Airbnb Booking Analysis
→
→
Why This Project?
One of the more challenging aspects of data analysis is answering questions when the available data doesn't perfectly align with what you'd ideally want. I chose to focus on Airbnb bookings because the data presented an interesting analytical challenge: distinguishing between listings that are actively booked and those that are simply unavailable due to owner preferences or part-time rental status. This required thoughtful assumptions, proxy measures, and a careful methodological approach.
Background
Many Airbnb listings rely on recent reviews to attract bookings, but it’s unclear whether newer reviews actually drive more reservations. This project estimates the effect of recent reviews on booking outcomes using observational data from Airbnb.
Data
I used a public Airbnb dataset that includes listing-level variables such as:
- review_scores_rating: Average guest rating
- number_of_reviews: Total number of reviews
- price: Nightly price
- last_review: Date of most recent review
I also used calendar-level variables such as:
Overall, it is important to know that the data is a snapshot in early March.
Methodology
Data decisions:
- Because the dataset doesn’t directly indicate when a booking has occurred, I had to infer bookings based on availability data. Specifically, I treated stretches of days where available = false as potential bookings. However, to reduce the risk of misclassifying listings that were unavailable for other reasons (e.g., the host pausing the listing or traveling) I excluded any unavailable streaks longer than 14 days.
- Because the dataset is a snapshot captured in early March, it does not include bookings made after that point. To work around this limitation, I created a binary outcome variable for April 2025: listings are marked as "booked" if they have at least one unavailable day during the month. This simplifies the outcome while helping control for potential confounding, such as certain types of listings being more likely to receive early bookings, by focusing only on whether any booking occurred, rather than the total number of nights booked.
6 variables were calculated:
- booked: Number of days the listing was unavailable (available = false) in April 2025
- booked_yes: Binary indicator (1 if booked > 0, otherwise 0)
- not_booked: Number of days the listing was available (available = true) in April 2025
- total_avail: Combined total of booked and not_booked days in April 2025
- prior_booked: Number of days the listing was unavailable in March 2025
- days_since_last_review90: Binary indicator (1 if last_review was within 90 days of April 1, 2025)
To control for confounding, we used propensity score matching with the following covariates:
- Review Score (review_scores_rating)
- Number of Reviews (number_of_reviews)
- Prior Bookings (prior_booked)
- Availability (total_avail)
- Price (price)
We estimated the average treatment effect on bookings using matched pairs and checked balance using standardized mean differences (SMDs).
Results
Covariate Balance
Before matching, there were substantial differences between listings that recently received a review and those that had not. For example, listings with recent reviews had significantly more prior bookings and reviews overall, and tended to be priced lower:
| Variable | No Recent Review | Recent Review |
|---|---|---|
| Prior Booked Days | 4.66 | 8.73 |
| Total Availability (April 2025) | 27.10 | 27.99 |
| Price | $475.71 | $266.97 |
| Review Score (Rating) | 4.81 | 4.84 |
| Number of Reviews | 28.79 | 85.51 |
After propensity score matching, these covariates were better balanced across groups. Standardized mean differences (SMD) for all covariates fell below the common threshold of 0.1, suggesting a good match.
| Variable | No Recent Review | Recent Review |
|---|---|---|
| Prior Booked Days | 9.02 | 8.73 |
| Total Availability (April 2025) | 27.65 | 27.99 |
| Price | $267.90 | $266.97 |
| Review Score (Rating) | 4.85 | 4.84 |
| Number of Reviews | 78.59 | 85.51 |
Treatment Effect Estimation
A logistic regression was used to estimate the effect of recent reviews on the probability of being booked in April 2025 (booked_yes).
Model Results (Logistic Regression)
- Recent Review Coefficient: 0.679
- Odds Ratio: ~1.97 (exp(0.679))
- p-value: < 0.001 (statistically significant)
Estimated Risk Difference
Listings with a recent review had a 15.8 percentage point higher probability of being booked in April 2025 compared to similar listings without a recent review.
Covariate Balance (Standardized Mean Differences)
| Covariate | SMD | Treated Mean | Control Mean | Pooled Std Dev |
|---|---|---|---|---|
| Prior Booked Days | -0.050 | 8.73 | 9.02 | 5.78 |
| Total Availability | 0.055 | 27.99 | 27.65 | 6.28 |
| Price | -0.003 | 266.97 | 267.90 | 292.92 |
| Review Score (Rating) | -0.030 | 4.84 | 4.85 | 0.27 |
| Number of Reviews | 0.060 | 85.51 | 78.59 | 115.18 |