Messaging to Improve Phone Survey Response Rates

How to cite

Glazerman, Steven; Dillon, Andrew; Karlan, Dean; Warren, Shana; Rosenbaum, Michael. 2020. "Messaging to Improve Phone Survey Response Rates." AEA RCT Registry. AEARCTR-0006111. https://doi.org/10.1257/rct.6111-2.1

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Steven Glazerman* , Andrew Dillon , Dean Karlan , Shana Warren , Michael Rosenbaum

* Primary investigator

Registry ID
AEARCTR-0006111
DOI
https://doi.org/10.1257/rct.6111-2.1
First registered
Last updated
Status
Completed
Trial dates
2020-06-15 – 2021-06-30
Countries
Bangladesh (), Burkina Faso (), Côte d'Ivoire (), Colombia (), Ghana (), Mexico (Mexico City), Philippines (), Rwanda (), Sierra Leone (), Zambia ()
JEL codes
C83
Registry record
http://www.socialscienceregistry.org/trials/6111

Abstract

Substantial literature on survey response rates focuses on framing and the appeal to altruism as a motivation for participating, with methods like pre-survey post-cards and letters to incentivize cooperation, with evidence coming primarily from the U.S. and Europe. More recently, there has been interest in mobile phone surveys in low and middle income countries, where the efficacy of methods for improving response rates is not as well known. This study randomizes the use of pre-survey text messages, whether to send them and which type of appeal to make. The study also randomizes the messaging used in the consent script, appealing alternatively to “researcher” or “government” as the motivating authority, in the first round of experiments, and appeals to efficacy of participation and to self-interest with reminders about monetary compensation. The experiment is conducted in 11 random-digit dial (RDD) surveys in 10 countries, with followup surveys in 5 of those surveys.

Intervention

The experiment varied two factors

Factor 1 is SMS text message sent to respondent prior to CATI interview call, with 3 possible levels:
S0 = No SMS
SG= SMS, appeal to "government"
SR= SMS, appeal to "researcher"

Factor 2 is appeal in the consent script, with 3 levels:
G = consent appeals to "government"
R = consent appeals to "researcher"
P = consent appeals to "policymaker"

In Colombia and Mexico, it was a 2x2 (no cases assigned to S0 or P)
In the other wave 1 countries it was a 3x1 (S, G, or P)
"Mixed message" cells (SG-R, SG-P, SR-G, SR-P, etc.) are not populated.
In Spanish-speaking countries, "policymaker" appeals are omitted because terminology is hard to distinguish from government.

In Wave 2 surveys, the treatment factor structure is altered and different messaging contrasts are planned. The pre-survey SMS message has the following appeal type for each treatment arm:
A1 Placebo/short message -- no specific appeal
A2 General Learning (Food access), says that the first wave of the survey was informative about food access
A3 General Learning (Household finances), says the first wave of the survey related to household finances
A4 Specific Learning (Food access), same as above but shared a statistic about food access
A5 Specific Learning (Household finances), same as above, but shared a statistic about household finances
B1 Self Interest, reminded respondents about the monetary incentive
B2 General Learning (Food access) + Self interest, same as A2, but included message about monetary incentive
B3 General Learning (Household finances) + Self interest, same as A3 but included message about monetary incentive
B4 Specific Learning (Food access) + Self interest = A4 + monetary incentive message
B5 Specific Learning (Household finances) + Self interest = A5 + monetary incentive message

Experimental design

Randomization is built into the SurveyCTO case management system. Respondents are randomized into treatment arms and are subsequently sent either an SMS with the assigned messaging or no SMS, and subsequently if the are contacted, are read a script that has one of the randomly determined appeals. The study uses paradata collected as part of the survey as well as survey responses.

Design details

Randomization method
Randomization programmed in SurveyCTO.
Randomization unit
Individual cases. Each case is a phone number for an individual.
Clusters
The survey is RDD and the number of working numbers cannot be known ahead of time with certainty. We estimate that we will have about 50,000 cases across countries/surveys (countries are strata, not clusters)
Observations
50000 cases
Arms
Varies by country/survey, but approximately 12000 cases in each treatment arm for the Wave 1 experiment. The Wave 2 experiment will be conducted with followups from the RDD survey. This sample will be substantially smaller, approximately 6,000 total spread over 10 treatment arms, with unequal cell sizes ranging from 8.3% (~500 cases) to 16.7% (~1,000 cases).
Primary outcomes
Survey consents and completions
Secondary outcomes
Contact rates (for SMS arms) Breakoffs Respondent attention Response distributions for key survey items: mask-wearing behavior and hand-washing behavior

Cited by

No citation data available for this registration yet.

Keywords