Why people cancel Netflix
A team survey project on churn. We wrote the questionnaire, collected responses through Qualtrics, and I ran the statistics in R and presented the results. The short answer: the paid-sharing crackdown hurts more than the recommendation algorithm does.
The question
Netflix has three things going on at once: people complain about the recommendation feed, the company started charging for shared passwords, and competitors keep adding content. We wanted to know which of those actually moves someone toward cancelling. Every item was a 0 to 10 scale, and the outcome was "How likely are you to cancel or downgrade in the next six months?"
What correlates with cancelling
I tested each item against cancellation intent with a Pearson correlation. Five of the ten came back significant at 5%. Feeling overwhelmed by choice or seeing repetitive titles did not matter on their own. Watching less because of that fatigue did.
Where opinion sits
One-sample t-tests against the scale midpoint of 5. The paid-sharing policy scored 3.0 on fairness, far below neutral (p < 0.0001), and people reported lower trust because of it. Switching intent averaged 7.4. Live sports and events, which Netflix has been pushing, landed below the midpoint.
Which feed layout people want
Respondents ranked five recommendation layouts from best to worst. Model 2 won clearly: mean rank 2.0 and first choice for 52% of people. A Friedman test says the rankings are not random (p < 0.0001), so the layout is worth testing as a fatigue fix.
Putting it together
A multiple regression with every item predicting cancellation intent explains 66% of the variance (adjusted R² 0.56, n = 44). Five predictors are significant: how much paid sharing changed the person's plan, how fair they think it is, willingness to switch, perceived quality versus competitors, and watching less because of fatigue. The biggest coefficient belongs to paid sharing.
Limits: a convenience sample of 44 people, so this points at what to test rather than what the whole subscriber base thinks. The regression has ten predictors on 44 rows, which is why I lean on the adjusted R² and the individual correlations, not just the headline number.
Tools
R for cleaning the Qualtrics export, the correlation and t-tests, the Friedman test and the regression. Python (matplotlib) for the presentation charts. The charts on this page are redrawn from the same survey data.