Relevance & Research Question

The general population travel survey is burdensome for the respondents as each
respondent should document a full-day trip, including the precise start times, end times, and locations. Predicting travel purposes automatically would be more beneficial than traditional travel surveys to overcome the burdensome.

This study aims to answer research question "How well can we predict the travel purpose using sensor data from a smartphone-based travel diary study?”

Methods & Data

CBS collected the data from November 2022 to February 2023 using the ODiN app and contains a total of 505 users. Administrative data, or demographic variables, from the CBS database is linked with the location data. Multiple bounding boxes with varying radiuses were determined from OSM for different tags associated with trip purposes. The data will be partitioned into training and testing sets with a ratio of 80:20 from the 4961 locations available.

This study used an Artificial Neural Network (ANN) and as comparison, the balanced accuracy of Random Forest (RF), Naive Bayes (NB), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGB) were evaluated. Weather data that has a potential to assist trip-purpose prediction was also included.

Results
The initial analysis was performed without OSM data. Later, OSM was included and treated as count data and percentages. ANN model with OSM data from 1 radius as count data and percentage was the most optimal model. The model's accuracy increased to over 70% for the overall model. ANN and SVM showed similar results while RF and XGB performed almost perfect. Unfortunately, NB was not good enough. The confusion matrix indicates that despite the accurate classification of classes, the limited number of observations, notably in sports and education, results in low accuracy. Regrettably, the inclusion of weather data did not enhance the model.

Added Value

Sensor data from a smartphone-installed travel diary application successfully predicts the purpose of a trip by utilising spatial and temporal patterns. Prediction models give more importance to trends that occur over time in specific locations. Some types of stops would be more important for making accurate predictions than others.

Relevance & Research Question

When applied to the survey research context, satisficing theory describes a range of behavioral strategies survey participants may use to reduce the cognitive effort required to answer questions conscientiously and truthfully (i.e., optimally). The resulting response error undermines data quality, affecting both reliability and validity of survey results. While general recommendations for preventing satisficing exist, they do not fully account for the complexity and individuality of the behavior. Therefore, satisficing remains a significant challenge to survey researchers. If we could predict which respondent is at risk of a certain type of satisficing behavior, we may be able to prevent satisficing using targeted interventions, particularly in a panel survey context where we have more information on satisficing types and their correlates over time. This study aims to explore the potential for such targeted approaches by investigating the robustness and predictiveness of survey satisficing in self-administered mixed-mode panels. The key research questions are:

1. Can distinct patterns of satisficing behavior be identified using latent class analysis (LCA)?
2. Do these patterns replicate across survey waves and modes?
3. Do respondent characteristics correlate with the identified satisficing pattern?
4. Can future satisficing behavior be predicted based on satisficing patterns exhibited in previous survey waves?

Methods & Data

The analyses base on the first three waves of the German Social Cohesion Panel (SCP), which is conducted jointly by the German Institute for Economic Research (DIW Berlin) and the Research Institute Social Cohesion (RISC) recruited in 2021. The SCP is a mixed-mode panel survey with participants self-selecting into either paper-and-pencil (PAPI) or web (CAWI) mode. The sample consists of 17,029 individuals nested in 13,053 households. I generated several indicators of satisficing behavior, including extreme and midpoint response selection, open and closed question nonresponse, speeding (in CAWI), and nondifferentiation (Nd) across item batteries. Nd is the tendency to select the same or similar response categories across a number of items, resulting in overall limited response variability. I measured nondifferentiation using the mean root of pairs method, distinguishing between unidimensional (weak Nd) and multidimensional or reverse-coded item batteries (strong Nd). To identify distinct patterns of satisficing behavior, I employed latent class analysis (LCA). To account for the hierarchical data structure, I used robust standard errors.

For each combination of survey wave (waves 1, 2, 3) and mode (PAPI, CAWI), I estimated six latent class models (assuming one to six latent classes). From the 36 models, I selected the six final models based on information criteria (BIC, CAIC, SABIC) and classification diagnostics (entropy, average posterior probabilities, odds of correct classification). Finally, I interpreted the six resulting latent class models regarding their homogeneity and distinctiveness.

Afterwards, i used multinomial logistic regression to investigate whether individual-level characteristics such as sociodemographic factors, predict class membership. Finally, I applied logistic regression to predict the most likely latent class membership in a wave by the estimated posterior class probability for the same latent class in a previous wave.

Results

Across the online survey mode, the LCA models identified three consistent latent classes that differ in their propensity to engage in certain satisficing strategies: The largest class are Optimizers. Counterintuitively, optimizers do not completely dispense with satisficing but exhibit comparably little and unspecific satisficing behavior. Optimizers may occasionally skip questions or apply nondifferentiation to reduce cognitive effort, sometimes speeding through questions.

ExtreMists are the second largest class. A typical ExtreMist is very likely to nonrespond to at least one closed as well as open question, reliably generating item missings. ExtreMist’s responses will likely be the highest or lowest extreme values of the given response scales. By providing an approximate answer, they circumvent the cognitive engagement that nuanced and differentiated responses necessitated.

Indifferents are the smallest class. Typical Indifferents can be identified through nondifferentiation with a tendency to select the midpoints of response scales. Furthermore, they are at risk to speed through the survey.

However, in the paper mode, models demonstrate variability in their global-level structures across waves. A fourth class of "Missers," exclusively emerging in the second wave's paper mode, was identified as primarily engaging in item nonresponse. Besides, no class of optimizers was identified in the paper mode of the third wave.

In the multinomial analyses, I found individual-level characteristics, such as education and income to be associated with class membership, with limited predictive power.

Regarding the robustness of class membership, the regression analyses revealed that the estimated posterior probabilities of belonging to a satisficing class in a previous wave were consistently significant predictors of belonging to that same class in the future, with odds ratios ranging from around 2 to 10. In some models, I found main effects of the survey mode as well as moderation effects of the survey mode on the effect of the past on the future satisficing strategy.

The effect sizes are consistently modest, with Nagelkerke R² values between 0.6 to 0.25.

Added Value

This study provides several important insights for survey research. First, it demonstrates the feasibility of using LCA to identify distinct typical satisficing patterns in self-administered mixed-mode surveys. Second, the identification of distinct satisficing patterns across survey waves and modes suggests the potential for targeted interventions to mitigate satisficing. Third, the finding that the global-level satisficing patterns replicate across survey waves for the online mode but not for the paper mode suggests that mode-specific approaches to targeted interventions may be necessary. Fourth, the moderate predictiveness of past satisficing behavior on future satisficing as well as the limited predictive power of individual-level characteristics indicate that while individual-level characteristics play a role, situational factors also substantially influence satisficing. This implies that targeted interventions should not rely solely on past behavior but should be enhanced with real-time data and the application of learning algorithms. Fifth, the finding that even optimizers have non-negligible risks of engaging in undesirable response behavior suggests that innovative prevention methods should target not only the extreme cases but also individuals who are more under the radar but potentially more approachable.