OLS random quantile sampling draws only ten residual values, shared across calls and variables #208

OLS._predict_quantile generates a fixed number of residual draws — ten — regardless of how many rows are being predicted, and rebuilds its generator from self.seed on every call. The imputed conditional distribution is therefore a ten-point discrete mixture, and repeated calls and separate target variables all receive the same ten values in the same row order.

Lower priority than #207 because this path is opt-in: random_quantile_sample defaults to False (ols.py:304), autoimpute() never reaches it (comparisons/autoimpute_helpers.py:373-377 always passes an explicit quantiles=[quantile]), and no caller inside the package sets it True. It is still wrong for anyone who does opt in.

Mechanism

# microimpute/models/ols.py:390
count_samples: int = 10
# microimpute/models/ols.py:426
rng = np.random.default_rng(self.seed)

Ten beta draws are mapped through norm.ppf and then assigned across all n test rows, so the realised moments of the imputed distribution are whatever those ten draws happened to be — not the fitted residual distribution.

Reproduction

import numpy as np, pandas as pd, logging
from microimpute.models.ols import OLSResults

o = OLSResults.__new__(OLSResults); o.seed = 42; o.logger = logging.getLogger("x")
v = o._predict_quantile(pd.Series(np.zeros(500)), 1.0, 0.5, True)
print(len(np.unique(v.values)))                                               # 10, not 500
print(np.allclose(v, o._predict_quantile(pd.Series(np.zeros(500)), 1.0, 0.5, True)))  # True

Ten distinct multipliers for 500 rows, and the second call is identical to the first.

Because the per-row standard error is proportional between two models fitted on the same X, the ratio of the two variables' shocks has exactly one distinct value across the test set — the two imputed variables are perfectly comonotonic. On data with a true residual correlation of −0.041, the imputed shocks correlate at 0.9999999999999999.

Suggested fix

Draw one variate per row (size=len(mean_preds)), and hold a single np.random.Generator on the results object, advanced across calls and variables, rather than reseeding from self.seed inside the method. qrf.py:273 already draws per row; OLS did not follow. The cross-variable half of the problem is the same root cause as #207.

Origin

Long-standing rather than a regression: eb5e88c, "quantile sampling for default imputation (quantreg and ols)", 7 May 2025. No comment suggests the count of ten is deliberate.

Found during a pre-JOSS-submission audit (#201), reproduced independently twice.