The Concept of Identification in Simultaneous Equations Models

Imagine you’re at a bustling fruit market. You see that when prices for mangoes are high, fewer people buy them, but farmers are happy to sell more. When prices are low, shoppers flock to the stalls, but farmers are less enthusiastic. You plot these price and quantity points in a notebook. At the end of the day, you have a cloud of dots. Now, a friend asks, “Can you draw the demand curve from that data?” You try, but then you pause. “Wait… am I tracing the demand curve, or am I tracing the supply curve?”

This simple market scenario captures the essence of one of the most fundamental challenges in econometrics: the identification problem. In many economic systems, variables aren’t just related in one direction; they are determined together, simultaneously. Price affects demand, but demand also affects price. When we just look at the data-the final equilibrium points-we can’t be sure which underlying relationship we are actually measuring. This is where the concept of identification comes in. It’s the detective work econometricians must do *before* they can even start estimating, to ensure the parameters they’re looking for can actually be found.

What is the identification problem, really? 🔗

At its core, the identification problem asks a simple question: can we find a unique value for each unknown parameter in our model, given the data we have? If the answer is no, our model is unidentified. This means we can’t be certain about the economic relationship we’re trying to estimate.

To understand this, we need to distinguish between two types of models:

• Structural Model: This is the “true” theoretical model, the one we really care about. It describes the underlying economic behaviour. For our market, it would be two separate equations:
  1. The Demand Equation: $Q_d = \alpha_0 + \alpha_1 P + \alpha_2 (\text{Income}) + u_1$
  2. The Supply Equation: $Q_s = \beta_0 + \beta_1 P + \beta_2 (\text{Rainfall}) + u_2$

  The parameters $\alpha_1$ (price elasticity of demand) and $\beta_1$ (price elasticity of supply) are the structural parameters. They are the “truth” we want to uncover.

• Reduced-Form Model: This is the model we can actually estimate from the data. It’s what we get when we solve the system for the endogenous variables (the variables determined *within* the system, like Price and Quantity). In this model, we would express Price and Quantity only as functions of the exogenous variables (the “outside” variables, like Income and Rainfall).

The identification problem is all about translation. We can always estimate the reduced-form parameters from our data. The real question is: can we use those estimated reduced-form parameters to work backward and find a single, unique set of structural parameters? If we can, the equation is identified. If we can’t (maybe no solution exists, or multiple solutions exist), it’s unidentified.

The curious case of observational equivalence 🔗

Why would an equation be unidentified? The most common reason is a sneaky problem called observational equivalence. This happens when two or more different theoretical structures (i.e., different sets of structural parameters) produce the exact same observable data. If the data looks identical, no amount of statistical analysis on *that data alone* can tell you which theory is the correct one.

Let’s go back to the market, but simplify it. Imagine we didn’t know about Income or Rainfall. Our model is just:

Demand: $Q = \alpha_0 + \alpha_1 P$
Supply: $Q = \beta_0 + \beta_1 P$

All we see is the data where $Q_d = Q_s$. We just see a scatter plot of equilibrium points. The problem is, this scatter plot could have been generated by a steep demand curve and a flat supply curve. Or it could have been generated by a flat demand curve and a steep supply curve. Both “theories” would produce the same cloud of data points. The models are observationally equivalent. Because we can’t distinguish between them, we can’t identify *either* equation. We are stuck.

The paradox of identification: How to solve the puzzle 🔗

This leads us to one of the most fascinating and counter-intuitive ideas in economics: the paradox of identification. You might think that to get a good estimate of a relationship, you need to include every variable that could possibly affect it. The paradox states the exact opposite: an equation in a simultaneous system is identified by the variables that are absent from it, but present elsewhere in the system.

This sounds like magic, but it’s not. It’s about finding a way to break the observational equivalence. We need something that will “hold one curve still” while “shifting the other one,” so we can trace its shape.

Let’s revisit our full model:

Demand: $Q_d = \alpha_0 + \alpha_1 P + \alpha_2 (\text{Income})$
Supply: $Q_s = \beta_0 + \beta_1 P + \beta_2 (\text{Rainfall})$

Now, let’s try to identify the demand curve. The “paradox” tells us to look at the variables *excluded* from this equation. The variable Rainfall is in the supply equation but *not* the demand equation. (This is a reasonable assumption; rainfall affects the harvest, but it doesn’t directly change your desire to buy mangoes).

Here’s how it works:

1. Changes in rainfall (say, a drought one year, great rain the next) will shift the supply curve.
2. But, because rainfall isn’t in the demand equation, the demand curve stays put.
3. As the supply curve shifts back and forth (due to the “exogenous” shock of rainfall), it creates new, different equilibrium points (Price, Quantity).
4. Crucially, all these new points lie along the same, stable demand curve.

We are no longer looking at a meaningless cloud. We are literally tracing out the demand curve. The “Rainfall” variable, by being *excluded* from the demand equation, acts as an instrumental variable that allows us to identify the demand curve’s parameters. The same logic applies to identifying the supply curve: the “Income” variable, which is in the demand equation but not the supply equation, shifts the demand curve around and allows us to trace the shape of the stable supply curve.

Checking the status: The conditions for identification 🔗

So, how do we know formally if our equations are identified? We don’t have to guess. Econometricians have developed formal conditions to check the identification status of any equation in a system. These rules are just a mathematical way of formalizing the “paradox” we just discussed. There are two conditions: the order condition and the rank condition.

The order condition: A quick counting rule 🔗

The order condition is a necessary “rule of thumb” for identification. It’s a simple counting exercise. If an equation fails this test, it is definitely unidentified. (If it passes, it *might* be identified, but we need to check the second rule).

Here is the rule, stated simply:

For an equation to be identified, the number of exogenous variables excluded from it must be greater than or equal to the number of endogenous variables included in it, minus one.

Let’s write that as a formula:
Let $M$ = total number of endogenous variables in the whole system.
Let $K$ = total number of exogenous variables in the whole system.
For a specific equation:
Let $m$ = number of endogenous variables *in this equation*.
Let $k$ = number of exogenous variables *in this equation*.
The number of excluded exogenous variables is $(K – k)$.
The number of included endogenous variables minus one is $(m – 1)$.

The Order Condition is: $(K – k) \ge (m – 1)$

This rule leads to three possible outcomes for our equation’s identification status.

The rank condition: The sufficient rule 🔗

While the order condition is a fast check, the “true” test is the rank condition. It is both necessary *and* sufficient. It’s mathematically more complex, involving the ranks of matrices built from the model’s parameters. In essence, the rank condition checks to make sure the excluded variables *actually do their job*. It confirms that the excluded variables truly are independent and have a real, non-zero impact in the *other* equations. If you had two excluded variables that were just perfect copies of each other (perfectly collinear), the order condition might pass, but the rank condition would fail, because you only have *one* unique piece of “shifting” information, not two. For most introductory purposes, econometric texts emphasize that the order condition is the first hurdle to clear.

What to do next: Identification status and estimation 🔗

The identification status of your equation isn’t just a label; it directly dictates what estimation method you are allowed to use. Running a simple Ordinary Least Squares (OLS) on a simultaneous equation will give biased and inconsistent results. You must use a method that accounts for the simultaneity, and your choice depends on the status.

Status 1: Under-identified 🔗

• Order Condition: $(K – k) < (m - 1)$
• Meaning: You have not enough information. There are not enough “shifters” (excluded exogenous variables) to trace out the curve.
• Example: Our first, simple model ($Q = a+bP$, $Q = c+dP$) with no exogenous variables. For the demand equation, $K-k = 0$ and $m-1 = 1$. Since $0 < 1$, it's under-identified.
• What to do: Nothing. You cannot get unique estimates for your structural parameters. Your model is fundamentally flawed. You must go back to your economic theory and find a valid exogenous variable to include in the system (but exclude from this equation).

Status 2: Exactly identified 🔗

• Order Condition: $(K – k) = (m – 1)$
• Meaning: You have just enough information. There is exactly one unique way to solve for the structural parameters from the reduced-form estimates.
• Example: Our full model ($Q_d = f(P, \text{Income})$, $Q_s = f(P, \text{Rainfall})$).
  • For the demand equation: $K-k = 1$ (Rainfall is excluded). $m-1 = 1$ (P and Q are endogenous, so $m=2$). Since $1 = 1$, it is exactly identified.
  • For the supply equation: $K-k = 1$ (Income is excluded). $m-1 = 1$. Since $1 = 1$, it is also exactly identified.
• What to do: You can use an estimation method called Indirect Least Squares (ILS). This involves estimating the reduced-form equations with OLS, and then using algebra to solve for the structural parameters.

Status 3: Over-identified 🔗

• Order Condition: $(K – k) > (m – 1)$
• Meaning: You have more than enough information. You have more “shifters” (instruments) than you strictly need to trace the curve.
• Example: Let’s add another variable, “Fertilizer Cost,” to the supply equation.
  Demand: $Q_d = f(P, \text{Income})$
  Supply: $Q_s = f(P, \text{Rainfall, Fertilizer Cost})$
  Now, for the demand equation: $K-k = 2$ (Rainfall and Fertilizer Cost are excluded). $m-1 = 1$. Since $2 > 1$, the demand equation is now over-identified.
• What to do: This is common and actually good! But you can’t use ILS, because the “extra” information means you’d get multiple, slightly different answers for your parameters. Instead, you must use a method that efficiently combines all this information, like the Two-Stage Least Squares (2SLS) estimator. 2SLS is the workhorse for estimating simultaneous equations in modern econometrics. A huge benefit of over-identification is that it allows you to run tests (like the Sargan test) to check whether your instruments are valid, which is a luxury you don’t have in the exactly identified case.

In the end, identification is the crucial first step that separates a meaningful econometric model from a simple statistical correlation. It forces us to think deeply about our economic theory *before* we ever touch the data, asking “What moves my world, and how can I be sure I’m measuring what I think I’m measuring?”

What do you think? Can you think of another real-world economic example (besides supply and demand) where two or more variables are determined simultaneously? How does the “paradox of identification” change the way you think about what it means to “leave a variable out” of a model?

References

1. https://press.princeton.edu/books/hardcover/9780691122806/econometric-analysis
2. https://ocw.mit.edu/courses/14-382-econometrics-spring-2017/resources/mit14_382s17_lec14/