the Global South. Whereas the adverse impacts of COVID-19 on mental health are shown to be global phenomena^18,19, previous related works among disabled people are limited to the Global North^20,21. Second, our findings, based on the nationally representative data, have external validity in the country. Existing works using representative survey data in the Global South focus on welfare and labor-market consequences among poor populations^22,23. To our knowledge, our study provides the first nationally representative evidence for the adverse associations of disability with psychological well-being during the COVID-19 pandemic in the Global South. Third, we consider both physical and non-physical disabilities. Most extant work considers disabilities in general, or either physical or mental disabilities. Fourth, we find distinct underlying mechanisms for physical and non-physical disabilities, adding to the literature on potential mediating pathways from disability to psychological well-being^21,24–26.

Results

Characteristics of people with disabilities

To construct disability measures for functioning difficulties not due to aging, we trimmed the original 2021 CSES sample (Table 1A; ‘Data’, ‘Measures’, and ‘Empirical design’ in Methods). The resulting analysis sample included 276 (1.9%) working-age adults (ages 20–59) with any disability (functional difficulties) not due to aging (59% were males). The gendered difference in disability depends on the type of disability: physical and non-physical disabilities, which are not mutually exclusive. On the one hand, physical disability was more common among males than females (1.6% vs. 0.6%) (Supplementary Fig. S1A), which reflects the post-conflict setting of Cambodia (‘Study setting’ in Methods), where physical disabilities were common among ex-combatants and landmine victims^27,28. Indeed, physical disabilities were more commonly caused by conflict than non-physical disabilities (18% vs. 6%), whereas over 50% of disabilities were due to disease (Supplementary Fig. S1B). On the other hand, among individuals with non-physical disability, 66% had seeing difficulties, followed by hearing (19%), psychological (13%), speaking (11%), and feeling (sensing) (11%) difficulties, without notable difference between males and females (these disability types are not mutually exclusive; Supplementary Figs. S1A and S2D). As such, non-physical disabilities other than seeing difficulties were relatively uncommon or underrepresented in our data. Among 276 individuals with any disability, 25% had severe difficulties; physical disabilities were severer than non-physical disabilities, and both physical and non-physical disabilities were severer among males than among females (Supplementary Fig. S1C). Given the limited variation in the severity of disabilities (especially non-physical ones), we did not distinguish disability severity in the remaining analysis.

Table 1.

Construction of analysis and matched samples for 2021 CSES.

	Treatment group					Control group
	Any disability	Physical disability	Non-physical disability	Non-physical disability excluding feeling and psychological difficulties	Seeing and hearing difficulties	No disability
	(1)	(2)	(3)	(4)	(5)	(6)
A. Analysis samples
Heads of households and spouses of household heads (base sample)	1669 (0.093)	545 (0.032)	1376 (0.078)	1327 (0.075)	1296 (0.073)	16,363
Focus on working-age adults (ages 20–59)	617 (0.042)	205 (0.014)	465 (0.032)	439 (0.030)	421 (0.029)	14,009
Drop individuals who reported mild difficulties only	381 (0.026)	160 (0.011)	252 (0.018)	233 (0.016)	217 (0.015)	14,009
Drop individuals with functioning difficulties caused solely by aging (analysis sample)	276 (0.019)	152 (0.011)	150 (0.011)	130 (0.009)	114 (0.008)	14,009
Heads of households (analysis sample)	192 (0.025)	122 (0.016)	91 (0.012)	79 (0.011)	65 (0.009)	7386
B. Within-village matched samples
Head and spouse	225					417
		128				260
			118			200
				106		179
					94	154
Head	129					228
		87				162
			59			97
C. Within-district matched samples
Head and spouse	263					1218
		144				734
			143			645
				126		586
					111	500
Head	175					685
		113				486
			82			311

The table shows the construction of analysis samples (A), within-village matched samples (B), and within-district matched samples (C) for any disability, physical disability, and non-physical disability, as well as restricted measures of non-physical disabilities (excluding feeling and psychological difficulties, and seeing and hearing difficulties), in the 2021 CSES. The number of observations in the treatment and control groups is shown. The prevalences of disability (proportion) are in parentheses. See ‘Measures’ in Methods for the definitions of the disability measures, and ‘Empirical design’ in Methods for matching designs (within-village matching and within-district matching).

Main analysis: psychological well-being

The histogram plots of subjective well-being and depression at the time of the survey (‘Measures’ in Methods) by disability status in the analysis samples show lower happiness and life satisfaction and higher depression among individuals with any disability than those without disabilities (according to the Kolmogorov-Smirnov test, the distributions are significantly different; Fig. 1). The differences are more pronounced among individuals with physical disabilities; the distributions of subjective well-being for individuals with non-physical disabilities do not differ significantly from those without disabilities.

Fig. 1.

Histogram plots of psychological well-being by disability status. The figure shows the histogram plots of psychological well-being for individuals without disabilities and individuals with any disability (A, D, G), physical disability (B, E, H), and non-physical disability (C, F, I) among household heads and spouses in the analysis samples from the 2021 CSES (constructed in Table 1). See ‘Measures’ in Methods for the definitions of disability and psychological well-being measures. P-values for the Kolmogorov-Smirnov (KS) test for the equality of distributions are reported.

In each analysis sample, we matched each treated individual (with disabilities) with controlled individuals (without disabilities) within villages and districts. Without information about the onset of disability (treatment timing) in the 2021 CSES, we used gender and age as pre-treatment covariates for matching. Although this restrictive matching design precludes causal interpretations, it enables us to better capture the treatment-control difference than an unmatched comparison²⁹. Our main estimates – the Ordinary Least Squares (OLS) estimates in the matched samples (Eqs. 1 and 2, ‘Empirical design’ in Methods) – are reported in Fig. 2 (the matching estimates without covariates controlled for are reported in Supplementary Table S1 and are similar). Compared to the within-village matched samples, the within-district matched samples better represent the analysis samples and, in turn, the (underlying) population, because they have a less trimmed treatment group (about 20% vs. 5% trimmed, Table 1; ‘Empirical design’ in Methods). At the same time, the composition of types of non-physical disabilities in both matched samples is very similar to the analysis sample (Supplementary Fig. S2E and F).

Fig. 2.

Main analysis: associations of disability with psychological well-being. The figure shows the estimated coefficient of the treatment variable (any disability, physical disability, or non-physical disability) in the OLS regression among household heads and spouses in the matched samples from the 2021 CSES (constructed in Table 1). See ‘Measures’ in Methods for the definitions of disability and psychological well-being measures, and ‘Empirical design’ in Methods for matching designs and regression. Strata size is used as weights. For within-village matching (WVM), individual covariates and village fixed effects are controlled for (Eq. 1). For within-district matching (WDM), individual covariates, village covariates, and district fixed effects are controlled for (Eq. 2). The individual covariates are gender, age, and age squared. The village covariates are urban (indicator), population (log), distance to provincial capital (log km), and distance to district capital (log km). The means for individuals without disabilities in the matched sample for any disability are in brackets (those in the matched samples for physical disability and non-physical disability reported in Supplementary Table S1 are similar). 95% confidence intervals based on robust standard errors clustered by household are shown.

Consistent with the descriptive results in the analysis samples, subjective well-being was lower, and depression was higher among individuals with any disability than those without disabilities; the results for happiness and life satisfaction are similar, and the results are similar between the within-village and within-district matching. For example, severe depression (dichotomized measure) among individuals with any disability was more common by about 13% or about 50% more common than among individuals without disabilities. The point estimates for physical disability are larger than those for non-physical disability, although the difference is not statistically significant. Although none of the within-village matching estimates for non-physical disability are statistically significant, the within-district matching estimates are statistically significant except for happiness. These main results for psychological well-being are robust to multiple inference, additional covariates controlled for, omitted variable bias, and migration (Supplementary Note 1, Figs. S3, S4, Table S2).

Mechanisms

It has been widely demonstrated that adverse economic and health shocks can reduce subjective well-being and harm mental health^30,31. Economic and health shocks were significant global events during the COVID-19 pandemic³². If disability exacerbated these shocks specific to the pandemic, the adverse associations of disability with people’s psychological well-being could be augmented through these shocks. We examined whether pandemic-related economic shock (index of nonemployment during the pandemic) and health shock (index of perceived risk of COVID-19 infection at the time of the survey) were linked to the adverse associations of disability with psychological well-being found above (‘Measures’ in Methods).

Preventive health behaviors were the key to reducing actual infection risk^33–35. Our measure of perceived infection risk captures people’s perception updated through their pandemic experience. We examined whether preventive health behaviors (index) during the pandemic before the survey were related to the associations of disability with health shock (ex-post perceived risk of infection), as well as psychological well-being, at the time of the survey (i.e., serial mediation through preventive behaviors) (‘Measures’ in Methods). This pathway, which builds on the timeline of these measures in our data, is the reverse of the pathway from ex-ante perceived infection risk (vulnerability) to preventive behaviors studied in the literature^36–39.

These mediation paths are shown in Fig. 3, where the treatment (disability), mediators, and outcome are placed along the timeline: before the pandemic, during the pandemic, and at the time of the survey. To examine these pathways, we employed structural equation modeling (SEM) for the corresponding four-equation system, which models economic shock, health shock, preventive behaviors, and psychological well-being (Eqs. 3.1–3.4, ‘Empirical design’ in Methods)^40,41. To obtain credible estimates of the associations in these pathways, we used the same matched samples as in the main analysis: within-village matched samples for any disability and physical disability (with similar results under the within-district matching) and within-district matched samples for non-physical disability (which yielded mostly statistically significant results above).

Fig. 3.

Mediation model. The figure is the path diagram for the mediation analysis (see ‘Empirical design’ in Methods). The coefficients of the four-equation system (Eqs. 3.1–3.4) corresponding to each path are shown. The treatment (disability), mediators, and outcome are placed along the timeline: before the pandemic (disability), during the pandemic (preventive behaviors, economic shock), and at the time of the survey (health shock, outcome).

The estimated coefficients are shown in Fig. 4. The patterns are mostly similar across psychological well-being measures. Compared to those without disabilities, economic shock was greater among individuals with any, physical, and non-physical disability, health shock was greater among those with non-physical disability (p < 0.1), and preventive behaviors were less common among those with any and non-physical disability. All individual preventive behaviors, based on which the index was constructed, were less common for those with non-physical disability, albeit statistically weak; this is especially so for mask-wearing and staying at home; the results for observed and self-reported mask-wearing are similar (Supplementary Fig. S5).

Fig. 4.

Mediation analysis: coefficients, direct, and total effects. The figure shows the estimated coefficients of the four-equation system using SEM shown in Fig. 3 (see ‘Empirical design’ in Methods for mediation analysis): (1) those of the treatment variable () – any disability, physical disability, or non-physical disability – in Eq. 3.1 for economic shock; (2) those of the preventive behaviors () and treatment variable (direct effect) () in Eq. 3.2 for health shock; (3) those of the treatment variable () in Eq. 3.3 for preventive behaviors; and (4) those of the three mediators – economic shock (), health shock (), and preventive behaviors () – and the treatment variable (direct effect) () in Eq. 3.4 for each outcome (happiness, life satisfaction, depression severity, and severe depression). The figure also shows the estimated total effects of the treatment variables, which are the sum of the direct effect and the indirect effects reported in Fig. 5. The sample consists of household heads and spouses in the matched samples from the 2021 CSES (constructed in Table 1). See ‘Measures’ in Methods for the definitions of disability measures and mediators, and ‘Empirical design’ in Methods for matching designs: within-village matching (WVM) and within-district matching (WDM). Strata size is used as weights. Individual covariates, village covariates, and province fixed effects are controlled for. See the notes to Fig. 2 for the individual and village covariates. 95% confidence intervals based on robust standard errors clustered by household are shown.

The adverse associations of the mediators with psychological well-being were distinct between physical and non-physical disabilities: for physical disability, the associations were significant for economic shock, not health shock; for non-physical disability, the associations were significant for health shock, not economic shock. For non-physical disability, preventive behaviors were favorably associated with life satisfaction and depression severity for health shock, not economic shock, and predicted lower health shock (ex-post risk perception).

The estimated adverse indirect effects of mediators reported in Fig. 5 are consistent; the estimated direct and total effects of disability are shown in Fig. 4. On the one hand, economic shock was a significant mediator for physical disability. For example, the indirect effect of economic shock (-0.18) accounted for over a quarter of the total effect (-0.68) for happiness. On the other hand, health shock was a mediator for non-physical disability. For example, the indirect effect of health shock (0.019) accounted for 11% of the total effect (0.18) for depression severity (p < 0.1). Preventive behaviors further mediated the health-shock pathway for non-physical disabilities, accounting for 17% of the total effect (0.19). The results for any disability combine these two sets of results for physical and non-physical disabilities.

Fig. 5.

Mediation analysis: indirect effects. The figure shows the estimated indirect effects of preventive behaviors on health shock () and those of economic shock, health shock, and preventive behaviors on each outcome (happiness, life satisfaction, depression severity, and severe depression) (, , and , respectively) based on the estimation results of the four-equation system (Eqs. 3.1–3.4) using SEM shown in Fig. 3 (see ‘Empirical design’ in Methods for mediation analysis). The estimated coefficients, along with the direct and total effects, are reported in Fig. 4. The sample consists of household heads and spouses in the matched samples from the 2021 CSES (constructed in Table 1). See ‘Measures’ in Methods for the definitions of disability measures and mediators, and ‘Empirical design’ in Methods for matching designs: within-village matching (WVM) and within-district matching (WDM). Strata size is used as weights. Individual covariates, village covariates, and province fixed effects are controlled for. See the notes to Fig. 2 for the individual and village covariates. 95% confidence intervals based on robust standard errors clustered by household are shown.

Heterogeneity analysis

Interpreting the main results for the aggregate measure of non-physical disability, combining sensory and mental disabilities, is not straightforward, and it may mask heterogeneity across different types of non-physical disabilities. We examined the association of psychological well-being and mediators with seeing, hearing, speaking, feeling, and psychological difficulties (with a prevalence of at least 10%, Supplementary Fig. S2D) in the within-district matched sample for non-physical disabilities (Supplementary Figs. S6 and S7). Whereas the results for seeing difficulties (the most common type) are largely similar to those for aggregate non-physical disability, distinct patterns appear for the remaining types as follows:

1. Both happiness and life satisfaction were considerably lower, and severe depression was relatively more common with psychological and feeling difficulties.

2. Economic shock was significantly larger with psychological, feeling, and speaking difficulties; health shock was considerably larger with psychological and feeling difficulties, albeit statistically weak.

3. Overall preventive behaviors (index) were less common with psychological difficulties. In particular, social distancing, staying at home, and avoiding gatherings were less common with psychological difficulties; social distancing and staying at home were less common with speaking difficulties; and mask-wearing and handwashing were less common with feeling difficulties.

4. Almost no significant differences appeared for hearing difficulties.

We repeated the main analysis of psychological well-being using two restricted measures of non-physical disabilities: (1) excluding feeling and psychological difficulties (for which we found strong heterogeneity), and (2) focusing on seeing and hearing difficulties (two major sensory disabilities). The construction of analysis and matched samples for these two measures is analogous (Table 1B and C). The estimation results based on the within-district matching become weaker compared to the original ones: for the first restricted measure, the point estimates for life satisfaction, depression severity, and severe depression decrease in a small magnitude, while maintaining statistical significance; for the second restricted measure, they further decrease, and the first two estimates lose statistical significance (Supplementary Table S1D and E).

Discussion

Associations of disability with psychological well-being

Disability was adversely associated with psychological well-being. This might have been more so for physical disability than for non-physical disability, which might reflect the difference in their severity. The results are consistent across happiness, life satisfaction, and depression, as well as between within-village and within-district matching, except for non-physical disability. Our findings are consistent with previous work before the COVID-19 pandemic^24,25,42. At the same time, our results are distinct from a previous quasi-experimental study in Cambodia that found no significant difference in happiness and life satisfaction between adults with landmine amputation and those without disabilities in a heavily mined province (Banteay Meanchey)²⁷; that is, amputation (severe physical disability) did not adversely affect adults’ subjective well-being before the pandemic. Although the sample of this previous work is not comparable to our nationally representative sample, our findings suggest that disabled people’s psychological well-being could become worse during the pandemic.

Economic and health shocks as mechanisms

The mediation results are consistent with the idea that economic and health shocks led to the adverse associations of disability with psychological well-being during the pandemic. Compared to people without disabilities, disabled people, both physical and non-physical, experienced a stronger economic shock (nonemployment). Economic shock was associated with lower psychological well-being among people with physical disability, partly because they lost employment during the pandemic. People with non-physical disability also experienced stronger health shock (perceived risk of infection) than people without disabilities, and health shock was associated with lower psychological well-being. This was likely partly because people with non-physical disability less commonly adopted preventive behaviors, which predicted lower risk perception (ex post), and this experience led to an increase in their perceived vulnerability. Overall, economic shock and health shock, serially mediated by preventive behaviors, showed distinct pathways for physical and non-physical disabilities, respectively.

Heterogeneity across non-physical disabilities

The results for aggregate non-physical disability, mainly driven by seeing difficulties, masked stronger patterns for other underrepresented disabilities. According to the descriptive analysis, psychological well-being was lower among individuals with feeling/psychological difficulties, though the strong association of psychological difficulties with depression might reflect the conceptual overlap of these two measures. Feeling/psychological difficulties were associated with a larger economic shock. Feeling (or sensing) difficulties prevented mask-wearing and handwashing, probably due to sensory or physiological constraints. Psychological difficulties prevented people from social distancing, staying at home, and avoiding gatherings, perhaps because it was difficult for them to practice these behaviors to secure the necessary help⁴³. These consistent patterns bolster the role of preventive behaviors constrained by non-physical disabilities as an underlying mechanism. The main analysis using the restricted measures of non-physical disabilities indicates the stronger associations of underrepresented disabilities, such as feeling/psychological difficulties, with psychological well-being, compared to seeing/hearing difficulties. At the same time, the results indicate that the original results were not primarily driven by feeling/psychological difficulties.

Limitations

The following limitations of our study are noted.

1. Using the cross-sectional CSES data, we cannot directly examine how the COVID-19 pandemic altered the associations of disability with psychological well-being; instead, we exploited pandemic-related shocks to provide indirect evidence.

2. The lack of information about disability onset restricted our matching design based on a few pre-treatment covariates, precluding causal interpretations. Since this design could weaken the representativeness of the estimates, especially for within-village matching, interpreting the generalizability of our findings also requires caution.

3. Because our economic shock measure does not necessarily capture employment loss caused by the pandemic (‘Measures’ in Methods), we cannot disentangle the extent to which its mediation effect was attributable to the pandemic; the mediation results may reflect the broad economic shock people experienced during the pandemic rather than pandemic-specific effects alone.

4. The mediation results of health shock for non-physical disability are not statistically strong, albeit marginally significant (p < 0.1). This can be because risk perception captured a relatively small share of the broad health shock specific to the pandemic. This can be also because the mediation results were mainly driven by underrepresented non-physical disabilities. With a small number of individuals with each type of non-physical disabilities (but seeing difficulties), the heterogeneity analysis for non-physical disabilities also has weak statistical power.

5. Although the available information about self-reported disability in the 2021 CSES is standard in national surveys⁴⁴(‘Measures’ in Methods), there is no standard way to construct disability measures for functioning difficulties not due to aging, and our measures might be noisy. In particular, our sample trimming could exclude some individuals with functioning difficulties not due to aging if they did not report real causes of disability other than aging (due to ignorance or stigma, for example) (‘Empirical design’ in Methods). This might weaken the efficiency of our estimates, especially for non-physical disabilities.

6. Our standard self-reported measures of psychological well-being (outcomes) and economic/health shocks and preventive behaviors (mediators) may contain measurement errors due to desirability bias, incomplete recall, stigma, and so forth. If measurement errors are uncorrelated with disability, they do not cause bias in our estimates, although they can reduce the efficiency of the estimates. In this case, our statistical inference can be considered conservative. Although measurement errors correlated with disability can cause bias, controlling for interviewer fixed effects through village fixed effects should help reduce such potential bias. Regarding self-reported preventive behaviors, similar results for observed and self-reported mask-wearing indicate that measurement errors were unlikely to be a significant concern.

Conclusion

Using nationally representative data in Cambodia during the COVID-19 pandemic, our study revealed that disability not due to aging was associated with lower subjective well-being and higher depression among working-age adults. These adverse associations tended to be stronger for physical disability than non-physical disability and were linked to pandemic-related economic and health shocks (nonemployment and ex post perceived risk of infection), respectively. Although preventive health behaviors could reduce perceived vulnerability, they were less common among people with non-physical disabilities, which may partly explain the associations among disability, higher perceived infection risk, and lower psychological well-being. At the same time, these results could mask the strong adverse patterns of underrepresented non-physical disabilities, such as feeling and psychological difficulties. Although caution about the limitations of the mediation analysis discussed above is needed, these results suggest that the pandemic might have worsened disabled adults’ well-being, as cautioned by researchers^9,10. The potentially intense vulnerability to the pandemic among disabled people in the Global South highlights the criticality of disability-inclusive policy response to sustain their psychological well-being. Our findings suggest the following implications for future pandemics.

First, for disability-inclusive policymaking, it is crucial to capture representative situations of people with disabilities and their heterogeneity. Although these can, in principle, be obtained from nationally representative data, national survey data rarely contain information about psychological well-being and pandemic-related shocks and behaviors, especially health-related ones. Our study, in collaboration with the 2021 CSES national statistics, exemplifies the utility of such information. At the same time, as people with underrepresented disabilities in national surveys, such as those with mental disabilities, can be particularly vulnerable, supplementary data for underrepresented groups is needed. When combined with broad health shock data, such data can also help better capture the mediating role of health shocks.

Second, disability-inclusive policies should address economic and health shocks as underlying mechanisms during both pandemic and non-pandemic periods. To tackle economic shocks, on the one hand, social protection programs, such as cash and food transfers tailored to people with disabilities, are needed^3,45. Whether such programs should be targeted toward specific disabilities depends on particular contexts. In Cambodia, people with physical disabilities and feeling/psychological difficulties can be a primary target. To strengthen the economic inclusion and resilience of people with disabilities, income-generating programs such as vocational training are called for, especially during non-pandemic periods. A randomized field experiment on vocational training for people with physical disabilities in Cambodia has shown its substantial positive impacts on labor-market outcomes before the COVID-19 pandemic²⁸. Similar findings have been obtained in a quasi-experimental study on vocational training for disabled ex-combatants in Rwanda⁴⁶.

To address health shocks, on the other hand, careful support for people with different non-physical disabilities is needed as safety nets and to enhance resilience, so they can better adjust or substitute preventive behaviors, such as mask-wearing, handwashing, and social distancing, and manage the perceived vulnerability during pandemics. Detailed supplementary data on underrepresented disabilities discussed above are crucial to designing specific interventions and programs. More broadly, more empirical work is needed to better understand vulnerable populations in the Global South during the COVID-19 pandemic^15–17, which is foundational for effective, inclusive, and equitable policymaking to sustain their well-being in the face of future pandemics.

Methods

Study setting

Cambodia is one of the poorest countries in Southeast Asia. Cambodia’s history has been one of civil conflict and instability for much of the last 50 years. The Cambodian conflict resulted in thousands of amputees^47,48; Cambodia is often considered the country with the highest prevalence of amputees in the world. In 2007, the Government of Cambodia signed the Convention on the Rights of Persons with Disabilities, which was ratified in 2012, and introduced an inaugural national disability law, the Law on the Protection and the Promotion of the Rights of Persons with Disabilities⁴⁹. It is estimated that the direct cost of disability is 19% of monthly household expenditure, and only 7% of the cost is covered by income support from the government (about 5 USD per month) and family sources⁵⁰. Constrained access to health care for disabled people has been a problem⁵¹. A high prevalence of psychological distress and post-traumatic stress disorder (PTSD) has been found among people with physical disabilities⁵².

Cambodia detected the first COVID-19 case on January 27, 2020. The epidemiological phases of Cambodia consisted of containment (January 27, 2020 – February 19, 2021), mitigation (February 20, 2021 – October 31, 2021), and full reopening (November 1, 2021-)⁵³. The government implemented various measures, such as quarantine, face coverings, hand hygiene, physical distancing, school closures, and border restrictions/closures, in containment phase 1, successfully containing the spread of the COVID-19 virus (no deaths due to the virus in phase 1). It experienced a large community outbreak at the beginning of mitigation phase 2, followed by a country-wide lockdown between April 15 and May 15, 2021 (2781 deaths in phase 2). After the rollout of the COVID-19 vaccination started on February 10, 2021, high vaccination coverage was quickly attained in phase 2 (93% coverage by September 15, 2021)^53–55. Since June 2020, the government has implemented the COVID-19 Cash Transfer Program for ID Poor Households, covering 19% of the total population (ID Poor is an official identification of poor households)⁵⁶.

Data

The CSES is a cross-sectional survey conducted by the National Institute of Statistics (NIS) of the Ministry of Planning in Cambodia. As designated official statistics, CSES provides essential information about living conditions in the country, such as poverty and labor force participation, which is widely used for policymaking. We collaborated with the NIS to design and implement the module on psychological well-being and COVID-19-related shocks in the 2021 CSES. The sampling frame was the register of villages and enumeration areas (EAs) in the 2019 General Population Census of Cambodia. Following the stratified sampling design, the 2021 CSES sampled 10,080 households in 1008 primary sampling units (PSUs, villages; i.e., 10 households per village) in 24 provinces and one capital, Phnom Penh (including 202 districts)⁵⁷. The sample is nationally representative. In-person interviews with the head of household (or the head’s spouse if the head was unavailable) were conducted at their homes from February 2021 to January 2022, which mainly corresponded to mitigation phase 2 during the pandemic period (84 PSUs per month). Non-response was nonexistent.

Measures

Exposure: disability

Consistent with the WHO International Classification of Functioning, Disability, and Health (ICF)⁵⁸, the 2021 CSES asked a question about self-reported functioning limitations for each household member: “Does ..[NAME].. have any of the following? [Enter the 3 most important] 1. Seeing difficulties; 2. Hearing difficulties; 3. Speaking difficulties; 4. Moving difficulties; 5. Feeling difficulties; 6. Psychological difficulties (strange behavior); 7. Learning difficulties; 8. Fits; 9. Other (specify); 10. Don´t know”. Feeling (or sensing) difficulties include numbness and hypersensitivity. For each functioning limitation, the survey also asked the difficulty level (mild, moderate, or severe) and its cause (19 options described in the notes to Supplementary Fig. S1, such as disease and old age). These questions differ from the Washington Group Short Set (WGSS) questionnaire, which captures six functional domains (seeing, hearing, walking, remembering/concentrating, self-care, communicating), used in the Cambodia Demographic and Health Survey (DHS) in 2021–2022, for example⁵⁹. We constructed three indicator measures, which are equal to 1 if the individual had any functioning difficulties (any disability), moving difficulties (physical disability), and any functioning difficulties other than moving difficulties (non-physical disability). Physical and non-physical disabilities are not mutually exclusive. This aggregate measure of non-physical disability captures both sensory and mental disabilities.

Outcome: psychological well-being

The survey collected information about the psychological well-being of the head of household and his/her spouse (if any) by asking three questions: “Taking all things together, how would you describe your degree of happiness? 1 = very unhappy 2 3 4 5 6 7 8 9 10 = very happy”; “All things considered, how satisfied are you with your life as a whole these days? 1 = totally dissatisfied 2 3 4 5 6 7 8 9 10 = totally satisfied”; and “How much did you experience the following over the past week?: worrying, difficulty sleeping, and low energy. 1 = not at all, 2 = a little, 3 = quite a bit, 4 = extremely”. We constructed ordered measures for happiness (10 scales), life satisfaction (10 scales), and depression (4 scales), as well as an indicator measure for severe depression (depression = 3 or 4) for each individual.

Mediator: COVID-19-related economic and health shocks

The following question was asked of the head of household: “Since March 2020, have you lost your job, stopped working, or stopped earning money? 1 = Yes, permanently, 2 = Yes, temporarily, 3 = No.” We constructed an indicator measure for any employment loss (permanent or temporary) during the pandemic – both the containment phase 1 and the mitigation phase 2 – for each household. The survey also asked questions about employment during the past 7 days – in phase 2 – for each household member (aged 5 years or older). We constructed an indicator measure for paid employment (including self-employment) for each individual. The combination of these two measures – experience of employment loss during the pandemic prior to the survey and employment status at the time of the survey – serves as a proxy for pandemic-related economic shocks, including temporal variation, though it does not necessarily capture employment loss caused by the pandemic. For example, among 228 household heads and 617 individuals without disabilities in the matched sample for any disability (Table 1B), 47.6% and 86.1% experienced employment loss and had paid employment, respectively (the proportions in the matched samples for physical and non-physical disability are similar).

We focused on pandemic-related health shocks captured by perceived risk of infection, using the following two questions asked of the head of household and his/her spouse (if any): “What do you think is the percent chance that you will be infected with COVID-19 during the next 12 months?”; and “What percent of people in this village do you think will be infected with COVID-19 in the next 12 months?” We constructed two proportional measures for own and community infection risks, which capture an individual’s perceived risk of infection at the time of the survey. For example, among 617 individuals without disabilities in the matched sample for any disability (Table 1B), the own and community infection risks were 7.6% and 9.0%, respectively, at means (the means in the matched samples for physical and non-physical disability are similar).

The following question about preventive health behaviors practiced during the pandemic before the survey was asked of the head of household: “Since March 2020, how have members of your household taken the following precautionary measures to prevent the infection of COVID-19? 1 = Never, 2 = Sometimes, 3 = Frequently. a. Use of mask, b. Handwashing, c. Keep social distance (at least 1 meter), d. Staying at home and avoid going out unless necessary, e. Avoid gatherings with many people.” We constructed five indicator measures for frequently adopting mask-wearing, handwashing, social distancing, staying at home, and avoiding gatherings. Additionally, interviewers recorded whether respondents wore a mask during interviews. We compared self-reported mask-wearing with this observed measure for a robustness check.

We constructed three indices (z-scores) for economic shock, health shock, and overall preventive behaviors by taking the first principal component of the two, two, and five corresponding individual measures (with inverse employment); the principal component analysis for economic/health shock and preventive behaviors was done among individuals and household heads, respectively, in the matched sample for any disability constructed below (Table 1A).

Empirical design

Sample trimming

Our analysis focused on heads of households and their spouses for whom the outcome measures (psychological well-being) are available. We focused on individuals’ own disabilities, without considering the disabilities of other household members. We considered people with any/physical/non-physical disability as treated individuals and people without disabilities as controlled individuals. When we examined physical and non-physical disability, we excluded people with non-physical disability only and people with physical disability only, respectively, from the sample. The prevalence of any, physical, and non-physical disability in these base samples was 9.3%, 3.2%, and 7.8%, respectively (Table 1A). Among individuals with non-physical disabilities, 85% and 28% had seeing and hearing difficulties, respectively (Supplementary Fig. S2A).

To construct disability measures for functioning difficulties not due to aging, we trimmed the base samples as follows (Table 1A). First, we focused on working-age adults (ages 20–59). The prevalence of any, physical, and non-physical disability significantly decreased (4.2%, 1.4%, and 3.2%, respectively), indicating the high prevalence of disabilities among the elderly. This was especially so for hearing difficulties among non-physical disabilities (Supplementary Fig. S2B). Second, we focused on moderate or severe functioning difficulties by excluding individuals who reported mild difficulties only (we did so for any, physical, and non-physical disability separately). It has been shown that measures based on severe difficulties are more reliably self-reported and less prone to measurement errors⁶⁰. The prevalence of any, physical, and non-physical disability further decreased (2.6%, 1.1%, and 1.8%, respectively); this was especially evident for seeing difficulties (Supplementary Fig. S2C). In this way, we excluded mild seeing difficulties typically corrected by wearing glasses, for example⁵⁰. Third, we dropped individuals with functioning difficulties caused solely by aging (we did so for any, physical, and non-physical disability separately). The prevalence of non-physical disability further decreased (1.9%, 1.1%, and 1.1%, respectively), indicating a high prevalence of non-physical disabilities caused by aging; again, this was especially notable for seeing difficulties (Supplementary Fig. S2D). In particular, among individuals with any disability caused solely by aging, 92% had moderate seeing or hearing difficulties (mostly seeing), and only 3% had severe ones. In this way, we excluded farsightedness due to aging (presbyopia) and hard-of-hearing due to aging, for example. The resulting analysis sample for any disability for individual-level measures consists of 14,285 individuals (79.2% of the base sample; 7578 heads of households, 6707 spouses), including 276 individuals and 192 household heads with any disability in the treatment group (1.9% and 2.5%, respectively); the analysis sample for household-level measures consists of 7578 households.

Matching

Without information about the onset of disability (treatment timing) in the 2021 CSES, we used gender and age as pre-treatment covariates for matching. Gender and age are likely to be correlated with psychological well-being. Gender and age were strong predictors of disabilities in the analysis samples, regardless of whether village fixed effects (dummies) were controlled for (Supplementary Table S3). Disabled males were more common than disabled females; this is especially so for physical disabilities, reflecting the post-conflict setting. The propensity to be disabled increased with age in an increasing manner. Whereas the age distribution of individuals without disabilities shows twin peaks at around age 35 and age 50, reflecting the demographic composition affected by the genocide under the Khmer Rouge⁶¹, the majority of individuals with disabilities were ages 45–60 (Supplementary Fig. S8)²⁷.

For each of the three analysis samples for any, physical, and non-physical disability, we matched each treated individual with controlled individuals within villages. Within-village matching (i.e., spatial blocking) helps obtain credible estimates of the associations for physical disabilities due to conflicts, especially landmines, because it fully controls for landmine contamination within the village, which determines local vulnerability to landmine accidents²⁷. Fully controlling for village heterogeneity that affects psychological well-being, within-village matching can also help better capture the associations for physical and non-physical disabilities caused by disease and other reasons. Specifically, we employed exact matching on gender, four age cohorts (ages 20–29, 30–39, 40–49, and 50–59), and village fixed effects (which subsume survey-month fixed effects and interviewer fixed effects). Given the small size of the strata defined by the matching variables and village fixed effects (e.g., 6406 strata in total in the analysis sample for any disability), we allowed different numbers of treated and controlled individuals within strata. The resulting matched samples include 225, 128, and 118 treated individuals (about 82%, 84%, and 79% of the analysis samples) and 417, 260, and 200 controlled individuals (in 215, 126, and 111 matched strata) for any, physical, and non-physical disability, respectively (Table 1B). Dropping unmatched strata, these trimmed samples do not represent the populations that the analysis samples represent. Age is much more balanced in the matched samples than in the unmatched samples (Supplementary Table S4, Fig. S8). Our matching analysis is close to matched case-control studies on disability⁴². For household-level measures, we conducted matching among household heads (Table 1B).

We obtained the difference in weighted means of outcomes (psychological well-being) between the treatment and control groups in the matched sample, using strata size as weights. This corresponds to the estimated coefficient of the treatment variable (an indicator measure for disability) in the weighted univariate regression⁶². Specifically, the weight for the treatment group is equal to one, and the weight for the control group is equal to , where and are the total number of matched observations of the treatment and control groups, respectively, and and are the number of matched observations of the treatment and control groups, respectively, in stratum s. Inference is based on robust standard errors clustered by household. This would be an unbiased estimator for the average treatment effect on the treated if the treatment (the onset of the disability) is independent of potential outcomes between treated and controlled individuals in the matched sample⁶³. Given our restrictive matching design based on a few pre-treatment covariates, we do not assume that this conditional independence (unconfoundedness) holds.

To reduce the bias, particularly due to the remaining imbalance in age, in the matched sample, we employed an OLS regression specification of the form,

1

where Outcome_ihv is the psychological well-being of individual i of household h in village v; Disability_i is an indicator measure for disability (any/physical/non-physical) (treatment variable) of individual i; X_i is a vector of individual covariates; θ_v is a vector of village fixed effects; and ε_ihv is an error term. Strata size is used as weights. The individual covariates are gender, age, and age squared. Among 642 individuals in the matched sample for any disability, for example, 45.0% were female, and the mean age was 48.0.

We matched treated and controlled individuals within districts as an alternative matching design. An advantage of within-district matching over within-village matching is that with a larger control group within districts, we can have a less trimmed treatment group that better represents the analysis sample while using finer age cohorts for better balance. Specifically, we employed exact matching on gender, eight age cohorts (ages 20–24, 25–29, 30–34, 35–39, 40–44, 45–49, 50–54, and 55–59), and district fixed effects (e.g., 4810 strata in total in the analysis sample for any disability). The resulting matched samples comprise 263, 144, and 143 treated individuals (about 95% of the analysis samples) and 1218, 734, and 645 controlled individuals (in 231, 137, and 127 matched strata) for any, physical, and non-physical disability, respectively (Table 1C). Age is more balanced than in the within-village matched samples (Supplementary Table S4, Fig. S8).

We employed an OLS regression specification of the form,

2

where Outcome_ihvd is the psychological well-being of individual i of household h in village v of district d; X_v is a vector of village covariates; ϕ_d is a vector of district fixed effects; and ε_ihvd is an error term; all other variables are the same as those defined above. Compared to Eq. 1, we additionally controlled village covariates within districts. The village covariates are urban (indicator), population (log), distance to provincial capital (log km), and distance to district capital (log km). In the matched sample for any disability, for example, among 1481 individuals, 44.2% were female, and the mean age was 48.2; among 576 villages, 36.3% were urban, the mean population was 1751, the mean distance to provincial capital was 39.4 km, and the mean distance to district capital was 12.7 km.

Mechanisms

Using the same matched samples as in the main analysis, the mediation model in Fig. 3 leads to the following four-equation system^40,41:

3.1

3.2

3.3

3.4

where Outcome_ihvp is the psychological well-being of individual i of household h in village v of province p; Economic shock_i, Health shock_i, and Preventive behaviors_h are potential mediators (z-score); φ_p is a vector of province fixed effects; and ε_ihvp is an error term; all other variables are the same as those defined above. Although we used the same matched samples as in the main analysis, we employed province fixed effects, rather than village or district fixed effects, to ensure common support for each mediator index (some mediators lacked variation within some villages or districts in the matched samples). The base results with province fixed effects controlled for are almost the same as the original results with village or district fixed effects controlled for, reported in Fig. 2 (Supplementary Fig. S3). We estimated Eqs. 3.1–3.4 jointly using SEM. Estimation is based on maximum likelihood, and inference is based on robust standard errors clustered by household. The indirect effect of preventive behaviors on health shock is ; those of economic shock, health shock, and preventive behaviors on outcome are , , and , respectively. The direct effects of disability on health shock and outcome are and , respectively. The total effects of disability on health shock and outcome are and , respectively.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization: Y.T. Investigation: Y.T., K.K., M.O. Methodology: Y.T., K.K. Formal analysis: Y.T. Visualization: Y.T. Funding acquisition: Y.T., K.K., M.O. Project administration: Y.T., K.K., M.O. Supervision: Y.T., K.K., M.O. Writing – original draft: Y.T. Writing – review & editing: Y.T., K.K., M.O.

Funding

This study was supported by grants from the Japan Society for the Promotion of Science (18H05312; 20K20332; 19H00590; 20K01610; 21K13676).

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethics declarations

As Cambodia’s national statistical office, the National Institute of Statistics conducted CSES as designated official statistics under the Statistics Law. As the Law (Article 33) requires respondents to participate in CSES, informed consent was effectively obtained from all subjects. The analysis of this study was approved by the Institutional Review Board of the University of Tokyo (#23–488) and performed in accordance with relevant guidelines and regulations.