Tail Quantiles and Extreme Weight Diagnostics

Why Look Beyond the Center of the Weight Distribution?

A weight distribution can look unremarkable near its center while still having a consequential upper tail. A few observations may carry weights far above the rest, or a sizeable fraction of observations may have moderately large weights. These patterns are different: a single upper quantile or a single maximum cannot describe both well. Upper quantiles and tail plots provide complementary summaries of how many weights are large and how large they are.

Here, \(W_i\) denotes the weight for analysis unit \(i\), for \(i=1,\ldots,n\). In a longitudinal analysis, the unit is typically a person and \(W_i\) may be that person's weight at a specified follow-up time. Interpret diagnostics at a clearly specified time and for a clearly specified weight definition. The summaries describe the observed weights; they do not, by themselves, identify why the tail looks as it does.

The strict inequality matters: a weight equal to \(t\) is not counted as an exceedance. Quantile conventions also matter in finite samples; software may interpolate between order statistics rather than use nearest rank. State the convention when reporting quantiles, and apply it consistently when comparing samples.

What Quantiles and Tail Plots Show

An upper quantile answers a magnitude question: how large is a weight near a specified upper rank? For example, \(\hat q_{0.99}\) is the observation at rank \(\lceil 0.99n\rceil\). A tail plot answers a prevalence question: for each threshold \(t\), what fraction of observations have weights strictly greater than \(t\)? A useful plot places \(t\) on the horizontal axis and \(\hat S(t)\) on the vertical axis. It is a nonincreasing step function, dropping when a threshold reaches an observed weight.

Under a continuous weight distribution with strictly increasing distribution function near its \(p\)-quantile \(q_p\), the population quantile is defined by \(F(q_p)=p\), and the fraction above it is \(1-p\). This identity can fail at atoms: if many weights equal \(q_p\), then \(P(W>q_p)\) can be smaller than \(1-p\). The empirical analogue also depends on ties. Thus a quantile and a plotted strict exceedance proportion should be read together, not assumed to match exactly.

How Sample Size Changes Tail Summaries

At a fixed population exceedance probability, larger samples tend to contain more tail observations. They also make the observed tail proportion less variable. This matters especially for very high quantiles: their ranks may correspond to only a handful of units in a modest sample.

To derive these statements, set \(I_i=I(W_i>t)\). Each \(I_i\) is Bernoulli with success probability \(p_t\), and independence of the weights implies independence of the indicators. Thus \(C(t)\) is the sum of \(n\) independent Bernoulli variables, so it is binomial. Since \(\hat S(t)=C(t)/n\), linearity of expectation gives \(E[\hat S(t)]=np_t/n=p_t\). Independence makes the covariances between distinct indicators zero, so

Finally, \(C(t)=0\) exactly when every indicator is zero. Independence gives \(P(C(t)=0)=(1-p_t)^n\). This proves the result. \(\square\)

This calculation also explains why a high sample quantile can be unstable in a small sample. For instance, the nearest-rank 99th percentile in a sample of 100 is the 99th ordered observation; in a sample of 1,000 it is the 990th. Those ranks represent very different numbers of observations in the upper tail. Reporting the percentile without \(n\), the quantile convention, and the corresponding upper-tail count can make the summary hard to assess.

By definition, \(F(q_p)=P(W\le q_p)=p\). The events \(W\le q_p\) and \(W>q_p\) are disjoint and together cover all possible values of \(W\). Therefore \(P(W>q_p)=1-P(W\le q_p)=1-p\). Continuity rules out a positive probability mass at the quantile; strict increase ensures the specified quantile is not ambiguous over an interval. With a discrete distribution, the complement identity still holds, but \(P(W\le q_p)\) need not equal \(p\), so the final value need not be \(1-p\). \(\square\)

Practical Reading, Comparisons, and Failure Modes

A useful summary pairs several upper quantiles with exceedance proportions at thresholds chosen for the analysis. Quantiles make it easy to compare magnitudes at selected ranks. A tail plot shows how the proportion changes across a range of thresholds and makes plateaus or abrupt drops visible. A histogram can give a broader view of the distribution, but its appearance depends on bin choices and may conceal a small number of very large observations. These summaries are complementary; none explains the cause of a tail pattern.

Use comparable definitions when comparing time points or analyses: the same type of weight, unit of analysis, quantile convention, and thresholds. If a tail plot is based on observed weights, its empirical steps are exact for those observations. Claims about the underlying population tail require additional assumptions, including an appropriate sampling model. The independent, identically distributed calculation above is a useful reference, but repeated records from the same person are not independent units. If weights are dependent across units or arise from a complex sampling design, the stated variance calculation may not apply as written.

• Do not infer a broad tail from one high quantile alone. Check nearby upper quantiles and how many observations exceed relevant thresholds.
• Do not treat a zero observed tail proportion as proof of no tail risk. In a finite sample, rare exceedances may simply be absent; the sample-size calculation shows why.
• Respect ties and strict thresholds. If many weights equal a threshold, they are not counted by \(\hat S(t)\). At the largest observed value, the strict exceedance proportion is zero.
• Do not interpret a tail shape as a diagnosis of cause. The pattern may prompt checks of weight construction, data, and histories, but does not establish model misspecification or a positivity violation.
• Do not compare summaries with different sample sizes or conventions without noting the difference. A percentile can correspond to very few observations, and interpolation rules can change its reported value.

These diagnostics have limits. They summarize the marginal distribution of the weights, not their relationship to outcomes or to specific covariate histories. Similar tail plots can arise for different reasons, and a benign-looking plot cannot verify the assumptions of a weighted analysis. Treat an unusual tail as a signal for further examination rather than a pass-or-fail test.