ml4qc Python Package Documentation

The ml4qc Python package offers a toolkit for employing machine learning technologies in survey data quality control. Among other things, it helps to extend the surveydata package and advance SurveyCTO’s machine learning roadmap. See further below for more about the research program and specifics on the package.

Installation

Installing the latest version with pip:

pip install ml4qc

Overview of research program

The working title and abstract for the overarching research program is as follows:

Can machine learning aid survey data quality-control efforts, even without access to actual survey data?

A robust quality-control process with some degree of human review is often crucial for survey data quality, but resources for human review are almost always limited and therefore rationed. While traditional statistical methods of directing quality-control efforts often rely on field-by-field analysis to check for outliers, enumerator effects, and unexpected patterns, newer machine-learning-based methods allow for a more holistic evaluation of interviews. ML methods also allow for human review to train models that can then predict the results of review, increasingly focusing review time on potential problems. In this research program, we explore the potential of ML-based methods to direct and supplement QC efforts across international settings. In particular, we look at the potential for privacy-protecting approaches that allow ML models to be trained and utilized without ever exposing personally-identifiable data — or, indeed, any survey data at all — to ML systems or analysts. Specifically, metadata and paradata, including rich but non-identifying data from mobile device sensors, is used in lieu of potentially-sensitive survey data.

We currently envision three phases to the work:

  • Phase 1: Foundation-building and early exploration

    • Goal 1: Establish a flexible codebase to serve as a platform for exploration and analysis

    • Goal 2: Test whether popular supervised prediction models work to predict quality classifications when trained on meta/paradata

    • Goal 3: Test whether popular unsupervised models can identify useful patterns in enumerator effects

  • Phase 2: Dig deeper across settings, refine tools and understanding

    • Goal 1: Tune models and features for supervised models, establish links between review processes, quality classifications, and effectiveness

    • Goal 2: Find useful structure for measuring and reporting enumerator effects, how to control out nonrandom sources of variation

    • Goal 3: For all models, establish requirements for training data (e.g., sample sizes)

  • Phase 3: Test the potential for transfer learning, further refine tools

    • Goal 1: Develop instrument-agnostic meta/paradata format and test the potential for transfer learning to enable useful results earlier in the launch of a new survey

    • Goal 2: Support continued scaling of experimentation across more settings with easy, production-ready tools

We are currently in Phase 1. See the “Examples” section below for early results.

Overview of Python package

The ml4qc package builds on the scikit-learn toolset. It includes the following utility classes for working with survey data:

  • SurveyML provides core functionality, including preprocessing, outlier detection, and cluster analysis

  • SurveyMLClassifier builds on SurveyML, adding support for running classification models and reporting out results

While SurveyMLClassifier supports a variety of approaches, the currently-recommended approach to binary classification is as follows:

  1. Do not reweight for class imbalances; use SurveyMLClassifier.cv_for_best_hyperparameters() to find the optimal hyperparameters for a given dataset, with neg_log_loss, neg_brier_score, or roc_auc as the CV metric to optimize. This will optimize for an unbiased distribution of estimated probabilities.

  2. Use a calibration_method (isotonic or sigmoid) to calibrate the estimated probability distribution.

  3. Almost always (and especially when classes are imbalanced), specify a non-default option for the classification threshold (and possibly threshold_value), as the default threshold of 0.5 is unlikely to be optimal. When in doubt, use threshold='optimal_f' to choose the threshold that maximizes the F-1 score.

This is essentially the approach used in the examples linked below.

When there are nonrandom aspects to interview assignment, it is also recommended to initialize SurveyMLClassifier with a list of control_features to control out as much of the nonrandom-assignment effects as possible. During pre-processing, control_features will be used in OLS regressions to predict each other feature value, and it will be the residuals that are used in further analysis. This can be particularly important in distinguishing enumerator effects from the effects of nonrandom assignment. See the CAPI2 example linked below.

Examples

This package is best illustrated by way of example. The following example analyses are available:

  • CATI1 analysis

    • Setting: CATI survey in Afghanistan

    • Review and quality classification strategy: Holistic review of all submissions

    • Supervised results: Precision for predicting rejected submission ranges from 10% to 43% (against a base rate of 3.8%)

    • Unsupervised results: Significant enumerator effects discovered and summarized

  • CATI2 analysis

    • Setting: CATI survey in Afghanistan

    • Review and quality classification strategy: Holistic review of all submissions

    • Supervised results: Precision for predicting rejected submission ranges from 20-50% (against a base rate of 4.7%), but wide variation due to very small training sample

    • Unsupervised results: Significant enumerator effects discovered and summarized, but not at cluster level

  • CATI3 analysis

    • Setting: CATI survey in Afghanistan

    • Review and quality classification strategy: All completed interviews classified the same, all incomplete interviews rejected

    • Supervised results: Because there are clear meta/paradata differences between complete and incomplete interviews, all ML models achieve 100% precision in predicting rejection

    • Unsupervised results: Very significant enumerator effects discovered and summarized

  • CAPI1 analysis

    • Setting: CAPI survey in Afghanistan

    • Review and quality classification strategy: Holistic review of all submissions

    • Supervised results: With only 5 rejected submissions, instead sought to predict “not approved as GOOD quality” with a base rate of 70% (resting almost completely on the distinction between a “GOOD” and an “OKAY” quality rating); none of the models succeed in predicting the distinction and it’s not clear that a larger sample size would help

    • Unsupervised results: Very significant enumerator effects discovered and summarized

  • CAPI2 analysis

    • Setting: CAPI survey in Ethiopia

    • Review and quality classification strategy: Submissions flagged with automated statistical checks at the question level, plus randomly-selected interviews, reviewed for individual responses in need of correction; those that require correction classified as “OKAY” (vs. “GOOD”) quality

    • Supervised results: Essentially no predictive power with any of the models

    • Unsupervised results: Even once many of the effects of nonrandom assignment are controlled out, there appear to be enumerator effects at the cluster as well as individual level

  • IPA analysis

    • External repository: Innovations for Poverty Action has an external repository where they are exploring these methods with their own datasets

Reference

Indices and tables