NHANES Mortality Covariate Analysis

Cohort framing

This run analyzes 11,171 rows and 80 covariates after excluding SEQN. The target mortstat is imbalanced but not extreme: 8,724 rows (78.1%) are 0 and 2,447 rows (21.9%) are 1. Overall data quality is reasonable at the table level, with 26 duplicate rows (0.23%) and an overall missing-cell rate of 9.2%, but that average hides a sharp difference between the lower-level and higher-level covariate bundles.

This file should also be read as a labeled subset rather than the full covariate source table. The export metadata for this dataset indicates that the wider feature source had 20,470 rows, while only 11,171 rows had valid mortality labels. That matters because the associations below describe the mortality-observed cohort, not everyone in the upstream NHANES feature export.

RIDAGEYR is the strongest single mortality signal in the run with eta-squared 0.381. It is not part of the l1 / l2 / l3 workbook bundles; it is the separately added age feature for the mortality-ready export. In practice it behaves like the anchor covariate against which the staged bundles should be judged.

Level hierarchy

l1: best baseline bundle

l1 is the cleanest starting point for staged modeling. Its six features have lower average missingness than l3 (7.0% vs 10.4%), and the strongest l1 mortality signal is BPXSY at eta-squared 0.124. That is materially stronger than the other core anthropometric features in this run and makes systolic blood pressure the clearest l1 candidate to carry incremental mortality information beyond age.

The main l1 structural issue is redundancy, not sparsity. BMXBMI and BMXWAIST are strongly correlated at 0.885, so they are likely to compete for similar body-composition signal in an interpretable model. Their distributions also show right tails (BMXBMI skew 1.50, max 130.21; BMXWAIST median 96.0 cm, max 175.0 cm), so robust scaling or outlier checks would matter even in the baseline bundle. BPXSY itself is right-skewed as well (skew 1.26, max 270), which is another reason to inspect its extreme tail before using coefficient-based models.

If the goal is a first-pass covariate model, l1 looks like the right place to start because it is small, comparatively clean, and clinically interpretable. The main caution is to avoid overcounting body-size information by treating BMXBMI and BMXWAIST as independent wins.

l2: strongest second-stage expansion

l2 is the best next layer after l1 because it is complete in this table: all five questionnaire / disease-history variables have 0% missingness. The activity-history block is the clearest l2 contributor to mortality separation: PAD200 scores 0.101, PAD320 0.0415, and PAD020 0.0328. DIQ010 also clears the report's weak-signal floor at 0.0217. At the threshold of eta-squared >= 0.02, l2 contributes four signals, which is much broader coverage than l1.

The caveat is that the workflow inferred these questionnaire variables as numeric because they are stored as integer-coded values in the CSV. That is visible in the profile overview, where 79 of 80 analyzed variables were treated as numeric. The effect sizes are still useful as a screening signal, but DIQ010, PAD020, PAD200, and PAD320 should be recoded as categorical before any final model comparison or coefficient interpretation. In other words, l2 looks genuinely useful, but the current ranking likely understates its category structure.

Given your priority order, l2 is the most defensible second-stage addition: it adds multiple mortality-relevant signals without adding the missingness and multicollinearity burden that comes with l3.

l3: highest upside, highest risk

l3 dominates the table numerically and statistically, but it is the hardest bundle to trust naively. It contains 67 laboratory, CBC, and dietary features, contributes 14 of the features with eta-squared >= 0.02, and includes the strongest non-age signal in the run: LBDSBUSI at 0.134. Other notable l3 mortality signals are LBDSCRSI (0.0566), LBXGH (0.0485), LBXRDW (0.0470), LBXSKSI (0.0452), and LBDSGLSI / DRXTKCAL (both 0.0375).

The problem is that l3 also carries almost all of the data-quality and redundancy burden. Its average missing rate is 10.4%, compared with 7.0% for l1 and 0% for l2. The worst missing columns are all l3: LBXSASSI (11.8%), LBXSATSI (11.8%), LBXSC3SI (11.8%), LBDSTPSI (11.2%), and LBDSGBSI (11.2%). The co-missing structure is nearly deterministic for several serum chemistry pairs, including LBXSASSI with LBXSATSI (joint missing rate 11.77%, Jaccard 0.999), LBXSATSI with LBXSC3SI (11.74%, 0.997), and LBDSTPSI with LBDSGBSI (11.19%, 1.000). That pattern looks much more like module-level collection availability than independent feature dropout.

l3 is also where almost all of the strongest linear correlations live. Among pairs with absolute correlation >= 0.75, 14 of 15 are l3-to-l3; the only non-l3 pair is BMXBMI vs BMXWAIST. The densest examples are DRXTMFAT vs DRXTTFAT (+0.978), LBXHCT vs LBXHGB (+0.974), LBXLYPCT vs LBXNEPCT (-0.938), DRXTSFAT vs DRXTTFAT (+0.937), and DRXTCARB vs DRXTKCAL (+0.890). This is exactly the kind of multicollinearity that can make a lab-heavy model look unstable unless the bundle is regularized or grouped before interpretation.

Some l3 distributions are also extremely heavy-tailed. LBDSCRSI has median 79.56, max 1573.52, and skew 16.0; LBDSBUSI has median 4.28, max 34.99, and skew 2.76. So even before feature selection, l3 is signaling a need for robust transforms, careful winsorization rules, or at least sensitivity checks on outlier handling.

For your stated preference order, the right interpretation is not that l3 is weak. It is that l3 is powerful but expensive: it offers the most incremental signal candidates, but it should be treated as a third-stage expansion only after the cleaner l1 and l2 bundles are established.

Recommended reading of the staged bundles

The evidence in this run supports a staged strategy:

• Start with age + l1 as the clean baseline, with special attention to BPXSY and to BMXBMI / BMXWAIST redundancy.
• Add l2 next, especially the physical-activity variables, because they are complete in this table and show consistent target separation.
• Add l3 last, and treat it as a structured module rather than a flat list of independent features because its missingness and correlation patterns are bundle-level, not column-level accidents.

If you want the next pass to stay explainable, the main pre-modeling checks should be:

• recode questionnaire fields such as DIQ010, PAD020, PAD200, and PAD320 as categorical rather than leaving them as numeric codes;
• compare l1 models that include either BMXBMI or BMXWAIST before using both together;
• run l3 sensitivity analyses with and without the heaviest-missing serum chemistry block, since several of those columns disappear together;
• prefer grouped regularization or feature-cluster selection inside l3 instead of interpreting raw coefficients from a fully expanded lab-and-diet model.