Graphs or formal tests

noetsi

Fortran must die
#1
I was considering this in the context of stationarity (unit roots) but it applies to many things.

I used to assume that formal tests were the way to go. Say the ADF for stationarity. But I was reading through an author I have a lot of respect for, and he focused on graphs to make these decisions not the formal tests. Of course there are well known problems with stationarity tests, but I am curious what people think. Are graphical methods or formal tests generally better for testing violations of assumptions in regression and time series.

I always thought tests were more scientific than looking at a graph of say residuals :p That they only used graphical methods before something better (and better computers) came along. But I am not so sure now.
 

Miner

TS Contributor
#2
Graphs are great, but you have to be careful. The human mind is geared to identify patterns, and can often see a pattern when none exists (i.e., pure randomness). However, I would never test for correlation without having plotted the data to see whether there was a useful relationship.
 

noetsi

Fortran must die
#3
I usually have the reverse problem. For instance graphs of residuals are supposed to show things like non-linearity. But I have never seen a real world example that did not look like a simple blob (I know text books show you these, I have not seen one outside text books).
 
#4
do you ever see 'visual inference' type things by hadley wickham? This discussion sort of reminds me of that, in that it is sort of trying to combine graphing and formal tests.