First, what's in the spreadsheet
Take an export of a year of bookings. Every row's a booking. Columns are a mixed bag, and they need sorting before any test:
- Categorical: class (SUV, economy, sports, luxury), gearbox, where it was picked up, who booked
- Ordinal: cancellation terms, strict through to flexible
- Continuous: total price, deposit, days rented, mileage limit
- Dates: when the booking was made and when the car was collected (one minus the other gives lead time)
A rental page already looks a lot like this table. People who rent luxury cars Dubai from suppliers see the price, deposit, mileage cap and cancellation rules of each offer next to each other, so every pick of offer A over offer B is a tiny vote that can be counted.
Describe it
Rental prices are never bell-shaped. Three Lamborghini weekends and the right tail goes on for miles; the mean then describes nobody. Median and IQR are the better pair. Histograms go per class, too, since a fleet-wide one mashes hatchbacks and Rolls-Royces into one blob. Categorical columns? Just count. If most people who hire a vehicle for two or three days go for an automatic, that's worth knowing before anyone opens R and starts building models.
Rental price distributions are right-skewed, which is exactly where median outperforms mean as a measure of centre. See our guide on mean vs median vs mode for a deeper explanation of when each measure applies.
Do tourists and residents actually choose differently?
A cross-tab of who booked against which class they took, followed by a chi-square test of independence, answers the question. Days rented is skewed, so Mann-Whitney U suits it better than a t-test. A tiny p-value is no reason to celebrate yet. Cramér's V comes next. With twenty thousand rows almost anything is "significant", and a car rental company only cares about a gap big enough to change which cars it buys.
Effect size matters: with twenty thousand rows, almost anything becomes "significant". Cramér's V tells you whether the difference is large enough to act on.
Predicting the choice
Did the customer take the no-deposit offer or not? A yes/no outcome like that calls for logistic regression. Days rented, car class, lead time and whether the booking came from overseas all make sensible predictors. Odds ratios are surprisingly easy to explain to a manager who hated maths at school: one more week of lead time, and the odds of picking flexible cancellation go up by so much. Four car classes instead of two outcomes? Multinomial regression, same idea.
For a full walkthrough of logistic regression, including how to interpret odds ratios, see our logistic regression guide with worked examples.
The calendar matters more than it seems
In Dubai, the rush comes roughly October to April. July? Plenty of people just stay indoors. A plot of daily bookings shows it without squinting; seasonal decomposition then splits trend from yearly cycle. Skip this step and every autumn brings the "discovery" that luxury car rental is booming. It isn't. It's just autumn.
Seasonal decomposition is a core technique in time series analysis. Our time series analysis page covers the methods used to separate trend, seasonality and noise in booking data.
Where rental data trips people up
- Selection bias. The file holds only those who booked, never the ones who saw the price and closed the tab.
- Tiny niches. Fifteen convertible bookings can't carry a percentage with any confidence.
- Cause and effect. Long rentals come with lower daily rates, but maybe the discount made people stay longer, not the reverse.
- Too many tests. Compare twenty pairs of groups and one will look significant by pure luck.
Running twenty tests at α = 0.05 gives roughly a one-in-three chance of at least one false positive. The Bonferroni correction is the standard fix. See our guide on Bonferroni correction for how to apply it.
Longer rentals coincide with lower daily rates, but the direction of cause and effect is not obvious from the data alone. Our guide on correlation vs causation explains why this distinction matters before drawing business conclusions.
So what comes of it?
Nothing here is beyond an intro stats course. The trick is one clear question (more SUVs before the season? which terms do people value?) and letting data from each rented car answer it.
For a broader view of how statistics drives business decisions, our business statistics guide covers the core methods used across industries.
- Descriptive: median and IQR for skewed prices; per-class histograms; frequency counts for categorical columns
- Inference: chi-square test of independence on cross-tabs; Mann-Whitney U for skewed continuous variables
- Effect size: Cramér's V after chi-square; essential with large row counts
- Regression: logistic for binary outcomes (deposit yes/no); multinomial for four-class outcomes
- Time series: seasonal decomposition to separate trend from cycle before reporting
- Pitfalls: selection bias, small-niche unreliability, confounded cause and effect, multiple comparisons
Column types at a glance
| Column type | Examples in rental data | Appropriate summary | Appropriate test |
|---|---|---|---|
| Categorical | Car class, gearbox, pickup location, booker type | Counts and proportions | Chi-square test of independence |
| Ordinal | Cancellation terms (strict to flexible) | Median rank | Mann-Whitney U or Kruskal-Wallis |
| Continuous (skewed) | Total price, deposit, days rented, mileage limit | Median and IQR | Mann-Whitney U |
| Derived continuous | Lead time (booking date minus collection date) | Median and IQR | Logistic regression predictor |