Why the Largest Weights Matter
A weighted analysis gives observations unequal contributions. If one observation has a weight far above the rest, it can carry a substantial fraction of the total weight, even when the sample itself is large. The maximum weight is a useful signal, but its raw numerical value does not by itself say how much of the weighted sample that observation represents. To assess concentration, compare each weight with the total.
This tutorial considers a fixed set of \(n\) observed, nonnegative weights \(w_1,\ldots,w_n\), with positive total weight. The weights may be unstabilized or stabilized; the basic concentration calculations are the same. They describe how total weight is distributed across observed individuals. They do not, by themselves, establish why a weight is large or what effect it has on a particular outcome estimate.
The normalized weights satisfy \(p_i\ge0\) and \(\sum_i p_i=1\). Thus \(p_{(1)}\) is the share carried by the single most heavily weighted observation, while \(C_k\) is the share carried collectively by the \(k\) largest weights. These shares are proportions, not additional weights used to refit the model.
What the Maximum and Top-k Shares Tell Us
A maximum weight of 20 can be modest if the total weight is 20,000, but substantial if the total is 40. The maximum share distinguishes these cases: they are \(20/20{,}000=0.001\) and \(20/40=0.5\). Similarly, the maximum alone can miss a cluster of moderately large weights. Reporting \(C_k\) for a few substantively chosen values of \(k\) shows whether concentration is due to one observation or several.
Worked Example: One Observation Carries Half the Weight
Suppose five observations have weights \(2,2,3,3,10\). Find the maximum share and the top-two share.
The total is \(S=2+2+3+3+10=20\). In decreasing order, the weights are \(10,3,3,2,2\). Therefore,
The largest-weight observation carries half of the total weight, and the two largest together carry 65%. This is direct evidence of concentration in the weighted sample. It does not tell us whether the weighted outcome estimate is substantially changed by those observations: that also depends on their outcome values and the estimator.
To prove the result, consider any subset of \(k\) weights. If it omits one of the \(k\) largest weights, it must include a weight no larger than the omitted one: there are only \(k-1\) other weights among the top \(k\). Replacing that included weight with the omitted weight cannot decrease the subset’s sum. Repeating such replacements yields the top-\(k\) subset, whose sum is therefore at least as large as the sum of every other \(k\)-observation subset. Dividing by the same positive total \(S\) proves the claim for shares. \(\square\)
Worked Example: Comparing Raw Maxima With Maximum Shares
Two analyses each have four weights. Analysis A has weights \(1,1,1,12\); analysis B has weights \(2,2,2,24\). Compare their raw maxima and their concentration.
For A, \(S_A=15\), so its maximum share is \(12/15=0.80\). For B, \(S_B=30\), so its maximum share is \(24/30=0.80\). The raw maximum in B is twice as large, but the normalized concentration is identical. In fact, every weight in B is twice its counterpart in A, so every normalized weight is unchanged.
This example shows why a raw maximum cannot be interpreted without its total. A common positive rescaling of all weights changes the numerical maximum but not the distribution of normalized weight across observations.
For the first claim, the rescaled total is \(aS\), and each rescaled normalized weight is \(aw_i/(aS)=w_i/S=p_i\). Sorting is unchanged because \(a>0\), so all \(C_k\) are unchanged.
For the second claim, any set of \(r\) observations has total normalized weight at most \(rm\), since each member’s share is no larger than \(m\). If those observations carry at least share \(q\), then \(rm\ge q\), which implies \(r\ge q/m\). Because \(r\) is an integer, \(r\ge\lceil q/m\rceil\). This is a necessary lower bound, not a guarantee that that many observations actually carry share \(q\). \(\square\)
Worked Example: How Many Observations Must Carry 80%?
In a sample, the largest normalized weight is \(m=0.25\). What is the smallest number of observations that could carry at least 80% of total weight?
If \(r\) observations carry at least 0.80, then \(r(0.25)\ge0.80\), so \(r\ge0.80/0.25=3.2\). Since \(r\) must be an integer, \(r\ge4\). Thus at least four observations are required. This bound does not establish that four actually suffice; for that, calculate \(C_4\) from the ordered weights.
For comparison, the general bound \(p_{(1)}\ge1/n\) follows because \(1=\sum_i p_i\le n p_{(1)}\). Hence a maximum share below \(1/n\) is impossible when the total weight is positive.
Calculating and Reporting Concentration
A transparent diagnostic begins by checking that the observed weights are finite and nonnegative and that \(S>0\). Then sort the weights, calculate \(p_{(1)}\), and calculate top-\(k\) shares for several values of \(k\). Report the sample size and the chosen values of \(k\), so a reader can tell how many observations are summarized. The particular \(k\) values should suit the sample and question; there is no universal cutoff at which a share proves that weights are unacceptable.
Confirm that weights are defined for the analyzed observations, are nonnegative, and have a positive finite total.
Divide each weight by \(S\), or sort the raw weights and divide their cumulative sums by \(S\).
Report the maximum share and selected top-\(k\) shares, with \(n\) and the \(k\) values stated.
Describe how much total weight is carried by the largest weights, then consider the estimator, outcomes, and model context before drawing further conclusions.
For example, one might report: “Among 500 observations, the largest weight represented 6% of total weight; the largest five represented 18%, and the largest ten represented 27%.” These statements quantify concentration without implying that any particular threshold is universally safe. A high share is a reason to investigate and to assess the implications for the analysis, not a standalone test of correctness.
Alternatives, Diagnostics, and Limitations
Tail quantiles, such as the 95th percentile, describe where a weight lies relative to the weight distribution. They can be useful alongside concentration summaries, but a quantile does not directly state what fraction of total weight the tail carries. The maximum share directly addresses the single largest contributor; \(C_k\) addresses the largest \(k\) contributors. Together they distinguish an isolated extreme weight from a broader concentration of large weights.
Several checks help prevent misleading conclusions:
- Do not compare raw maxima without considering scale. Multiplying every weight by the same positive constant changes the maximum but leaves normalized shares unchanged.
- Do not rely on the maximum alone. Several weights can be moderately large even if no single observation has an overwhelming share; inspect \(C_k\) as well.
- Do not treat a concentration summary as a model check with a pass/fail threshold. A high share reveals concentration, not its cause. It does not by itself prove misspecification, positivity failure, or bias.
- Do not equate weight share with outcome influence. An observation’s role in a particular estimate depends on the outcome values and the estimator, not only on its weight.
- Investigate invalid or unexpected weights. Negative, undefined, or nonfinite values are not covered by these definitions and may signal a data-processing or weight-construction problem. If all weights are zero, the normalized shares are undefined.
These summaries also have finite-sample limitations. They describe the observed weighted sample, not the population’s underlying weight distribution. A small maximum share does not prove that the weighting model is correct, and a large share does not show whether the analysis is biased. The earlier material on weight diagnostics and on tail quantiles provides related checks, but none of these summaries replaces assessment of the estimand, model assumptions, or outcome analysis.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Weights are \(1,2,2,5\). Calculate the total, maximum share, and top-two share.
- Why can two analyses with different raw maximum weights have exactly the same concentration?
- If the maximum normalized weight is 0.10, what is the minimum number of observations that could carry at least 60% of total weight?
- What does a top-three share of 0.70 establish, and what does it not establish about the outcome estimate?
- Why is a top-\(k\) share at least as large as the share carried by any other specified group of \(k\) observations?
Solutions. (1) The total is \(S=1+2+2+5=10\). The maximum share is \(5/10=0.50\); the top-two share is \((5+2)/10=0.70\). (2) If all weights in one analysis are a common positive multiple of the corresponding weights in the other, the multiplier cancels between each weight and the total, leaving all normalized shares unchanged. (3) If \(r\) observations carry at least 0.60, then \(0.10r\ge0.60\), so \(r\ge6\). At least six are needed, though six may not suffice in the observed sample. (4) It establishes that the three largest weights collectively carry 70% of total weight, and that no other three observations carry more. It does not establish the estimate’s bias, model validity, or outcome sensitivity. (5) Replacing any included weight outside the largest \(k\) with an omitted larger weight cannot reduce the group’s total; repeating this gives the top-\(k\) group.