Municipal-Level Estimation of Tourism Perception: A Machine Learning-Based Approach to Small Area Estimation
Understanding the perception and acceptance of tourism among local residents is crucial for sustainable tourism management. Austria’s new “Tourism Acceptance” survey and its methodological framework includes a machine learning model estimating the level of tourism acceptance in Austria’s municipalities.
Drawing on a sample of approximately 12,000 respondents per year, the Tourism Acceptance survey asks participants about their opinions on tourism’s impact on and significance for the economy, labor market and leisure activities in their place of residence, as well as whether they think the number of tourists is too low, excessive or acceptable. Tourism intensity and its impact on the local population varies in different regions of Austria. Consequently, there is a need to monitor tourism acceptance and perception on a smaller regional level than the federal province. However, interviewing a representative sample in each region or municipality in Austria is neither affordable nor feasible. Therefore, we have developed a Small Area Estimation Model, which builds on possible tourism indicators that can influence tourism acceptance in the municipalities.
To estimate the perception of the entire population, we employ an XGBoost model, leveraging auxiliary data to impute the response variable for non-surveyed citizens on a unit level. We do this by linking each respondent's answer to auxiliary administrative data, including demographic information (age, sex, place of residence, income), employment sector (NACE classification), and municipal-level data (tourism-related profits, number of overnight stays per capita). These auxiliary variables are available for the entire country’s population. The XGBoost model is then trained on this data to predict a person’s answer for the question “How do you perceive the number of tourists in your place of residence?”. We predict the answers of residents not included in the survey. This allows us to aggregate both the predicted estimates and, if applicable, the actual survey responses of all residents within the same municipality, providing a single acceptance estimate per municipality. Additionally, we use a bootstrap approach to estimate the errors, and construct confidence intervals. This is done by calculating the standard errors from 1000 model predictions per person from 1000 bootstrap weights. The confidence intervals per municipality are then constructed through a bootstrap-based simulation: for each individual, a simulated probability is drawn from a truncated normal distribution with the estimated probability as the mean and the bootstrap-derived standard deviation. This process is repeated 1,000 times, generating a distribution of regional estimates, and the 2.5th and 97.5th percentiles of these estimates define the 95% confidence interval.
This approach parallels methods used in Small Area Estimation (SAE), extending the analysis beyond the sample to provide comprehensive estimates. The survey results offer valuable insights into tourism acceptance, highlighting spatial variations and underlying socio-economic factors influencing residents' perspectives on tourism.
We provide a toy example using synthetic data to showcase the entire process in TourismAcceptance_ImputationBeyondTheSample.rmd/html.