Song removal and recorded use: ranks

Results

A logistic classification model relates song-use annual ranks to the removal decisions for the 2025 edition. The repertoire contains 557 songs: 79 removed and 478 kept. The model is fitted to the complete repertoire, and all results below describe that full-data fit.

At a removal-probability threshold of 0.2824, the model identifies 56 of the 79 removed songs (70.9%) and correctly classifies 447 of the 478 kept songs. It classifies 31 kept songs as removed and 23 removed songs as kept. Overall, 54 classifications differ from the recorded decisions.

Persistent low ranking has the largest standardized coefficient. A one-standard-deviation increase in the fraction of historical years spent in the bottom 79 multiplies fitted removal odds by 1.57, conditional on the other predictors. Worse recent and historical mean ranks are also associated with greater removal probability.

Long-term rank trend (+0.076) and recent change in rank (-0.026) carry smaller coefficients than the popularity and bottom-79 measures. Recent and historical ranks are correlated (0.94), so coefficients should be interpreted jointly.

Fit and classification

The threshold maximizes F1 across the 557 fitted outcomes, with the higher threshold breaking an exact tie. The reported precision, recall, F1 and classification counts use that threshold. The probability-ranking measures use the continuous fitted scores.

Classifying every song as kept would give 85.8% accuracy and identify none of the removals. A constant score has average precision 0.142. Among the 79 highest fitted removal probabilities, 51 songs were actually removed (64.6%). This fixed-size ranking calculation is separate from the F1 threshold; song keys break probability ties.

Contribution of the other four predictors

To measure what rank levels and trends contribute beyond persistent low popularity, compare the five-predictor model with logistic regression using only historical bottom-79 frequency. Both models are fitted to all 557 songs, standardize their inputs, and use the same L2 regularization parameter, C = 0.01. Each model’s classification threshold maximizes its own full-data F1.

Adding recent mean rank, historical mean rank, long-term rank trend and recent rank change raises average precision from 0.611 to 0.651: an increase of 0.040, or 6.5% relative to the one-feature model. ROC AUC rises from 0.891 to 0.921. Brier score falls from 0.089 to 0.080, a 10.3% reduction in mean squared probability error.

The improvement in classification is modest. The one-feature model identifies 60 of the 79 removals; the full model identifies 56 at its own F1 threshold. Kept songs flagged as removed number 47 for the one-feature model and 31 for the full model. The net reduction in classification errors is 12. At the song level, 16 classifications become correct and 4 become incorrect.

This comparison measures the four predictors’ combined contribution to the descriptive fit. It does not allocate the gain among the individual predictors. The precision–recall curve and probability-error measures capture improvements beyond the particular classification thresholds.

Comparison statistics, one-feature model parameters and per-song results; comparison figure in SVG format.

Data and predictors

Annual song-use statistics are drawn from singings using the 1991 Denson Edition as their primary book. The analysis uses 1995–2019 for historical measures and 2022–2024 for recent measures. Years 2020–2021 are omitted because of pandemic disruption. The 2025 figures and Overall columns are excluded because they include use after removal decisions could have been made or announced.

Larger rank numbers indicate lower popularity. Positive rank trends and recent changes indicate worsening relative standing. Source ranks are competition ranks: tied songs share the best position in their tied group.

The bottom-79 cutoff follows the 79 removals in this repertoire. It denotes the lowest 79 rank positions among 557 songs (14.18%), or rank 479 and worse. If a tied group crosses the boundary, fewer than 79 songs can qualify in that year. This feature does not constrain the model to classify exactly 79 songs as removed.

Model specification

The outcome is removal from the 2025 edition: removed = 1 and kept = 0. The model is logistic regression with an intercept and an L2 penalty, using regularization parameter C = 0.01. Each of the five predictors is standardized using its mean and standard deviation across the 557 songs. No class weighting is applied. Fitted probabilities are obtained by applying the logistic function to the intercept plus the weighted standardized predictors.

A positive coefficient corresponds to higher fitted removal odds; a negative coefficient corresponds to lower odds, holding the other predictors fixed. Correlation among predictors limits how separately their coefficients can be interpreted.

Where use and decisions diverge

The following songs have the lowest fitted removal probabilities among actual removals, or the highest fitted probabilities among retained songs. These cases indicate decisions that the available use statistics represent less closely.

Rank sparklines show annual popularity from 1995 (left) to 2024 (right). All use the same scale: rank 1 at the top, rank 557 at the bottom. The gap omits 2020–2021; the dot and number show the 2024 rank. Hover over the line to see individual annual ranks. These are observed ranks, not fitted probabilities.

Removed songs with the lowest fitted probabilities

Kept songs with the highest fitted probabilities

Song-level results

Download all 557 results and predictor values. The CSV preserves the source song-key order. The table starts with the highest fitted removal probabilities. Click a column heading to sort; click again to reverse the order. Song keys follow page order, with top (t) before bottom (b). The rank-trend column sorts by 2024 rank.

Rank sparklines show annual popularity from 1995 (left) to 2024 (right). All use the same scale: rank 1 at the top, rank 557 at the bottom. The gap omits 2020–2021; the dot and number show the 2024 rank. Hover over the line to see individual annual ranks. These are observed ranks, not fitted probabilities.