2 Separation Problems

Separation occurs in discrete choice models when a linear combination of one or more independent variables perfectly predicts a category of the outcome variable (Albert and Anderson Reference Albert and Anderson1984). There are two main reasons why separation occurs: (1) at least one of the parameters is infinite or (2) the true parameters are finite, but perfect prediction occurs as an artifact of a particular DGP and realized sample. We focus only on the latter case. Here, separation can be thought of as a finite-sample problem: if enough additional data is collected, the problem disappears.

In cases like these, where the true parameters are finite, separation creates theoretical and practical problems. To understand these problems, consider a sample where a single predictor perfectly predicts a category of the outcome variable. In such a situation, the sample log-likelihood function is monotonic in the estimate of that predictor’s parameter (i.e., better fit can always be found by moving the estimate further from zero). As Albert and Anderson (Reference Albert and Anderson1984) show, because of the monotonicity, there is no unique MLE that solves the first order conditions. Instead, the log-likelihood converges to an asymptote as the estimate goes to $\pm \infty $ , depending on the true parameter’s sign.

Regarding the estimator’s theoretical finite-sample properties, recall that bias is defined based on the expected value of the MLE (i.e., the average MLE over possible samples), and consider a DGP where separation is plausible in any given realized sample. In these situations, the expected value of the MLE includes samples where the estimate is $\pm \infty $ . Therefore, the estimator’s moments are undefined.

Concerning practical problems in estimation, separation leads to numerically computed estimates and standard errors that are much larger than the truth.Footnote ³ Because of the monotone log-likelihood, the numerically obtained MLE will tend to be (i) much larger in magnitude than the true parameter and (ii) a function of the optimization software’s numeric tolerance (Zorn Reference Zorn2005). To put this another way, while the true MLE is infinite, numerical optimizers will return a finite estimate that is typically much larger than the true parameter. Additionally, because a unique MLE does not exist, tests based on asymptotic results are likely misleading as a unique MLE is a standard regularity condition for these results.

These inflated estimates may not be of major concern, if the standard errors also inflate enough to prevent type-1 errors; however, there is no guarantee that this will be the case. In our replication study below, where separation is detected, some null hypotheses are rejected only when the separation problem is ignored but not once it is corrected. While it is impossible to say which decision is correct, the presence of separation suggests that the former is more suspect than the latter. Additionally, inflated standard errors raise the prospect of type-2 errors and under-powered studies. In our simulations, we find that separation can severely affect power, and in Section B.5 of the Supplementary Material, we show an example where both type-1 and type-2 errors can increase when separation is present but goes uncorrected.

Two further complications emerge in moving from the binary to multinomial outcomes. First, because there are more categories in the outcome, samples need to be larger in order for the threat of separation by chance to disappear. For example, with one binary regressor and a binary outcome, we just need enough observations for every box in the cross tabulation to be filled. As the number of outcomes increases, this task requires more observations. Second, common implementations of multinomial models (e.g., Stata or R) provide neither warnings of possible separation nor make any attempt to identify problematic regressors.

Moving to the strategic setting introduces two more complications. First, standard visual diagnostics are less informative. Specifically, common rules-of-thumb ask analysts to look for estimates that are implausibly large, while this can be an important red flag, it is often difficult to know exactly how big is too big. This determination is clouded in the strategic context where the scale parameter is not always fixed to 1 like it is in ordinary logits and probits. In fact, the scale parameters sometimes contain another player’s estimated choice probabilities (e.g., Signorino and Tarar Reference Signorino and Tarar2006) or are estimated as free parameters (e.g., Whang, McLean, and Kuberski Reference Whang, McLean and Kuberski2013), making the context of “too big” difficult to pin down. Second, strategic models contain interdependent and endogenous parameters by construction. When separation leads to inflated estimates in one player’s utility function, this inflation can spill over into estimates of that player’s conditional choice probability, which then affects the estimation of other players’ utility functions. Analyzing strategic interdependence is a main motivator of structural modeling, but care must be taken to minimize biases that may cascade up a game tree.

2.1 Separation Corrections

With logits and probits, the primary existing solutions to the separation problem involve PL estimation (Zorn Reference Zorn2005). Penalization requires the analyst to impart some extra-empirical information (i.e., information from outside the data) to induce numerical stability in the optimization routine. We want to choose information that encapsulates our belief that the coefficient estimates should not be too large. From a Bayesian perspective, penalization is a type of prior belief where the true parameters are unlikely to be huge for any particular variable. As Gelman et al. (Reference Gelman, Jakulin, Pittau and Su2008) put it, the key idea is that large changes on the logit/probit scale (typically 5 or more) are very rare and the penalty/prior should reflect this understanding (Reference Gelman, Jakulin, Pittau and Su2008, 1361). In most cases, this information takes the form of a Jeffreys prior penalty term that is maximized when the parameters are all zero, although others propose penalty terms based on the Cauchy with median 0 and scale 2.5 (Gelman et al. Reference Gelman, Jakulin, Pittau and Su2008) or log- $F(1,1)$ (Greenland and Mansournia Reference Greenland and Mansournia2015).Footnote ⁴ All of these penalties pull the estimates away from $\pm \infty $ and toward 0.

Before deriving the BR strategic estimators, we first describe the model. Consider the extensive-form deterrence game in Figure 1. There are two actors, A and B, each of whom has two actions $y_i \in \{0,1\}$ for $i \in \{A,B\}$ . At the start of the game, each player receives private information in the form of an action-specific shock $\varepsilon _i(y_i)$ . Each shock reflects private information that i has regarding her payoff for taking action $y_i$ .

Figure 1

Standard two-player deterrence game.

After receiving her information, A acts. If A chooses $y_A=0$ , the game ends at the status quo ( $SQ$ ). However, if A challenges B by taking action $y_A=1$ , then B responds by either backing down to A’s challenge by taking action $y_B=0$ (ending the game at $BD$ ) or standing firm against A by taking action $y_B=1$ (ending the game at $SF$ ). When the game ends at outcome $o \in \{SQ, BD, SF\}$ , players receive a payoff equal to $U_i(o)+\varepsilon _i(y_i)$ . This payoff contains a deterministic component: $U_i(o)$ representing a commonly known and observable payoff to each player and a stochastic component: $\varepsilon _i(y_i)$ , which is the privately known cost/benefit to player i for taking action $y_i$ .

The solution concept for this game is quantal response equilibrium (QRE). At the QRE, B chooses 1 if $U_B(SF)+\varepsilon _B(1)> U_B(BD)+\varepsilon _B(0)$ , which can be described as

$$ \begin{align*} y_B &= \mathbb{I}\left[U_B(SF) - U_B(BD) + \varepsilon_B(1) -\varepsilon_B(0)>0 \right], \end{align*} $$

where $\mathbb {I}[\cdot ]$ is the indicator function. Likewise, A chooses 1 if

$$\begin{align*}y_A =\mathbb{I}\left[(1-\Pr(y_B=1)) U_A(BD) + \Pr(y_B=1)U_A(SF)- U_A(SQ) + \varepsilon_A(1)- \varepsilon_A(0)> 0\right]. \end{align*}$$

To transform this game into an empirical model, we need to (i) specify the deterministic portion of the utilities in terms of observed data and (ii) assume a distribution for the action-specific shocks. For exposition, consider the following specification:

$$ \begin{align*} U_B(SF) &= X_{B} \beta,\\ U_B(BD) &= 0,\\ U_A(SQ) &= X_{SQ}\; \alpha_{SQ,}\\ U_A(BD) &= X_{BD} \;\alpha_{BD,}\\ U_A(SF) &= X_{SF}\; \alpha_{SF,}\\ \Pr(y_B=1) &= p_B = F_B\left(U_B(SF)\right), \\ \Pr(y_A=1) &= p_A = F_A\big((1-p_B)U_A(BD) + p_BU_A(SF)- U_A(SQ)\big), \end{align*} $$

where $F_i$ is the distribution that describes $\varepsilon _i(1)-\varepsilon _i(0)$ . Our goal is to estimate the parameters $\theta =(\alpha ,\beta )$ using D observations of actors playing this game. Standard practices estimate $\theta $ in one of two ways: a full information maximum likelihood (FIML) estimator or a two-step from Bas, Signorino, and Walker (Reference Bas, Signorino and Walker2008) called statistical backward induction (SBI).

2.1.1 Statistical Backward Induction

The SBI procedure is as follows:

1. 1. Using only observations, where $y_A=1$ , regress $y_B$ on $X_B$ using a logit or probit (depending on $F_B$ ) to produce $\hat {\beta }^{SBI}$ . Estimate $\hat {p}_{B}^{SBI} = F_B(X_B \hat {\beta }^{SBI})$ .

2. 2. Regress $y_A$ on $Z^{SBI} = \begin {bmatrix} -X_{SQ}& X_{BD}(1-\hat {p}_{B}^{SBI}) & X_{SF}(\hat {p}_{B}^{SBI})\end {bmatrix}$ using a logit or probit (depending on $F_A$ ) to produce $\hat {\alpha }^{SBI}$ .

Note that, because each step is a binary choice model, the MLE for $\hat {\theta }$ solves

(1)

where $d =1,\ldots , D$ indexes each observed play of this game.

Because this approach relies on two distinct binary outcome models, standard PL-based solutions apply. Let $L_B(\beta \mid y)$ and $L_A(\alpha \mid y)$ be the objective functions in Equation (1), then the bias-reduced SBI (BR-SBI) estimates are

(2)

where g is the logged penalty function. If the penalty is a density function (e.g., Cauchy or log- $F$ ), then g is the logged density function, while if g is the Jeffreys prior penalty, then

$$\begin{align*}g(\cdot)=\frac{1}{2}\log(\det(I(\cdot))),\end{align*}$$

where I is the estimated Fisher information matrix calculated using the Hessians of the uncorrected log-likelihoods. Firth (Reference Firth1993, 36) suggests that standard errors for $\hat {\beta }^{BR-SBI}$ can be estimated using $I\left(\hat {\beta }^{BR-SBI}\right)^{-1}$ . This means that standard errors for $\hat {\alpha }^{BR-SBI}$ can be estimated using common two-step maximum likelihood results.

2.1.2 Full Information ML

The SBI estimator is easily implemented, but this ease comes at the cost of statistical efficiency. The FIML maximizes a single log-likelihood function that re-computes the choice probabilities at every step in the optimization process. Because the theoretical model has a unique equilibrium, the FIML is consistent and asymptotically efficient.

Using the above parameterization, the FIML estimates maximize the log-likelihood:

(3) $$ \begin{align} L(\theta \mid y) &= \sum_{d=1}^D \left\{ \mathbb{I}(y_{{A,d}}=0)\log(1-p_{A,d}) \right. \nonumber\\ &\qquad\quad + \mathbb{I}(y_{{A,d}}=1) \mathbb{I}(y_{{B,d}}=0) \log(p_{A,d} \cdot (1-p_{B,d}))\\ & \quad\qquad + \left.\mathbb{I}(y_{{A,d}}=1) \mathbb{I}(y_{{B,d}}=1) \log(p_{A,d} \cdot p_{B,d}) \right\}, \nonumber \end{align} $$

and the bias-reduced FIML estimates are given as

(4)

If g is the logged Jeffreys prior, then the Hessian of Equation (3) needs to be computed at each step in the numeric optimization process: a non-trivial task. Alternatively, Cauchy or log- $F$ penalties can also be used. We provide an extension to R’s games package called games2 that allows analysts to fit the BR-FIML with Jeffreys prior, Cauchy(0, 2.5) or log- $F(1, 1)$ penalties.

In choosing among these three penalties, we point out some pros and cons. The main advantages of the Jeffreys prior are that it is widely used and implemented for binary outcome models; as such, the BR-SBI with Jeffreys prior can be easily fit using existing software. For the FIML, however, the Jeffreys prior requires that the Hessian be negative definite at every guess of the parameter values. This requirement always holds with logits and probits but can fail in more complicated likelihoods. When the logged Jeffreys prior does not exist, density-based penalties based on the Cauchy or log- $F$ distributions provide easy-to-use alternatives. Additionally, the density-based penalties perform best in simulations. In particular, the log- $F$ penalty performs very well, although all three offer vast improvements over the uncorrected methods. Furthermore, Beiser-McGrath (Reference Beiser-McGrath2020) finds that the Firth correction can be problematic in the kind of large-N, rare-events data that dominate international relations. Specifically, he finds that the Jeffreys prior penalty can produce estimates that are in different directions from the original results, implying that this penalty may do more than just shrink the estimates. Separation-induced inflation is always away from zero, so sign changes are concerning. Given this finding, the density-based penalties may be preferred, but we recommend that analysts consider multiple penalties where possible to ensure that the corrections are not dependent on the specific penalty.

3 Performance

We now consider Monte Carlo experiments to compare the BR-SBI and BR-FIML estimators given by Equations (2) and (4), respectively, to their unpenalized counterparts. The experimental setup is presented in Figure 2, where we consider four parameters. The $\beta $ parameters and the variable $X_B$ characterize B’s payoffs, whereas the $\alpha $ parameters and $X_A$ form A’s payoffs. Regressors $X_A$ and $X_B$ are i.i.d. Bernoulli(0.5), whereas the values of $\alpha $ and $\beta $ are chosen to induce separation. In the interest of space, we present the simplest experiment here, while additional and more realistic simulations are deferred to the Supplementary Material.

Figure 2

Monte Carlo version of the two-player deterrence game.

Our main simulation considers a sparse model where separation is likely to emerge in the data recording B’s choice of 0 or 1. Let B’s choice be given by

$$ \begin{align*} y_B &= \mathbb{I}\left[ -1 + 4 X_B + \varepsilon_B(1)> 0+ \varepsilon_B(0) \right]\\ &=\mathbb{I} \left[ -1 + 4 X_B + \varepsilon_B(1) - \varepsilon_B(0) > 0\right]. \end{align*} $$

Each error term is i.i.d. standard normal, such that $p_B = \Phi \left (\frac {-1 +4 X_B}{\sqrt {2}}\right )$ . Note that a large, but not unreasonable, coefficient on $X_B$ will ensure that in most samples $y_B=0$ only when $X_B=0$ .

The DGP for player A is

$$ \begin{align*} y_A &= \mathbb{I}\left[-2.5 (X_Ap_B) + \varepsilon_A(1)> 1.5 + \varepsilon_A(0) \right]\\ &= \mathbb{I}\left[-1.5 -2.5 (X_Ap_B) + \varepsilon_A(1) - \varepsilon_A(0) > 0 \right]. \end{align*} $$

In terms of Figure 2, the parameters of interest are $\alpha _0 = 1.5$ , $\alpha _1 = -2.5$ , $\beta _0=-1,$ and $\beta _1=4$ . We repeat the Monte Carlo experiment 5,000 times with samples of size $D=500$ and keep the results where the lp-diagnostic detects separation between $X_B$ and ending the game at outcome $BD$ . In cases where the lp-diagnostic does not detect separation, the results are nearly identical across estimators. As with many applications of strategic probits, the status quo is the most common outcome (about 90% of observations), whereas $BD$ and $SF$ each emerge about 5% of the time. This means that the first step of the SBI typically has about 50 observations to use.

Before considering the simulation results, one additional point is worth mentioning. Recall that the expected value for $\hat {\beta }_1$ is undefined for the uncorrected estimators. As such, the observed estimates are whatever values get “close enough,” such that the optimization software issues a successful convergence code. In other words, the numeric estimates produced by the ordinary SBI and FIML estimators reflect a type of regularization: They will be closer to the zero (and the truth) than the true MLE of $\pm \infty $ , but in ways that are highly dependent on algorithm and tolerance choices.

3.1 Parameter Estimates

The Monte Carlo results are reported in Table 1. The first thing to note is that the BR techniques makes a noticeable and positive impact on both the point estimates and their precision. This translates into substantial decreases in the multivariate root-mean-squared error (RMSE). For both the SBI and the FIML, the PL approach helps when separation is present. The BR-FIML (log-F) has the smallest RMSE of all the estimators considered, while also having the least bias in estimating $\beta _1$ .

Table 1

Monte Carlo results when separation is present in Player B’s decision.

Note: SD refers to the standard deviation of estimates produced by the simulation. SE refers to the standard errors produced by each estimator averaged over simulations. True standard errors are estimated using Hessian curvature at the true parameter values and the data within each simulation and then averaging over simulations. Power refers to the proportion of simulations where the null hypothesis is correctly rejected, and coverage refers to the proportion of simulations where the 95% confidence interval contain the true value.

Second, we see that the FIML estimators tend to outperform their SBI counterparts. One reason for this is that the FIML is a system estimator and will be more efficient by construction. However, it is also worth noting that the separation-induced inflation is worst in the unpenalized SBI and that while the FIML still exhibits bias, its RMSE is about 3/4 that of the SBI. These differences emerge in part, because the SBI is less efficient by construction, but they are mostly due to differences in their default fitting algorithms.Footnote ⁶

3.4 Choice Probabilities

Moving beyond the point estimates, $\hat {p}_B$ plays a key part in fitting the model, particularly for the SBI. As such, we want to know if any of these corrections have negative consequences on estimating $p_B$ . In Table 2, we consider the statistical properties of $\hat {p}_B$ . Because $X_B$ is binary, there are only two values that $p_B$ can take on, making it easy to break down this analysis by $X_B$ . There are three important takeaways from these results. First, the BR-FIMLs are more biased when estimating $p_B$ when $X_B=0$ , this result matches Rahman and Sultana (Reference Rahman and Sultana2017) who finds that BR correction in the parameters can sometimes make bias in predicted probabilities worse. Second, despite this bias when $X_B=0$ , the BR estimators offer modest improvements in RMSE when $X_B=0$ and substantial improvements in both bias and RMSE when $X_B=1$ . These latter results are unsurprising given the inflation in $\hat {\beta }_1$ . Third, when combining the results, we see that the three BR-FIMLs are most preferred from a RMSE perspective, despite having more bias when $X_B=0$ . The bias and RMSE improvements they offer when $X_B=1$ offset these concerns in this experiment.

Table 2

Bias and RMSE in estimating $p_B$ .

4 Application: Deterrence in Interstate Disputes

We now reexamine results from Signorino and Tarar (Reference Signorino and Tarar2006) who study deterrence in interstate disputes using data on 58 crises between 1885 and 1983. The players in this game are an aggressor and defender state. The aggressor (A) decides between attacking a protégé state of the defender (B) or preserving the status quo. If A chooses the latter, the game ends, but if A chooses the former, then the defender can either protect its protégé or back down. The dependent variable takes on three values: status quo, attack-war, and attack-back down. Section C of the Supplementary Material contains descriptions of the independent variables and the model specification. We start by applying the lp-diagnostic to the data. The diagnostic results are reported in Table 3, where four of five checks provide evidence of separation.Footnote ⁹

Table 3

Checking for separation in Signorino and Tarar (Reference Signorino and Tarar2006).

Note: The Z variables are transformed using estimates of $p_B$ from the unpenalized estimators.

Compounding the separation problem is the issue of fitting a complicated strategic model to a relatively small sample. In replicating these results, we found that the determinant of the FIML information matrix is negative at many steps in the optimization process, making the logged Jeffreys prior penalty term undefined. As a result, we use the log- $F$ penalty as it does not rely on the curvature of the baseline log-likelihood and performed well in simulations. The BR-SBI continues to use the Jeffreys prior penalty here as the probit objective function does not have the same complexity as the FIML, the penalty always exists, and it remains the most common choice for binary-outcome models. Beyond these difficulties, we also note that fitting a 21-parameter strategic model with 58 observations is a demanding proposition. Nonetheless, this example provides us with a clear case where separation is present.

The results are presented in Table 4. Fitting the ordinary SBI produced severe numerical instability; as such, the estimates and standard errors are the means and standard deviations from a non-parametric bootstrap where we discard results beyond $\pm 50$ to keep everything on roughly the same scale across the estimators. The fact that we even had to consider this approach with the SBI is a warning against using an uncorrected model. There are slight differences between the replicated FIML and published results, which we attribute to slight differences in software implementation.

Table 4

Signorino and Tarar replication.

Note: Standard errors in parentheses (Model 6 is bootstrapped).

What is most striking about the results in Table 4 is that while many of the point estimates have the same sign across all four estimators, some results that were significant in the Signorino and Tarar (Reference Signorino and Tarar2006) analysis are no longer significant at traditional levels. Additionally, we note that the estimates and standard errors on the uncorrected SBI are incredibly large despite the precautions we took to make the estimates appear more reasonable. Combining this observation with the lp-diagnostic results provides us with good reason to suspect that a BR estimator may be more appropriate. Indeed, the two BR estimators largely agree with each other in terms of magnitude and sign in 18 out of 21 estimates, although in the BR-SBI case fewer estimates are statistically significant at standard levels. This difference may result from the relative inefficiency of the two-step estimator.

The fact that a few estimates change signs across the estimators is an interesting puzzle. Specifically, cases where signs are different across the ordinary-SBI and BR-SBI are unexpected. Correcting for separation is not supposed to change the direction of an estimate, although Beiser-McGrath (Reference Beiser-McGrath2020) notes that this can happen in some binary-outcome cases with the Jeffreys prior penalty. He finds this to be the case in large, rare-events data. Here, however, we see sign flips in small, even-event data, and it also occurs with the density-based penalties. These unexplained sign flips may suggest that there may be some heretofore unknown issues with BR estimation in (very) small samples. In some exploratory simulations, we find that sign flips can happen in small, highly collinear samples like this one, but we cannot be certain that collinearity is causing the sign flips here. Future work should spend more time on this puzzle as it is very unusual to see signs change when applying PL methods.

In examining player B’s (the defender’s) utility function, Signorino and Tarar (Reference Signorino and Tarar2006) find that the defender is more likely to support its protégé if B has nuclear weapons, if the protégé imports a lot of arms from B, and if there was a past, but unresolved, crisis between the defender and the aggressor (Reference Signorino and Tarar2006, 593). Our analysis concurs with these results in terms of sign, but only the effect of nuclear weapons remains significant at the 5% level. The overall decrease in coefficient magnitudes is consistent with a separation problem. The changes in significance suggest that some original findings resulted from separation-induced inflation in the point estimates that exceeded the inflation in the standard errors. Many of these findings may, of course, still be true, but we cannot reject these null hypotheses with these data once we correct for separation.

The uncorrected SBI is the most conservative model here: it rejects no hypotheses and, as such, makes no type-1 errors. In contrast, we may suspect that the uncorrected FIML is guilty of some type-1 errors, making the SBI, and its extreme results, a safe choice for cautious analysts. However, this protection against type-1 errors comes at the cost of power. Based on the simulations in Table 1 and in Section B of the Supplementary Material, we find that the uncorrected SBI has almost no power to identify effects on coefficients where separation is a concern. Analysts can weigh their own acceptance for type-1 and type-2 errors, but we find that the BR estimators present a good balance between these two concerns.

To better demonstrate these numeric issues and illustrate how the BR corrections work, we consider the profiled log-likelihood of the FIML, the BR-FIML, and the BR-SBI for the coefficient on military alliance in B’s utility function. We focus on this variable as the uncorrected coefficient estimate of about 13 (against a scale of $\sqrt {2}$ ) is suggestive of a separation problem. The profiling procedure fixes the value of a single coefficient and refits the model. Repeating this procedure at many values demonstrates the model’s sensitivity to changes in this estimate. For a well-behaved problem, we would expect a classic upside-down U shape with a maximum at the estimated parameter value. The profiled results are shown in Figure 3. Specifically, for the ordinary FIML (top pane) there appears to be a local maxima at the estimate, but model fit can be improved by increasing this estimate past positive 20. Put another way, while the estimate is a local maximum, it is not the global maximum; “better” fit can be found at estimates further toward $\infty $ . This push toward $\pm \infty $ is the classic sign of the separation problem. Looking at the two BR profiles we see that, at least in the range considered, the estimates are at the maximum. Note that the BR-SBI has a flattish section at the right-hand end of the plot, however, this drops off quickly if we explore past this region, and we find no reason to suspect that there are better log-likelihood values beyond the range presented here.

Figure 3

Profiled log-likelihood on the coefficient associated with how a military alliance affects B’s decision to intervene.