Missing values

When a procedure reads data, "missing" covers two things:

• system-missing: an empty cell (. / never entered);
• user-missing values: values declared on a variable with MISSING VALUES (-1, 99, 999, or a range such as 11 THRU 14).

Both are treated as missing by every procedure. This page states each rule, so one dataset cannot end up with contradictory Ns across procedures.

The five rules

Rule	Content
A	A single-variable procedure looks at that variable alone: its system-missing and user-missing values are both invalid.
B	Multi-variable procedures align variables case by case. Correlations use pairwise deletion, the PSPP / SPSS default (each pair drops its own missing cases); procedures that need complete cases (paired t-test, paired permutation, simple regression, K-means, reliability, factor analysis) use row-wise deletion, keeping only cases valid on every variable involved.
C	Weighted statistics require both the value and the weight to be valid; either one missing drops the case.
D	With a grouping variable, cases missing on the grouping variable or the analysis variable enter no group (a user-missing value of the grouping variable is not a group of its own); group labels use the displayed value, value labels first.
E	Categorical procedures (frequencies, crosstabs, FREQUENCIES, bar charts) count user-missing values as a missing category rather than a value; crosstab rows and columns exclude them.

Procedure by procedure

Procedure	Rule
Descriptives, frequencies, histogram, box plot, error bar, Q-Q plot	A
One-sample t-test, normality test, ROC	A
Independent-samples t-test, one-way / two-way ANOVA, ANCOVA, Kruskal-Wallis, ordered-alternative trend, two-sample permutation test	D (missing on the grouping or the analysis variable both drop the case) + A
Paired t-test, Wilcoxon signed rank, paired permutation	B (complete pairs)
Pearson / Spearman correlation, partial correlation	B (pairwise)
Simple / multiple / stepwise regression, logistic regression, factor analysis, K-means, reliability	B (row-wise)
Weighted descriptives and other weighted procedures	C
Frequencies, crosstabs, CTABLES, bar charts	E
Multiple response sets (MRSETS /FREQUENCIES, /CROSSTABS)	A + E: a variable is tested for "selected" on its own, counts are unweighted, and "cases with a response" counts cases with at least one valid response

The raw cell values (user-missing ones included) stay in the dataset: the data editor, exports and "find cases" show them, and only procedures apply the rules above. Dataset.rawColumn returns raw values, Dataset.numericColumn returns the analysis values with user-missing values blanked out.

Difference from PSPP

PSPP also excludes user-missing values by default but offers a per-procedure MISSING=INCLUDE switch that treats them as ordinary values. This app always excludes them and does not offer the switch: one consistent rule is harder to misread. To really use 99 as a value, RECODE it first (or drop the MISSING VALUES declaration).

Everything else follows PSPP 2.1.2, and test/pspp_parity_test.dart checks it against expectations shipped with PSPP itself (AUTORECODE, RANK, ONEWAY, pairwise CORRELATIONS, NPAR TESTS Mann-Whitney) on every CI run.

Defect and conventions fixed in v1.4.0

The review and the parity work turned up one alignment defect and two convention mismatches — all of the "the numbers do not match the textbook / PSPP" kind:

1. Case misalignment (defect): some multi-variable procedures filtered each column separately and then truncated by position, which misaligned the variables (a case valid on pre but missing on post shifted the two columns by one). All procedures now read case-aligned data per the table above, with regression tests.
2. Levene's test now centres on the group mean (PSPP / SPSS convention). It previously used the group median (Brown-Forsythe), which does not reproduce the Levene table of PSPP or SPSS; SPSS's ONEWAY / T-TEST / GLM all use the mean.
3. Mann-Whitney U now includes the tie correction in its asymptotic variance, σ² = n1·n2·[(n+1)/12 − Σ(t³−t)/(12·n·(n−1))], with p = 2·Φ(z) (PSPP's "Asymp. Sig. (2-tailed)"). The continuity-corrected Z / p (SPSS, R correct = TRUE) remain available as extra['zContinuityCorrected'] / extra['pContinuityCorrected'].

See the statkit API for the implementations and Analysis for the procedures.