Why Stabilized Weights Can Still Have Heavy Tails

Why Stabilization Does Not Guarantee Well-Behaved Weights

A stabilized treatment weight is a ratio: its numerator describes a simpler treatment mechanism, while its denominator describes the probability or density of the observed treatment under the conditioning history. A numerator can make weights less variable in many settings, but it cannot ensure that the denominator is safely bounded away from zero. If an observed treatment is very unlikely under the denominator model, the ratio can still be enormous.

This matters because an analysis can be sensitive to a small number of observations even when most weights look ordinary. A heavy right tail can make the variance very large or infinite, undermine large-sample approximations, and make estimated effects depend strongly on a few observations. The issue is not restricted to a poorly chosen numerator: the denominator probabilities, the fitted treatment model, and multiplication across decision times can each contribute.

Definition: Let $A_t$ be treatment at decision time $t$, let $\bar A_{t-1}$ denote prior treatment history, and let $H_t$ denote the observed history used in the denominator treatment model. A stabilized weight through time $K$ has the form $$ SW_K=\prod_{t=0}^{K} \frac{q_t(A_t\mid \bar A_{t-1},V)} {g_t(A_t\mid \bar A_{t-1},H_t)}, $$ where $q_t$ is the numerator probability or density and $g_t$ is the denominator probability or density. For continuous treatment, the factors are densities evaluated at the observed treatment. A factor can be large when its denominator is small relative to its numerator; stabilizing the numerator alone imposes no general upper bound on that ratio.

The relevant positivity condition is that, on histories occurring with positive probability, treatment values represented by the numerator mechanism also have positive probability under the denominator mechanism. Positivity in this support sense does not mean that all denominator probabilities are uniformly large. A denominator can be strictly positive everywhere and still become arbitrarily close to zero over histories that occur. In practice, sparse histories and fitted probabilities near zero can create the same numerical problem even when the underlying probabilities are not known.

A Mean of One Can Hide a Heavy Tail

A simple binary-treatment model makes the distinction between a finite mean and a stable distribution explicit. Suppose $X\sim \operatorname{Uniform}(0,1)$, and, conditional on $X=x$, treatment is assigned with probability $p(x)=x^\alpha$, where $\alpha>0$. Focus on treated individuals. The stabilized numerator is the marginal treatment probability $q=P(A=1)$.

Result: In this model, the treated-unit stabilized weight $W=q/p(X)$ has conditional mean one. This does not imply that its variance is finite.

To prove the mean statement, first calculate $q$:

q=P(A=1)=\int_0^1 x^\alpha\,dx=\frac{1}{\alpha+1}.

The conditional density of $X$ among treated individuals follows from Bayes' rule. Since the marginal density of $X$ is 1 on $(0,1)$,

f_{X\mid A=1}(x)=\frac{P(A=1\mid X=x)f_X(x)}{P(A=1)} =\frac{x^\alpha}{q}=(\alpha+1)x^\alpha,\qquad 0<x<1.

Therefore, using $W=q/X^\alpha$,

E[W\mid A=1] =\int_0^1 \frac{q}{x^\alpha}(\alpha+1)x^\alpha\,dx =q(\alpha+1)\int_0^1 1\,dx=1.

The cancellation in this calculation explains why an average weight of one is compatible with very large individual weights. The mean accounts for the frequency of low-probability histories; it says nothing by itself about how variable their weights are. $\square$

Worked Example: A Polynomial Tail Despite Stabilization

Continue with $p(x)=x^\alpha$ and $q=1/(\alpha+1)$. For a treated individual, $W=q/X^\alpha$. Find the upper-tail probability and determine when the variance exists for $\alpha=2$.

For $w\ge q$, the event $W>w$ is equivalent to $$ \frac{q}{X^\alpha}>w \quad\Longleftrightarrow\quad X<\left(\frac{q}{w}\right)^{1/\alpha}. $$

Using the treated conditional density derived above, set $x_0=(q/w)^{1/\alpha}$. Then $$ P(W>w\mid A=1)=\int_0^{x_0}(\alpha+1)x^\alpha\,dx =x_0^{\alpha+1} =\left(\frac{q}{w}\right)^{(\alpha+1)/\alpha}. $$

This is a polynomial, rather than exponentially declining, tail. More generally, for $r>0$, $$ E[W^r\mid A=1] =q^r(\alpha+1)\int_0^1 x^{\alpha-\alpha r}\,dx. $$ The integral is finite exactly when $\alpha-\alpha r>-1$, or $r<1+1/\alpha$. With $\alpha=2$, the tail exponent is $3/2$, and moments exist only for $r<3/2$. The mean ($r=1$) is finite, but the second moment ($r=2$) and hence the variance are infinite. The numerator is stabilized at the fixed value $q=1/3$; the heavy tail arises because $p(x)$ approaches zero as $x$ approaches zero.

How Small Denominators Arise in Practice

The continuous example isolates a general mechanism: some histories make the observed treatment very unlikely according to the denominator model, while the numerator remains appreciable. The same mechanism occurs with binary or multivalued treatment. A ratio need not have a large numerator in absolute terms; a small enough denominator is sufficient.

Worked Example: A Sparse History Produces a Large Ratio

Suppose a particular history has denominator probability $g(A_t=1\mid H_t)=0.004$, while the corresponding numerator probability is $q(A_t=1\mid \bar A_{t-1},V)=0.20$. For an individual with that history who receives treatment, the time-specific factor is $$ \frac{0.20}{0.004}=50. $$

The denominator probability is small, not zero, so this factor does not by itself demonstrate a literal violation of support positivity. It does show how weak practical overlap can produce a large contribution from one observed treatment. If the history is rare, there may be too little data to estimate its treatment probability reliably; if it is common enough to occur, an observed treated individual there can still receive substantial weight. A stabilized numerator has not removed the effect of the small denominator.

A second route is model behavior. For example, a fitted logistic denominator model can assign probabilities near zero to some covariate patterns. This can occur when the fitted linear predictor is very negative, including in sparse data or near separation. The resulting ratios can be extreme even if the model returns probabilities strictly between zero and one. Fitted probabilities are estimates, so a very small fitted value can also be noisy: small changes in model fitting or the data may substantially change the corresponding weight.

Worked Example: A Fitted Probability Near Zero

Suppose a logistic denominator model gives a fitted linear predictor of $-8$ for a history, so that the fitted probability of the observed treatment is $$ \hat g=\operatorname{expit}(-8)=\frac{1}{1+\exp(8)}\approx 0.000335. $$ If the numerator probability for that treatment is $0.20$, the estimated factor is $$ \frac{0.20}{0.000335}\approx 597. $$

For comparison, if the same numerator is divided by a fitted denominator of $0.10$, the factor is $0.20/0.10=2$. The large change comes from the denominator, not from a change in stabilization. The value 597 is a direct calculation, not evidence by itself that the fitted model is correct or that the true probability equals $0.000335$. It signals a history for which the estimated weight is highly sensitive to the treatment model and available data.

Longitudinal Products Compound the Problem

In a longitudinal analysis the cumulative weight is a product. Thus, several factors that are individually unremarkable can combine into a large value. Conditional dependence between factors is common because histories evolve over time, so a simple independent-factors calculation should not be assumed to describe real data. Still, a basic moment identity shows precisely why products can be consequential when independence does hold.

Result: If $R_1,\ldots,R_m$ are independent nonnegative random variables, then for any $r>0$, $$ E\left[\left(\prod_{j=1}^m R_j\right)^r\right] =\prod_{j=1}^m E[R_j^r], $$ where either side may be infinite.

To see this, write the expectation as an integral with respect to the joint distribution. Independence makes that distribution the product of the marginal distributions, and the integrand factors as $\prod_j R_j^r$. Tonelli's theorem applies because the integrand is nonnegative, including when one or more integrals are infinite. The multiple integral therefore factors into the product of the marginal integrals. This identity is not a claim that longitudinal weight factors are independent; it makes clear that, even under independence, moment behavior of a product depends on the moments of every component. $\square$

Worked Example: Two Moderate Factors Become a Large Cumulative Weight

Consider a treated trajectory with two observed treatment factors. At the first time, the numerator probability is $0.20$ and the denominator probability is $0.04$, giving a factor of $$ \frac{0.20}{0.04}=5. $$ At the second time, conditional on the observed history, the numerator probability is $0.25$ and the denominator probability is $0.0625$, giving $$ \frac{0.25}{0.0625}=4. $$ The cumulative weight is $$ SW_1=5\times4=20. $$

Equivalently, the product of the numerator probabilities is $0.20\times0.25=0.05$, and the product of the denominator probabilities is $0.04\times0.0625=0.0025$. Their ratio is $0.05/0.0025=20$, matching the product of the two factors. Neither individual factor is in the hundreds, but repeated ratios amplify the weight. With more decision times, moderate factors can continue to compound; dependence among histories may make particular sequences of small denominators more or less likely.

Diagnostics, Alternatives, and Limits

Weight diagnostics should examine the distribution of the cumulative weights, not just whether their mean appears plausible. Useful checks include the maximum and upper quantiles, a histogram or other tail-sensitive summary, the number of observations carrying a large share of the total weight, and the component factors at each time. In a longitudinal setting, inspecting factors by time can help distinguish one severely small denominator from repeated moderate ratios. These checks describe the estimated weights; they do not establish that the treatment models are correctly specified or that positivity holds.

When extreme weights appear, investigate the histories and treatment patterns that produced them. Check for sparse cells, coding errors, inappropriate treatment-model conditioning, and fitted probabilities near zero. Compare model fits that are scientifically defensible, paying attention to whether the same observations remain influential. A more stable numerical fit does not by itself solve limited overlap: the data may simply contain little information about the relevant treatment choices in some histories.

Unstabilized weights use a simpler numerator, often one, and can be more variable in ordinary finite samples. Stabilized weights may reduce variation by incorporating a numerator treatment mechanism, but the denominator still controls the possibility of a large ratio. A different denominator model can change estimated weights, yet is not automatically better: the model must reflect the intended conditioning history and be supported by data. Weight truncation is another response to extreme values, but it changes the weights and raises separate questions about bias and the resulting estimand; observing a heavy tail does not itself determine an appropriate cutoff.

There are also limits to what a weight summary can establish. A finite sample cannot verify that a population variance is finite, and a modest observed maximum does not rule out very small probabilities in histories rarely sampled. Conversely, one large estimated weight may reflect model instability rather than a truly extreme population weight. Weight diagnostics therefore need to be interpreted alongside treatment-model specification, overlap, sample size, and the scientific setting.

Common pitfall: “The numerator is stabilized, so the weights cannot be extreme.” This confuses the numerator with the ratio. For example, a numerator of $0.20$ divided by a denominator of $0.004$ gives a factor of $50$; multiplying factors across time can increase it further. The correction is to inspect the denominator probabilities, the observed treatment histories, and the cumulative product.

Check Your Understanding

Answer each question, then compare with the solutions.

If a numerator probability is $0.15$ and the denominator probability for the observed treatment is $0.005$, what is the factor?
In the model $p(x)=x^\alpha$, why can the treated-unit weight have mean one while its variance is infinite?
Two consecutive weight factors are $3$ and $6$. What is their cumulative product, and what does this illustrate?
Does a fitted denominator probability of $0.0002$ prove a structural positivity violation? Explain.
What is one reason to inspect time-specific factors as well as the cumulative weight?

Solutions. (1) The factor is $0.15/0.005=30$. (2) The conditional mean is one because the treated covariate distribution is proportional to $p(x)$, which cancels the inverse probability in the weight; for $\alpha\geq1$, the second moment condition $2<1+1/\alpha$ fails, so the variance is infinite. (3) The product is $3\times6=18$, illustrating how several factors combine in a longitudinal cumulative weight. (4) No. It is an estimated probability and indicates an extreme fitted ratio may result, but does not by itself establish the true probability or a structural support failure. (5) It can reveal whether an extreme cumulative value is driven by one near-zero denominator or by repeated moderate factors.