Synthetic Data in Practice - Case Studies

In our first article on the topic of "Quantitative Synthetic Data," we looked beyond the hype to explore the concept of synthetic data and its potential. Now, we move from theory to practice. This article examines four powerful use cases—Imputation, Augmentation, Expansion, and Creation—that demonstrate how this technology can be applied to solve specific, everyday challenges in market research.
We start with one of the most common issues: missing data.

Imputation refers to the targeted replacement of missing values in a dataset. In any real-world data collection, gaps are unavoidable—they arise from survey dropouts, participant fatigue, or simply the refusal to answer sensitive questions (item non-response). While it may seem straightforward to simply delete cases with missing values, this approach can significantly reduce the statistical power and, more critically, lead to substantial biases if the non-responses are not entirely random. Simple methods, such as replacing missing values with the mean or median, may fill the gaps but often create new issues by distorting the natural variance of the data and weakening the relationships between the variables.

A far more sophisticated solution is offered by modern machine learning-based imputation. Here, an ML model is trained on the complete data and learns the complex, multivariate relationships between all variables. Based on this understanding, it then predicts the most likely value for each missing entry on a case-by-case basis. This preserves the original data structure, including its variance and correlations, ensuring that subsequent analyses are both stable and valid.

Augmentation is a sophisticated generative technique developed to tackle the critical challenge of insufficient sample sizes, particularly within specific subgroups of a dataset. In market research, the “small N problem” frequently arises, where a central study segment is represented by too few respondents or there is an imbalance between target audiences. This data scarcity renders standard statistical analyses unreliable, making it practically impossible to draw robust conclusions.

The core objective of augmentation is to generate new, synthetic data points of such high quality that they are statistically indistinguishable from those of the genuinely underrepresented group. This extends far beyond simple methods like duplicating datasets. Instead, the model learns not only the characteristics of each variable individually but also the complex, multivariate relationships and dependencies among all variables simultaneously. It learns, for example, how age, geographic location, and expressed attitudes collectively influence purchasing behavior within this specific group.

The augmentation process isolates the actual respondents of the target audience or the entire dataset and trains a generative model exclusively on their data. After training, this model can be used to draw new data points from the learned distribution. These are not mere duplicates of the original data but entirely novel combinations of features that follow the complex patterns learned by the model. Each synthetic case is thus a plausible, statistically consistent new member of the subgroup. Combining the augmented data with the original dataset enables more meaningful and robust multivariate analyses that would otherwise not be possible.

This powerful technique enables researchers to go beyond respondents' mere answers and gain deeper, more strategic insights. This enrichment can be achieved through two distinct and valuable approaches. The first approach generates new variables by uncovering latent structures within the existing dataset. The second approach links the dataset with an external information source to add new knowledge.

Internal Expansion – Generating New Variables from Existing Data

The most significant insights of a dataset often do not lie in the answers to a single question, but in the complex interplay of multiple responses. Internal expansion analyzes the patterns and interrelationships among several existing variables to create a new, composite variable. This new variable typically represents a higher-level concept or latent construct that would be difficult or impossible to measure directly.