We have $n$ observations of human performance before and after a training program. We have a random sample of $n$ individuals and assume population distributions are normal.
We thus have: $x_{1,t},\ x_{2,t},\ ...,\ x_{n,t}$ and $x_{1,t+1},\ x_{2,t+1},\ ...,\ x_{n,t+1}$. It is clear from the example that ${\rm Cov}(x_t,x_{t-1}) \neq 0$.
What is the convention and reasonable assumptions on the variance of the difference $d = x_{t} - x_{t+1}$?
$$\widehat{\rm var}(d) = \widehat{\rm var}(x_t)+\widehat{\rm var}( x_{t+1}) - 2\widehat{\rm cov}(x_t,x_{t+1}) ?$$
Or assume constant variance (homoscedasticity) over time?
$$\widehat{\rm var}( d) = 2\widehat{\rm var}(x) - 2\widehat{\rm cov}(x_t,x_{t+1}) ?$$
Is it correct to simply take the variance of $d$? How do you recommend start thinking (rigorously) about a problem like this?