When reading through https://www.math.lsu.edu/~smolinsk/Quant_Interview_Prep.pdf, I saw the question "Discuss correlation matrices with panel data"

How would you do a correlation matrix with panel data and why/how is this different to whatever the "normal" situation would be?


Consider the case of a simple linear regression using panel data. So, for example, assume we are trying to figure out how age and education affect individuals' income. Also, forget the whole log-linear regression part. To keep this simple, I'm just going to use a simple linear regression. In essence, I want to find the coefficients for the following equation:

$$ income = \beta_0 + \beta_{educ}\cdot educ + \beta_{age}\cdot age +\epsilon$$

where $\epsilon$ follows some random distribution (typically, normal, centered around zero).

For this example, assume we have surveyed the same set of individuals multiple times over the course of a few years.

Panel data is typically modeled assuming that there is an individual-specific term that dictates the individual-level panel effect. So the equation that describes your model would be something like:

$$ income_{n,t} = \beta_0 + \beta_{educ}\cdot educ_{n,t} + \beta_{age}\cdot age_{n,t} + c_n + \epsilon_{n,t}$$

where $n$ is the index for each individual ($n=1,...,N$), $t$ is the index for each surveying period ($t=1,...,T$), and $\epsilon_{n,t}$ is assumed to be i.i.d. across individuals and time (i.e., independently distributed across all dimensions). Furthermore, $c_n$ is typically assumed to be either fixed or normally distributed across individuals according to some distribution. Regardless of it being fixed or random, notice how $c_n$ is constant within each individual over time. This is supposed to capture everything that is "special" about that individual and that sets them apart from their peers (note: you can also add a $d_t$ term to represent year-specific effects, but I'm omitting that just to keep things simple). But as I mentioned before, the $c_n$ term is time-invariant (i.e., it remains constant over the multiple $t$ survey periods). Therefore, this approach makes it impossible to teem out the effect of other time-invariant characteristics, such as race and gender.

That's where the correlation matrix comes in. Modeling panel data using a correlation matrix relies on assuming the following equation:

$$ income_{n,t} = \beta_0 + \beta_{educ}\cdot educ_{n,t} + \beta_{age}\cdot age_{n,t} + \epsilon_{n,t}$$

where, for each individual, $\epsilon_{n,t}$ is normally distributed across the multiple $t$ years following some correlation matrix $\boldsymbol{\Sigma}$. Stated clearly:

$$ \boldsymbol{\epsilon}_n = (\epsilon_{n,1},\epsilon_{n,2},...,\epsilon_{n,T}) \sim MultiVariateNormal(\boldsymbol{0},\boldsymbol{\Sigma})$$

where $\boldsymbol{0}$ is a $(T \times 1)$ vector of zeros and $\boldsymbol{\Sigma}$ is a $(T \times T)$ correlation matrix.

When you assume the error term is distributed this way, you can finally add time-invariant characteristics to the model and it will still be estimable using likelihood maximization. Furthermore, the correlation terms within $\boldsymbol{\Sigma}$ are also estimable.

I hope this clears up a bit of the confusion around modeling panel data.

  • $\begingroup$ "you can finally add time-invariant characteristics to the model and it will still be estimable using likelihood maximization. " this bit it not totally clear to me $\endgroup$ – Trajan May 8 '20 at 7:02
  • $\begingroup$ and what the normal situation be (thats not panel data)? $\endgroup$ – Trajan May 8 '20 at 7:02
  • $\begingroup$ What you are calling "the normal situation" (i.e., a case without panel data) is what is usually called "cross sectional data". Here, we just assume that income follows that first equation I wrote and that the error term is independently and identically distributed (iid) across all observations. In clearer terms, it's a case where I only have one observation per individual, not multiple observations per individual. In more concrete terms, a cross-sectional dataset would contain a snapshot of a bunch of people's information at one specific point in time. $\endgroup$ – Felipe D. May 12 '20 at 21:06
  • $\begingroup$ As for the time-invariant part: some characteristics can change over time, such as an individual's age, weight, education level, etc. Other, however, cannot: race, eye color, height (at least after a certain age). When using panel data, the "classic" panel data approach (i.e., not using correlation matrices) makes it impossible to estimate the effects of these time-invariant variables. However, if you are using panel data with the correlation matrix approach, you can estimate the effect of time-invariant variables such as race. $\endgroup$ – Felipe D. May 12 '20 at 22:24
  • $\begingroup$ If you have panel data but want to use a non-panel approach (i.e., purely cross-sectional data), you can also estimate the effect of these time-invariant effects. However, the estimates of this cross-sectional approach on panel data will likely be biased because there will be some sort of correlation in the error terms of some of the observations. $\endgroup$ – Felipe D. May 12 '20 at 22:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.