Urban Heat Island Fieldwork Under Time Pressure
Thirty-six geography students survey an urban heat island in a city of 600,000. They use six handheld temperature sensors and two anemometers. Sampling runs at 07:00 and 14:00 on two cloudless days in July along three 4 km transects: a park, a low-rise residential area and a dense commercial core. Each site is measured for 5 minutes. Results show a maximum 3.8°C difference between park and core at 14:00. Problems: one sensor was dropped and reads 0.6°C high in an ice-bath check; wind speed varies from 1.2 to 4.5 m/s; four sites are inaccessible; and one transect was surveyed only at 14:00. The group wants to claim a statistically significant heat island.
1. Assess the reliability and validity of the data. Distinguish between random error, systematic error and sampling bias.
Read the worked answer
Reliability is weaker than the group assumes. The 0.6°C sensor drift is about 16% of the 3.8°C difference, and it is systematic, not random: if the faulty sensor was used in the core, it inflates the result. Wind varying from 1.2 to 4.5 m/s affects sensible heat exchange, so a 5-minute reading may not represent the site. The missing morning transect and four inaccessible sites create sampling bias, not random error.
Validity is also limited. The park-core difference may reflect land use, but also traffic, shade, altitude and time of day. Two cloudless days give no estimate of day-to-day variability, so a significance test would treat repeated sites as independent when they are spatially clustered.
I would separate three issues: sensor drift (systematic), instrument response and wind (random or site-specific), and missing transects (bias). The first can be corrected by calibration, the second by longer averaging, but the third cannot be fixed by statistics alone.
2. Recommend the statistical test and graphical presentation you would use, given the data problems. Explain why a simple t-test on all sites would be inappropriate.
Read the worked answer
A simple t-test on all sites would be inappropriate because sites are nested within transects and land-use classes, and because one transect has no morning data. Treating 36 sites as independent replicates inflates the effective sample size and ignores spatial autocorrelation.
Use a paired comparison only where the same site has both morning and afternoon readings, after correcting the faulty sensor. For the park-core difference, use a non-parametric test such as Mann-Whitney if the data are not normal, or a linear mixed model with land-use class as a fixed effect and transect as a random effect if the design is balanced enough.
Graphically, show:
- Boxplots of temperature by land-use class and time of day, with individual points.
- A transect profile with distance on the x-axis and temperature on the y-axis, so the reader sees the gradient rather than just a mean difference.
- A wind-speed covariate plot to show whether the hottest sites were also the calmest.
Report confidence intervals, not just p-values.
3. Redesign the fieldwork to produce defensible evidence with the same equipment and two field days. State the trade-offs you accept.
Read the worked answer
With the same equipment and two days, I would accept fewer sites but better replication. Design: six sites per land-use class, three visits per site per day, and two sensors per site rotated between land-use classes. That gives 36 sites but with paired sensors and repeated measures.
Use 10-minute averages at each visit, not 5-minute spot readings. Measure wind at each site at the same height as the temperature sensor. Complete the morning and afternoon runs for all three transects; if time is short, drop one residential transect rather than leave a whole time period missing.
The trade-off is spatial coverage: fewer unique sites than the original 36, but a balanced design that can estimate sensor error and wind effects. I would also add one fixed reference station in the park for the full two days, so drift can be corrected. This is more defensible than a larger, unbalanced sample.