SPB Git

Repository Public

AnovaTools.swift 13.7 KB 196 lines swift

1import Foundation2import SheetCore3import Statistics4import AnalysisKit56enum AnovaSupport {7    /// Localized source names of the engine's ANOVA rows.8    static func source(_ s: String, rowName: String? = nil, colName: String? = nil) -> String {9        switch s {10        case "Between Groups": L("Between groups")11        case "Within Groups": L("Within groups")12        case "Rows": rowName ?? L("Rows")13        case "Columns": colName ?? L("Columns")14        case "Interaction": L("Interaction")15        case "Error": L("Error")16        case "Within": L("Within cells")17        case "Total": L("Total")18        default: s19        }20    }2122    static func table(_ r: AnovaResult, rowName: String? = nil, colName: String? = nil) -> ReportTable {23        var rows: [[Value]] = [[.text(L("Source of variation")), .text(L("Sum of squares")), .text(L("df")), .text(L("Mean square")), .text("F"),24                                .text(L("p-value")), .text(L("η²")), .text(L("Partial η²"))]]25        for row in r.rows {26            rows.append([.text(source(row.source, rowName: rowName, colName: colName)), .finite(row.sumOfSquares), .finite(row.df), .finite(row.meanSquare),27                         .finite(row.f), .finite(row.pValue), .finite(row.etaSquared), .finite(row.partialEtaSquared)])28        }29        return ReportTable(title: L("Analysis of variance"), rows: rows,30                           columnFormats: [1: .fixed(4), 2: .integer, 3: .fixed(4), 4: .fixed(4), 5: .pValue, 6: .fixed(4), 7: .fixed(4)],31                           note: L("p-values are upper-tail probabilities of the F distribution."))32    }3334    static func groupTable(_ groups: [(label: String, count: Int, mean: Double, variance: Double)], title: String) -> ReportTable {35        var rows: [[Value]] = [[.text(L("Level")), .text(L("n")), .text(L("Mean")), .text(L("Variance")), .text(L("Standard deviation"))]]36        for g in groups { rows.append([.text(g.label), .number(Double(g.count)), .finite(g.mean), .finite(g.variance), .finite(g.variance.squareRoot())]) }37        return ReportTable(title: title, rows: rows, columnFormats: [1: .integer])38    }3940    static func effectBlocks(_ row: AnovaRow, name: String, dfError: Double, alpha: Double, null: String, locale: Locale) -> [ResultBlock] {41        guard let p = row.pValue, let f = row.f else { return [] }42        return [.interpretation(L("%@: F(%@, %@) = %@. ", name, Fmt.df(row.df, locale: locale), Fmt.df(dfError, locale: locale), Fmt.fixed(f, 4, locale: locale))43                                + Interpret.decision(p: p, alpha: alpha, null: null, locale: locale))]44    }45}4647/// One-way ANOVA with Tukey HSD.48enum OneWayAnovaTool {49    static let descriptor: ToolDescriptor = {50        var t = ToolDescriptor(51            id: "compare.anova-oneway", title: L("One-Way ANOVA"), group: Common.groupCompare, systemImage: "chart.bar.doc.horizontal",52            summary: L("Compares the means of several independent groups, with Tukey HSD pairwise comparisons."),53            fields: [.range("data", L("Data range"))] + GroupFields.fields() + [54                .toggle("tukey", L("Tukey HSD pairwise comparisons"), true),55                .alpha(L("Significance level α")),56            ],57            info: [58                L("F = mean square between groups / mean square within groups with (k − 1, N − k) degrees of freedom; η² = SS between / SS total."),59                L("Tukey HSD uses the Tukey–Kramer adjustment for unequal group sizes; adjusted p-values and simultaneous intervals come from the studentized range distribution with the family-wise level 1 − α."),60                L("Blank cells, text and error values are treated as missing."),61            ],62            run: run)63        t.demo = { _, v in v.set("data", range: "Groups!A1:C31"); v.set("groupCols", columns: [0, 1, 2]) }64        return t65    }()6667    static func run(_ input: ToolInput) throws -> ToolResult {68        let alpha = try Common.alpha(input.values)69        let frame = try input.frame()70        let (groups, missing) = try GroupFields.read(input)71        let labels = groups.map(\.label)72        let r = try Inference.oneWayANOVA(groups.map(\.values), labels: labels)73        guard let between = r.row("Between Groups"), let within = r.row("Within Groups"), let p = between.pValue, let f = between.f else {74            throw ToolError(L("The analysis of variance could not be computed."))75        }76        var blocks: [ResultBlock] = [77            .metrics([Metric("F", f, .fixed(4)), Metric(L("Degrees of freedom"), text: Fmt.df(between.df, locale: input.locale) + ", " + Fmt.df(within.df, locale: input.locale)),78                      Metric(L("p-value"), p, .pValue, detail: L("(upper tail)")), Metric(L("Effect size (η²)"), between.etaSquared ?? .nan, .fixed(4))]),79            .table(AnovaSupport.table(r)),80            .interpretation(Interpret.decision(p: p, alpha: alpha, null: L("all group means are equal"), locale: input.locale)),81            .table(AnovaSupport.groupTable(r.groups.map { ($0.label, $0.count, $0.mean, $0.variance) }, title: L("Groups"))),82            .chart(Common.boxPlot(groups, title: L("Distribution by group"), yTitle: L("Value"), locale: input.locale)),83        ]84        if input.values.bool("tukey"), groups.count >= 2 {85            let t = try Inference.tukeyHSD(groups.map(\.values), labels: labels, level: 1 - alpha)86            var rows: [[Value]] = [[.text(L("Comparison")), .text(L("Mean difference")), .text("q"), .text(L("Adjusted p-value")),87                                    .text(L("Lower bound")), .text(L("Upper bound")), .text(L("Decision"))]]88            for c in t.comparisons {89                rows.append([.text(c.groupA + " − " + c.groupB), .finite(c.difference), .finite(c.q), .finite(c.pValue), .finite(c.lower), .finite(c.upper),90                             .text(c.pValue < alpha ? L("reject equality") : L("fail to reject equality"))])91            }92            blocks.append(.table(ReportTable(title: L("Tukey HSD pairwise comparisons"), rows: rows,93                                             columnFormats: [1: .fixed(4), 2: .fixed(4), 3: .pValue, 4: .fixed(4), 5: .fixed(4)],94                                             note: L("Simultaneous %@ confidence intervals; critical q = %@ with %d groups and %@ error degrees of freedom.",95                                                     Fmt.percent(t.level, 1, locale: input.locale), Fmt.fixed(t.criticalQ, 4, locale: input.locale), t.groupCount,96                                                     Fmt.df(t.dfError, locale: input.locale)))))97        }98        blocks.append(.assumptions(r.assumptions, title: L("Assumptions")))99        if let note = GroupFields.missingNote(missing) { blocks.append(.note([note])) }100        return ToolResult(title: L("One-Way ANOVA"), subtitle: L("%@ — %d groups", frame.label, groups.count), blocks: blocks, sheetName: L("ANOVA"))101    }102}103104/// Two-way ANOVA without replication (table) or with replication (balanced, long format).105enum TwoWayAnovaTool {106    static let descriptor: ToolDescriptor = {107        var t = ToolDescriptor(108            id: "compare.anova-twoway", title: L("Two-Way ANOVA"), group: Common.groupCompare, systemImage: "square.grid.2x2",109            summary: L("Tests the effects of two factors, without replication (one value per cell) or with balanced replication."),110            fields: [111                .choice("design", L("Design"), [FieldOption("table", L("Without replication (table of values)")),112                                                FieldOption("replicated", L("With replication (value and two factor columns)"))]),113                .range("data", L("Data range")),114                .column("value", L("Value column"), visible: { $0.choice("design") == "replicated" }),115                .column("factorA", L("Factor A (rows)"), numeric: false, visible: { $0.choice("design") == "replicated" }),116                .column("factorB", L("Factor B (columns)"), numeric: false, visible: { $0.choice("design") == "replicated" }),117                .alpha(L("Significance level α")),118            ],119            info: [120                L("Without replication: every numeric column is a level of the column factor and every row a level of the row factor; the interaction cannot be separated from the error."),121                L("With replication: fixed-effects F tests against the within-cell mean square; the design must be balanced (the same number, at least 2, of observations in every cell)."),122                L("Unbalanced designs are not supported (they need Type II/III sums of squares): use the regression tool with categorical dummies instead."),123            ],124            run: run)125        t.demo = { _, v in v.set("design", choice: "table"); v.set("data", range: "Groups!H1:K4") }126        return t127    }()128129    static func run(_ input: ToolInput) throws -> ToolResult {130        let v = input.values131        let alpha = try Common.alpha(v)132        let frame = try input.frame()133        let result: AnovaResult134        var rowLevels: [String], colLevels: [String]135        var rowName = L("Rows"), colName = L("Columns")136        var subtitle = frame.label137        if v.choice("design") == "replicated" {138            let value = try input.column("value"), a = try input.column("factorA"), b = try input.column("factorB")139            guard Set([value.index, a.index, b.index]).count == 3 else { throw ToolError(L("Choose three different columns.")) }140            let la = a.labels, lb = b.labels141            rowLevels = []; colLevels = []142            var cells = [[String]: [Double]]()143            var skipped = 0144            for (i, x) in value.numbers.enumerated() {145                guard !x.isNaN, !la[i].isEmpty, !lb[i].isEmpty else { skipped += 1; continue }146                if !rowLevels.contains(la[i]) { rowLevels.append(la[i]) }147                if !colLevels.contains(lb[i]) { colLevels.append(lb[i]) }148                cells[[la[i], lb[i]], default: []].append(x)149            }150            guard rowLevels.count >= 2, colLevels.count >= 2 else { throw ToolError(L("Each factor needs at least two levels.")) }151            let grid = rowLevels.map { r in colLevels.map { cells[[r, $0]] ?? [] } }152            let sizes = Set(grid.flatMap { $0.map(\.count) })153            guard sizes.count == 1, let reps = sizes.first, reps >= 2 else {154                throw ToolError(sizes.count == 1155                                ? L("Every cell holds a single observation: use the design without replication.")156                                : L("The design is unbalanced (cells hold between %d and %d observations). Unbalanced two-way ANOVA is not supported; use the regression tool with categorical dummies.", sizes.min() ?? 0, sizes.max() ?? 0))157            }158            result = try Inference.twoWayANOVA(replicated: grid)159            rowName = a.name; colName = b.name160            subtitle += " — " + value.name + " ~ " + a.name + " × " + b.name161            if skipped > 0 { subtitle += " · " + L("%d rows skipped", skipped) }162        } else {163            let numeric = frame.columns.filter(\.isNumeric)164            guard numeric.count >= 2 else { throw ToolError(L("The table needs at least two numeric columns.")) }165            let labelCol = frame.columns.first { !$0.isNumeric && $0.textCount > 0 }166            guard numeric.allSatisfy({ $0.missingCount == 0 }) else { throw ToolError(L("The table contains missing values; two-way ANOVA without replication needs one value in every cell.")) }167            colLevels = numeric.map(\.name)168            rowLevels = (0..<frame.rowCount).map { r in labelCol.map { $0.labels[r] }.flatMap { $0.isEmpty ? nil : $0 } ?? L("Row %d", r + 1) }169            result = try Inference.twoWayANOVA((0..<frame.rowCount).map { r in numeric.map { $0.numbers[r] } })170            if let labelCol { rowName = labelCol.name }171        }172        let errorRow = result.row("Within") ?? result.row("Error")173        let dfError = errorRow?.df ?? .nan174        var blocks: [ResultBlock] = [.table(AnovaSupport.table(result, rowName: rowName, colName: colName))]175        if let r = result.row("Rows") {176            blocks += AnovaSupport.effectBlocks(r, name: rowName, dfError: dfError, alpha: alpha, null: L("the means of all levels of %@ are equal", rowName), locale: input.locale)177        }178        if let c = result.row("Columns") {179            blocks += AnovaSupport.effectBlocks(c, name: colName, dfError: dfError, alpha: alpha, null: L("the means of all levels of %@ are equal", colName), locale: input.locale)180        }181        if let i = result.row("Interaction") {182            blocks += AnovaSupport.effectBlocks(i, name: L("Interaction"), dfError: dfError, alpha: alpha, null: L("there is no interaction between the two factors"), locale: input.locale)183        }184        let a = rowLevels.count185        let groups = result.groups.enumerated().map { i, g in (i < a ? rowLevels[i] : colLevels[i - a], g.count, g.mean, g.variance) }186        blocks.append(.table(AnovaSupport.groupTable(Array(groups.prefix(a)), title: L("Levels of %@", rowName))))187        blocks.append(.table(AnovaSupport.groupTable(Array(groups.dropFirst(a)), title: L("Levels of %@", colName))))188        blocks.append(.chart(ChartBlock(title: L("Level means"),189                                        spec: ChartKit.columns(categories: groups.map(\.0), series: [(L("Mean"), groups.map(\.2))], xTitle: "", yTitle: L("Mean")),190                                        summary: L("Bars of the mean of each level: %d levels of %@ and %d levels of %@.", a, rowName, colLevels.count, colName), height: 220)))191        blocks.append(.assumptions(result.assumptions, title: L("Assumptions")))192        return ToolResult(title: v.choice("design") == "replicated" ? L("Two-Way ANOVA with Replication") : L("Two-Way ANOVA without Replication"),193                          subtitle: subtitle, blocks: blocks, sheetName: L("Two-way ANOVA"))194    }195}196
Files main
Menu

Theme