+1 please. Thanks.

Hope you read, share, cite. Here is the open-access link to the paper: doi.org/10.1016/j.ec... (14/14)

We thank handling editor Audra Bowlus, anonymous referees, and colleagues whose constructive comments helped us improve the paper. We also thank my excellent research associates Raied Arman & Farhana Kabir. (13/14)

We are also grateful to the WEE-DiFine initiative of BIGD. This project was a derivative of a separate project funded by them. (Check out ongoing RFPs of WEE-DiFine and WEE-Connect: bigd.bracu.ac.bd/all-projects/) (12/14)

Regression or population parameter estimates without accounting for statement framing might be biased and the researcher will have no way of addressing this bias. Randomizing statement framing allows credible estimation of bounds (a la partial identification). (11/14)

Researchers can also estimate population parameters as bounds. They can report the mean value of each statement by different framing and report the bound on the value of the parameter. (10/14)

Researchers can control statement framing in regression. Interaction with covariates allows obtaining two estimates of the relationship between a variable of interest and an attitude measure. These two estimates can be used as bounds on the correlation of interest. (9/14)

If all respondents receive the same framing, framing effects become a systematic feature of the data and cannot be accounted for in any analysis. To limit this systematic bias, we propose that future surveys randomize the framing of statements across respondents. (8/14)

We show that the conditional mean of each aggregated index for HH heads is significantly different between treatment and control groups. For PCA, not only does the magnitude vary, but also the sign switches. We find similar results using education status as a covariate. (7/14)

Building on Dunsch et al. (2018), we demonstrate the possible consequences of framing effects using two commonly used standard techniques of aggregating multiple responses: (i) Kling index (ii) PCA. (6/14)

Like Dunch et al (2018), we find a significant framing effect: responses vary |7| to |21| pp for the same statements between different framings. Framing effects persist among both positively and negatively framed questions. (5/14)

We randomly ask one of the two versions of the module: in one version, the first three statements are positively framed, and the following three are negatively framed, and vice versa in the second version. (4/14)

We build on this work and show that even with mixing the framing of statements there remains a significant framing effect. In a baseline survey of a study on mobile money, we ask respondents six statements to measure their attitudes towards mobile money. (3/14)

Asking “Is X easy to use” and “Is X not difficult to use” do not yield symmetric results. Dunch et al (2018) show that framing affects responses and suggest mitigating what they call “acquiescence bias” by mixing statement framing. (2/14)