>h]SrSSKrSSKJr SSKJr SSKJr SSK J r J r J r J r Sr"SS \5rSS jrSS jrS rS rg)z| Rank based methods for inferential statistics Created on Sat Aug 15 10:18:53 2020 Author: Josef Perktold License: BSD-3 N)stats)rankdata) HolderTuple)_tconfint_generic_tstat_generic_zconfint_generic_zstat_genericc*[R"U5n[R"U5n[U5n[U5nUS:XdUS:Xa [S5e[R"X45nUR S:a[R "[SU5nO [U5nUSUnXRSnUR S:a9[R "[SU5n[R "[SU5n O[U5n[U5n XgX4$)aCompute midranks for two samples Parameters ---------- x1, x2 : array_like Original data for two samples that will be converted to midranks. Returns ------- rank1 : ndarray Midranks of the first sample in the pooled sample. rank2 : ndarray Midranks of the second sample in the pooled sample. ranki1 : ndarray Internal midranks of the first sample. ranki2 : ndarray Internal midranks of the second sample. rone sample has zero lengthN)npasarraylen ValueError concatenatendimapply_along_axisr) x1x2nobs1nobs2 x_combinedrankrank1rank2ranki1ranki2s rC:\Users\julio\OneDrive\Documentos\Trabajo\Ideas Frescas\venv\Lib\site-packages\statsmodels/stats/nonparametric.pyrankdata_2samprs( BB BB GE GE zUaZ566)J""8Q ; # %LE LE$$Xq"5$$Xq"5""  ''cT\rSrSrSrS SjrS SjrSrS SjrSSjr SS jr S r g)RankCompareResultEzResults for rank comparison This is a subclass of HolderTuple that includes results from intermediate computations, as well as methods for hypothesis tests, confidence intervals and summary. NcUnUcSnURU- n[R"URUR- 5nUR SLa [ XVX#5$[XVURUU5$)a{ Confidence interval for probability that sample 1 has larger values Confidence interval is for the shifted probability P(x1 > x2) + 0.5 * P(x1 = x2) - value Parameters ---------- value : float Value, default 0, shifts the confidence interval, e.g. ``value=0.5`` centers the confidence interval at zero. alpha : float Significance level for the confidence interval, coverage is ``1-alpha`` alternative : str The alternative hypothesis, H1, has to be one of the following * 'two-sided' : H1: ``prob - value`` not equal to 0. * 'larger' : H1: ``prob - value > 0`` * 'smaller' : H1: ``prob - value < 0`` Returns ------- lower : float or ndarray Lower confidence limit. This is -inf for the one-sided alternative "smaller". upper : float or ndarray Upper confidence limit. This is inf for the one-sided alternative "larger". rF) prob1r sqrtvarnobsuse_trrdf)selfvaluealpha alternativep0diffstd_diffs rconf_intRankCompareResult.conf_intMsoD :BzzB77488dii/0 :: $TUH H$TTWWe%02 2r c 2Un[R"URUR- 5nUR(d[ UR X4USS9upVSnO%[UR X4URUSS9upVSn[UUURUS9nU$)atest for superiority probability H0: P(x1 > x2) + 0.5 * P(x1 = x2) = value The alternative is that the probability is either not equal, larger or smaller than the null-value depending on the chosen alternative. Parameters ---------- value : float Value of the probability under the Null hypothesis. alternative : str The alternative hypothesis, H1, has to be one of the following * 'two-sided' : H1: ``prob - value`` not equal to 0. * 'larger' : H1: ``prob - value > 0`` * 'smaller' : H1: ``prob - value < 0`` Returns ------- res : HolderTuple HolderTuple instance with the following main attributes statistic : float Test statistic for z- or t-test pvalue : float Pvalue of the test based on either normal or t distribution. r)r0normalt) statisticpvaluer* distribution) r r&r'r(r)r r%rr*r) r+r,r.r/r1statpvdistrress rtest_prob_superior$RankCompareResult.test_prob_superior{s>77488dii/0zz%djj" +,.HDE%djj"&1;HDED!#!WW',  r crURUSS9nURUSS9n[R"URUR:[5nSn[ [R "UURUR/5[R "XSRUR/5UUUS9nU$)atest of stochastic (non-)equivalence of p = P(x1 > x2) Null hypothesis: p < low or p > upp Alternative hypothesis: low < p < upp where p is the probability that a random draw from the population of the first sample has a larger value than a random draw from the population of the second sample, specifically p = P(x1 > x2) + 0.5 * P(x1 = x2) If the pvalue is smaller than a threshold, say 0.05, then we reject the hypothesis that the probability p that distribution 1 is stochastically superior to distribution 2 is outside of the interval given by thresholds low and upp. Parameters ---------- low, upp : float equivalence interval low < mean < upp Returns ------- res : HolderTuple HolderTuple instance with the following main attributes pvalue : float Pvalue of the equivalence test given by the larger pvalue of the two one-sided tests. statistic : float Test statistic of the one-sided test that has the larger pvalue. results_larger : HolderTuple Results instanc with test statistic, pvalue and degrees of freedom for lower threshold test. results_smaller : HolderTuple Results instanc with test statistic, pvalue and degrees of freedom for upper threshold test. larger)r.smallerz7Equivalence test for Prob(x1 > x2) + 0.5 Prob(x1 = x2) )r7r8results_largerresults_smallertitle)r>r rr8intrchooser7)r+lowuppt1t2idx_maxrEr=s rtost_prob_superior$RankCompareResult.tost_prob_superiorsT $ $Sh $ ?  $ $Si $ @**RYY2C8IBIIg/1||R\\.J%L"$7YY 4J!K)+*, % r cVURX4S9upVXU--nXU--nUS:aXpXx4$)aNconfidence interval of a linear transformation of prob1 This computes the confidence interval for d = const + slope * prob1 Default values correspond to Somers' d. Parameters ---------- const, slope : float Constant and slope for linear (affine) transformation. alpha : float Significance level for the confidence interval, coverage is ``1-alpha`` alternative : str The alternative hypothesis, H1, has to be one of the following * 'two-sided' : H1: ``prob - value`` not equal to 0. * 'larger' : H1: ``prob - value > 0`` * 'smaller' : H1: ``prob - value < 0`` Returns ------- lower : float or ndarray Lower confidence limit. This is -inf for the one-sided alternative "smaller". upper : float or ndarray Upper confidence limit. This is inf for the one-sided alternative "larger". )r-r.r)r2) r+constsloper-r.low_pupp_prHrIs rconfint_lintransf#RankCompareResult.confint_lintransfsBF}}5}J em#em# 19xr cUc URn[RRU5[R "S5-$)a Cohen's d, standardized mean difference under normality assumption. This computes the standardized mean difference, Cohen's d, effect size that is equivalent to the rank based probability ``p`` of being stochastically larger if we assume that the data is normally distributed, given by :math: `d = F^{-1}(p) * \sqrt{2}` where :math:`F^{-1}` is the inverse of the cdf of the normal distribution. Parameters ---------- prob : float in (0, 1) Probability to be converted to Cohen's d effect size. If prob is None, then the ``prob1`` attribute is used. Returns ------- equivalent Cohen's d effect size under normality assumption. )r%rnormppfr r&)r+probs reffectsize_normal#RankCompareResult.effectsize_normals22 <::Dzz~~d#bggaj00r c Sn[R"UR5nURcUR 5upVOURnUR n[R"U5n[R "URUS95nURSS:a URnURn [R"[R"UR55n [R"U5nUc'[[U55V s/sHn SU -PM nn UV s/sHn SU -PM n n Sn SSKJn U"XXXx4Xx2) %sz-Probability sample 1 is stochastically larger)summary_params)ynamexnamer)rEr-)r atleast_1dr%r8r>r7 atleast_2dr2shapeTr)r&var_probrangerstatsmodels.iolib.summaryr_)r+r-rar`effectr7r8pvaluescir)sdiixname2rEr_summs rsummaryRankCompareResult.summary2s2$tzz* ;;  $ 7 7 9 Iv[[FI--' ]]4==u=5 6 88A;?B  ]]2774==1 2MM), =*/F *<=*8s "E7E#)N皙? two-sided)g?rt)rWrsrt)N)rsN) __name__ __module__ __qualname____firstlineno____doc__r2r>rMrTr[rp__static_attributes__rrr rr"r"Es1,2\3j9v:>&1(T1:,r r"c[R"U5n[R"U5n[U5n[U5nX4-nUS:XdUS:Xa [S5e[ X5upgp[R "USS9n [R "USS9n [R "USS9n [R "U SS9n [R "[R"Xh- U - U -S5SS9nXS- -n[R "[R"Xy- U - U -S5SS9nXS- -nX4-X- -nUX4-[R"X>-XO--5--nU(a[R"X>-XO--S5n[R"X>-S5US- - nU[R"XO-S5US- - - nUU- nS[RR[R"U5U5-nO8S[RR[R"U55-nSnXU- S-- nXU- S-- nUU- UU- -nUUU- UU- --nXS-S- - U- nXS-S- - U- n[S0SU_S U_S U_S U_S U_S U_SU_SU_SU_SU_SU_SU _SU _SU_SU_SUS-S- _SUS-S- _SU_SU_6$)u Statistics and tests for the probability that x1 has larger values than x2. p is the probability that a random draw from the population of the first sample has a larger value than a random draw from the population of the second sample, specifically p = P(x1 > x2) + 0.5 * P(x1 = x2) This is a measure underlying Wilcoxon-Mann-Whitney's U test, Fligner-Policello test and Brunner-Munzel test, and Inference is based on the asymptotic distribution of the Brunner-Munzel test. The half probability for ties corresponds to the use of midranks and make it valid for discrete variables. The Null hypothesis for stochastic equality is p = 0.5, which corresponds to the Brunner-Munzel test. Parameters ---------- x1, x2 : array_like Array of samples, should be one-dimensional. use_t : boolean If use_t is true, the t distribution with Welch-Satterthwaite type degrees of freedom is used for p-value and confidence interval. If use_t is false, then the normal distribution is used. Returns ------- res : RankCompareResult The results instance contains the results for the Brunner-Munzel test and has methods for hypothesis tests, confidence intervals and summary. statistic : float The Brunner-Munzel W statistic. pvalue : float p-value assuming an t distribution. One-sided or two-sided, depending on the choice of `alternative` and `use_t`. See Also -------- RankCompareResult scipy.stats.brunnermunzel : Brunner-Munzel test for stochastic equality scipy.stats.mannwhitneyu : Mann-Whitney rank test on two samples. Notes ----- Wilcoxon-Mann-Whitney assumes equal variance or equal distribution under the Null hypothesis. Fligner-Policello test allows for unequal variances but assumes continuous distribution, i.e. no ties. Brunner-Munzel extend the test to allow for unequal variance and discrete or ordered categorical random variables. Brunner and Munzel recommended to estimate the p-value by t-distribution when the size of data is 50 or less. If the size is lower than 10, it would be better to use permuted Brunner Munzel test (see [2]_) for the test of stochastic equality. This measure has been introduced in the literature under many different names relying on a variety of assumptions. In psychology, McGraw and Wong (1992) introduced it as Common Language effect size for the continuous, normal distribution case, Vargha and Delaney (2000) [3]_ extended it to the nonparametric continuous distribution case as in Fligner-Policello. WMW and related tests can only be interpreted as test of medians or tests of central location only under very restrictive additional assumptions such as both distribution are identical under the equality null hypothesis (assumed by Mann-Whitney) or both distributions are symmetric (shown by Fligner-Policello). If the distribution of the two samples can differ in an arbitrary way, then the equality Null hypothesis corresponds to p=0.5 against an alternative p != 0.5. see for example Conroy (2012) [4]_ and Divine et al (2018) [5]_ . Note: Brunner-Munzel and related literature define the probability that x1 is stochastically smaller than x2, while here we use stochastically larger. This equivalent to switching x1 and x2 in the two sample case. References ---------- .. [1] Brunner, E. and Munzel, U. "The nonparametric Benhrens-Fisher problem: Asymptotic theory and a small-sample approximation". Biometrical Journal. Vol. 42(2000): 17-25. .. [2] Neubert, K. and Brunner, E. "A studentized permutation test for the non-parametric Behrens-Fisher problem". Computational Statistics and Data Analysis. Vol. 51(2007): 5192-5204. .. [3] Vargha, András, and Harold D. Delaney. 2000. “A Critique and Improvement of the CL Common Language Effect Size Statistics of McGraw and Wong.” Journal of Educational and Behavioral Statistics 25 (2): 101–32. https://doi.org/10.3102/10769986025002101. .. [4] Conroy, Ronán M. 2012. “What Hypotheses Do ‘Nonparametric’ Two-Group Tests Actually Test?” The Stata Journal: Promoting Communications on Statistics and Stata 12 (2): 182–90. https://doi.org/10.1177/1536867X1201200202. .. [5] Divine, George W., H. James Norton, Anna E. Barón, and Elizabeth Juarez-Colunga. 2018. “The Wilcoxon–Mann–Whitney Procedure Fails as a Test of Medians.” The American Statistician 72 (3): 278–86. https://doi.org/10.1080/00031305.2017.1305291. rr axisg@r rWNr7r8s1s2var1var2r'rfrrr(mean1mean2r%prob2somersd1somersd2r*r)rr)r rrrrmeansumpowerr&rr6sfabsrXr")rrr)rrr(rrrrmeanr1meanr2meanri1meanri2S1S2wbfndf_numerdf_denomr*r8rrrfr'r%rs rrank_compare_2indeprasJ BB BB GE GE =D zUaZ566#1"#9 E& WWU #F WWU #Fggf1%Gggf1%G &07:C@q IB!)OB &07:C@q IB!)OB =FO ,DU]bggej5:&=> >>D 88EJ3S988EJ, :BHHUZ-;;  UWWZZt b11UZZ]]266$<00  ! !D ! !Du te|+H $,- .C qyAo % .E qyAo % .E  t F r b "& -1 7: &. $) 16 =A $*  28   $)   16   ',ai!m   ?Dai!m  !# +0 r cp[R"U5n[R"U5nUR5UR5pTX- nX- n[R"S/U45R SS9n[R"S/U45R SS9n USSUSS-S- n U SSU SS-S- n X-R5n X-R5n U S-U-R5U S-- nU S-U-R5U S-- nXU- - XU- - -nXE-nUU-nX-U-XB- - nX-U-XR- - nUU-S-US-US- - US-US- - -- n[ S0SS_SS_S S_S S_S U_S U_S U_SU_SU_SU_SU_SS_SS_SU _SU _SU S-S- _SU S-S- _SU_SU_6nU$)a Stochastically larger probability for 2 independent ordinal samples. This is a special case of `rank_compare_2indep` when the data are given as counts of two independent ordinal, i.e. ordered multinomial, samples. The statistic of interest is the probability that a random draw from the population of the first sample has a larger value than a random draw from the population of the second sample, specifically p = P(x1 > x2) + 0.5 * P(x1 = x2) Parameters ---------- count1 : array_like Counts of the first sample, categories are assumed to be ordered. count2 : array_like Counts of the second sample, number of categories and ordering needs to be the same as for sample 1. ddof : scalar Degrees of freedom correction for variance estimation. The default ddof=1 corresponds to `rank_compare_2indep`. use_t : bool If use_t is true, the t distribution with Welch-Satterthwaite type degrees of freedom is used for p-value and confidence interval. If use_t is false, then the normal distribution is used. Returns ------- res : RankCompareResult This includes methods for hypothesis tests and confidence intervals for the probability that sample 1 is stochastically larger than sample 2. See Also -------- rank_compare_2indep RankCompareResult Notes ----- The implementation is based on the appendix of Munzel and Hauschke (2003) with the addition of ``ddof`` so that the results match the general function `rank_compare_2indep`. rr}r NrurWr7r8rrrrr'rfrrr(rrr%rrrr*r)rr)r rrrcumsumr")count1count2ddofr)rrfreq1freq2cdf1cdf2cdfm1cdfm2r%rrrrfr(r'vn1vn2r*r=s rrank_compare_2ordinalrsj`ZZ F ZZ F::<5 NE NE >>A3, ' . .A . 6D >>A3, ' . .A . 6D!"XSb !Q &E !"XSb !Q &E ]   !E ]   !E 1Hu  ! ! #eQh .D 1Hu  ! ! #eQh .D % (==H =D /C , %, /C , %, /C )a36UQY/#q&EAI2FF GB  d 4 D T !% ,0 69 %- #( 05 <@ #'  /3   #(   05   &+QY]   >CQY]   " */ C Jr c8URUR5$)a Probability indicating that distr1 is stochastically larger than distr2. This computes p = P(x1 > x2) for two continuous distributions, where `distr1` and `distr2` are the distributions of random variables x1 and x2 respectively. Parameters ---------- distr1, distr2 : distributions Two instances of scipy.stats.distributions. The required methods are cdf of the second distribution and expect of the first distribution. Returns ------- p : probability x1 is larger than x2 Notes ----- This is a one-liner that is added mainly as reference. Examples -------- >>> from scipy import stats >>> prob_larger_continuous(stats.norm, stats.t(5)) 0.4999999999999999 # which is the same as >>> stats.norm.expect(stats.t(5).cdf) 0.4999999999999999 # distribution 1 with smaller mean (loc) than distribution 2 >>> prob_larger_continuous(stats.norm, stats.norm(loc=1)) 0.23975006109347669 )expectcdf)distr1distr2s rprob_larger_continuousrPsT == $$r cn[RRU[R"S5- 5$)a Convert Cohen's d effect size to stochastically-larger-probability. This assumes observations are normally distributed. Computed as p = Prob(x1 > x2) = F(d / sqrt(2)) where `F` is cdf of normal distribution. Cohen's d is defined as d = (mean1 - mean2) / std where ``std`` is the pooled within standard deviation. Parameters ---------- d : float or array_like Cohen's d effect size for difference mean1 - mean2. Returns ------- prob : float or ndarray Prob(x1 > x2) rW)rrXrr r&)ds rcohensd2problargerr}s#6 ::>>!bggaj. ))r )T)r T)rznumpyr scipyr scipy.statsrstatsmodels.stats.baserstatsmodels.stats.weightstatsrrrr rr"rrrrrrr rrsP .)(XY YxXvQh*%Z*r