Correction: Polling India via regression and post-stratification of non-probability online samples

Roberto Cerina, Raymond DuchPLoS 2024 · DOI · OpenAlex

Abstract: This article presents a formal correction to a 2021 study published in PLOS ONE (DOI: 10.1371/journal.pone.0260092). The correction addresses a layout error where figures were incorrectly swapped, while reaffirming the study's core findings: that modern statistical techniques can effectively leverage "messy" online data to predict election outcomes in the complex political landscape of India.

Introduction

Traditional "gold standard" polling—face-to-face interviews based on random probability sampling—is becoming prohibitively expensive and logistically difficult, especially in a country as vast and diverse as India. Consequently, researchers are increasingly turning to non-probability online samples from platforms like Amazon Mechanical Turk (AMT) or Facebook.

However, these digital samples are notoriously biased, often over-representing younger, wealthier, and more educated individuals. This research explores whether these "convenience" samples can be salvaged using Multilevel Regression and Post-stratification (MRP) to produce reliable national and regional election forecasts.

Methodology: The MRP Framework

The authors employ a two-stage statistical approach to correct for demographic imbalances:

  1. Multilevel Regression: Researchers build a model to estimate the relationship between individual characteristics—such as age, education level, caste, income, and region—and voting intentions. This allows them to estimate how a specific "type" of person is likely to vote, even if that group is small in the raw sample.
  2. Post-stratification: These estimates are then "mapped" onto the actual distribution of the Indian population using a "stratification frame" (census-style data from the 2014 National Election Study). If the online sample has too many college graduates, the model mathematically "shrinks" their influence to match their true proportion in the general population.

Visualizing the Bias

A key insight of the paper is the systematic nature of bias in online samples. By plotting the marginal distributions of various demographic groups in the samples against the true population, the researchers can visualize exactly where the discrepancies lie. For example, individuals with "No Formal Education" are consistently under-represented in digital samples, while certain age categories or regions may be over-represented.

Comparison of demographic marginals between the general population and various survey samples (NES 2014, AMT, and Facebook).

Key Results

The quantitative results demonstrate that MRP-corrected online samples performed remarkably well in predicting the 2019 Indian General Election:

  • Turnout: The model predicted a national turnout mean of 0.789, closely aligning with observed trends.
  • Seat Projections: For the 2019 election, the model estimated the NDA would win 328 seats (actual: 354) and the UPA would win 81 seats (actual: 99).
  • Stability and Regional Accuracy: The model's estimates for major alliances remained relatively stable in the weeks leading up to the election, filtering out daily campaign "noise." Furthermore, the correlation between predicted and observed behavior at the zone level was high, particularly for the NDA ($cor = 0.94$), indicating that the model captured regional dynamics effectively.

Limitations and Future Work

While MRP is a powerful tool, it is not a silver bullet. The accuracy of the final prediction is heavily dependent on the quality of the "stratification frame." If the underlying census data is outdated, the post-stratification will be flawed. Furthermore, the model struggles with the "Other" category, as predicting the fragmented performance of smaller, regional parties remains a significant statistical challenge.

Future research may involve refining these frames with more recent census data or exploring "dynamic" MRP models that can better account for late-breaking shifts in voter sentiment that online samples might pick up faster than traditional polls.

Summary updated

Abstract

[This corrects the article DOI: 10.1371/journal.pone.0260092.].
BibTeX
@inproceedings{cerina2024correction,
  author = {Roberto Cerina and Raymond Duch and Raymond M Duch},
  title = {{Correction: Polling India via regression and post-stratification of non-probability online samples}},
  booktitle = {PLoS},
  year = {2024},
  doi = {10.1371/journal.pone.0315139},
}

Loading related content...

Figure