Semi-Automated Non-response Detection for Surveys

SANDS (Semi-Automated Non-response Detection for Surveys) is an open-access AI tool developed by the National Center for Health Statistics to help researchers and survey administrators detect non-response in open-ended survey text. The model is designed to be used in conjunction with human reviewers to quickly partition a large volume of text for manual review.

To quickly get started, follow the model card or the instructions below. Please review the sections on Uses, Misuses and Out-of-scope, and Risks, Limitations, and Biases before applying the model to real survey data.

To quickly get started, follow the model card or the detailed instructions on Getting Started Section.

This model is a fine-tuned version of the supervised SimCSE BERT base uncased model. It was introduced at AAPOR 2022 at the talk Toward a Semi-automated item nonresponse detector model for open-response data. The model is uncased, so it treats important, Important, and ImPoRtAnT the same.

• Developed by: National Center for Health Statistics, Centers for Disease Control and Prevention
• Model Type: Text Classification
• Language(s): English
• License: Apache-2.0

Parent Model: For more details about SimCSE, we encourage users to check out the SimCSE Github repository, and the base model on HuggingFace. The model was fine-tuned on 3,000 labeled open-ended responses from RANDS during COVID 19 Rounds 1 and 2. The base SimCSE BERT model was trained on BookCorpus and English Wikipedia.

Training procedure

• Learning rate: 5e-5
• Batch size: 16
• Number training epochs: 4
• Base Model pooling dimension: 768
• Number of labels: 5

To use SANDS, first install python. Using a package manager, install pandas and the transformers module:

> pip install transformers pandas

Once you’ve installed the modules, the following code illustrates how to download the model, and parse a fixed set of responses:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import pandas as pd

# Load the model
model_location = "NCHS/SANDS"
model = AutoModelForSequenceClassification.from_pretrained(model_location)
tokenizer = AutoTokenizer.from_pretrained(model_location)

# Create example responses to test
responses = [
    "sdfsdfa",
    "idkkkkk",
    "Because you asked",
    "I am a cucumber",
    "My job went remote and I needed to take care of my kids",
]

# Run the model and compute a score for each response
with torch.no_grad():
    tokens = tokenizer(responses, padding=True, truncation=True, return_tensors="pt")
    output = model(**tokens)
    scores = torch.softmax(output.logits, dim=1).numpy()

# Display the scores in a table
columns = ["Gibberish", "Uncertainty", "Refusal", "High-risk", "Valid"]
df = pd.DataFrame(scores, columns=columns)
df.index.name = "Response"
print(df)

The code should output the following table

This model is intended to be used on survey responses for data cleaning to help researchers filter out non-responsive responses or junk responses to aid in research and analysis. The model will return a score for a response in 5 different categories: Gibberish, Refusal, Uncertainty, High Risk, and Valid as a probability vector that sums to 1.

Response types

• Gibberish: Nonsensical response where the respondent entered text without regard for English syntax. Examples: ksdhfkshgk and sadsadsadsadsadsadsad
• Refusal: Responses with valid English but are either a direct refusal to answer the question asked or a response that provides no contextual relationship to the question asked. Examples: Because or Meow.
• Uncertainty: Responses where the respondent does not understand the question, does not know the answer to the question, or does not know how to respond to the question. Examples: I dont know or unsure what you are asking.
• High-Risk: Responses that may be valid depending on the context and content of the question. These responses require human subject matter expertise to classify as a valid response or not. Examples: Necessity or I am a cucumber
• Valid: Responses that answer the question at hand and provide an insight to the respondents thought on the subject matter of the question. Examples: COVID began for me when my children’s school went online and I needed to stay home to watch them or staying home, avoiding crowds, still wear masks

.

The model has been trained to specifically identify survey non-response in open ended responses where the respondent taking the survey has given a response but their answer does not respond to the question at hand or providing any meaningful insight. Some examples of these types of responses are meow, ksdhfkshgk, or idk. The model was fine-tuned on 3,000 labeled open-ended responses to web probes on questions relating to the COVID-19 pandemic gathered from the Research and Development Survey or RANDS conducted by the Division of Research and Methodology at the National Center for Health Statistics. Web probes are questions implementing probing techniques from cognitive interviewing for use in survey question design and are different than traditional open-ended survey questions. The context of our labeled responses limited in focus on both COVID and health responses, so responses outside this scope may notice a drop in performance.

The responses the model is trained on are also from both web and phone based open-ended probes. There may be limitations in model effectiveness with more traditional open ended survey questions with responses provided in other mediums.

This model does not assess the factual accuracy of responses or filter out responses with different demographic biases. It was not trained to be factual of people or events and so using the model for such classification is out of scope for the abilities of the model.

We did not train the model to recognize non-response in any language other than English. Responses in languages other than English are out of scope and the model will perform poorly. Any correct classifications are a result of the base SimCSE or Bert Models.

To investigate if there were differences between demographic groups on sensitivity and specificity, we conducted two-tailed Z-tests across demographic groups. These included education (some college or less and bachelor’s or more), sex (male or female), mode (computer or telephone), race and ethnicity (non-Hispanic White, non-Hispanic Black, Hispanic, and all others who are non-Hispanic), and age (18-29, 30-44, 45-59, and 60+). There were 4,813 responses to 3 probes. To control for family-wise error rate, we applied the Bonferroni correction was applied to the alpha level (α < 0.00167).

There were statistically significant differences in specificity between education levels, mode, and White and Black respondents. There were no statistically significant differences in sensitivity. Respondents with some college or less had lower specificity compared to those with more education (0.73 versus 0.80, p < 0.0001). Respondents who used a smartphone or computer to complete their survey had a higher specificity than those who completed the survey over the telephone (0.77 versus 0.70, p < 0.0001). Black respondents had a lower specificity than White respondents (0.65 versus 0.78, p < 0.0001). Effect sizes for education and mode were small (h = 0.17 and h = 0.16, respectively) while the effect size for race was between small and medium (h = 0.28).

As the model was fine-tuned from SimCSE, itself fine-tuned from BERT, it will reproduce all biases inherent in these base models. Due to tokenization, the model may incorrectly classify typos, especially in acronyms. For example: LGBTQ is valid, while LBGTQ is classified as gibberish.

Model and code, including source files and code samples if any in the content, are released as open source under the Creative Commons Universal Public Domain. This means you can use the code, model, and content in this repository except for any offical trademarks in your own projects.

Open source projects are made available and contributed to under licenses that include terms that, for the protection of contributors, make clear that the projects are offered “as-is”, without warranty, and disclaiming liability for damages resulting from using the projects. This model is no different. The open content license it is offered under includes such terms.