Consider a panel with data $(y_{i,t})$ for $i\in\{1,...,N\}$ and $t\in\{1,...,T\}$.

Problem: I want to know how the cross-sectional average of $y$ behaves in an average time period.

My approach:

  • For each $t$, I calculate the cross-sectional average of $y_{i,t}$: $\bar{y}_t=\frac{1}{N}\sum\limits_i y_{i,t}$.
  • I take the time series mean of that series: $\bar{\bar{y}}=\frac{1}{T}\sum\limits_t \bar{y}_t$.
  • The standard error of $\bar{\bar{y}}$ is $\frac{\sigma}{\sqrt{T}}$, where $\sigma=\sqrt{\frac{1}{T-1}\sum\limits_t (\bar{y}_t-\bar{\bar{y}})^2}$.


  • The standard error of $\bar{\bar{y}}$ takes the variation across time periods into account, but not the cross-sectional variation. If there is large variation across entities but little variation across time, wouldn't my standard error be misleading? Indeed, the standard errors of the cross-sectional means (depending on the sample size $N$) are irrelevant for the standard error of the time series average.

Consider the extreme case: every time period is identical: $y_{i,t}=y_{i,t'}$ for all $t,t'\in\{1,...,T\}$. Then, the standard error of $\bar{\bar{y}}$ would be zero, regardless of how dispersed $y_{i,t}$ is in the cross-section.

Instead, I would almost like to calculate the time series mean of the standard errors of $\bar{y}_t$ to get a feeling for how much $y_{i,t}$ varies cross-sectionally in the average time period?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.