This quiz probes robust techniques to identify and handle extreme values in MMM datasets.
IQR method flags outliers if a data point lies beyond 1.5×IQR above Q3 or below ______.
mean
Q1
mode
median
Lower threshold identifies extreme low values outside normal spread. It’s robust to non‑normal distributions.
Winsorising at 99th percentile replaces extreme highs with the value at ______ percentile.
95th
99th
90th
50th
This caps influence of outliers while preserving rank ordering. It reduces variance without deleting data.
Cook’s distance highlights influential observations combining leverage and ______.
residual size
standard error
cookie
CTR
Large residuals at high leverage disproportionately affect coefficient estimates. Monitoring Cook’s D safeguards stability.
2025 Robyn defaults to winsorise spends at +/- ______ SD in log space.
1
0.5
2
5
Clamping extreme log spends prevents single week spikes from skewing adstock parameter fit.
Before winsorising, analysts should understand if spikes are real events or ______ errors.
data
DNS
cookie
creative
Legitimate promo bursts carry signal; blindly trimming could erase true effects. Investigate cause first.
MAD (median absolute deviation) thresholding is preferred over z‑score when data are ______.
uniform
non‑normal
censored
perfect
MAD uses median, resisting distortion from heavy tails more than mean‑SD z‑score.
Winsorising revenue but not spend can bias ROI because variance reduction is applied ______.
equally
never
asymmetrically
randomly
Applying to one side only distorts ratio metrics; symmetric treatment maintains consistency.
Outlier removal logs should capture threshold, count, and ______ of modified points.
dns
pixel
font
IDs
Documentation enables audit and potential rollback if decisions are questioned later.
When outliers are frequent, switch to robust regression such as ______ loss.
MAE
MSE
cross‑entropy
Huber
Huber loss blends L1 and L2, reducing sensitivity to large residuals without manual trimming.
Extreme value theory suggests setting cutoff where tail distribution fits a ______ model.
Gaussian
GPD (generalized pareto)
Poisson
Gamma
Fitting heavy‑tail allows statistically justified threshold selection rather than arbitrary percentile.
Starter
Review the basics.
Solid
Nice work—refine the details.
Expert!
Exceptional command of the topic.
