Pre-Analysis Plan: A field experiment providing information on polls in South Africa 2019 study
How to cite
Kreft, Brynde; Orkin, Kate. 2026. "Pre-Analysis Plan: A field experiment providing information on polls in South Africa 2019 study." AEA RCT Registry. AEARCTR-0017627. https://doi.org/10.1257/rct.17627-1.0
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
This is the analysis plan for an experiment showing election polls to potential voters in multiple locations in South Africa. The study complements a previous experiment (Orkin, 2016).
This study is a survey experiment conducted as part of a nationally-representative opinion
poll of 1,300 adults by the research company Citizen Surveys. The survey is conducted the
month before the May 2019 elections. I conduct two separate experiments in two groups of
provinces: the competitive and uncompetitive. In the former group, where the election for
the provincial assembly is close, respondents are randomly allocated to 3 equal groups: a
control group and two treatment arms. The two treatment groups are each read an opinion
poll in which the election is predicted to be close; the two polls vary only in which party
is leading the race. In the latter group of provinces, where the incumbent African National
Congress (ANC) party is almost certain to win an outright majority, the respondents are also
split into a control and two treatment groups. Treated individuals hear a hypothetical poll,
with the two treatments varying the margin of ANC’s lead over the remaining parties.
At the time of lodging this analysis plan, the treatment assignment has been received from
the polling company but has been stored by a research assistant, Marta Grabowska, in a
separate, encrypted location the PIs cannot access. Marta has constructed the variables of interest
in order to check for data quality issues, but conducted no analysis of treatment effects.
Intervention
The study comprises two experiments, each with three treatment arms.
Experiment 1:
There are three treatment groups in this experiment, the control group and two treatment
arms. The control group receive no treatment: they just proceed to the follow-up module.
Individuals assigned to treatment received one of two information messages. The messages
each give two pieces of information: an opinion poll for the upcoming provincial race, and
that there are many undecided voters who may swing the outcome of the provincial election.9
The opinion polls differ in the two treatments. Specifically:
Message 1: The poll puts the DA at 40% of vote share, the ANC at 38% and the EFF at
10%.
Message 2: The poll puts the ANC at 40% of vote share, the DA at 38% and the EFF at
10%.
Experiment 2:
There are three treatment groups in this experiment, the control group and two treatment
arms. Messages are similar in structure to experiment 1, except the opinion polls used are
different:
Message 1: The poll puts the ANC at 70% of vote share, with the DA and EFF at 15%
each.
Message 2: The poll puts the ANC at 60% of vote share, with the DA and EFF at 15%
each.
Experimental design
The study comprises two experiments.
1. Experiment 1: This experiment is conducted in the provinces where the public polls predicted the election would be close. There are two polls, both of which say the provincial election is close. They vary only which of the two parties, ANC or the DA, is just ahead. The third party, the Economic Freedom Fighters (EFF), is far behind.
2. Experiment 2: This experiment is conducted in the provinces where the ANC is likely to win a large outright majority. The two treatment messages both give polls which state the ANC is likely to win the provincial race. The polls vary how much the ANC is predicted to be leading by. The DA and EFF are tied in second place.
Design details
Cited by
No citation data available for this registration yet.