Fixed effect regression that is not panel data

"Thus, in the fixed effect models, if the heterogeneity is fixed over time, this unobserved heterogeneity can be controlled. This heterogeneity is removable from the data by differencing, for instance, any time invariant components of the model can be taken away by taking a first difference."

We work for an organization controlling for unobserved heterogeneity (unobserved variables) through fixed effect models who do not use data at different points in time. Other than these facts and they use dummies for each state we have few details of what they do.

Anyone know how you control for unobserved heterogeneity if you don't have data at different times.
I remain baffled how you run fixed effects without multiple observations on the dummies for states they use. It seems like that is an unavoidable requirement for fixed effects - but they say they have no panel data.

More baffling to me is that in theory fixed effects deal with all omitted variable issues (ones that are time invariant anyhow). But if that is true why do you add variables to your model, they actually add 41 :p

We want to show them that variables not in their model matter. But how do you do that when they can argue given the fixed effect that variables outside their model are inherently controlled for. :p

Of course if fixed effects works the way people say, why would you ever not use fixed effects regression when you have data at more than one point of time since it eliminates all bias...
Last edited:
This seems important....but I don't really understand it. We just look to see the impact of unstandardized variables mainly interval in nature. For what it is worth we only care about the impact on our customers and we have our customers whole data set.

"Because it reduces concerns that omitted variables drive any associations between dependent and independent variables, many researchers use linear fixed effects regression. This estimator reduces the variance in the independent variable and narrows the scope of a study to a subset of the overall variation in the data set. While some researchers readily acknowledge these features, many studies can benefit from more specificity in the description of the variation being studied and the consideration of counterfactuals that are plausible given the variation being used for estimation (e.g., a typical within-unit shift in X)."

From the discussion he uses this seems to impact the estimates, although only the variance is logically reduced which I don't really understand.

untitled (
Yow this is what makes me wonder if I can run statistics. If professional people get it wrong what chance to amateurs like me have. :p

"So what if I told you … panel data, or data with two dimensions, such as repeated observations on multiple cases over time, is really not that complicated. In fact, there are only two basic ways to analyze panel data, which I will explain briefly in this piece, just as every panel dataset has two basic dimensions (cases and time). However, when we confuse these dimensions, bad things can happen. In fact, one of the most popular panel data models, the two-way fixed effects model–widely used in the social sciences–is in fact statistical nonsense because it does not clearly distinguish between these two dimensions. This statement should sound implausible to you–really?, but it’s quite easy to demonstrate, as I’ll show you in this post."

[A little later]
In short: there are an untold number of analyses of panel data affected by an issue that is almost impossible to identify because R and Stata obscure the problem. Thanks to multi-collinearity checks that automatically drop predictors in regression models, a two-way fixed effects model can produce sensible-looking results that are not just irrelevant to the question at hand, but practically nonsense. Instead, we would all be better served by using simpler 1-way fixed effects models (intercepts on time points or cases/subjects, but not both).

What Panel Data Is Really All About | Robert Kubinec

Notice he did not say SAS messes it up, probably because he never runs SAS. :p
In comparing random effect models to fixed effect models Allison argues

"The more parsimonious model (the random effects model in this case) will lead to more efficient estimates, but those estimates might be biased if the restrictions of the model are incorrect. The less parsimonious model (the fixed effects model) is less prone to bias, but at the expense of greater sampling variability."

We have the only population we care about our own. So we are primarily interested in avoiding bias, we don't even do statistical test normally. So that means the fixed effects model is better for us right? :p