How to test for differences with multiple observations per subject under the same (level-2) condition? I'm trying test for the difference in a numeric outcome (WearTime) that was measured 7 times (7 consecutive days) for each subject under the same condition (Season) that was measured at the participant level. The question I want to test is: "Does Season affect daily WearTime?" This is not a primary outcome, but more of a check for going forward to determine if Seasonality needs to be accounted for in future analyses.
Data structure: 700obs, 5 variables


*

*SubjectID: 100 individuals 

*ObservationDay: 1 to 7; Day 1 refers to
1st dat etc. Every subject has 7 days with values 

*Date: Date of ObservationDay; 7 days in sequence spread throughout a 
year 

*Season: Fall/Winter/Spring/Summer; Season when observation period 
began based on Date from ObservationDay 1 for that subject

*WearTime: 0-1440 minutes; How long an electronic device was worn by a 
participants on each ObservationDay


My instinct for testing this is a mixed model ANOVA approach with Season as an independent fixed-effects factor, and subject ID as a random effects factor.
e.g. code from R (as this is the notation I am familiar with)
library(lme4)
library(car)
Anova(lmer(formula = WearTime ~ Season + (1 | Subject), data = df)

But if WearTime is measured at level-1 (per day, per subject), should Season be classified as a level-2 (per subject) variable?
Something more along the lines of:
WearTime ~ (1 | Season)

So my questions to you is which interpretation of the Season variable is correct? As a level-1 or 2 variable? And if the latter how would I go about testing for a difference by Season (as when I try to apply car::Anova to the above model I get an error, so I assume this is not a valid approach)? Just want to make sure I'm using the appropriate statistical approach before wrestling with code.
Thanks in advance!
 A: It does not matter what level a variable belongs to. The software doesn't care. At least, all the mixed effects software that I am aware of does not care, and you are using equation formats commonly used in mixed effects software. 
So, the heart of your question seems to be how to specify the random effects. You have repeated measures within subjects, so ... + (1 | Subject) is appropriate. If you don't specify random intercepts for subjects then you won't be controlling for the non-independence of observations within each subject. 
So, the question is between:
WearTime ~ Season + (1 | Subject)

or 
WearTime ~ 1 + (1 | Season) + (1 | Subject)

Since your research question is specifically about the "affect" of Season, and since you can't really consider the sample of seasons in your dataset as being taken from a larger population of seasons, I would favour the first case where you treat Season as a fixed effect and your interest centres on the estimated regression coefficients for it. 
In the second case you treat season as random (crossed random effects) and your interest centres on the estimated variance for it - but that makes much less sense to me. 
You may also also want to include the date of each observation, perhaps as the number of days from baseline and see if it interacts with season. So you could have:
WearTime ~ Season * Days + (1 | Subject)

or even
WearTime ~ Season * Days + (Season * Days | Subject)

provided that the data supports such a random structure.
It's also common to consider nonlinearities for time, which you could introduce with higher order terms, or even better, splines.
