Mode and Shelter Choice Planning During Evacuation: A Multinomial Logistic Regression Analysis of COVID-19-Induced Migration in India

Abstract

Background: The COVID-19 pandemic triggered unprecedented mobility disruptions worldwide as governments imposed strict lockdowns to contain the spread of the virus. In India, prolonged restrictions severely affected economic activity, particularly for migrant workers, leading to a large-scale and unplanned exodus from urban employment centres to native places. This sudden population movement undermined containment efforts and contributed to the spatial diffusion of infections. Understanding evacuees’ behavioural responses during such crises is therefore critical for effective emergency logistics and evacuation planning. Methods: This study examines the determinants of transport mode and shelter choice decisions made by migrants during the COVID-19-induced evacuation in India. Using primary survey data, a multinomial logistic regression model is developed to analyze how socio-economic characteristics influence evacuees’ choices of travel mode and shelter type. Results: The results reveal significant heterogeneity in decision-making, highlighting the role of economic vulnerability and accessibility constraints in shaping evacuation behaviour. Conclusions: The findings offer actionable insights for policymakers and emergency planners to design inclusive evacuation strategies, improve crisis-responsive transportation planning, and enhance shelter provisioning in future pandemics or large-scale disruptions. The study contributes to the logistics and humanitarian operations literature by providing empirical evidence on evacuation behaviour under public health emergencies.

1. Introduction

In December 2019, COVID-19 emerged in Wuhan, China, and spread all over the world in a very short span of time. Consequently, the World Health Organization (WHO) declared COVID-19 as a pandemic on 11 March 2020 [1]. As a result, countries worldwide implemented lockdowns to prevent the disease outbreak, pausing almost all economic activities. Employees of both organized and unorganized sectors are affected by such a regulatory move. Employees from the organized sector either worked from home or were fired from their jobs, while many in the unorganized sector lost their jobs, particularly in India. As per the report from CMIE, around 10 million Indians lost their jobs during the second wave of COVID-19 [2]. However, no official data from the government can be found [3]. Indian migrant workers have faced multiple hardships with the shutdown of factories and workplaces. Due to this, the masses found little purpose for staying at their work location. In India, it has been estimated that around 37% of the total population are found to be migrants, and an exodus of these migrants began after a few weeks of lockdown [4]. This exodus compromised the idea of lockdown and ended up spreading the disease to several locations. With food shortage and uncertainty about the future, there is a large-scale evacuation of migrants. Due to the surge in demand of reaching their native place, the transportation system countered several constraints while serving these migrants. Government restrictions, increased monitoring and policing, and limited modes of transportation resulted in several mishaps for the migrants. It has been reported that 884 deaths were due to non-COVID-19 reasons during lockdown [5]. It has been mentioned that some deaths (around 50%) were due to transport-related issues. In particular, the reasons were walking (non-availability of transport mode), an accident during travelling and suicides due to an inability to go home. Another critical issue that emerged from these deaths is the lack of basic necessities in the shelters provided by the government [4]. Lack of facilities (transport availability and appropriate shelter in this case) can be attributed to limited understanding of migrants’ behaviour and associated factors. Such behaviour analysis can help decision-makers make the evacuation process efficient and help them to devise better planning for such events.

To make the evacuation process efficient, it is required to incorporate all four steps of the disaster cycle, namely prepare, mitigate, response and recover. However, it can be observed from the migrants’ exodus that these lockdown restrictions are not planned properly [6]. There was a lack of evidence-based decision-making during this event. In response to such an exodus, central and state governments later arranged transport modes for them. Overall, there is a need to understand migrants’ behaviour to make the evacuation process efficient. Pandemic-induced evacuation of economically disadvantaged populations is still empirically understudied, despite an increasing amount of research on evacuation behaviour during natural catastrophe [7]. Lockdowns were ineffectual in numerous areas of India due to the significant “crisis of mobility” caused by the COVID-19 outbreak, as migrant workers frantically tried to return home by all means possible [8]. During the initial lockdown wave, almost 43.3 million interstate migrants went back to their home states, with millions using non-traditional methods. Low-wealth mobile groups are consistently underrepresented in evacuation behaviour models, despite the fact that income, employment status, and resource access are known to impact emergency responses. Furthermore, natural hazard scenarios have been the main emphasis of multinomial logistic frameworks in catastrophe contexts [9], with no extension to public health emergency evacuations. Accordingly, this paper attempts to understand the migrants’ behaviour during the pandemic. In this study, evacuation behaviour of migrants is expressed in terms of the transport mode used for evacuation and shelter planning. Migrants’ choice of transport mode and shelter is influenced by multiple factors, including the type of shelter, age, income, and duration of stay (detailed more in Section 3). Once the underlying relationship of mode choice and shelter planning is analyzed, it can help decision-makers for future crises. Therefore, the key research objectives (Ros) of this study are now stated: (RO1) identify socio-economic determinants of transport mode choice during COVID-19 evacuation; (RO2) examine factors shaping shelter type selection; (O3) derive actionable policy recommendations for inclusive emergency planning. The key contributions of this study are as follows:

• One of the first empirical studies to jointly model transport mode and shelter choice among migrants during pandemic-induced displacement in India.
• Application of multinomial logistic regression to examine evacuation behaviour across socio-economic strata in a developing-country context.
• Empirical evidence bridging humanitarian logistics, emergency transport planning, and public health policy for low-income and informal-sector migrants.
• Actionable insights for policymakers to target evacuation resources toward the most vulnerable migrant segments.

The rest of this paper is structured as follows: Section 2 explains the review of the previous study to determine the evacuees’ behaviour. The materials and methods are explained in Section 3. The multinomial logistics regression model is developed, and results are proposed in Section 4 including discussion and insights of developed models. Conclusions and future work are proposed in Section 5.

2. Literature Review

Evacuation planning is an established research domain that attracts scholars from multiple disciplines, including engineering, management, computer science, mathematics, transportation studies, and environmental science. The interdisciplinary nature of this field reflects the complex challenges associated with large-scale population movement during emergencies, which involve infrastructure planning, behavioural decision-making, logistical coordination, and risk management, among many other critical activities [10,11]. Over the past few decades, considerable research has focused on developing analytical models, simulation frameworks, and optimization techniques to support effective evacuation planning and disaster response. These approaches include traffic simulation models, routing optimization algorithms, and behavioural modelling techniques aimed at improving evacuation efficiency and minimizing risk during crises [12,13].

The majority of evacuation literature focuses on sudden-onset hazards such as hurricanes, floods, wildfires, earthquakes, and nuclear accidents, where immediate physical threats necessitate rapid relocation of affected populations [14]. In such contexts, previous literature examines operational aspects of evacuation systems, including traffic management, route optimization, fleet allocation, shelter location planning, and emergency communication strategies [15]. Simulation-based approaches have been widely used to analyze crowd dynamics and traffic flow during evacuation, while optimization models assist in designing efficient routing and shelter allocation strategies for humanitarian logistics operations [16].

Beyond operational planning, a substantial body of research has focused on predicting human behaviour during evacuation, as evacuee decisions significantly influence evacuation efficiency and congestion levels. Studies have applied a wide range of modelling techniques to capture evacuation behaviour, including statistical models, simulation methods, machine learning algorithms, and optimization-based frameworks [13,17]. Among these approaches, agent-based simulation (ABS) has emerged as one of the most widely used methods for modelling human behaviour in evacuation scenarios because it represents each evacuee as an autonomous agent capable of interacting with the environment and other agents [18]. This approach allows researchers to capture emergent phenomena such as crowd congestion, panic behaviour, and route selection under uncertainty. In evacuation contexts, agent-based models have been used to simulate crowd dynamics during emergencies, accounting for spatial interactions among evacuees and their environment [19]. Recent methodological developments have further integrated optimization techniques with agent-based epidemic simulations. For instance, Akopov [20] proposed a hybrid multi-swarm particle swarm optimization algorithm to improve the calibration and solution efficiency of agent-based epidemiological models. More integrated simulation approaches such as system dynamics (SD) and agent-based modelling (ABM) have further expanded the analytical toolkit for studying complex crisis scenarios. SD models capture feedback relationships between epidemiological processes, policy interventions, and social responses. For example, Jia, Li, and Fang [21] developed a SD framework to analyze COVID-19 prevention and control strategies, demonstrating how policy measures dynamically influence infection spread and healthcare capacity over time.

In the context of epidemics and pandemics, researchers have employed dynamic epidemiological models to study disease spread and its interaction with population behaviour. One of the most widely used frameworks is the susceptible–infected–recovered (SIR) model, which divides the population into three compartments—susceptible (S), infected (I), and recovered (R)—and models disease transmission through a system of differential equations. The SIR model and its extensions have been widely used to forecast infection trajectories and evaluate the potential impact of public health interventions during pandemics [22,23].

Despite these methodological advances, most evacuation behaviour studies continue to focus on immediate evacuation scenarios associated with natural disasters or technological hazards. Comparatively fewer studies examine evacuation behaviour during prolonged or slow-onset crises, such as public health emergencies or pandemics, where mobility decisions are influenced not only by immediate safety concerns but also by economic vulnerability, mobility restrictions, and uncertainty regarding future conditions. The COVID-19 pandemic generated unprecedented disruptions to mobility and livelihoods worldwide. In India, lockdown measures triggered large-scale reverse migration of workers from urban centres to their native places, highlighting the need to extend traditional evacuation frameworks to incorporate socio-economic constraints, public health risks, and migration dynamics [24].

Within the evacuation literature, relatively few studies have specifically examined transport mode choice behaviour during emergency relocation. Duan et al. [25] employed a multinomial logit model to analyze determinants influencing evacuation mode selection in Xi’an, emphasizing the role of spatial accessibility and household characteristics during short-notice evacuation scenarios. Similarly, Suman et al. [26] utilized multinomial logistic regression to explore how improvements in public transport systems may encourage shifts from private to public transport during evacuation. Borowski and Stathopoulos [27] examined the potential role of ride-sourcing services in facilitating small-scale urban evacuation through a survey-based case study. Table 1 summarizes this recent literature and also presents their shortcomings or outcomes.

Few studies have explored related aspects of evacuation decision-making and mobility behaviour. Rahman and Baker [28] analyzed mode-switching behaviour using multinomial logistic regression in the context of infrastructure development, while McCaffrey et al. [29] investigated factors influencing evacuation decisions during wildfires in the United States. Toledo et al. [30] also examined travel behaviour during wildfire evacuation events. In hurricane evacuation contexts, Bian et al. [31] and Sadri et al. [32] applied nested logit models to estimate evacuation mode choice demand and analyze traveller behaviour. Additionally, Yin et al. [33] identified determinants affecting vehicle choice during hurricane evacuations, while Liu et al. [34] proposed a framework to analyze household gathering behaviour and mode choice decisions during no-notice evacuations. Comprehensive reviews further highlight emerging research directions in evacuation modelling, including the integration of advanced technologies, improved behavioural modelling techniques, and consideration of vulnerable populations with specific mobility or medical needs during evacuation [10].

Despite the substantial body of literature on evacuation planning and modelling, several important gaps remain. First, most existing studies focus on evacuation behaviour in the context of sudden-onset natural disasters, whereas comparatively limited attention has been given to evacuation dynamics during prolonged public health crises such as pandemics. Pandemic-induced mobility differs fundamentally from traditional disaster evacuation because decisions are shaped not only by immediate safety concerns but also by economic vulnerability, mobility restrictions, and uncertainty regarding disease transmission [35,36]. Second, while numerous studies examine evacuation decision-making or transport mode choice independently, relatively few have investigated the joint decision-making process involving both transport mode and shelter selection, which are inherently interconnected during crisis relocation. Third, although simulation-based approaches such as ABM and SD have advanced the understanding of epidemic spread and crowd dynamics, empirical studies that estimate behavioural determinants of evacuation decisions using primary survey data remain limited, particularly in developing-country contexts. The large-scale reverse migration observed in India during the COVID-19 lockdown presents a unique opportunity to examine evacuation behaviour under pandemic conditions.

Table 1. Recent studies addressing evacuation.

Reference	Study Region	Context/Hazard	Methodology	Limitation/Research Gap
Borowski & Stathopoulos [27] examine the role of ride-sourcing platforms in evacuation transportation decisions	USA	Urban emergency evacuation	Discrete choice model using stated preference survey	Focuses on ride-sourcing adoption rather than broader evacuation mobility or shelter decisions
Phiophuead & Kunsuwan [37] identify socio-demographic and travel-related determinants affecting evacuation mode choice	Thailand	Flood and landslide evacuation	Binary logistic regression	Focuses only on government vs. private vehicles and does not consider shelter decisions
Urbane et al. [38] examine optimal evacuation mode split and role of information in evacuation behaviour	Not location-specific	Rapid-onset disasters (tsunami and wildfire)	Simulation and empirical comparison	Focuses on simulation outcomes rather than empirical behavioural modelling
Thakur et al. [39] analyse socio-demographic and risk-related determinants affecting evacuation decisions	Auckland, New Zealand	Volcanic eruption evacuation	Logistic regression with stated preference survey	Focuses primarily on evacuation decisions rather than detailed mode and shelter choice interactions
Xu et al. [40] investigate transportation mode choice during return phase after evacuation	USA	Post-disaster return mobility	Logistic regression, machine learning, causal inference	Focuses on post-disaster return rather than evacuation decisions themselves
Lu et al. [41] explore heterogeneous travel mode choices during unexpected transport disruptions	Xi’an, China	Urban transport disruptions	Mixed logit model	Focuses on transit disruption rather than disaster evacuation contexts
This Study	India	Mode choice during pandemic	Multinomial logistic regression	Addresses evacuation choice in prolonged crisis such as pandemic

Table 1. Recent studies addressing evacuation.

Reference	Study Region	Context/Hazard	Methodology	Limitation/Research Gap
Borowski & Stathopoulos [27] examine the role of ride-sourcing platforms in evacuation transportation decisions	USA	Urban emergency evacuation	Discrete choice model using stated preference survey	Focuses on ride-sourcing adoption rather than broader evacuation mobility or shelter decisions
Phiophuead & Kunsuwan [37] identify socio-demographic and travel-related determinants affecting evacuation mode choice	Thailand	Flood and landslide evacuation	Binary logistic regression	Focuses only on government vs. private vehicles and does not consider shelter decisions
Urbane et al. [38] examine optimal evacuation mode split and role of information in evacuation behaviour	Not location-specific	Rapid-onset disasters (tsunami and wildfire)	Simulation and empirical comparison	Focuses on simulation outcomes rather than empirical behavioural modelling
Thakur et al. [39] analyse socio-demographic and risk-related determinants affecting evacuation decisions	Auckland, New Zealand	Volcanic eruption evacuation	Logistic regression with stated preference survey	Focuses primarily on evacuation decisions rather than detailed mode and shelter choice interactions
Xu et al. [40] investigate transportation mode choice during return phase after evacuation	USA	Post-disaster return mobility	Logistic regression, machine learning, causal inference	Focuses on post-disaster return rather than evacuation decisions themselves
Lu et al. [41] explore heterogeneous travel mode choices during unexpected transport disruptions	Xi’an, China	Urban transport disruptions	Mixed logit model	Focuses on transit disruption rather than disaster evacuation contexts
This Study	India	Mode choice during pandemic	Multinomial logistic regression	Addresses evacuation choice in prolonged crisis such as pandemic

Addressing these gaps, the present study develops a multinomial logistic regression framework to analyze the determinants of transport mode and shelter choice decisions among migrants during the COVID-19-induced evacuation in India, thereby contributing empirical insights to the literature on evacuation behaviour and crisis mobility planning. In addition to the empirical studies discussed above, evacuation decision-making can be explained through Random Utility Theory (RUT), which underpins discrete choice models such as the multinomial logit model used in this study [42]. According to RUT, individuals select the alternative that maximizes their perceived utility under existing economic and accessibility constraints. Accordingly, the explanatory variables used in this study capture key determinants of evacuation behaviour. Socio-demographic characteristics (age, gender, marital status, education) reflect heterogeneity in mobility needs and risk perception during emergencies [43]. Economic variables (employment type and household income) represent financial capacity and vulnerability during crisis relocation. Accessibility-related factors, including vehicle ownership and current commuting mode, determine the feasible transport options available during evacuation. Residential context and social conditions, such as living arrangements and duration of stay, influence access to support networks and shelter alternatives, while previous disaster experience captures preparedness and risk awareness in evacuation decision-making [44]. Together, these variables provide the theoretical basis for analyzing transport mode and shelter choice decisions during evacuation.

3. Materials and Methods

3.1. Study Area and Data

The Delhi national capital region was selected as the case study for this research. Delhi is a densely populated megacity with 6 satellite cities housing a large number of migrants at almost every stratum of economic classification. The target respondents were suitable for this study as most of them have a native place of belonging and desire to go there in case of crisis. Most of these people reside within the national capital region of Delhi. Respondents were recruited using a purposive sampling approach targeting migrants who had been residing in Delhi and undertook or planned to undertake evacuation during the COVID-19 lockdown of 2020. Inclusion criteria required respondents to (a) be currently or formerly residing in Delhi as a migrant and (b) have experienced or anticipated a need for evacuation during the lockdown period. A pilot test of the questionnaire was conducted with 25 respondents prior to full deployment to check clarity and reliability of items. Then, a structured questionnaire was administered to migrants residing in Delhi who had evacuation experience during the COVID-19 lockdown. Data were collected between April 2020 and October 2020 through an online survey platform. A total of 300 questionnaires were distributed, of which 119 valid responses were received, yielding a response rate of 39.66%. After removing options with fewer than 10 responses (following the minimum cell-size convention for stable multinomial logistic regression estimation [45]), the final analytical sample comprised 119 respondents with 4 transport mode categories and 4 shelter type categories.

3.2. Outcome Variables

Evacuation mode and shelter type were chosen as the outcome variables for this study to investigate the evacuees’ behaviour during disease outbreak. Respondents were asked to indicate the type of transport mode they prefer for evacuation and the shelter type they want to use. Eight types of transport modes were identified for evacuation: bus, carpool, train, car, barge, aircraft, personal bike and ride sourcing. However, for barge, ride sourcing, carpool and train, only one, two, three and eight responses were given respectively. Thus, they were removed from analysis. For the shelter type, five options were given: public shelter, second residence, portable vehicles or customized vehicles, hotels and hospitals. Out of the five shelter types, hospitals received only 8 responses. Any option of mode that received less than 10 responses was removed from the analysis. Thus, a total of 4 evacuation transport modes and 4 shelter types were evaluated in this study.

3.3. Explanatory Variables

A comprehensive literature review was performed to identify the factors influencing the evacuees’ behaviours. A total of 12 factors were analyzed in this study: age group, gender, marital status, employment type, education level, previous disaster experience, transport mode preference for day-to-day commute, vehicle ownership, residing with, duration of stay at present residence and annual family income. The list of explanatory variables and their classification is presented in

Table 2. Classification of variables with coding.

Explanatory Variable	Classification	Coding
Age group	<15	1
	15–30	2
	31–45	3
	46–60	4
	Above 60	5
Gender	Female	1
	Male	2
Marital Status	Single	1
	Married	2
	Widowed	3
Employment Type	Full-Time	1
	Part-Time	2
	Student	3
	Unemployed	4
	Retired	5
	Self-Employed	6
	Other	7
Education Level	Below Senior Secondary	1
	Senior Secondary	2
	Graduate	3
	Post-Graduate	4
	Doctoral	5
Previous Disaster Experience	Yes	1
	No	2
Current mode preference for daily commute	Car	1
	Bus	2
	Carpool	3
	Train or metro	4
	Motorcycle	5
	Cycle	6
	Walk	7
	Others	8
Vehicle Ownership	Yes	1
	No	2
Residing with	Family	1
	Friend	2
	Alone	3
Duration of Stay	Less than 6 months	1
	6 month–2 years	2
	2 years–5 years	3
	5 years–10 years	4
	More than 10 years	5
Annual Family Income	Less than 2 lacs	1
	2–5 lacs	2
	5–10 lacs	3
	More than 10 lacs	4
Transport mode	Bus	1
	Car	2
	Aircraft	3
	Personal bike	4
Shelter types	Public shelter	1
	Second Residence	2
	Hotels	3
	Portable or customized vehicles	4

.

A total of 11 explanatory variables spanning demographic, socio-economic, and behavioural dimensions were considered, along with the outcome variables—transport mode and shelter type. The portions of respondents per variable are presented in Figure 1 and Figure 2.

3.4. Analytical Method

Multinomial logistic regression (MLR) was employed to model evacuees’ transport mode and shelter type choices. MLR is appropriate when the dependent variable is nominal with more than two unordered categories—a condition met by both outcome variables in this study (four transport modes; four shelter types) [45]. The method has been applied in comparable evacuation studies: Duan et al. [25] used MNL for mode selection during emergency evacuation; Rahman and Baker [28] applied MNL to mode switch behaviour; and Bian et al. [31] used nested logit, a related framework for hurricane evacuation mode and destination choice. The general mathematical form of the multinomial logistic regression model is expressed as

$$U_{jn}={\beta }_{0j}+{\beta }_{1j}{X}_{1n}+{\beta }_{2j}{X}_{2n}+\dots +{\beta }_{nj}{X}_{mn}$$

(1)

where

• $U_{jn}$ = utility function for individual n for mode j;
• ${\beta }_{0j}$ = alternative specific constant for mode j;
• ${\beta }_{1j}, {\beta }_{2j}, {\beta }_{3j}, \dots {\beta }_{nj}$ = regression coefficients associated with explanatory variables;
• $X_{1n}, X_{2n}, \dots , X_{mn}$ = explanatory variables m for individual n.

3.5. Hypothesis Development

Prior studies have established that socio-economic factors such as income, vehicle ownership, and employment status significantly influence transport mode decisions during disaster evacuations [31,32]. This hypothesis tests whether similar dynamics hold in the context of a pandemic-induced displacement event among urban migrants in India. Based on that, the following hypotheses are proposed to examine the relationship between socio-economic characteristics of migrants and their evacuation decisions:

H0₁: 

There is no significant relationship between the socio-economic characteristics of migrants and their transport mode choice during COVID-19-induced evacuation.

HA₁: 

There is a significant relationship between the socio-economic characteristics of migrants and their transport mode choice during COVID-19-induced evacuation.

Shelter choice during crises has been shown to be shaped by demographic and socio-economic characteristics including age, gender, household composition, and education level [34,46]. This hypothesis tests whether such characteristics similarly determine shelter preference among migrant workers evacuating during a public health emergency. This results in the following hypothesis:

H0₂: 

There is no significant relationship between the socio-economic characteristics of migrants and their shelter type preference during COVID-19-induced evacuation.

HA₂: 

There is a significant relationship between the socio-economic characteristics of migrants and their shelter type preference during COVID-19-induced evacuation.

4. Results and Discussion

Evacuation mode and shelter type were chosen as the outcome variables for this study to investigate the evacuees’ behaviour during disease outbreak. Two models were analyzed using SPSS software (Version 20) to understand evacuees’ behaviour.

4.1. Mode Choice Model

The transport mode choice model evaluates how socio-economic characteristics of migrants influence their preference for one of four evacuation transport modes, bus, car, aircraft, and personal bike, with bus serving as the reference category. To assess whether the explanatory variables collectively improve prediction over a baseline model, the likelihood ratio test was performed. As shown in

Table 3. Model fitting information (mode choice model).

Model Fitting Information
Model	Model Fitting Criteria	Likelihood Ratio Tests
	−2 Log Likelihood	Chi-Square	df	Sig.
Intercept Only	239.733
Final	175.540	64.193	36	0.003

, the intercept-only model yielded a −2 log likelihood of 239.733, which reduced to 175.540 in the final model, producing a chi-square value of 64.193 with 36 degrees of freedom (p = 0.003). Since this value is well below the 0.05 threshold, the final model significantly outperforms the null model and the alternate hypothesis HA1 is accepted—confirming a statistically significant relationship between the socio-economic characteristics of migrants and their transport mode preference during the COVID-19 evacuation.

The goodness-of-fit statistics further validate the model. As presented in

Table 4. Goodness-of-fit (mode choice model).

Goodness-of-Fit				Pseudo R-Square
	Chi-Square	df	Sig.	Cox and Snell	0.431
Pearson	277.083	300	0.825	Nagelkerke	0.490
Deviance	175.540	300	1.000	McFadden	0.268

, the Pearson chi-square (p = 0.825) and Deviance (p = 1.000) significance values both exceed 0.05, confirming that the model adequately represents the observed data. The pseudo R² values: Cox and Snell = 0.431, Nagelkerke = 0.490, and McFadden = 0.268, indicate that the socio-economic variables explain approximately 43–49% of the variance in transport mode choice, which is considered satisfactory for behavioural discrete choice models of this nature. In terms of classification accuracy,

Table 5. Classification of variables for mode choice.

Classification
Observed	Predicted
	Bus	Car	Aircraft	Personal Bike	Percent Correct
Bus	4	5	0	2	36.4%
Car	2	63	1	6	87.5%
Aircraft	0	8	3	1	25.0%
Personal Bike	1	11	0	7	36.8%
Overall Percentage	6.1%	76.3%	3.5%	14.0%	67.5%

shows that the model correctly classified 67.5% of all cases. Car was predicted with the highest accuracy (87.5%), accounting for 76.3% of all predicted observations, while personal bike (36.8%), bus (36.4%), and aircraft (25.0%) were predicted with comparatively lower accuracy, likely reflecting smaller cell sizes and greater socio-economic heterogeneity among respondents selecting these modes.

4.2. Shelter Choice Model

The shelter choice model examines the socio-economic determinants of migrants’ preference among four shelter types, public shelter, second residence, hotels, and portable or customized vehicles, with portable or customized vehicles serving as the reference category. As shown in

Table 6. Model fitting information (shelter choice model).

Model Fitting Information
Model	Model Fitting Criteria	Likelihood Ratio Tests
	−2 Log Likelihood	Chi-Square	df	Sig.
Intercept Only	247.326
Final	172.595	74.731	33	0.000

, the likelihood ratio test yielded a chi-square value of 74.731 with 33 degrees of freedom (p = 0.000), indicating an even stronger overall model fit than the mode choice model and confirming acceptance of the alternate hypothesis HA2—that a statistically significant relationship exists between socio-economic characteristics and shelter type preference.

The goodness-of-fit statistics reported in

Table 7. Goodness-of-fit (shelter choice model).

Goodness-of-Fit				Pseudo R-Square
	Chi-Square	df	Sig.	Cox and Snell	0.466
Pearson	333.310	318	0.266	Nagelkerke	0.533
Deviance	172.595	318	1.000	McFadden	0.302

show that the Deviance significance value of 1.000 comfortably exceeds the 0.05 threshold, confirming adequate model fit, while the Pearson significance value of 0.266 also remains above the threshold. The pseudo R² values: Cox and Snell = 0.466, Nagelkerke = 0.533, and McFadden = 0.302—are marginally higher than those of the mode choice model, indicating that socio-economic variables explain shelter preference with comparatively greater power, accounting for approximately 47–53% of variance. Regarding classification,

Table 8. Classification of variables for shelter choice.

Classification
Observed	Predicted
	Public Shelter	Second Residence	Hotels	Portable or Customized Vehicles	Percent Correct
Public Shelter	6	8	0	0	42.9%
Second Residence	1	72	1	2	94.7%
Hotels	1	12	5	1	26.3%
Portable or Customized Vehicles	0	5	0	5	50.0%
Overall Percentage	6.7%	81.5%	5.0%	6.7%	73.9%

shows that the shelter choice model achieved an overall accuracy of 73.9%, outperforming the mode choice model. Second residence was predicted with the highest accuracy (94.7%), while portable or customized vehicles (50.0%), public shelter (42.9%), and hotels (26.3%) were predicted with progressively lower accuracy. The low predictive accuracy for hotels suggests considerable heterogeneity among hotel-preferring respondents, indicating that this preference is driven by factors beyond the socio-economic variables captured in this study.

4.3. Discussion

The parameter estimates from both models, presented in

Table 9. Parameter estimates for mode choice.

Parameter Estimates
Mode a		B	Std. Error	Wald	df	Sig.	Exp(B)	95% Confidence Interval for Exp(B)
								Lower Bound	Upper Bound
car	Intercept	1.944	2.297	0.716	1	0.397
	Age	0.999	0.823	1.474	1	0.225	2.716	0.541	13.621
	Durations	0.194	0.397	0.240	1	0.625	1.214	0.558	2.644
	Income	0.849	0.359	5.586	1	0.018	2.337	1.156	4.725
	Education level	−0.210	0.466	0.204	1	0.651	0.810	0.325	2.019
	Emp. type	0.147	0.296	0.248	1	0.619	1.159	0.649	2.069
	Daily commute mode	−0.188	0.202	0.868	1	0.352	0.829	0.558	1.230
	Residing with	−0.886	0.526	2.837	1	0.092	0.412	0.147	1.156
	Marital status	−1.394	1.137	1.502	1	0.220	0.248	0.027	2.306
	[Gender = 1.00]	1.046	1.209	0.749	1	0.387	2.846	0.266	30.402
	[Gender = 2.00]	0 b			0
	[prev. disas. exp = 1.00]	−0.629	1.432	0.193	1	0.660	0.533	0.032	8.826
	[prev. disas. exp = 2.00]	0 b			0
	[Vehicle ownership = 1.00]	−0.099	0.858	0.013	1	0.908	0.905	0.168	4.866
	[Vehicle ownership = 2.00]	0 b			0
aircraft	Intercept	−2.314	3.289	0.495	1	0.482
	Age	0.710	1.009	0.496	1	0.481	2.035	0.282	14.708
	Durations	0.186	0.486	0.146	1	0.702	1.204	0.464	3.123
	Income	0.993	0.425	5.466	1	0.019	2.700	1.174	6.209
	Education level	0.475	0.648	0.537	1	0.464	1.607	0.452	5.722
	Emp. type	−0.083	0.378	0.048	1	0.827	0.921	0.439	1.931
	Daily commute mode	−0.347	0.266	1.708	1	0.191	0.707	0.420	1.189
	Residing with	−0.093	0.654	0.020	1	0.887	0.911	0.253	3.284
	Marital status	−1.652	1.383	1.427	1	0.232	0.192	0.013	2.883
	[Gender = 1.00]	−0.234	1.599	0.021	1	0.884	0.792	0.034	18.170
	[Gender = 2.00]	0 b			0
	[prev. disas. exp = 1.00]	1.465	1.633	0.804	1	0.370	4.326	0.176	106.206
	[prev. disas. exp = 2.00]	0 b			0
	[Vehicle ownership = 1.00]	−0.238	1.113	0.046	1	0.830	0.788	0.089	6.978
	[Vehicle ownership = 2.00]	0 b			0
personal bike	Intercept	3.172	2.716	1.364	1	0.243
	Age	1.426	0.918	2.414	1	0.120	4.161	0.689	25.143
	Durations	0.357	0.447	0.636	1	0.425	1.429	0.595	3.433
	Income	−0.203	0.456	0.199	1	0.655	0.816	0.334	1.993
	Education level	−0.452	0.517	0.764	1	0.382	0.636	0.231	1.754
	Emp. type	−0.354	0.331	1.144	1	0.285	0.702	0.367	1.343
	Daily commute mode	0.151	0.238	0.403	1	0.525	1.163	0.730	1.853
	Residing with	−1.063	0.626	2.886	1	0.089	0.345	0.101	1.178
	Marital status	−2.216	1.339	2.739	1	0.098	0.109	0.008	1.504
	[Gender = 1.00]	0.452	1.368	0.109	1	0.741	1.572	0.108	22.967
	[Gender = 2.00]	0 b			0
	[prev. disas. exp = 1.00]	0.102	1.530	0.004	1	0.947	1.107	0.055	22.195
	[prev. disas. exp = 2.00]	0 b			0
	[Vehicle ownership = 1.00]	0.771	0.969	0.633	1	0.426	2.162	0.324	14.431
	[Vehicle ownership = 2.00]	0 b			0

^a The reference category is bus. ^b This parameter is set to zero because it is redundant.

and

Table 10. Parameter estimates for shelter choice.

Parameter Estimates
Shelter Type a		B	Std. Error	Wald	df	Sig.	Exp(B)	95% Confidence Interval for Exp(B)
								Lower Bound	Upper Bound
public shelter	Intercept	8.847	5.111	2.997	1	0.083
	Income	−0.722	0.533	1.833	1	0.176	0.486	0.171	1.381
	Durations	−1.094	0.539	4.113	1	0.043	0.335	0.116	0.964
	Emp. type	0.073	0.399	0.034	1	0.854	1.076	0.492	2.353
	Daily commute mode	−0.464	0.333	1.946	1	0.163	0.629	0.328	1.207
	Prev. disas. Exp.	−1.080	1.720	0.394	1	0.530	0.340	0.012	9.883
	Age	4.798	2.250	4.547	1	0.033	121.286	1.474	9981.222
	Residing with	1.540	0.859	3.213	1	0.073	4.667	0.866	25.149
	Education level	−2.577	0.826	9.725	1	0.002	0.076	0.015	0.384
	Marital status	−2.086	1.677	1.547	1	0.214	0.124	0.005	3.323
	[Gender = 1.00]	−4.389	1.649	7.082	1	0.008	0.012	0.000	0.315
	[Gender = 2.00]	0 b			0
	[Vehicle ownership = 1.00]	−1.406	1.403	1.005	1	0.316	0.245	0.016	3.830
	[Vehicle ownership = 2.00]	0 b			0
second residence	Intercept	2.523	4.065	0.385	1	0.535
	Income	0.053	0.391	0.019	1	0.892	1.055	0.490	2.271
	Durations	−0.567	0.404	1.971	1	0.160	0.567	0.257	1.252
	Emp. type	0.020	0.333	0.004	1	0.953	1.020	0.531	1.959
	Daily commute mode	−0.680	0.271	6.302	1	0.012	0.507	0.298	0.861
	Prev. disas. exp	1.129	1.446	0.609	1	0.435	3.093	0.182	52.660
	Age	5.368	2.130	6.352	1	0.012	214.393	3.299	13,934.631
	Residing with	0.676	0.698	0.937	1	0.333	1.966	0.500	7.731
	Education level	−1.633	0.719	5.156	1	0.023	0.195	0.048	0.800
	Marital status	−2.908	1.414	4.228	1	0.040	0.055	0.003	0.873
	[Gender = 1.00]	−1.758	1.062	2.744	1	0.098	0.172	0.022	1.380
	[Gender = 2.00]	0 b			0
	[Vehicle ownership = 1.00]	−0.328	1.049	0.098	1	0.754	0.720	0.092	5.629
	[Vehicle ownership = 2.00]	0 b			0
hotels	Intercept	−2.371	4.924	0.232	1	0.630
	Income	0.332	0.419	0.627	1	0.428	1.394	0.613	3.169
	Durations	−0.975	0.470	4.308	1	0.038	0.377	0.150	0.947
	Emp. type	−0.067	0.378	0.031	1	0.860	0.936	0.446	1.962
	Daily commute mode	−0.739	0.289	6.564	1	0.010	0.477	0.271	0.841
	Prev. disas. exp	2.379	1.827	1.697	1	0.193	10.799	0.301	387.393
	Age	4.673	2.187	4.564	1	0.033	107.034	1.471	7787.809
	Residing with	0.633	0.755	0.704	1	0.401	10.884	0.429	8.267
	Education level	−1.082	0.773	1.955	1	0.162	0.339	0.074	1.544
	Marital status	−1.901	1.490	1.628	1	0.202	0.149	0.008	2.771
	[Gender = 1.00]	−2.534	1.197	4.483	1	0.034	0.079	0.008	0.828
	[Gender = 2.00]	0 b			0
	[Vehicle ownership = 1.00]	0.057	1.168	0.002	1	0.961	1.059	0.107	10.450
	[Vehicle ownership = 2.00]	0 b			0

^a The reference category is portable or customized vehicles. ^b This parameter is set to zero because it is redundant.

, respectively, reveal a clear divergence in the socio-economic drivers of transport mode versus shelter choice. In the mode choice model, income emerged as the sole statistically significant predictor. It positively and significantly influenced both car choice (B = 0.849, p = 0.018, Exp(B) = 2.337, 95% CI [1.156, 4.725]) and aircraft choice (B = 0.993, p = 0.019, Exp(B) = 2.700, 95% CI [1.174, 6.209]) relative to bus. This indicates that with each unit increase in the income category, the odds of choosing car over bus increase by approximately 2.3 times and aircraft by approximately 2.7 times. In practical terms, migrants with higher annual family income demonstrated a consistent and statistically robust preference for private or premium transport over bus which is consistent with hurricane evacuation studies [31], where income and vehicle access have been established as the primary determinants of private mode preference. No other variable reached significance in the mode choice model, establishing income as the singular economic gateway to transport choice during the COVID-19 exodus.

The shelter choice model (Table 10) reveals a substantially richer set of significant predictors. Age was a strong positive predictor of public shelter (B = 4.798, p = 0.033, Exp(B) = 121.286), second residence (B = 5.368, p = 0.012, Exp(B) = 214.393), and hotels (B = 4.673, p = 0.033, Exp(B) = 107.034), all relative to portable vehicles. The notably large odds ratios indicate that older migrants have substantially higher odds of preferring any fixed shelter category over portable vehicles, likely due to established social ties, family networks at native places, and greater risk aversion to improvised arrangements. Education level was a significant negative predictor of public shelter (B = −2.577, p = 0.002, Exp(B) = 0.076) and second residence (B = −1.633, p = 0.023, Exp(B) = 0.195), suggesting that more highly educated migrants actively avoid institutional shelter options and lean toward portable or self-organized alternatives which reflects greater resourcefulness and preference for autonomy.

Gender emerged as one of the most policy-critical findings in the study. Female respondents showed dramatically reduced odds of choosing public shelter (B = −4.389, p = 0.008, Exp(B) = 0.012) and hotels (B = −2.534, p = 0.034, Exp(B) = 0.079) relative to portable vehicles. The Exp(B) of 0.012 for public shelter implies that female migrants have approximately 98% lower odds of selecting public shelter compared to male migrants—a finding that quantitatively corroborates well-documented concerns about safety, privacy, and dignity barriers in Indian emergency shelters [4,46]. This result has direct implications for shelter policy: gender-sensitive design and management of public shelters must be treated as an empirical necessity rather than an optional welfare provision. Duration of stay was a significant negative predictor of both public shelter (B = −1.094, p = 0.043, Exp(B) = 0.335) and hotels (B = −0.975, p = 0.038, Exp(B) = 0.377), indicating that longer-term Delhi residents are less likely to resort to institutional sheltering, possibly because stronger local networks and familiarity with the city enable more self-organized, flexible arrangements. Marital status negatively predicted second residence choice (B = −2.908, p = 0.040, Exp(B) = 0.055), suggesting that married migrants face greater logistical constraints in relocating to a second residence, possibly owing to the complexity of family-level coordination during crisis. Finally, daily commute mode significantly reduced the odds of both second residence (B = −0.680, p = 0.012, Exp(B) = 0.507) and hotel choice (B = −0.739, p = 0.010, Exp(B) = 0.477), pointing to the relevance of habitual mobility behaviour in shaping crisis shelter decisions, a dimension that has received limited attention in prior evacuation literature [10].

Notably, income did not significantly predict any shelter category, in sharp contrast to its central role in mode choice. This divergence suggests that transport and shelter decisions are governed by distinct dimensions of vulnerability: income directly determines access to premium or private transport, while shelter decisions are more strongly structured by social and demographic factors—age, gender, education, and household composition. This has important policy implications: targeted transport support for low-income migrants (such as subsidized bus capacity) and targeted shelter improvements for specific demographic groups (gender-sensitive facilities, age-appropriate provisions) should be designed as complementary but distinct policy interventions, rather than as a single undifferentiated emergency response. The overwhelming dominance of second residence as the preferred and most accurately predicted shelter option (94.7%) further reflects the critical role of social capital and family networks in evacuation destination choice [34], underscoring that government-provided public shelters are perceived as a last resort, and that improving their uptake requires directly addressing the barriers of safety, gender, and accessibility identified in this study.

4.4. Managerial and Policy Implications

The findings of this study offer several actionable implications for emergency planners, transport authorities, and policymakers. Since income is the primary determinant of transport mode choice, authorities should ensure dedicated and adequately capacitated bus services for low-income migrants during any future lockdown or large-scale crisis event, especially those earning below ₹2 lakhs annually, as this segment has the least access to private or premium transport alternatives. The strong preference for second residence as a shelter destination suggests that governments should consider forward-staging of essential supplies, medical support, and relief materials along major interstate migration corridors. This can be done majorly for states including Uttar Pradesh, Bihar, and Rajasthan rather than concentrating resources at urban transit points. The near-complete avoidance of public shelters by female migrants demands urgent attention: shelter planners must incorporate gender-sensitive design, including separate sanitation facilities, secure enclosures, and female staff deployment, to improve uptake among women. Finally, since longer-duration Delhi residents tend to self-organize rather than use institutional shelters, a pre-registered migrant database, capturing socio-economic profiles, native place, and household composition, would enable authorities to anticipate demand. This would further facilitate planning targeted interventions and avoid the reactive, ad hoc response that characterized India’s handling of the COVID-19 migrant evacuation.

5. Conclusions

COVID-19 outbreaks have spread quickly around the world, resulting in lockdowns across the globe to reduce contact rates. During lockdown, many lost jobs, food, and other basics. Masses fled to their hometowns and other safe places. Tourists, students, migrant labourers, or employees who lost their jobs in India have limited options, which is complex for both the government and evacuees. This study investigates evacuee behaviour in Delhi, focusing on evacuation decisions regarding shelter and transport mode choice. The study adopted multinomial logistic regression to analyze the choices of respondents.

This study provides empirical evidence that socio-economic characteristics significantly determine both transport mode and shelter choices of migrants during pandemic-induced evacuation. Income is the strongest predictor of mode choice; higher-income migrants strongly prefer car and aircraft over bus, while gender and education level are critical determinants of shelter preference. These findings confirm that evacuation behaviour is not uniform and that one-size-fits-all emergency logistics planning will systematically disadvantage low-income and female migrants. The results of this study suggest leveraging several modes of transportation to maintain social distancing and other basic requirements to avoid the transmission of COVID-19. According to population demographics, future evacuation planners can derive the appropriate fleet plan using the outcomes of this study. This study can give direction to policymakers for better planning considering socio-economic variables of migrants during such a disease outbreak in future. As an example, for emergency transport planners, subsidized or reserved bus capacity for migrants in the lowest income bracket (<2 lakhs annual income) should be incorporated into pandemic evacuation protocols. For shelter authorities, public shelters must address safety perceptions, particularly for female evacuees, since the results show women strongly avoid public shelter options. For policymakers, a pre-registered migrant database in Delhi (and other NCR cities) would enable needs-based resource allocation during future crises

This study has several limitations. First, the sample is restricted to Delhi NCR; results may not generalize to other Indian cities with different demographic profiles. Second, data were collected via an online survey, potentially underrepresenting migrants with limited digital access, precisely the most economically vulnerable group. Third, the cross-sectional design prevents causal inference. Fourth, the sample size, while adequate for MNL estimation, limits subgroup analyses. Future research should employ larger, multi-city samples and consider longitudinal designs to track behavioural change across pandemic phases.