Suppose I sample 100 frames from each of ten films. I now have 1,000 rows of colour measurements.

I still have ten films.

Frames from the same movie share lighting, grading, locations, and many other choices. Treating every frame as an independent example of a film can make the dataset look more informative than it is.

This is related to pseudoreplication: the analysis fails to account for the structure that connects observations. See this discussion of the unit-of-analysis problem.

Start with the claim

If the question concerns differences between movies from different periods, the movie is the unit I want to compare.

One approach is to calculate a summary for each film and compare those summaries. A hierarchical model offers another approach when within-film variation matters.

Neither approach makes the choice of films representative. That is a separate sampling question.

More measurements can still help

Additional frames can improve the estimate for one film, especially when scenes vary. They do not provide the same information as additional independently sampled films.

A film with many sampled frames should not silently receive more influence just because its extraction produced more rows.

Keep the unit visible

For Blue Movies, the workflow separates frame measurements from movie-level analysis. That distinction belongs in the explanation, not only in the code.

The same question appears elsewhere: several answers from one person, many crops from one photograph, or repeated requests from one customer.

Before trusting a sample size, ask what each row represents—and which rows share a source.