Assumptions of statistic models

trinker

ggplot2orBust
#1
I was watching [Hadley's dplyr video](https://www.youtube.com/watch?v=8SGif63VW6E) and he said a statement that made me think:

Hadley Wickham said:
statistical models ``don’t fundamentally surprise you''
[YOUTUBE]8SGif63VW6E#t=134[/YOUTUBE]

I asked for clarification via Twitter and he said:

Hadley Wickham said:
models make assumptions that by their nature, can't question...also note that it's never possible to test assumptions; you either assume them or not
I'm curious what others think about these comments. Do people agree?

Is he saying the model is a coat that either fits the data or not; whereas visualization is asking the data what type of coat will fit?
 

Jake

Cookie Scientist
#2
Here's how I interpret it. Take the example of a simple regression model that assumes normal errors. There is nothing in the model itself that tells us whether that assumption is reasonable for a given set of data--nothing in the parameter estimates or predicted values, for example. Of course, we can try to verify this assumption through other means, such as a QQ plot or density plot of the errors. But these are things that are outside of the regression model itself.