No going back: COVID-19 disease threat perception and male migrants' willingness to return to work in India

Abstract

This paper explores the causal link between the likelihood of re-migration to cities and the perceived threat of contracting COVID-19 using novel data on male reverse migrant workers in India. We find that reverse-migrants who believe there is a significant chance of contracting COVID-19 display a significantly lower likelihood of returning to their urban workplaces, regardless of their duration of migration. On the other hand, longer-duration migrants display a lower perceived chance of contracting COVID-19 than shorter-duration migrants. We also contribute to the migration literature by linking behavioural attributes to the decision to migrate. We find that more impatient individuals display a heightened belief regarding contracting COVID-19 and a higher projected likelihood of returning to work. Finally, we find that while both loss and risk-averse individuals have a lower projected likelihood of returning to urban workplaces, only loss-averse individuals perceive that their chance of contracting COVID-19 is lower.

1. Introduction

The first wave of the COVID-19 pandemic in India triggered a lockdown that had a severe impact on migrant workers who work primarily in the informal sector of Indian urban centres and prompted a reverse migration to their villages of origin (Andrade, 2020; Mukhra et al., 2020; Dandekar and Ghai, 2020; Mander, 2020; Srivastava, 2020b; Jitendra, 2020; Sengupta and Jha, 2020; Jesline et al., 2021). Barker et al. (2020) and Upadhyay et al. (2020) find that rural locations and small towns with a high incidence of outmigration suffered disproportionately as a result of the COVID-19 crisis and the ensuing reverse migration.

Given the importance of remittances to the rural economy (Barker et al., 2020) and the enormous impact of migrant enterprise and work on the economies of Indian urban centres (Mahendra Dev and Sengupta, 2020), it is important to study the factors that determine the projected likelihood of return by such reverse migrants to their original places of work after the relaxation of lockdown norms. However, the COVID-19 threat continued to persist post lockdown, particularly in urban India, making the return to urban centres a potential health hazard (Andrade, 2020; Mishra et al., 2020). Thus, we hypothesize that the perceived threat of contracting the illness would critically affect this re-migration decision. Moreover, behavioural factors like attitudes to uncertainty and impatience can also contribute to migration decisions but are traditionally understudied in the migration literature (Banerjee and Duflo, 2020; Goldbach and Schluter, 2018). Using responses from a telephonic survey conducted with male reverse migrants in April-May 2020, we study if the projected likelihood of return for a migrant is causally dependent on the perceived chance of contracting COVID-19. Further, we investigate empirically how these likelihoods are determined by factors such as duration of migration, demographic characteristics, behavioral factors, and access to information and resources.

As we use data collected using a survey questionnaire, our two main variables of interest: the perceived chance of contracting COVID-19 and the likelihood of return, maybe endogenously determined due to the presence of random measurement errors that may bias our simple OLS estimates downwards (Bound et al., 2001). Therefore, we employ a two-stage least squares (2SLS) regression framework and use the number of sources of information on COVID-19 and recall of COVID-19 prevention strategies as instrumental variables (IVs) for the chance of contracting COVID-19. From our 2SLS estimation, we find that reverse-migrants who believe there is a significant chance of contracting COVID-19 display a significantly lower likelihood of returning to their urban workplaces, regardless of their duration of migration. On the other hand, longer-duration migrants display a lower perceived chance of contracting COVID-19 as compared to shorter-duration migrants. Having a higher recall of prevention strategies is associated with a lower perceived probability of contracting the virus. Overall, better-established individuals, i.e., landowners and individuals with larger families, perceive a higher threat of contracting COVID-19.

We also contribute to the migration literature by linking behavioral attributes to the decision of migration as "measurement of subjective phenomena - attitudes, values, perceptions, intentions - could play an important role in gaining a deeper understanding of migration choices and behavior" (Fawcett 1985, p.9). We find that more impatient individuals display a heightened belief regarding contracting COVID-19 and a higher projected likelihood of return to work. Finally, we find that loss-averse individuals perceive that their chance of contracting COVID-19 is lower, while both loss and risk-averse individuals have a lower projected likelihood of returning to city.

The main policy implication of our finding is that central and state governments need to continue to implement and regulate COVID-19 safety features in urban workplaces. Moreover, they need to implement programs and interventions aimed at mitigating the perceived risk of COVID-19 contraction among migrant workers, such as targeted information campaigns, vaccination drives, adoption of preventive health behaviors, improving health care facilities and providing health insurance for COVID-19 for migrant workers. The remainder of this paper is organized as follows. Section 2 describes our novel data, and our empirical strategy. In Section 3, we present results from our multivariate analysis. Section 4 details certain robustness measures we undertake to establish our findings, and Section 5 discusses the potential policy directions based on our findings.

2. Data and variables of interest

Data used in this study come from individual telephonic surveys with male migrants from three states in India – Bihar, Uttar Pradesh (UP), and West Bengal (WB).1 This mode of data collection is becoming increasingly popular and has been previously used by Maffioli (2020) to survey respondents during an Ebola epidemic in Liberia in 2014 and more recently by Allen et al. (2021) to study the economic and educational impact of COVID-19 in Mozambique.2

Over April and May 2020, we collect data in two phases: 24th April −7th May (phase 1) and 8th May to 31st May (phase 2), yielding 495 total responses.3 Our sample comprises entirely of males. The decision to use only males is adopted to make both data collection and inference easier for us. First, males constitute the overwhelming bulk of migrant workers in India (Census, 2011). Furthermore, particularly among poorer households in rural India, cell phone ownership is mostly by males who also get primary access if there is only mobile phone to be shared among household members (Mohan et al., 2020; Scott et al., 2021). Thus, we realized that attempting to find a gender-balanced sample to contact by cell phone would be difficult and potentially lead to self-selection biases, where the women in our sample may be socioeconomically distinct from their male counterparts.

Our dataset and sample are novel on two counts: First, we collect information from a sample of migrant workers that returned to their place of origin following the announcement of lockdown to contain an outbreak of a disease for which there was no vaccine or cure at the time. Since our surveying began approximately a month after when the lockdown was declared, our data also allow us to study the effects of variables such as access to sources of information or relevant COVID-19 knowledge on a livelihood decision of paramount importance for a newly exposed population.4 Second, in addition to COVID-19-specific data, we also systematically collect detailed individual-level information from our sample of migrants that includes data on social networks and behavioural attributes, which are not available in current public datasets. Both of these factors would aid in framing more effective and potentially targeted information dissemination policies in the case of other future outbreaks. Deconstructing and understanding the behavior of individuals early on in a new outbreak situation is important. Dasgupta et al. (2020) observe that countries such as Taiwan and Hong Kong who, in the early 2000s had outbreaks of SARS (Severe Acute Respiratory Syndrome), used learnings from these to better tackle the COVID-19 outbreak than countries who had no recent experience with epidemics.

2.1. Empirical strategy

Our empirical strategy aims to explain the average likelihood of migrants' willingness to return to the city [binary variable: no (0), yes (1)] as a function of their perceived risk of COVID-19 infection [binary variable: zero risk (0), positive risk (1)], duration of migration, socio-demographic, behavioural, informational and social capital related factors.5 In the literature, migrants have been classified by their duration of migration into short duration, seasonal or temporary migrants, semi-permanent migrants and permanent migrants, depending on the amount of time they spend in a year at their work location away from their native town or village (Korra 2011; Keshri and Bhagat 2013; Panda and Mishra 2018; Dandekar and Ghai 2020). In general, shorter-term migrants outnumber those migrating for longer periods and are more economically and socially vulnerable (Banerjee and Duflo 2007; Korra 2011; Panda and Mishra 2018; Srivastava 2020a). Furthermore, there is often less emphasis on understanding short-term migrants and their behavior (Coffey et al., 2015 is a notable exception). For this study, we classify temporary or short-term, semi-permanent or medium-term, and permanent or long-term migrants as those individuals who stay away from their place of origin for less than 5 months, 5 to 10 months, and greater than 10 months, respectively. This differs somewhat from the definitions of internal migrants in India used over a decade ago by the 64th round (2007–08) of the NSS consumer survey. In the latter, binary classification is used, and short-term migrants are those who migrate for less than 6 months. However, our sample largely represents economic migrants who were most vulnerable to the COVID-19 shock, and it is not likely to be representative of Indian internal migrants (or even migrant workers) in general. We thus prefer to use a finer classification for our data collection.

The socio-demographic factors we capture in section 6 of the questionnaire include education, number of dependents in the household, religion, marital status, monthly income, land owned, bank account ownership, and size of current outstanding loans taken (if any). These factors are likely to be associated with the (return) migration decision and have been studied in the literature at least since Banerjee and Bucci (1995). From the literature on decision-making in health economics, indicators of lower socioeconomic status (SES) are often associated with less "rational" decision-making, specially studied in the context of scarcity and constrained cognitive function (Mani et al., 2013; Sheehy-Skeffington, 2020).

The behavioural factors we use are derived from survey measures and are defined as binary variables. We measure risk and loss aversion by using lotteries similar to Tversky and Kahneman's (1981) classic binary choice experiment. The measure of risk aversion takes the value of 1 if an individual prefers the sure gain and 0 if they prefer the risky one. An individual is deemed loss averse if he prefers the sure return in the gain domain and the risky return in the reflected loss domain. Time preference is studied by defining a variable impatience which takes a value of 1 if the individual shows a preference for immediate monetary rewards rather than future gains, and 0 if they do not.6 Using scales from the World Values Survey (Inglehart et al., 2014), we also collect information on participants' subjective health and well-being (a higher score implies a higher satisfaction with life).

We also identify unique sources of information related to COVID-19 that respondents use and the number of unique prevention strategies they can recall. Sources of information include other people, television, Internet and social media, newspapers, and health workers. Respondents can choose one or more from this set. To identify the number of unique prevention strategies that respondents have heard mentioned on media and can recall, researchers provide a list of strategies such as social distancing, wearing a mask, hand washing, following government guidelines regarding lockdown or curfew, and staying at home. The list of strategies comprises those which were being disseminated by the government of India and other non-governmental health organizations at the time. As with sources of information, respondents pick one or more of these.

How individuals react to disease-related information is important to study, particularly for outbreaks of hitherto unseen viruses such as COVID-19, for which neither individual beliefs about infection rate and mortality nor personal and social norms of prevention had been established at the time of our study.7 In such a situation, sources, volume and nature of initial messaging are crucial in determining disease-specific risk perception as well as attitudes towards prevention, as Azlan et al. (2020) note in a survey study in Malaysia set a few weeks earlier and at a similar point in the pandemic as our study. The sensitivity of COVID-19 risk perception to the provision of information is seen in Malesza and Kaczmarek (2021), who obtain that having access to more information about COVID-19 increases COVID-19-related anxiety among Polish respondents. On the other hand, using a questionnaire among COVID-19 patients in Wuhan, China, Zhong et al. (2020) find that COVID-19 threat perception is inversely related to knowledge regarding COVID-19. These differences in information-driven perception could be on account of differential emphasis on negative information, as documented by Hamza et al. (2020), who analyze news articles on the internet. In general, differential COVID-19 risk perception has been found among US individuals who use Fox News from those who use CNN (Ciancio et al., 2020). Furthermore, the effect of the spread of misinformation in manipulating individual COVID-19 risk assessments in unverified media such as WhatsApp, Facebook, or Twitter has also been documented, which may falsely cause an individual to either overestimate or underestimate risks related to COVID-19 infection (Farooqui, 2020; Lee et al., 2021; Mahmood et al., 2021).

Finally, we construct a measure of social capital adapted from Grootaert et al. (2004), taking a simple average of the number of unique network nodes identified by the respondents when seeking help regarding a job, travel, or borrowing money. Our definition of social capital and network is more aligned with what Sabatini (2009) describes as informal networks of weak ties (Granovetter, 1973). The higher the social capital index score, the larger the respondent's social network. As Aldrich (2011) suggests, having access to social capital and a wider social network can be important mechanisms through which those affected by covariate shocks (e.g., a pandemic) can better cope with its adverse effects. A higher index of social capital defined in this way can provide an individual and his family with much-needed insurance in their hometown (or the city they work in) to smoothen short-term household consumption.

3. Results

Our main dependent variables are willingness to return to the city post-lockdown and COVID-19 threat perception, which we specify as binary variables in the earlier section. In addition, duration of migration acts as an important moderator variable. Precise definitions for all variables used in our inferential analyses are given in Table B.2 of Appendix B. We compare averages of demographics and other variables related to the levels of these three variables in a univariate way in Tables B.3, B.4 and B.5 in Appendix B. Appendix D presents a descriptive analysis of our three variables of interest and the demographics and individual attributes associated with their levels.

3.1. Instrumental variables

In this section, we outline a multivariate estimation framework. The first outcome of interest is a migrant's willingness to return to the city post lockdown, expressed as a dummy variable [yes = 1/ no =0]. The second is the perceived threat of contracting COVID-19 expressed in binary terms [Positive chance = 1/No chance = 0].

A simplistic way of obtaining determinants for our main relationship of interest would be to estimate an ordinary least squares (OLS) regression with the likelihood of return as the dependent variable on the perceived threat of contracting COVID-19 and independent variables such as duration of migration and other demographic and behavioural attributes. However, a causal relationship between risk perception and the likelihood of return to urban workplaces framed using survey data may be subject to random measurement error in COVID-19 threat perception (Bound et al., 2001). This non-systematic measurement error in the COVID-19 threat perception results in adding to the white noise error term in the regression and causes the covariate to be correlated with the stochastic error creating endogeneity and violating a crucial assumption of classical linear regression (Bound et al., 2001; Angrist and Krueger, 2001; Murray, 2006; Pischke, 2007; Angrist and Pischke, 2008). As a result of this, the OLS coefficient estimate on the endogenous regressor: COVID threat perception, is likely to be biased downwards. In such a situation where we are unable to perform a randomized experiment, instrumental variables (IVs) may offer us a way of lowering this bias (Angrist and Krueger, 2001; Murray, 2006; Angrist and Pischke, 2008).

An instrument is a variable that is correlated with the troublesome (endogenous) regressor but does not otherwise have a statistical relationship with the dependent variable (Murray, 2006). In our case, we propose as instruments the number of unique sources of information on COVID-19 and the number of unique prevention strategies mentioned by a respondent. As we discuss in Section 2.1, informational variables may prove crucial in shaping an individual's perceived disease threat for an emerging epidemic with no recent precedent. This conceptual framework is also informed by the fact that disease risk perception is strongly determined by access to information, increasing the relevance of our instruments. For example, according to Smith (2006), the most important lesson from the SARS epidemic was the importance of effective communication with the public, which was of utmost importance in shaping public risk perceptions. More generally, Bandura (1977) also emphasizes the importance of several sources of information in reinforcing health-related expectations. Furthermore, disease risk perception is also seen to be strongly associated with health beliefs (for example, regarding COVID-19 appropriate behavior). Tong et al. (2020) apply the health belief model (Abraham and Sheeran, 2005) and show that perceived susceptibility to COVID-19 infection is strongly correlated with adherence to COVID-19 precautionary measures.8

3.2. Linear probability model (LPM): two-stage least squares (2SLS) estimation

The model we employ is specified below:

$$COVI{D}_{is}={\gamma }_0+{\gamma }_{1}Inf{o}_{is}+{\gamma }_{2}Pre{v}_{is}+\sum _{j=1}^J{\gamma }_{j}Be{h}_{ijs}+{\gamma }_j\sum _{j=1}^J{X}_{ijs}+{ϵ}_{is}$$ (1)

$$Retur{n}_{is}={\beta }_0+{\beta }_{1}COVI{D}_{is}+{\beta }_{2}Mi{g}_{is}+{\beta }_{3}COVID\times Mi{g}_{is}+\sum _{j=1}^J{\alpha }_{j}Be{h}_{ijs}+\sum _{j=1}^J{\gamma }_j{X}_{ijs}+{\epsilon }_{is}$$ (2)

In equation (i), $Retur{n}_{is}$is the willingness to return for the i^th migrant residing in the s^th state. $COVI{D}_{is}$is the disease threat perception of COVID-19. $Mi{g}_{is}$ is a categorical variable that indicates duration of migration. The specification we use interacts the Mig and COVID variables. $Be{h}_{is}$ is a vector of behavioural parameters related to risk aversion, loss aversion, impatience and subjective well-being and; ${\mathit{X}}_{is}$is a vector of household and individual-level characteristics (religion, years of education, marital status, social capital, etc.) and ${\eta }_{ihs}$is the error term.9 In Eq. (2), there are two COVID-19-related variables: the number of unique sources of information on COVID-19 ($Inf{o}_{is}$) and the number of unique prevention strategies ($Pre{v}_{is}$) mentioned by a respondent. As discussed in the previous sub-section, these are used as instruments in our estimation. In the two-stage least squares procedure that we employ (Angrist and Krueger, 2001; Murray, 2006), Eq. (1) is first estimated, and a linear predictor for COVID-19 threat ($\widehat{COVI{D}_{is})}$is obtained. This predictor should, in theory, not be correlated with the error term ${\eta }_{is}$ of Eq. (2) and is now used in lieu of the reported COVID-19 threat perception in the second stage estimation, which returns us an unbiased estimate of COVID-19 threat on the likelihood of return.

In order for our identification strategy to hold true, we must impose restrictions on the correlation between the two error terms ${\eta }_{is}$ and ${ϵ}_{is}$ and test for the validity of the identification strategy. For our instruments to be valid, we check to see whether or not the COVID-19-related informational factors affect the likelihood of returning to the city only via the perceived threat of contracting the disease. Thus, the decision to return to the city might be associated with the knowledge of COVID-19-related issues, but having valid instruments ensures that this effect is only via the perceived COVID-19 disease threat. In Section 3.1, we have discussed the relevance of our instruments, i.e., how informational variables determine COVID-19 threat perception. Furthermore, given that these informational variables are restricted to COVID-19, they may overlap with social networks (as these may be sources of said information), but it is highly unlikely that migration decisions would be directly based on the number of information sources or the recall of preventive measures, allowing us to exclude them from the endogenous equation (ii). Furthermore, even though random errors may be prevalent for our proposed instruments, as they are for all responses to our survey, these are unlikely to be correlated with each other improving our instrument validity.10 The coefficient on an endogenous regressor from an IV regression may be interpreted as a local average treatment effect (LATE) and was first introduced to the economics literature by Imbens and Angrist (1994). In our case, it may be interpreted as the impact on the likelihood of return of the part of COVID-19 threat perception that our instruments can identify.

3.3. LPM: results and discussion

3.3.1. Diagnostic tests for the validity of our IV specification

Table 1 presents the results from the linear probability model of migrants' willingness to return to the city and COVID-19 disease threat perception.11 We run a range of tests to establish the validity and robustness of the IV specification following Schmidheiney (2021) and Andrews et al. (2019). The Hausman Specification Test given below the coefficient estimates in Table 1 is significant at the 1% level, indicating that we do have significant endogeneity among the COVID-19 threat perception and the likelihood of return. The F-statistic from the first-stage regression provides information on whether we are dealing with weak instruments. A weak instrument is one which is weakly correlated with the endogenous regressor (Murray, 2006). In general, an F-statistic greater than 10 is considered a rule of thumb in defining non-weak instruments (Staiger and Stock, 1997; Stock and Yogo, 2005). The corresponding F-statistic from our linear probability model (LPM) provides an F-stat (2, 398) of 2.55 with a p-value of 0.07, indicating that we are indeed dealing with weak instruments. Additional tests using -weakivtest- in Stata (Pflueger and Wang, 2015) show that the Montiel-Pflueger effective F-statistic is the same value as that from our LPM. When compared with the Stock-Yogo Weak ID F-test critical values, it is further indicated that since our F-stat is not close to exceeding even 25% critical values, the instruments are weak. However, identification tests suggest that the system of equations is neither underidentified (Kleibergen-Paap rk LM statistic = 5.81, p = 0.05) nor overidentified (Hansen J-test = 1.447, p = 0.23).12 As Andrews et al. (2019) suggest, having weak instruments does not mean that there cannot be robust inference from the IV framework. We run a range of tests to establish whether our endogenous regressor (the chance of contracting COVID-19) is significant in the main equation, which provides a range of test statistics on whether we can still reasonably infer that the endogenous regressor coefficient is significantly different from zero (Anderson-Rubin Wald test = 17.34; Anderson-Rubin Chi-square test = 36.43; and Stock-Wright LM S statistic = 38.08, all significant at the 1% level). Furthermore, as we have a single endogenous regressor, we perform additional tests as described in the supplementary material of Andrews et al. (2019), each of which provides robust inference in the presence of a weak instrument. The results of the test are unambiguous: there is a clear rejection of the null hypothesis that the coefficient of the endogenous regressor is the same as zero. In particular, the most reliable of these, the conditional LR test (Andrews and Stock, 2007), has a test statistic of 32.11 and is significant at the 1% level. Additional tests such as the Anderson-Rubin (test statistic = 33.45, p-value < 0.01) and the Lagrange Multiplier K (test statistic = 4.44, p-value = 0.035) test also suggest the same. We report heteroscedasticity-robust standard errors and include state fixed effects in all estimations.13 We first discuss the factors associated with greater disease threat perception in our sample of migrant respondents.

Table 1.

OLS and 2SLS-IV estimates.

	First-stage	OLS	IV-2SLS
			Second stage
Variables	Contract COVID-19	Return to city	Return to city
	(1)	(2)	(3)
No. of unique sources of info on COVID-19	0.00674
	(0.0275)
No. of unique prevention strategies identified	−0.113⁎⁎
	(0.0519)
Migrant (5–10 months)	−0.125⁎⁎	0.213⁎⁎	0.459⁎⁎
	(0.0531)	(0.0863)	(0.178)
Migrant (> 10 months)	−0.250⁎⁎⁎	0.251⁎⁎⁎	0.0189
	(0.0664)	(0.0945)	(0.225)
(Predicted) Chance of contracting COVID-19		−0.0219	−1.659⁎⁎⁎
		(0.0773)	(0.519)
Interaction terms
(Predicted) COVID-19 Chance X Migrant (5 −10 months)		−0.185*	−0.929⁎⁎⁎
		(0.0970)	(0.226)
(Predicted) COVID-19 Chance X Migrant (> 10 months)		−0.130	−0.542*
		(0.121)	(0.303)
Behavioural factors
Subjective well-being	−0.0286⁎⁎	0.0240*	−0.0443⁎⁎
	(0.0132)	(0.0138)	(0.0214)
Risk aversion	−0.0293	−0.0982	−0.119*
	(0.0815)	(0.0643)	(0.0661)
Impatience	0.0945*	0.0840*	0.238⁎⁎⁎
	(0.0549)	(0.0484)	(0.0690)
Loss aversion	−0.137⁎⁎	0.0916*	−0.232⁎⁎⁎
	(0.0635)	(0.0488)	(0.0879)
Socio-demographic factors
Social capital index	−0.170*	0.0734	−0.284⁎⁎⁎
	(0.0879)	(0.0671)	(0.0985)
Years of education	−0.00108	−0.0149	−0.0235*
	(0.0159)	(0.0139)	(0.0130)
Religion (base: Hindu)	−0.110*	0.00203	−0.205⁎⁎⁎
	(0.0639)	(0.0497)	(0.0755)
Married	−0.0537	−0.102⁎⁎	−0.227⁎⁎⁎
	(0.0503)	(0.0464)	(0.0506)
Dependents	0.0818⁎⁎⁎	−0.0710⁎⁎	0.101⁎⁎
	(0.0243)	(0.0286)	(0.0473)
Land owned	0.0301⁎⁎	−0.0393⁎⁎⁎	0.0289
	(0.0138)	(0.0116)	(0.0187)
Standardized monthly income	0.0248	0.0294	0.0802⁎⁎⁎
	(0.0284)	(0.0212)	(0.0242)
Amount of loan taken (INR)	1 × 10−7	−7 × 10−7⁎⁎⁎	−1 × 10−7
	(6 × 10−7)	(2 × 10−7)	(2 × 10−7)
Own bank account	0.103	−0.0867	0.189⁎⁎
	(0.0711)	(0.0688)	(0.0942)
Constant	0.758⁎⁎⁎	0.437⁎⁎	1.658⁎⁎⁎
	(0.174)	(0.173)	(0.431)
Observations	418	418	418
R-squared	0.187	0.215	0.291
Hausman Specification Test		21.01⁎⁎⁎
Anderson-Rubin likelihood ratio test		1.23
First-stage F-stat / Sanderson-Windmeijer test		2.55*
Kleibergen-Paap rk LM statistic for under identification		5.80*
Weak identification Cragg-Donald Wald F-stat		3.02*
Hansen J-statistic for overidentification		1.45
Weak instruments robust inference tests
Anderson-Rubin Wald F-stat		17.34⁎⁎⁎
Stock-Wright LM S statistic		38.08⁎⁎⁎
Conditional LR test		32.11⁎⁎⁎
LM K test statistic		4.44⁎⁎
Anderson-Rubin chi-square test statistic		33.05⁎⁎⁎

^⁎⁎⁎p<0.01.

^⁎⁎p<0.05.

^⁎p<0.1

Note: Results from OLS and two-stage least squares regressions (linear probability models) reported. All estimations include state fixed effects. Robust standard errors in parentheses.

3.3.2. Perceived chance of contracting COVID-19

Column (1) of Table 1 provides estimates for the perceived likelihood of contracting COVID-19. We find that both medium and long-term migrants, i.e., those who migrate for 5 to 10 months and over 10 months in a year, respectively, perceive a lower risk of contracting COVID-19 as compared to short-term (less than 5 months) migrants. Given that this effect survives after controlling for socioeconomic, behavioural, and COVID-19 information-related factors informs us that there are other effects related to the duration of migration that alter disease-specific risk perception independent of these factors. A possible explanation could be greater familiarity with the city environment for the longer duration migrants vis-à-vis those who migrate for shorter durations. A bias arising from familiarity with a product or context could lead individuals to disregard more objective information leading to faulty decision-making (Kahneman, 2003). In the context of financial products, Wang et al. (2011) obtain that financial products that are more familiar appear to investors to be inherently less risky. It is also likely that long-term migrants display stronger place attachment to the urban centres of their work, which may, in turn, lower their risk perception relative to short-term migrants (e.g., Scannell and Gifford, 2010).

A higher rating on the subjective well-being scale is associated with having a lower perceived chance of contracting the disease, which is in line with literature on subjective well-being and risk perceptions in other domains (Guven and Hoxha, 2015). Loss-averse individuals have 13.7 percentage points lower perceived chance of contracting COVID-19 as compared to others. Following Kahneman and Tversky (1979), loss-averse individuals display risk-seeking behavior over the domain of losses. As contracting COVID-19 would entail significant losses for individuals, those identified as loss-averse, according to theory, should take on more risk, which may make them perceive a lower infection probability than the non-loss-averse. Redelmeier & Shafir (2020) posit that loss aversion could be closely linked to maintaining (or expecting to maintain) the status quo, which could manifest in lower support for strict COVID-19 interventions (Hameleers, 2020). We conjecture that perhaps the latter could correlate with a lower perceived chance of contracting COVID-19 as we see here and leave it to future research to examine this systematically. Finally, respondents who self-report as impatient (i.e. preferring immediate gains to later larger gains) have an almost 9.5 percentage point higher perceived chance of contracting COVID-19 than others, though this effect is only marginally significant.

There is a strong positive correlation between land owned and disease threat perception. Individuals with one more dependent in their household perceive an 8 percentage point higher chance of contracting COVID-19. This appears consistent with the results from Akesson et al. (2020), who find that individuals living with children tend to significantly overestimate COVID-19 mortality risk.

We now turn to the effect of COVID-19-related factors on the perceived chance of contracting COVID-19. Here, we find that having one more source of information related to the coronavirus is not significantly associated with the self-perceived chance of contracting the infection. On the other hand, being familiar with one more COVID-19 prevention strategy is associated with approximately 11.3 percentage points lower perceived chance of contracting COVID-19. This result is particularly interesting as it appears that COVID-19 messaging that provides more strategies for prevention and hence their recall makes individuals feel safer with respect to being infected. We conjecture that perhaps the provision of messaging that reveals more about COVID-19 prevention helps citizens allay fear in a situation of great uncertainty, as Finset et al. (2020) note. Furthermore, individuals with higher social capital in terms of a network (i.e., more individuals they can rely on for borrowings, jobs, or travel) report, on average, a lower perceived chance of contracting COVID-19, though this effect is only significant at the 10% level. This is likely to be on account of optimal insurance mechanisms via weak social ties and therefore underestimates disease threat (Sabatini, 2009; Aldrich, 2011). We also perform OLS (first-stage) regressions for our interacted endogenous terms [COVID-19 chance X Migrant (5 −10 months) and COVID-19 chance X Migrant (> 10 months)] and report estimates in Table B.10 in Appendix B.

3.3.3. Impact on willingness to return

Column (2) of Table 1 gives us OLS estimates for our endogenous equation (ii). These estimates are obtained without accounting for endogeneity. We observe that the coefficient on our endogenous regressor, i.e., the perceived likelihood of contracting COVID-19, is small in absolute value and not statistically significant. From column (3) of Table 1, which is the second stage of our 2SLS estimation, we see that a higher perceived chance of contracting COVID-19 is negatively and significantly associated with the willingness to return to the city, independent of the duration of migration. This can be interpreted as an apprehension of contracting COVID-19 once they return to the city, as COVID-19 cases in India have been particularly concentrated in urban areas. The fear of contracting COVID-19 is the single largest reason given for reverse migration, according to a survey conducted in 179 districts in May-June 2020 (Sahoo and Bhunia, 2021). Comparing our OLS and 2SLS estimates for the threat of contracting COVID-19 suggests that the OLS estimate is downward biased because of endogeneity. Thus, we do not discuss coefficients of other (non-endogenous) regressors from our OLS model as these are likely to be incorrectly estimated because of the downward bias in the coefficient for our endogenous regressor.

From column (3), relative to the base category, i.e., workers who migrate for less than 5 months, those with migration stints between 5 and 10 months in a year are more likely to return to their urban workplaces. For long-term (> 10 months) migrants, the likelihood is also higher, but this effect is not statistically significant. As discussed in Section 3.3.2, longer-term migrants may display higher familiarity and attachment to urban workplaces than temporary migrants, lowering their perceived disease threat and thereby increasing their likelihood of return. Further, individuals who display stronger place attachment are also more likely to return to their place of work, despite a natural hazard risk (Bonaiuto et al., 2016).

Our interactions between the predicted COVID-19 perception and migration duration allow us to define slopes for the three different durations. These slopes,$\frac{d\left(Return\right)}{d(\widehat{COVID)}},$ which are all negative, represent how the likelihood of return diminishes as the predicted COVID-19 likelihood increases. Using coefficients significant at least at the 10% level, and for a unit increase in predicted COVID-19 threat probability, the likelihood of return is the highest for short-term migrants, followed by that of medium-term migrants and long-term migrants.

In terms of behavioural factors, we observe that migrants who have a higher level of subjective well-being are less likely to be willing to return to the city. We also find that loss-averse individuals are almost 14 percentage points less willing to return to their urban workplaces. Additionally, risk-averse individuals are almost 12 percentage points less willing to return though this effect is only marginally significant. This displayed risk and loss aversion towards migration appears to be in line with behavioural explanations of migration decisions by Banerjee and Duflo (2020, p. 40), who state that 'migration is a plunge into the unknown.' Similar to our finding, Jaeger et al. (2010) find that in Germany, less risk-averse individuals are more likely to migrate. On the other hand, Conroy (2007) shows that more risk-averse individuals migrate from Mexico to the United States. Finally, migrants who are more impatient have a 24 percentage point higher self-reported willingness to return to the city post the lifting of lockdown and travel restrictions. This is in line with Nowtony (2010), who shows both theoretically and empirically that more impatient individuals have a higher propensity to migrate.

There are some key socio-demographic factors associated with the (un)willingness to return to the city. Individuals who are more educated display a lower projected willingness to return, though this coefficient is significant only at the 10% level. Workers from religious minority groups also display a lower willingness to return compared to Hindus. In addition, being married is negatively associated with the likelihood of returning to the city. On the other hand, having dependents, a higher standardized monthly income and a savings bank account are positively associated with the stated likelihood of return. Finally, a higher level of social capital is associated with a lower likelihood of returning to the city. Thus, we find that the result from our two group comparisons, i.e., a higher level of social capital is associated with a higher likelihood of return, changes when we control for endogeneity and other factors in our regression analysis. It is clear that our result with respect to the impact of social capital from our regression analysis is more reliable as it takes into account the heterogeneity arising from moderating variables.

4. Robustness checks

4.1. Alternative estimation framework: bivariate probit

Given that both variables of interest are binary in nature, we specify a probit model for estimating the relationship between socio-demographic factors, behavioural parameters, and variables related to information and social capital and the outcome variables. However, as we have discussed in Section 3, in the context of linear models, standard probit estimates may similarly be biased because of the presence of random measurement error for the COVID-19 threat perception. Thus, as before with our linear specification, we argue that both the willingness to return and disease threat perception are jointly determined and potentially endogenous. Following Chatterji et al. (2004), we use the bivariate probit model to allow for correlated errors between disease threat perception and willingness to return to the city post lockdown.14 Our model specification is similar to that used for 2SLS, except that we interact the categories of duration of migration (<5 months, 5–10 months and > 10 months) with perceived COVID-19 threat (no chance, significant chance). This is because, unlike the first stage of 2SLS, which computes a real-valued predictor for COVID-19 threat, which is interacted with migration duration, here, the COVID-19 threat variable is a binary indicator necessitating 6 separate interaction categories.

The Wald test for correlated errors indicates that the errors are highly correlated (p-value < 0.01), so the bivariate probit model is appropriate in this context. However, as Filippini et al. (2018) report, a zero-correlation parameter in the bivariate probit model may mask an underlying recursive data-generating process.15 Thus, to check for the robustness of the bivariate probit model, we run a recursive bivariate probit model. The results are reported in Table B.6 (Appendix B). This also allows us to present average treatment effects (Coban, 2021) for each type of migrant. Second, we test the validity of the exclusion restrictions using the Wald test for joint significance following Reichman et al. (2004) on the non-interaction model using an instrumental variables probit estimation. We find that the two instruments we use (number of unique sources of information and number of prevention strategies identified) are jointly non-significant in predicting the likelihood of returning to the city. The resulting Chi-squared is 4.60, and the p-value is 0.10.16

From column (1) of Table 2 we find, as with our linear specification in column (1) in Table 1, that longer duration migrants (between 5 and 10 months and greater than 10 months) show a lower perceived chance of contracting COVID-19, as compared to short term migrants. Our instruments: number of unique sources of information and the number of prevention strategies identified are associated with positive and negative signs, respectively, with the perceived chance of contracting COVID. The signs for the latter are consistent with corresponding estimates for the 2SLS estimation. Subjective well-being, impatience and loss aversion are respectively negative, positive (at the 10% level) and negative in their association with the likelihood of return. These, too, are consistent with estimates from 2SLS, with sizes that are comparable. As in our linear formulation, having dependents and owning land are associated with higher COVID-19 threat perception.

Table 2.

Bivariate probit estimation: Return to city and likelihood of COVID-19.

VARIABLES	Contract COVID-19	Return to city
	(1)	(2)
COVID-19 related factors
No. of unique sources of info on COVID-19	0.0418⁎⁎⁎
	(0.0162)
No. of unique prevention strategies identified	−0.130⁎⁎⁎
	(0.0281)
Migrant (5–10 months)	−0.0947*
	(0.0518)
Migrant (> 10 months)	−0.225⁎⁎⁎
	(0.0673)
Interaction Model
No chance X Migrant (5 −10 months)a		−0.00476
		(0.0480)
No chance X Migrant (> 10 months)		−0.0148
		(0.0494)
Positive chance X Migrant (< 5 months)		−0.423⁎⁎⁎
		(0.0271)
Positive chance X Migrant (5 −10 months)		−0.435⁎⁎⁎
		(0.0327)
Positive chance X Migrant (> 10 months)		−0.350⁎⁎⁎
		(0.0241)
Behavioural factors
Subjective well-being	−0.0241*	0.00148
	(0.0129)	(0.00854)
Risk aversion	−0.0224	−0.0946*
	(0.0780)	(0.0523)
Impatience	0.109⁎⁎	0.0772⁎⁎
	(0.0475)	(0.0331)
Loss aversion	−0.161⁎⁎⁎	0.0177
	(0.0615)	(0.0459)
Socio-demographic factors
Social capital index	−0.146*	−0.0370
	(0.0763)	(0.0453)
Years of education	−0.0157	−0.00554
	(0.0148)	(0.0103)
Religion (base: Hindu)	−0.132⁎⁎	−0.0296
	(0.0557)	(0.0305)
Married	−0.0223	−0.0797⁎⁎
	(0.0451)	(0.0321)
Dependents	0.0854⁎⁎⁎	−0.00195
	(0.0272)	(0.0196)
Land owned	0.0253⁎⁎	−0.0200⁎⁎
	(0.0124)	(0.00938)
Standardized monthly income	0.0387*	0.0187
	(0.0235)	(0.0157)
Amount of loan taken (INR)	−3 × 10−9	−3 × 10−6⁎⁎⁎
	(5 × 10−7)	(9 × 10−7)
Own bank account	0.125	−0.00509
	(0.0759)	(0.0447)
Observations	418
Adapted Hosmer-Lemeshow goodness-of- t-test	217.55⁎⁎⁎
Recursive Bivariate probit Wald test	332.22⁎⁎⁎
Wald comparison test for serially correlated errors (recursive)	6.45⁎⁎
Wald test of serially correlated errors	12.80⁎⁎

^⁎⁎⁎p<0.01.

^⁎⁎p<0.05.

^⁎p<0.1.

^aNote: The base category is No chance X Migrants (< 5 months). All estimations include state fixed effects. Marginal effects at sample mean reported. Robust standard errors in parentheses.

The second stage coefficients in column (3) of Table 1 (2SLS) may be compared with the corresponding average marginal effects on the likelihood of return from bivariate probit in column (2) of Table 2. Our regression reports five out of the six interactions between perceived COVID-19 threat and duration of migration in column (2) of Table 2. As in our linear estimation, we find that independent of the migration duration, a worker who perceives a greater than zero COVID-19 threat is less likely to return than workers with a zero COVID-19 threat perception. Testing our three marginal effects against each other that there is a statistically significant difference only between those of medium-term and long-term migrants, where medium-term migrants with a positive COVID-19 threat perception are more unwilling to return than long-term migrants.17 Our results with respect to these interactions are thus different from those in our linear specification, where for a given level of (predicted) COVID-19 threat perception, the long-term migrant groups appear more unwilling to return. We also compute average treatment effects (ATE) by migrant duration to provide comparisons with the LPM. We find that all three are statistically significant at the 1% level and a similar pattern for the effect of chance of contracting COVID-19 on willingness to return to the city. The ATEs are as follows: for migrants less than 5 months (ATE = −0.50, S.E. = 0.123, p-value < 0.01), migrants between 5 and 10 months (ATE = −0.573, S.E. = 0.055, p-value < 0.01), and migrants greater than 10 months in a typical year (ATE = −0.55, S.E. = 0.061, p-value < 0.01). Thus, the recursive model aligns with the overall finding that for a given chance of contracting COVID, longer-term migrants are somewhat less reluctant to return to the city.

With respect to behavioural factors, we find that, as with our linear specification, risk aversion is associated with a reduction in the likelihood of return (at the 10% level of significance), while impatience is associated with an increase in the stated likelihood of return. We find that individuals who are married and landowning display a lower stated likelihood of return. Our results with respect to landholding are thus different between our bivariate probit and linear IV specification, with land holding being insignificant in the latter. On the other hand, the variables years of education, religion, number of dependents, monthly income and own bank account are significant at least at the 10% level in our linear specification but are insignificant here.

In the 2SLS formulation, the linear predictor used for COVID-19 threat perception reflects only the part of the variation in the self-reported COVID-19 threat perception that is explained by our instruments. On the other hand, the bivariate probit specification simultaneously finds maximum likelihood estimates for the two equations and uses actual responses for the perceived COVID-19 likelihood. Accordingly, the first stage predictor of threat perception, $\widehat{COVID}$ that is used in the endogenous equation takes the form of a continuous variable in 2SLS, whereas in our bivariate probit estimation, COVID threat perception is binary. It is not, therefore, surprising that IV and bivariate probit specifications using the same data are documented to provide different estimates by Altonji et al. (2005). Chiburis et al. (2012) derive asymptotic results and present simulations comparing bivariate probit and linear IV estimators with a binary dependent variable such as ours and find notable differences in IV and bivariate probit estimates for samples under 5000. The latter recommends reporting both IV and probit results when they differ. Chiburis et al. (2012) also recommend a goodness-of-fit score test that can help detect misspecifications of the bivariate probit model. Accordingly, as reported in Table 2, the goodness-of-fit test results (adapted Hosmer-Lemeshow goodness-of-fit t-test) reject the null hypothesis that the bivariate probit is the appropriate model, further strengthening the case for our linear IV specification. Accordingly, when discussing insights from this study, we adopt a conservative approach and, as a rule of thumb, present our significant LPM estimate rather than that from bivariate probit when the latter is significant and in the same direction or insignificant. When there is a sign conflict as with the effect of interactions of duration of migration and COVID-19 threat perception, we refrain from discussing this effect more generally in the introduction and conclusion, even though we present both sets of results in their relevant sub-sections as Chiburis et al. (2012) recommend.

We also estimate an additional variation of our model where we use a state-wise seven-day rolling average of the daily confirmed cases separately for the two phases of our data collection. We then compute an unweighted average of these two rolling averages and use them in lieu of state fixed effects. The results are qualitatively similar to those presented in Table 2 and can be found in Table B.7 in Appendix B.

4.2. Alternative measures of disease threat

To test whether our findings are robust to the measurement and framing of disease threat and COVID-19 risk perception, we run ordered endogenous probit models with three other variables on which data was collected using our questionnaire. The specifications we run are analogous to what we use in Section 3 for our bivariate probit estimation. In lieu of COVID-threat perception, the related perception variables we use are:

• (a)How painful do you think COVID-19 infection can be? (increasing ordered response, 5 levels, question 8 of the questionnaire)
• (b)How severe or serious can the COVID-19 infection be for your health? (increasing ordered response, 5 levels, question 6)
• (c)Can you stop or prevent infection with the COVID-19 virus? (binary variable, Y/N, question 9)

These results can be found in Table B.8 in Appendix B. In general, the results are similar to our main findings, with a negative association between disease threat variables and willingness to return to the city. Consistent with the results presented in Table 2, we find that the largest negative associations between COVID disease threat and willingness to return were found for those who migrate for a relatively shorter duration (5–10 months).

4.3. Heterogeneity analysis

4.3.1. Duration of migration

Tables B.11and B.12 in Appendix B reports the results of heterogeneity analysis by the duration of migration using the LPM and bivariate probit model respectively. These results are broadly consistent with the results from bivariate probit estimates (Table 2). In particular, for migrants who typically spend between 5 and 10 months in the city, an increase in perceived disease threat is associated with a sharp reduction in the willingness to return to the city post-lockdown. Since the number of respondents within each category is different, estimates and identification are underpowered for the category of migrants who spend more than 10 months in the city. The Sanderson-Windmeijer first stage χ² and F statistics are significant for the less than 5 months and 5–10 month duration regressions, indicating that we do not have under-identification or weak identification with our instruments. Furthermore, the second stage Sargan-Hansen χ² is insignificant for these regressions showing that they are not over-identified.

4.3.2. Socioeconomic factors

Table B.9a in Appendix B reports the heterogeneous effects analysis by marital status, land owned, and whether or not respondents had a loan outstanding at the time of the survey. Table B.9b in Appendix B reports heterogeneous effects by reported monthly income. It would appear that the strong (negative) association between the perceived chance of contracting COVID-19 and the likelihood of return is driven by the unmarried, small landholding (less than 1 acre) and those not currently indebted. Regarding behavioural factors, the negative association between risk aversion and willingness to return to the city is driven by the married, those with small land holding and individuals without any outstanding loans at the time of the survey. The unmarried sample mainly drives the higher subjective well-being associated with a lower likelihood of returning to the city. From Table B.9b, it appears that the negative association between the likelihood of contracting COVID-19 for medium-term migrants and the likelihood of return is driven mainly by the poorest migrant workers (income < INR 10,000 per month). In addition, the negative impact of the perceived chance of contracting COVID-19 on return likelihood is highest for the poorest.

5. Conclusion and policy implications

Using novel data for reverse migrants who migrate from three major cities in India to go back to their villages in the wake of a nationwide COVID-19-related lockdown, we explore the causal link between the likelihood of re-migration to cities post-lockdown and the perceived threat of contracting COVID-19. We also explore the effects of demographic, socioeconomic, behavioural and information-related attributes on the jointly determined disease threat and the likelihood of return variables. We find that longer-term migrants perceive a lower disease threat as compared to more short-term migrants. Irrespective of migration duration, individuals who perceive a non-zero risk of contracting COVID-19 have a significantly lower stated likelihood of returning to their urban work centres. With respect to socio-demographics, we find that having dependents and owning more land increases the threat perception of COVID-19. With respect to behavioral factors, we find that individuals who display more impatience have a higher COVID-19 risk perception yet display a higher stated likelihood of returning to the city. On the other hand, loss-averse individuals display a lower disease threat perception.

Our results predict that the perceived risk of contracting COVID-19 significantly affects migration decisions. This is consistent with the reverse-migration observed during the second wave of the pandemic in India, which had seen no national lockdown, but partial lockdowns imposed by state governments from April - August 2021 (Kaushal and Kumar, 2021), as well as with the subsequent return of many migrant workers to their pre-pandemic urban workplaces since then. The latter can be attributed to a reduction in the COVID-19 risk perception among migrant workers, which may have occurred both due to a decline in the number of cases since the peak of the second wave in India and due to a reduction in the intensity of the infection with the emergence of the Omicron variant.

Our results suggest that the general policy direction related to the impact of a pandemic on migration should not only focus on improving the canonical supply side of health infrastructure such as increasing the provision of test kits, hospital beds, oxygen cylinders and ventilators for the migrant workers but also on the behavioral aspects of managing the perceived risk of disease contraction among them. As the risk of contracting COVID-19 is a very important determinant of migrant workers not wanting to return to their urban workplaces, the main policy implication of our finding is that the central and state governments can facilitate this return by regulating COVID-19 safety features in urban workplaces and lowering the perceived risk of COVID-19 contraction among the migrant workers. In the short run, this can be done through IEC (Information, Education and Communication) campaigns and vaccination drives. Studies show that information spread specifically through social media plays an important role in influencing COVID-19 risk perception (Wang et al., 2021). Therefore, it is crucial to spread accurate information through mass media and social media and adopt messaging that reduces COVID-19 risk perception. As various agencies provide information and misinformation to citizens, government and private entities in the development sector need to regulate this by providing accurate and updated information on COVID-19. Compensating for informational biases, provision and management of accurate and timely information by authorities may thus be crucial tools to support rapid healthcare mobilization and promote public compliance with rules pertaining to further waves of COVID-19 or outbreaks of other diseases. Dasgupta et al. (2020) document that perception of risk and concomitant action are influenced in a path-dependent manner for "procedurally rational" agents. Thus, administrators and citizens of countries and communities that have historically experienced similar disease outcomes tend to be much more ready with a public response to combat a new outbreak as compared to others. Future research that refines our understanding of the influence of behavioural and informational factors on threat perception and mobility in the wake of a pandemic will allow nations to explicitly incorporate this wisdom into more efficient public information dissemination policies.

COVID-19 risk perceptions are also positively correlated with preventive health behaviours such as mask-wearing, handwashing and social distancing. (Schneider et al., 2021). Thus, policymakers can focus on promoting the adoption of preventive health behaviours such as the 'Mask nahin toh tokenge, Corona ko rokenge (If you don't have a mask, you won't stop Corona)' campaign driven by the Behavioural Insights Unit of India, NITI Aayog. Studies also show that increased vaccination reduces COVID-19 risk perception (Jia et al., 2022). Therefore, policymakers should focus on increasing the vaccination rate among migrant workers and specifically on the booster dose, as less than 10% of Indians have received it till August 2022.18

In the long run, improving health care facilities and providing health insurance for COVID-19 can help to mitigate risk perceptions among the migrant workers. Over the past two years, COVID-19 has resulted in an increase in health care costs for individuals and families. To address this fact, many companies in the organized sector have now started providing COVID-19 insurance to their employees. However, migrant workers mostly belong to the unorganized sector and therefore do not receive such employee benefits. Thus, for them, the government can consider providing insurance coverage against COVID-19. This may contribute to the reduction in the risk perception of COVID-19 and increase the willingness of migrants to return to their pre-pandemic locations with more confidence. Finally, the risk perception of COVID-19 can increase due to a shortage of test kits, hospital beds, oxygen cylinders, ventilators etc., as was witnessed in India during the peak of the second wave in 2021. Therefore, the risk perception related to COVID-19 may be lowered with further improvements in the health infrastructure and the healthcare system in general.

Declaration of Competing Interest

None

Data availability

• Data will be made available on request.