Extreme Weights in Biologic Treatment of Arthritis

Why a Few Arthritis Treatment Decisions Can Matter

In a rheumatoid arthritis cohort, biologic treatment is not assigned at random. Disease activity helps shape decisions: clinicians may be more likely to initiate or continue a biologic for patients with high activity, while patients with low activity may be less likely to start one. In a marginal structural analysis, inverse-probability weights account for these observed treatment patterns. But if a treatment decision is rare for a particular disease-activity history, the corresponding inverse probability can be large. A small number of such decisions can then have substantial influence on a marginal estimate.

The diagnostic task is not simply to find the largest number in a weight column. It is to trace that number back through the treatment history: which decision contributed it, what disease activity and prior treatment preceded that decision, and whether the recorded history and fitted treatment probabilities are credible.

Longitudinal Weights and Rare Decisions

Definition: Let \(A_t\) denote biologic treatment at decision time \(t\), with \(A_t=1\) for biologic treatment and \(A_t=0\) otherwise. Let \(L_t\) denote measured disease activity and other covariates available before that decision. Write \(H_t=(\bar A_{t-1},\bar L_t)\) for the observed treatment and covariate history used to model \(A_t\). The treatment probability is \(g_t(A_t\mid H_t)=P(A_t\mid H_t)\). For a patient observed through time \(K\), the unstabilized treatment weight is \(W=\prod_{t=0}^{K}1/g_t(A_t\mid H_t)\). A decision with a small probability for the treatment actually received contributes a large factor to \(W\).

This weight targets the treatment history actually observed for each patient: a patient who received treatment with probability 0.05 contributes a factor of \(1/0.05=20\) at that decision. In a stabilized weight, each denominator factor is multiplied by a numerator probability that conditions on a less detailed treatment history, commonly prior treatment:

SW=\prod_{t=0}^{K} \frac{P(A_t\mid \bar A_{t-1})} {P(A_t\mid \bar A_{t-1},\bar L_t)}.

The numerator can reduce overall weight scale, but it does not make a rare decision in the denominator disappear. For a fixed treatment regime, numerator factors are constant across patients following that regime, so they cancel from its normalized weighted mean.

The causal interpretation of these weights requires assumptions. Consistency means that a patient's observed outcome equals the potential outcome under the treatment history that patient received. Sequential exchangeability means that, at each decision, conditional on the measured history \(H_t\), treatment is independent of the potential outcomes under the treatment regimes being compared. Positivity means that each treatment choice being compared has positive probability for histories in the target population. The treatment-probability model must also represent the relevant conditional probabilities adequately. A large estimated weight can signal sparse support or model problems; it does not by itself identify which assumption has failed.

Worked Example: Tracing a Weight Through Two Decisions

Suppose that at baseline, 70% of patients with high disease activity and 10% with low activity receive a biologic. At the next decision, among patients previously treated, only 4% of those with low current activity receive biologic treatment. For illustration, suppose the marginal baseline probability of treatment is 0.40, and among previously treated patients the numerator probability of treatment at the next decision is 0.60.

Consider a patient with high baseline activity who receives treatment, then has low activity and receives treatment again. The denominator probabilities for the two observed decisions are 0.70 and 0.04. Thus the unstabilized weight is:

W=\frac{1}{0.70}\times\frac{1}{0.04} =1.4286\times25 \approx35.71.

The second decision is the main source of the large weight: biologic treatment was unusual for a patient with low activity at that point. The stabilized weight is:

SW=\frac{0.40}{0.70}\times\frac{0.60}{0.04} \approx0.5714\times15 \approx8.57.

Stabilization lowers the scale in this example, but the low-activity treatment decision still drives the large contribution. The history also illustrates why looking only at a patient's baseline characteristics can miss the source of a longitudinal extreme weight.

What a Large Weight Does to a Marginal Estimate

A weighted mean gives each observation influence in proportion to its weight. The following identity makes the consequence of an unusually large weight explicit. Suppose the current observations have total weight \(S_w\) and weighted mean \(\bar Y_w\). Add one observation with outcome \(y\) and weight \(w\). The new mean is:

\bar Y_{w,\mathrm{new}} =\frac{S_w\bar Y_w+wy}{S_w+w} =\bar Y_w+\frac{w}{S_w+w}(y-\bar Y_w).

The change is therefore the distance from the old mean, multiplied by the new observation's share of total weight. If \(y\) differs substantially from the previous mean and \(w\) is large relative to \(S_w\), the estimate can move substantially. This is an exact finite-sample identity, not an approximation.

Worked Example: A Rare Treated Patient Shifts the Weighted Mean

Imagine 40 patients: 20 with high activity and 20 with low activity. Of the high-activity patients, 16 receive a biologic; of the low-activity patients, one receives it. The observed treatment probabilities are therefore \(16/20=0.80\) for high activity and \(1/20=0.05\) for low activity. The treated patients have outcome 6 on average in the high-activity group, while the one treated low-activity patient has outcome 10.

Among treated patients, the unstabilized weights are \(1/0.80=1.25\) for each high-activity patient and \(1/0.05=20\) for the low-activity patient. The weighted treated mean is:

\frac{16(1.25)(6)+1(20)(10)} {16(1.25)+1(20)} =\frac{120+200}{20+20} =8.

The unweighted mean among these 17 treated patients is:

\frac{16(6)+10}{17}=\frac{106}{17}\approx6.24.

The low-activity patient represents 20 units of total weight, equal to the combined 20 units for all 16 high-activity treated patients. Consequently, that patient's outcome has substantial influence on the weighted result. Here the weighted mean also reflects equal total weight for the two activity groups, as intended by the inverse probabilities in this constructed example.

The Kish effective sample size summarizes weight concentration as \((\sum_i w_i)^2/\sum_i w_i^2\). For these treated patients, it is:

\frac{40^2}{16(1.25^2)+20^2} =\frac{1600}{425}\approx3.76.

There are 17 treated people, but the concentration of weights makes this summary resemble fewer than four equally weighted observations. Effective sample size is a diagnostic summary, not a replacement sample size for inference.

A Weighted-Mean Identity and Its Conditions

Theorem: For a single treatment decision, let \(L\) be the measured pre-treatment history, \(A\in\{0,1\}\), and \(Y^a\) the potential outcome under treatment \(a\). Suppose consistency holds, \(Y^a\) is conditionally independent of \(A\) given \(L\), and \(0<P(A=a\mid L)<1\) almost surely in the target population. If the relevant expectations exist, then \(E[I(A=a)Y/P(A=a\mid L)]=E[Y^a]\) and \(E[I(A=a)/P(A=a\mid L)]=1\).

The first identity explains the target of a normalized inverse-probability weighted mean; the second explains why the denominator weight totals the target population in expectation. Here \(I(A=a)\) equals 1 when treatment \(a\) was received and 0 otherwise.

To prove the first identity, condition on \(L\). By iterated expectation, consistency on the event \(A=a\), and conditional exchangeability:

\begin{aligned} E\left[\left.\frac{I(A=a)Y}{P(A=a\mid L)}\right|L\right] &=\frac{P(A=a\mid L)E[Y\mid A=a,L]}{P(A=a\mid L)}\\ &=E[Y^a\mid A=a,L]\\ &=E[Y^a\mid L]. \end{aligned}

Taking expectations over \(L\) gives \(E[Y^a]\). For the second identity, condition on \(L\) again:

E\left[\left.\frac{I(A=a)}{P(A=a\mid L)}\right|L\right] =\frac{P(A=a\mid L)}{P(A=a\mid L)}=1.

Taking expectations gives 1. In longitudinal data, the same cancellation is applied successively at each decision under sequential exchangeability and positivity. When probabilities are very small, the cancellation still holds algebraically, but finite-sample estimates can be highly variable because a few observed decisions carry large weights. \(\square\)

Worked Example: Finding the Activity Pattern in a Treatment Model

Suppose a fitted model for biologic use at one decision is:

\operatorname{logit}P(A=1\mid L)=-3+3.5I(L=\text{high}).

For low activity, the fitted log-odds are \(-3\), so the fitted probability is:

P(A=1\mid L=\text{low})=\frac{e^{-3}}{1+e^{-3}}\approx0.0474.

For high activity, the fitted log-odds are \(0.5\), giving:

P(A=1\mid L=\text{high})=\frac{e^{0.5}}{1+e^{0.5}}\approx0.6225.

A treated patient with low activity therefore receives an unstabilized factor of approximately \(1/0.0474=21.10\). An untreated patient with high activity receives a factor of approximately \(1/(1-0.6225)=2.65\). The model points to a particular pattern: low activity followed by biologic use is much rarer than high activity followed by biologic use. The reverse treatment decision, withholding biologic treatment at high activity, is also less common than withholding it at low activity.

These calculations are a starting point for review, not proof that the model is correct or that the causal comparison is supported. Check whether the activity measure was recorded before treatment, whether treatment coding is consistent, and whether the model's probabilities agree with the observed histories.

Diagnostics, Alternatives, and Limitations

A useful diagnostic follows each extreme weight back to the decision-level probabilities that formed it. Review the largest weights alongside treatment history, current disease activity, prior biologic use, and the number of patients with similar histories. Compare observed treatment frequencies with fitted probabilities within clinically meaningful activity and treatment-history groups. Plotting weights or their logarithms, and reporting upper percentiles as well as the maximum, can reveal whether a single record or a broader group is responsible.

Also inspect the effective sample size and the weighted distribution of disease activity. If a few patients dominate, a marginal estimate may be sensitive to their outcomes. If predictions are unexpectedly near zero or one, check the treatment model, variable coding, timing of measurements, and sparse combinations of history variables. A high maximum weight alone does not tell you whether the source is a genuine rare decision, a data error, or model extrapolation.

Stabilized weights are a common alternative to unstabilized weights and can reduce scale, but stabilization does not guarantee low variability or fix a near-zero denominator probability. An outcome-regression approach models the outcome conditional on covariates; the parametric g-formula models covariate and outcome distributions under specified treatment histories. These approaches have different modeling requirements. Comparing estimates can be informative, but agreement does not establish that the treatment model or causal assumptions are correct.

Most importantly, do not treat a large weight as an automatic instruction to alter the analysis. Any decision to modify weights changes the analysis and can trade variance for bias. The later material on truncation addresses those choices directly. Here the priority is to identify which treatment decisions generate the extremes and assess how much those observations influence the marginal estimate.

Common Mistake

Wrong reasoning: “The stabilized weights are smaller than the unstabilized weights, so the rare-treatment problem has been solved.”
Correction: Stabilization changes the numerator as well as the scale. Inspect the denominator probability at each decision: in the two-decision example, the 0.04 probability remains the source of the large factor. Also check the weight distribution, effective sample size, and influential patient histories.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A patient receives biologic treatment with estimated probability 0.08 at one decision. What is the unstabilized weight factor for that decision?
In the worked two-decision example, which decision contributes most to the unstabilized weight, and why?
Why can one patient's outcome substantially change a weighted marginal mean?
Does a large estimated weight by itself prove that positivity fails? Give two other possible explanations for an extreme weight.

Solutions. (1) The factor is \(1/0.08=12.5\). (2) The second decision contributes the factor \(1/0.04=25\), which is much larger than the baseline factor \(1/0.70\approx1.43\); it corresponds to biologic use despite low activity. (3) The exact change after adding an observation is \(w/(S_w+w)\) times the difference between its outcome and the previous weighted mean. A large weight gives that observation a large share of total weight. (4) No. An extreme weight can reflect a genuinely rare treatment decision, an incorrectly coded or mistimed variable, or an inadequately specified treatment-probability model, among other possibilities. The histories and fitted probabilities must be checked.