Improving Migrant Tracking Through Household Incentives: Evidence from a Randomized Controlled Trial in Thailand and Vietnam

This study investigates how low-cost, network-informed strategies can reduce attrition in follow-up surveys of rural-to-urban migrants. Attrition and non-contact are persistent challenges not only in longitudinal panel studies, but also in cross-sectional data collection—especially when targeted respondents are mobile, difficult to locate, or sampled via intermediaries. In such cases, the sampled population may diverge from the interviewed population due to failure to reach individuals after the sampling stage, introducing potential bias, reducing external validity, and decreasing the final sample size—ultimately undermining the statistical power of the analysis. This issue is particularly acute in surveys of mobile or displaced populations, such as internal migrants, refugees, seasonal workers, and residents of informal settlements.

We explore whether leveraging the social ties between migrants and their rural origin households can improve response rates in tracking surveys. Migrants are often still connected to their households of origin, both socially and financially. These connections present an opportunity: if households are motivated to support researchers in reaching migrants, this may improve the success rate of follow-up data collection.
The study also has broader implications for all data collection efforts that rely on network-based sampling approaches such as snowball sampling and respondent-driven sampling (RDS), where respondent cooperation and accurate referral are essential for reaching eligible individuals.
To test this idea, we implement two types of interventions targeting the migrants’ rural households:

1. A conditional incentive, disbursed ex-post: Households are informed that they will receive a monetary reward if the researchers are able to successfully reach and interview the migrant. This incentive is conditional on both the household providing accurate contact information and the migrant ultimately completing the follow-up survey. The goal is to encourage households to actively assist in identifying valid contact details and encouraging the migrant’s cooperation.

2. An unconditional communication incentive, disbursed ex-ante: In a separate intervention, households receive a small airtime top-up for their mobile phone. They are encouraged to use this airtime to notify the migrant about the upcoming survey. This intervention aims to overcome common communication barriers, such as widespread reluctance to answer calls from unknown numbers due to concerns about scam calls and fraud. By prompting households to proactively inform migrants about the survey, the airtime credit helps migrants anticipate the follow-up contact, increasing their likelihood of participation.


Both interventions engage households as active partners in tracking migrants, testing whether small, network-based nudges can cost-effectively improve follow-up success. This research contributes to the survey methodology and social networks literature by exploring how family ties and informal communication can support participation. While implemented in a panel survey of internal migrants in Southeast Asia, the findings are likely generalizable to other settings and populations where social networks play a key role in locating and motivating respondents, including cross-sectional surveys that rely on third-party contact or referral.
We target household heads who, in the most recent TVSEP wave (2024), reported that at least one household member had migrated to an urban area.

The implementation follows a three-stage process. First, rural households are contacted by phone to confirm the migrant’s current location and provide updated migrant contact information. Households in treatment groups are informed about the different financial incentives during this call. Second, migrants are contacted using the phone numbers collected from their households. A randomly selected subset of migrants whose households are eligible for the ex-post incentive are informed about the financial reward available to their household upon survey completion. All migrants are then asked whether they are willing to participate in the follow-up survey. Finally, enumerators will perform face-to-face interviews with those migrants who agreed to be interviewed.
As the migrant data collection is embedded within the long-term TVSEP panel study, we can leverage historical household and migrant panel data to explore heterogeneity in intervention effects and examine how past characteristics influence survey participation.

Our survey sample will include approximately:
- 1115 migrants from 729 households in three rural provinces of Thailand (Nakhon Phanom, Ubon Ratchathani, and Buriram), representing 34.8% of the 2024 Thai household sample
- 697 migrants from 490 households in three rural provinces of Vietnam (Ha Tĩnh, Thừa Thiên Huế, and Đắk Lắk), representing 22.3% of the 2024 Vietnamese household sample

Because the experiment is being conducted in both Thailand and Vietnam—two contexts with historically different response rates and data quality among both household and migrant respondents—we are also able to descriptively compare intervention effects across settings. This allows us to explore whether the interventions are particularly effective in contexts where tracking and data quality have traditionally been more challenging, or whether they primarily reinforce already strong survey performance. While not designed for formal cross-country comparisons, these descriptive insights can inform how such strategies might be prioritized or tailored to different implementation environments.

Both Thailand and Vietnam are also settings where concerns about phone- and online-based fraud are widespread. We thus expect that some households might be hesitant to share their migrant member’s contact information, particularly when the conversation involves financial incentives. To address this, calls will be conducted by official university research teams using institutional phone lines (fixed lines), and enumerators will clearly identify themselves and their institutional affiliation at the start of each conversation. Despite these safeguards, the prevalence of scam-related mistrust may still limit the willingness of some households to disclose contact details. Importantly, the financial incentive could affect trust in both directions: it may encourage cooperation by signaling legitimacy and offering a clear benefit, but it could also raise suspicion among households wary of fraudulent schemes. We will explore whether the effectiveness of the interventions varies with households’ prior exposure to scam calls, as reported in the 2024 TVSEP wave, to better understand how trust and perceived legitimacy shape participation in this context.
An additional aspect of interest is whether, even if fewer households initially share contact information, those that perceive the research team as trustworthy (“non-scammers”) are more likely to follow the survey team’s request to notify the migrant about the upcoming interview. Such proactive communication may help reduce attrition in the second stage by ensuring that migrants are less likely to mistake the enumerators’ calls for potential scams.

The sampling stage, and thus experimental intervention is planned to start on the 10th of September 2025 in Thailand and 1st of October 2025 in Vietnam. Data collection (and thus our outcomes of interest) will be carried out between 10/2025 and 01/2026 in both Thailand and Vietnam.

Migration is widely recognized as a socially embedded process, shaped and sustained by interpersonal networks (Bashi 2007). Migrants often maintain economic and social ties between origin and destination (Levitt and Glick Schiller 2004; Lubbers et al. 2020). Yet, conventional migration surveys have largely overlooked the potential of these networks for improving data collection and sample representation (Merli et al. 2022).
A central challenge in migration research is high attrition and non-contact rates, which undermine the validity of longitudinal and tracking surveys. Internal migrants, as mobile populations, are especially prone to systematic loss and bias. For instance, in the 2010 wave of the TVSEP Migrant Tracking Survey, only 960 of 2,700 migrant-linked households were successfully interviewed —a tracking response rate of just 35%. This mirrors findings from economic field studies highlighting the difficulty of tracking highly mobile individuals over time (McKenzie & Mistiaen 2009; Barham, Macours & Maluccio 2017).
Existing studies leveraging social networks in migration research have primarily focused on using these networks to define and recruit the initial sample. This includes conventional approaches such as respondent-driven sampling (RDS) and snowball sampling, as well as newer methods like Network Sampling with Memory (NSM), which refine recruitment and the construction of representative samples through network ties (Mouw and Verdery 2022; Merli et al. 2022). However, these methods focus mostly on the sampling phase—i.e., identifying and reaching eligible respondents through their social ties.
In contrast, our experiment does not use networks to construct the sample; rather, we leverage existing social networks to support participation and minimize attrition within a predefined sample of migrants. In line with Ghimire et al. (2019), who emphasize the value of engaging origin households to improve migrant traceability and response, our approach focuses on leveraging these networks not for initial recruitment but to enhance follow-up survey participation and retention. Our focus is on assessing whether activating and incentivizing intermediaries within migrants’ social environments—specifically, family members in migrants’ rural origin households—can improve contact and response rates during longitudinal follow-up. Thus, we shift the use of social networks from sample generation to the enhancement of ongoing participation and retention, testing whether network-informed interventions can cost-effectively reduce attrition and strengthen survey data quality in hard-to-track mobile populations.
We test the effectiveness of conditional incentives and communication nudges delivered directly to network members—specifically: (1) small non-monetary rewards (e.g., phone credit top-ups/ mobile money) to encourage household members to proactively contact the migrant before survey staff attempt follow-up; (2) monetary incentives offered to the migrant’s home household, contingent on the migrant’s survey completion; and (3) a combination of both. We will also implement a migrant-level information treatment, informing the migrant about the conditional incentive offered to their household. This strategy targets persistent participation barriers—like unreachable numbers, reluctance to answer unfamiliar calls, and low motivation—by leveraging trusted social ties to facilitate re-engagement. It is grounded in evidence that respondents, especially those from stigmatized or mobile groups, are more likely to participate and respond truthfully when approached via acquaintances or friends rather than through random contact (Mouw and Verdery, 2022), and that cooperation with the survey team, which is also required from the social network side in case of a tracking survey, may be enhanced by monetary incentives (Singer and Ye, 2013).
Our experiment contributes to several strands of the literature. A large body of experimental research shows that monetary incentives—whether unconditional (prepaid) or conditional (post-completion)—significantly increase survey response rates across modes and contexts, including hard-to-reach groups. Meta-analyses (e.g., Singer 2002; Singer & Ye 2013) consistently find that (1) cash incentives are generally more effective than non-cash gifts; (2) prepaid incentives (given before participation) outperform conditional incentives; (3) incentives disproportionately increase participation among groups otherwise under-represented (e.g., minorities, lower-education groups) and in panel studies, thereby mitigating attrition and sample bias. All of these studies focus on incentives handed out to the respondent himself.
Behavioral economics explains that incentives work by activating motivations like reciprocity, compensation for time, and overcoming reluctance. Offering rural households a conditional, ex-post monetary reward for successfully facilitating the migrant’s interview targets the “compensation for time” channel, acknowledging the effort in providing updated contact details or arranging meetings. An unconditional, ex-ante phone-credit top-up/ mobile money transfer appeals to reciprocity, encouraging households to notify migrants before researchers contact them. This approach aligns with findings in development and experimental economics showing that peer referrals and network-based rewards tied to outcomes significantly boost participation and cooperation (Banerjee et al., 2013; Cai, de Janvry & Sadoulet, 2015). Here, we apply these insights to survey participation rather than program enrollment.
Finally, beyond improving response rates, understanding who is more likely to remain trackable across treatment groups allows us to assess potential biases in survey attrition. Non-contact often disproportionately excludes the most vulnerable or socially isolated migrants, distorting both descriptive findings and causal inference. In addition, in a context where phone-based scams are common, patterns of mistrust toward unfamiliar numbers may themselves contribute to selective non-contact or non-sharing of migrant contact details. Migrants or households with prior exposure to scam calls may be systematically less likely to engage with survey enumerators. If such skepticism correlates with socioeconomic status, or remittance behavior, it could generate non-random attrition that disproportionately excludes particular subgroups. By leveraging information on households’ reported exposure to scam calls in the 2024 TVSEP wave, we will examine whether mistrust-related selectivity accounts for part of the observed differences in response rates and attrition across treatment groups.



Intervention/ Treatment Arms

In the past TVSEP Migrant Tracking Surveys, we lost migrant respondents because (R1) rural households did not share the accurate phone number with us; (R2) migrants were not willing to pick up the phone of an unknown phone number, (R3) migrants were not willing to sacrifice time to participate in the interview.
This study uses a 2x2 factorial design at the household calling stage, complemented by a migrant-level information treatment, to test strategies for reducing attrition in migrant follow-up surveys:

1. an unconditional, ex-ante communication nudge in the form of a small phone credit top-up/ mobile money top-up to encourage the household to notify the migrant ahead of survey contact,
2. a conditional, ex-post monetary incentive paid to the migrant’s rural household contingent on survey completion,
3. a migrant-level treatment informing them of household incentives, in case they were announced to the household.

Both the unconditional nudge and the conditional ex-post monetary incentive are intended to address barriers R1, R2, and R3. These interventions encourage households to provide accurate contact details and to reach out to the migrant before our team calls. The motivations differ—reciprocity in the case of the ex-ante nudge, and monetary self-interest in the case of the ex-post incentive. In addition, both aim to increase the chances that households will persuade migrants to participate, making it more likely that migrants agree to spend time on the interview.
Half of the migrants in household treatment groups that receive the conditional ex-post payment will also be exposed to an information treatment during the migrant phone call. This treatment informs them that their household will receive monetary compensation upon completing the survey. The information is provided immediately before asking whether they are willing to schedule a meeting with our survey team. This design allows us to examine the impact of the information treatment on both the likelihood of agreeing to participate and the subsequent actual survey take-up. Implemented during the migrant recruitment phase, the information treatment specifically targets R3 by enhancing the migrant’s motivation to take part in the interview.
The interventions are implemented at both the household and migrant calling stages.

2025-09-10

2026-01-31

Our primary measures of interest are:

1. Household provision of valid migrant contact information
2. Number of call attempts per migrant/ household until the contact is achieved
3. Successful scheduling of the migrant interview
4. Migrant survey completion

We will construct the primary outcomes as follows:

1. Household provision of valid migrant contact information will be coded as a binary variable indicating whether the household confirms the migrant’s current location and either (i) confirms the phone number recorded during the 2024 household survey, or (ii) provides an updated phone number of the migrant during the household call. Where sample size permits, we will additionally distinguish between (a) provision of any contact information and (b) provision of valid contact information, defined as cases in which the migrant was successfully reached using the number provided by the household.

2. Number of call attempts per migrant or household until contact is achieved will be constructed as a count variable based on paradata recorded by enumerators. It captures the total number of outbound call attempts (up to 10 attempts) made to either the household or the migrant before successful contact is established. Missed calls, unanswered calls, and disconnected numbers will all be included in the count. This variable provides an operational measure of the effort required to reach respondents and serves as an indicator of how the interventions affect contact efficiency.
For households, this measure records call attempts to confirm whether the migrant remains absent from the household and to obtain updated contact information. While the household-level treatment (information about the financial incentive) is provided only after contact is established, we will still analyze call-attempt frequency at the household level. Because randomization occurs at the village level, households may exchange information about the survey, and prior treatment of other households in the same village could influence responsiveness or willingness to answer calls. For migrants, we expect the number of call attempts to be influenced by prior contact between the migrant and their rural household, including information flows related to the TVSEP migrant survey and its financial incentive structure.

3. Successful scheduling of the migrant interview will be coded as a binary variable indicating whether the migrant agrees to a scheduled interview time and location during the migrant call. We will further distinguish between (i) initial agreement to schedule an interview and (ii) actual follow-through, defined as the migrant appearing and commencing the interview at the agreed time.

4. Migrant survey completion will be coded as a binary variable indicating whether the migrant completes the entire Migrant questionnaire. Completion determines eligibility for the household-level ex-post financial incentive, which is conditional on full survey completion.

1. Response quality indicators, e.g. data entry errors; number of answered questions/ completeness of key survey sections
2. Migrant-reported prior notification by household of TVSEP survey call
3. Timing of survey completion
4. Heterogeneity of intervention effects and their effect on sample representativeness

1. Response quality indicators will be constructed from survey paradata to capture the completeness and reliability of responses. Measures include the number of questions answered, completion of key survey sections, detectable data entry errors, and self-reported satisfaction with this year’s Migrant survey. These indicators provide additional information on how the interventions affect not just participation, but also the quality of the data collected.

2. Migrant-reported prior notification by their household. This binary variable indicates whether the migrant reports prior notification of the survey call, allowing us to assess how effectively household-level communication interventions foster early awareness.

3. Timing of survey completion will be measured as the number of days between the first successful migrant contact and final survey completion.

4. Finally, we will explore heterogeneity of intervention effects, examining whether the interventions have different impacts across subgroups, such as country, gender, household income brackets, education, prior experience with scam calls, trust levels, the strength of social ties between the household and migrant (proxied by length of absence and frequency of interactions), and prior engagement with the TVSEP panel—proxied by how long the household has participated. The heterogeneity analysis will shed light on whether and how the interventions shape the composition of respondents who complete the survey, providing insights into potential selection effects. By examining which types of migrants are more or less likely to be reached and complete the survey under different incentive structures, we can better understand how financial and network-based interventions may influence the representativeness of survey samples. This has broader implications for designing tracking surveys and other studies of mobile or hard-to-reach populations, as it informs how interventions may systematically include or exclude certain subgroups.

The experiment follows a 2×2 factorial design at the household level, complemented by a secondary migrant-level information treatment.
At the household level, rural origin households with at least one urban migrant are randomly assigned to one of four treatment groups:

1. Control (no incentive) = C
2. Ex-ante nudge only = T1
3. Ex-post incentive only = T2
4. Combined treatment (both nudge and incentive) = T3

Randomization is conducted at the village level to avoid potential spillover effects or perceived inequities between households living nearby. All households within the same village therefore receive a consistent treatment message.
During the migrant contact stage, we introduce an additional migrant-level information treatment among the groups eligible for the ex-post incentive (T2 and T3). For a random subset of these migrants, enumerators explicitly inform them about the incentive that their household will receive upon successful interview completion. This nested design enables assessment of whether directly communicating the incentive to migrants amplifies household-level effects on participation.
The study thus enables identification of (i) the separate and joint impacts of ex-ante and ex-post incentives on survey participation, and (ii) the additional effect of providing information directly to migrants.

The experiment follows a 2×2 factorial design at the household level, complemented by a migrant-level information treatment.

Household-level treatments:
Households are randomly assigned to one of four groups based on two dimensions: the provision of an ex-ante nudge and the use of ex-post monetary compensation. The household-level randomization assigns one-sixth of households to C, one-sixth to T1, one-sixth to T2, and one-half to T3.
1. Control group (C): No ex-ante nudge, no ex-post compensation. This group represents the baseline condition.
2. T1: Receives an ex-ante unconditional nudge (30,000 VND / 40 THB)
3. T2: Receives ex-post monetary compensation (30,000 VND / 40 THB) upon completion of the migrant survey
4. T3: Receives both an ex-ante unconditional nudge and ex-post monetary compensation.

Migrant-level information treatment:
Within each household treatment, migrants are further randomized to either receive an information treatment or not. This creates the following migrant-level subgroups:
- T2 group migrants: Half receive no migrant-level information (T2-i), and half receive the information treatment (T2+i).
- T3 group migrants: Half receive no migrant-level information (T3-i), and half receive the information treatment (T3+i).

The proportions correspond to the overall household-level assignment, so one-twelfth of all households are T2-i, one-twelfth are T2+i, one-fourth are T3-i, and one-fourth are T3+i. For additional details on the treatment groups and sample sizes, please refer to the PDF file provided in the appendix.

During the first sampling stage for the 2025 Migrant Tracking Survey, we will contact all households in the TVSEP panel reporting at least one urban migrant in Ho Chi Minh City, Hanoi, and Bangkok (approximately 1800 household-migrant links). Control group households will simply be asked to provide migrant phone numbers. Households in T1 will receive the ex-ante unconditional phone credit/ mobile money nudge aimed at encouraging them to inform migrants proactively of our call to fix an interview meeting. T2 households will be offered a conditional monetary incentive payable after the migrant completes the survey, targeting effort in sharing accurate contact information and motivating the migrant. T3 households receive both interventions.

Randomization is implemented at the village level in order to minimize potential spillover effects between treatment groups and to reduce the likelihood of confusion or perceived inequity among survey participants. This design ensures that all households within a given village receive a consistent incentive message. Because Treatment 3 (T3) is expected to be the most effective at reducing migrant attrition, and given that the primary objective of the study is to maximize the number of completed migrant surveys, we employ unequal allocation across treatment arms. Specifically, 50 percent of respondents are assigned to T3, while 17 percent are assigned to each of the remaining groups (Control, T1, and T2). Randomization occurs prior to sampling, and balance on observable household characteristics will be verified. Given the likelihood of some non-response—either to our phone calls or to outreach by village heads—the final analytic sample is expected to be somewhat (and potentially even substantially) smaller than the initial target.

This factorial design allows independent assessment of the main effects of each intervention and their interaction, providing an additional evaluation of whether combined incentives produce synergistic gains in follow-up survey participation.
In addition, during the second stage for the 2025 Migrant tracking survey, we will contact the migrants to confirm their location in the major urban centers where our data collection takes place; and to fix a meeting time and location. During this migrant calling stage, we will inform 50% of the migrants whose household was drawn into T2 or T3 of the compensation scheme that their household will receive. Randomization happens at the migrant level.

To evaluate the effectiveness of these measures, we will track attrition at every stage: if the phone number was not shared, if the migrant did not answer calls, if the migrant was unwilling to schedule an interview, or if they failed to attend or complete the interview. Additionally, we will collect self-reported data from migrants on whether they were contacted by their household in advance.

Computer-based
- Ex-ante randomization for household-level treatment group assignment
- Within tablet randomization for migrant-level information treatment

Randomization of household-level treatments was implemented at the village (cluster) level. Villages were first randomly ordered and then sequentially assigned to treatment arms to achieve the target total numbers of households across groups, according to the experimental design target size (Control, T1, T2, T3). This procedure ensures target household counts while maintaining a random assignment order. The random seed and resulting assignment were recorded for reproducibility.

Randomization for the migrant information treatment was implemented at the individual level, within the Survey Program (Survey Solutions).

Yes

211 villages in Thailand
191 villages in Vietnam

1115 migrants from 729 households in Thailand 697 migrants from 490 households in Vietnam We anticipate some attrition and non-response due to factors such as scam-call concerns in both countries and frequently changing mobile phone numbers. These challenges may particularly affect the first household calling stage, where the main treatment arms are implemented. We expect that any attrition will occur at random and be roughly proportionally distributed across treatment groups.

Based on the unequal treatment allocation described above, we expect the following approximate sample sizes at the household–migrant link level:
- Control group (C): 186 observations in Thailand and 116 in Vietnam.
- T1 (ex-ante nudge only): 186 in Thailand and 116 in Vietnam.
- T2 (ex-post compensation only): 186 in Thailand and 116 in Vietnam.
- T3 (both ex-ante nudge and ex-post compensation): 557 in Thailand and 349 in Vietnam.

At the migrant level, where an additional information treatment is applied among T2 and T3 households, we expect approximately:
- No information treatment (T2-i and T3-i combined): 371 observations in Thailand and 232 in Vietnam.
- Information treatment (T2+i and T3+i combined): 372 observations in Thailand and 233 in Vietnam.

These figures reflect the expected distribution given the randomization proportions (1/6, 1/6, 1/6, and 1/2 across household treatment arms, with an even split across information treatment conditions within T2 and T3).

The experiment is designed to reduce attrition in follow-up surveys of rural-to-urban migrants. To assess whether our household- and migrant-level interventions affect survey participation, we calculated minimum detectable effect sizes (MDEs) for our main binary outcome: migrant survey take-up (yes/no). The calculations are based on the pooled sample across Thailand and Vietnam, accounting for unequal treatment group sizes at the household level (Control: 17%, T1: 17%, T2: 17%, T3: 50%) and clustering at the village level. Power calculations are based on the pooled sample across Thailand and Vietnam, which includes 1,812 migrants in total (302 each in the Control, T1, and T2 groups, and 906 in T3). The randomization occurs at the household level, with clustering at the village level (average cluster size ≈ 4.6). To account for clustering, we applied design effects corresponding to intra-cluster correlations (ICCs) of 0.01, 0.10, and 0.30. These yield design effects of 1.04, 1.46, and 2.39, respectively. Under a two-sided test with 5% significance and 80% power, the resulting minimum detectable effects (MDEs) for our main binary outcome (survey take-up, yes/no) range between approximately 7–12 percentage points under low clustering (ICC = 0.01) and 13–20 percentage points under high clustering (ICC = 0.30). MDEs are smaller when comparing the Control group to the larger treatment arm (T3) and larger for comparisons involving smaller arms (T1, T2), reflecting unequal group sizes. Assuming a lower baseline take-up rate (20%) yields slightly smaller MDEs compared to a 50% baseline, due to reduced variance. Overall, the study is well-powered to detect policy-relevant changes in migrant survey participation across a realistic range of clustering and baseline participation rates. Further details of the assumptions, formulae, and step-by-step calculations are provided in the accompanying Power Calculations Appendix (PDF).