>hXSrSSKrSSKrSSKJrJrJr SSKJ r Sr Sr Sr Sr S rS#S jrS$S jrS%S jrS&S jrS'SjrS(SjrS(Sjr"SS5r"SS\5r"SS\5r"SS\5r"SS\5r"SS\5r"SS\5r"SS \5r"S!S"\5r\"5R>r \"5R>r!\"5R>r"g))aDStatistical power, solving for nobs, ... - trial version Created on Sat Jan 12 21:48:06 2013 Author: Josef Perktold Example roundtrip - root with respect to all variables calculated, desired nobs 33.367204205 33.367204205 effect 0.5 0.5 alpha 0.05 0.05 power 0.8 0.8 TODO: refactoring - rename beta -> power, beta (type 2 error is beta = 1-power) DONE - I think the current implementation can handle any kinds of extra keywords (except for maybe raising meaningful exceptions - streamline code, I think internally classes can be merged how to extend to k-sample tests? user interface for different tests that map to the same (internal) test class - sequence of arguments might be inconsistent, arg and/or kwds so python checks what's required and what can be None. - templating for docstrings ? N)statsoptimizespecial)brentq_expandingc0[R"XU5$Nrnctdtrxdfncs jC:\Users\julio\OneDrive\Documentos\Trabajo\Ideas Frescas\venv\Lib\site-packages\statsmodels/stats/power.pynct_cdfr's >>"! $$c6S[R"XU5- $Nr r s rnct_sfr+s w~~ba( ((rc0[R"XX05$rrncfdtrr dfndfdrs rncf_cdfr/s >>#B **rc6S[R"XX05- $rrrs rncf_sfr3s w~~c. ..rc0[R"XX05$r)rncfdtri)qrrrs rncf_ppfr"7s ??3R ++rc UnUcUS- nUS;aUS- nOUS;aUnO [S5eSnUS;a[RRXc5n[R "[R "U55(a[RnO#[XU[R"U5-5nUS;a[RRXc5n [R "[R "U 55(a[RnU$U[XU[R"U5-5- nU$) zCalculate power of a ttest r two-sided2s@smallerlarger8alternative has to be 'two-sided', 'larger' or 'smaller'rr%r&r*r%r&r)) ValueErrorrtisfnpanyisnannanrsqrtppfr) effect_sizenobsalphar alternativedalpha_pow_crit_uppcrit_lows r ttest_powerr@;s A z AX)) - -() ) D3377;;v* 66"((8$ % %66D("''$-8D4477;;v* 66"((8$ % %66D K GH!BGGDM/: :D KrcUnUS;aUS- nOUS;aUnO [S5eSnUS;aZ[RRU5n[RR X[ R "U5-U- - 5nUS;a][RRU5nU[RRX[ R "U5-U- - 5- nU$)aCalculate power of a normal distributed test statistic This is an generalization of `normal_power` when variance under Null and Alternative differ. Parameters ---------- effect size : float difference in the estimated means or statistics under the alternative normalized by the standard deviation (without division by sqrt(nobs). nobs : float or int number of observations alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. alternative : string, 'two-sided' (default), 'larger', 'smaller' extra argument to choose whether the power is calculated for a two-sided (default) or one sided test. The one-sided test can be either 'larger', 'smaller'. r$r'r(r+rr,r- r.rnormr0sfr1r5r6cdf) r7r8r9r:sigmar;r<r=crits r normal_powerrHas, A)) - -() ) D33zz~~f%zz}}TbggdmOE$99:44zz~~f%  t oe&;;<< KrcUnUcUnUS;aUS- nOUS;aUnO [S5eX4- nSn US;a][RRU5n [RR X-U[ R "U5-U- - 5n US;a`[RRU5n U [RRX-U[ R "U5-U- - 5- n U $)aCalculate power of a normal distributed test statistic This is an generalization of `normal_power` when variance under Null and Alternative differ. Parameters ---------- diff : float difference in the estimated means or statistics under the alternative. nobs : float or int number of observations alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. std_null : float standard deviation under the Null hypothesis without division by sqrt(nobs) std_alternative : float standard deviation under the Alternative hypothesis without division by sqrt(nobs) alternative : string, 'two-sided' (default), 'larger', 'smaller' extra argument to choose whether the power is calculated for a two-sided (default) or one sided test. The one-sided test can be either 'larger', 'smaller'. Returns ------- power : float r$r'r(r+rr,r-rB) diffr8r9std_nullstd_alternativer:r;r< std_ratior=rGs rnormal_power_hetrNs@ A")) - -() )*I D33zz~~f%zz}}T-rwwt}_>?@44zz~~f%  t/ @ AB B KrcUcUn[RRU5n[RRU5n[R"Xc-XT-- S5U- S-nU$)aexplicit sample size computation if only one tail is relevant The sample size is based on the power in one tail assuming that the alternative is in the tail where the test has power that increases with sample size. Use alpha/2 to compute the one tail approximation to the two-sided test, i.e. consider only one tail of two-sided test. Parameters ---------- diff : float difference in the estimated means or statistics under the alternative. power : float in interval (0,1) power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. Note: alpha is used for one tail. Use alpha/2 for two-sided alternative. std_null : float standard deviation under the Null hypothesis without division by sqrt(nobs) std_alternative : float standard deviation under the Alternative hypothesis without division by sqrt(nobs). Defaults to None. If None, ``std_alternative`` is set to the value of ``std_null``. Returns ------- nobs : float Sample size to achieve (at least) the desired power. If the minimum power is satisfied for all positive sample sizes, then ``nobs`` will be zero. This will be the case when power <= alpha if std_alternative is equal to std_null. r)rrCr0r1maximum)rJpowerr9rKrL crit_powerrGn1s rnormal_sample_size_one_tailrUscR"&J ::>>% D **T_z'CCQ G   B Irc|US- nX- n[RRX%U5n[XuX`S-U-5nU$)zpower for ftest for one way anova with k equal sized groups nobs total sample size, sum over all groups should be general nobs observations, k_groups restrictions ??? rrPrfr0r) r7r8r9k_groupsr df_numdf_denomrGr=s rftest_anova_powerr\sB\FH 77;;uh /D $q.4*? @D Krc|X!peUS-Xe-U--n[RRX5U5n[XXg5n U $)aCalculate the power of a F-test. Parameters ---------- effect_size : float The effect size is here Cohen's ``f``, the square root of ``f2``. df2 : int or float Denominator degrees of freedom. This corresponds to the df_resid in Wald tests. df1 : int or float Numerator degrees of freedom. This corresponds to the number of constraints in Wald tests. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. ncc : int degrees of freedom correction for non-centrality parameter. see Notes Returns ------- power : float Power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. Notes ----- changed in 0.14: use df2, df1 instead of df_num, df_denom as arg names. The latter had reversed meaning. The sample size is given implicitly by ``df2`` with fixed number of constraints given by numerator degrees of freedom ``df1``: nobs = df2 + df1 + ncc Set ncc=0 to match t-test, or f-test in LikelihoodModelResults. ncc=1 matches the non-centrality parameter in R::pwr::pwr.f2.test ftest_power with ncc=0 should also be correct for f_test in regression models, with df_num (df1) as number of constraints and d_denom (df2) as df_resid. rPrW) r7df2df1r9nccrZr[rrGr=s r ftest_powerrasGXH a8,s2 3B 77;;uh /D $ -D KrcrXU-U--n[RRX1U5n[XaX%5nU$)aqCalculate the power of a F-test. Based on Cohen's `f^2` effect size. This assumes df_num : numerator degrees of freedom, (number of constraints) df_denom : denominator degrees of freedom (df_resid in regression) nobs = df_denom + df_num + ncc nc = effect_size * nobs (noncentrality index) Power is computed one-sided in the upper tail. Parameters ---------- effect_size : float Cohen's f2 effect size or noncentrality divided by nobs. df_num : int or float Numerator degrees of freedom. This corresponds to the number of constraints in Wald tests. df_denom : int or float Denominator degrees of freedom. This corresponds to the df_resid in Wald tests. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. ncc : int degrees of freedom correction for non-centrality parameter. see Notes Returns ------- power : float Power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. Notes The sample size is given implicitly by ``df_denom`` with fixed number of constraints given by numerator degrees of freedom ``df_num``: nobs = df_denom + df_num + ncc Set ncc=0 to match t-test, or f-test in LikelihoodModelResults. ncc=1 matches the non-centrality parameter in R::pwr::pwr.f2.test ftest_power with ncc=0 should also be correct for f_test in regression models, with df_num (df1) as number of constraints and d_denom (df2) as df_resid. rW)r7rZr[r9r`rrGr=s rftest_power_f2rc:s>j 6)C/ 0B 77;;uh /D $ -D Krc>\rSrSrSrSrSrSrSrS Sjr S r g) PowerizzSStatistical Power calculations, Base Class so far this could all be class methods c URRU5 [SSSSSSSSS9UlS S KJn U"[5UlS Hn[S S S9UR U'M SHn[SS S9UR U'M SHn[SSS9UR U'M SHn[SSS9UR U'M g)Ng{Gz?g$@g333333?g333333?r )r7r8r9rRnobs1ratiorZr[r) defaultdict)r8rirZr[r'gI@low start_upp)r[?)rj:0yE>rPr9g-q=g?)rmupp)__dict__updatedict start_ttp collectionsrk start_bqexp)selfkwdsrkkeys r__init__Power.__init__s T"$S$'s!%'!  ,&t,:C$(R3$?D  S !;C$(R3$?D  S ! C$(TQ$?D  S !C$(U $BD  S !rc[er)NotImplementedError)ryargsrzs rrR Power.powers!!rcNURS5nUR"U0UD6U- $)NrR)poprR)ryrrzpower_s r_power_identityPower._power_identitys)'"zz4(4(611rc ^^^TR5VVs/sH up#UbM UPM snnm[T5S:wa [S5eTSmTS:XaTS TR"S0TD6$TSS:Xa>SSKnSSKJn UR"SU5 TS:XaTS $TS :XaTS$[S 5eSTlUUU4S jnTRTnTRTn /n [U4SS0U D6upSn U RU 5 SnU (dW R(aSnO[ R""U5(d0[$R&"XgSS9upnnUSnU RU5 OSnSnU RS/5 US:Xa[ R("U5S:aSnOITS;aA[$R*"USSSS9un nUR(aSOSnU RU5 OSnUS:XdSSKnSSKJnJn UR"UU5 U R1SU5 U TlW $s snnf![a& S nSSKnSS KJ n UR"ST3U5 GNf=f![a Sn U RS5 GNf=f)a solve for any one of the parameters of a t-test for t-test the keywords are: effect_size, nobs, alpha, power exactly one needs to be ``None``, all others need numeric values *attaches* cache_fit_res : list Cache of the result of the root finding procedure for the latest call to ``solve_power``, mainly for debugging purposes. The first element is the success indicator, one if successful. The remaining elements contain the return information of the up to three solvers that have been tried. Nrz%need exactly one keyword that is NonerrRr7)HypothesisTestWarningz"Warning: Effect size of 0 detectedr9zACannot detect an effect-size of 0. Try changing your effect-size.c>UTT'TR"S0TD6nT=RS- slTRS:a [S5e[R"U5(a[R $U$)Nriz possible endless loop (500 NaNs))r_counter RuntimeErrorr1r3inf)r fvalr{rzrys rfuncPower.solve_power..funcs`DI''/$/D MMQ M}}s""#EFFxx~~vv  r?) ValueWarningz'Warning: using default start_value for full_outputTFrfvecg-C6?)r9rRr7rpG?)ConvergenceWarningconvergence_docr)itemslenr.rRwarningsstatsmodels.tools.sm_exceptionsrwarnrrvKeyErrorrrxrappend convergedr1r3rfsolveabsbrentqrrinsert cache_fit_res)ryrzkvrrr start_valuerfit_kwdsfit_resvalresfailedsuccessinfodictiermsgrrrrr{s`` @r solve_powerPower.solve_powers,!JJL6LSQAqL6 s8q=DE E!f '>W ::%% %  ! #  M MM>@U Vg~G}$g~G}$ !dee   Y..-K##C( !'K$K(KHCF NN3  CMMG88K((*2//$FJ+L'sC 'x(v&axBFF4L4/;;%__T49=?FC#$;;aAGNN1%G!|  ! MM/+= > q'"$ }7N YK  D MMCC5I< X  Y !F NN4  !s. H-H-?H3 #I&3,I#"I#&JJNc "SSKJn SSKJn U R U5upSSKJn U RRn SnSnUS:XaU "[U55n[R"SS[U55Vs/sH nU "U5PM nn[U5H5unnUR"UX$40UD6nURUUXUUS U-S 9 S nM7 OUS ;aU "[U55n[R"SS[U55Vs/sH nU "U5PM nn[U5H6unnUR"UUU40UD6nURUUXUUS U-S 9 SnM8 OhUS;aWU "[U55n[U5H6unnUR"UUU40UD6nURUUXUUS U-S 9 SnM8 O [S5eUcSnUR!W5 UR#U5 UR%SS9 U $s snfs snf)a Plot power with number of observations or effect size on x-axis Parameters ---------- dep_var : {'nobs', 'effect_size', 'alpha'} This specifies which variable is used for the horizontal axis. If dep_var='nobs' (default), then one curve is created for each value of ``effect_size``. If dep_var='effect_size' or alpha, then one curve is created for each value of ``nobs``. nobs : {scalar, array_like} specifies the values of the number of observations in the plot effect_size : {scalar, array_like} specifies the values of the effect_size in the plot alpha : {float, array_like} The significance level (type I error) used in the power calculation. Can only be more than a scalar, if ``dep_var='alpha'`` ax : None or axis instance If ax is None, than a matplotlib figure is created. If ax is a matplotlib axis instance, then it is reused, and the plot elements are created with it. title : str title for the axis. Use an empty string, ``''``, to avoid a title. plt_kwds : {None, dict} not used yet kwds : dict These remaining keyword arguments are used as arguments to the power function. Many power function support ``alternative`` as a keyword argument, two-sample test support ``ratio``. Returns ------- Figure If `ax` is None, the created figure. Otherwise the figure to which `ax` is connected. Notes ----- This works only for classes where the ``power`` method has ``effect_size``, ``nobs`` and ``alpha`` as the first three arguments. If the second argument is ``nobs1``, then the number of observations in the plot are those for the first sample. TODO: fix this for FTestPower and GofChisquarePower TODO: maybe add line variable, if we want more than nobs and effectsize r)utils)rainbowNrrPr8rzes=%4.2F)lwr9colorlabelzNumber of Observations)z effect sizer7eszN=%4.2Fz Effect Sizerqr9zdepvar not implementedz Power of Testz lower right)loc)statsmodels.graphicsrstatsmodels.graphics.plottoolsr create_mpl_axmatplotlib.pyplotpyplotcmDark2rr1linspace enumeraterRplotr. set_xlabel set_titlelegend)rydep_varr8r7r9axtitleplt_kwdsrzrrfigpltcolormap plt_alpharcolorsiiirrRxlabelns r plot_powerPower.plot_powersd /:%%b)'66<<  f S-.F+-;;q#s;?O+PQ+Pahqk+PFQ#K0B 2t;d;e$Rj RA1 1 < <SY'F+-;;q#s4y+IJ+Iahqk+IFJ"4A ;5ADA Ur$Rj A ?& )  !SY'F"4A ;5ADAu$Rj A ?  ) 56 6 =#E f U m $ ?RKs :H H )rrrxrv)r8NNg?NNN) __name__ __module__ __qualname____firstlineno____doc__r|rRrrr__static_attributes__rrrrerezs. C,"2tlAE=AZrrec>^\rSrSrSrSSjrSU4SjjrSrU=r$) TTestPoweripzKStatistical Power calculations for one sample or paired sample t-test c[XX4US9$)a&Calculate the power of a t-test for one sample or paired samples. Parameters ---------- effect_size : float standardized effect size, mean divided by the standard deviation. effect size has to be positive. nobs : int or float sample size, number of observations. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. df : int or float degrees of freedom. By default this is None, and the df from the one sample or paired ttest is used, ``df = nobs1 - 1`` alternative : str, 'two-sided' (default), 'larger', 'smaller' extra argument to choose whether the power is calculated for a two-sided (default) or one sided test. The one-sided test can be either 'larger', 'smaller'. . Returns ------- power : float Power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. r r:r@)ryr7r8r9r r:s rrRTTestPower.powerus@;e'24 4rc&>[TU]UUUUUS9$)a solve for any one parameter of the power of a one sample t-test for the one sample t-test the keywords are: effect_size, nobs, alpha, power Exactly one needs to be ``None``, all others need numeric values. This test can also be used for a paired t-test, where effect size is defined in terms of the mean difference, and nobs is the number of pairs. Parameters ---------- effect_size : float Standardized effect size.The effect size is here Cohen's f, square root of "f2". nobs : int or float sample size, number of observations. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. power : float in interval (0,1) power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. alternative : str, 'two-sided' (default) or 'one-sided' extra argument to choose whether the power is calculated for a two-sided (default) or one sided test. 'one-sided' assumes we are in the relevant tail. Returns ------- value : float The value of the parameter that was set to None in the call. The value solves the power equation given the remaining parameters. *attaches* cache_fit_res : list Cache of the result of the root finding procedure for the latest call to ``solve_power``, mainly for debugging purposes. The first element is the success indicator, one if successful. The remaining elements contain the return information of the up to three solvers that have been tried. Notes ----- The function uses scipy.optimize for finding the value that satisfies the power equation. It first uses ``brentq`` with a prior search for bounds. If this fails to find a root, ``fsolve`` is used. If ``fsolve`` also fails, then, for ``alpha``, ``power`` and ``effect_size``, ``brentq`` with fixed bounds is used. However, there can still be cases where this fails. )r7r8r9rRr:superr)ryr7r8r9rRr: __class__s rrTTestPower.solve_powers-vw"{;?[TU]UUUUUUS9$)asolve for any one parameter of the power of a two sample t-test for t-test the keywords are: effect_size, nobs1, alpha, power, ratio exactly one needs to be ``None``, all others need numeric values Parameters ---------- effect_size : float standardized effect size, difference between the two means divided by the standard deviation. `effect_size` has to be positive. nobs1 : int or float number of observations of sample 1. The number of observations of sample two is ratio times the size of sample 1, i.e. ``nobs2 = nobs1 * ratio`` alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. power : float in interval (0,1) power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. ratio : float ratio of the number of observations in sample 2 relative to sample 1. see description of nobs1 The default for ratio is 1; to solve for ratio given the other arguments it has to be explicitly set to None. alternative : str, 'two-sided' (default), 'larger', 'smaller' extra argument to choose whether the power is calculated for a two-sided (default) or one sided test. The one-sided test can be either 'larger', 'smaller'. Returns ------- value : float The value of the parameter that was set to None in the call. The value solves the power equation given the remaining parameters. Notes ----- The function uses scipy.optimize for finding the value that satisfies the power equation. It first uses ``brentq`` with a prior search for bounds. If this fails to find a root, ``fsolve`` is used. If ``fsolve`` also fails, then, for ``alpha``, ``power`` and ``effect_size``, ``brentq`` with fixed bounds is used. However, there can still be cases where this fails. r7rir9rRrjr:rryr7rir9rRrjr:rs rrTTestIndPower.solve_power0hw"{Xl[TU]"S0UD6 g)Nr)ddofrr|)ryrrzrs rr|NormalIndPower.__init__Ts   4 rcpURnUS:aX$-nSSX&- - SXv- - -- nOX&- n[XX5S9$)aCalculate the power of a z-test for two independent sample Parameters ---------- effect_size : float standardized effect size, difference between the two means divided by the standard deviation. effect size has to be positive. nobs1 : int or float number of observations of sample 1. The number of observations of sample two is ratio times the size of sample 1, i.e. ``nobs2 = nobs1 * ratio`` ``ratio`` can be set to zero in order to get the power for a one sample test. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. ratio : float ratio of the number of observations in sample 2 relative to sample 1. see description of nobs1 alternative : str, 'two-sided' (default), 'larger', 'smaller' extra argument to choose whether the power is calculated for a two-sided (default) or one sided test. The one-sided test can be either 'larger', 'smaller'. Returns ------- power : float Power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. rro)r:)rrH) ryr7rir9rjr:rrr8s rrRNormalIndPower.powerXsOFyy 19KEel+bEL.AABD[TU]UUUUUUS9$)a>solve for any one parameter of the power of a two sample z-test for z-test the keywords are: effect_size, nobs1, alpha, power, ratio exactly one needs to be ``None``, all others need numeric values Parameters ---------- effect_size : float standardized effect size, difference between the two means divided by the standard deviation. If ratio=0, then this is the standardized mean in the one sample test. nobs1 : int or float number of observations of sample 1. The number of observations of sample two is ratio times the size of sample 1, i.e. ``nobs2 = nobs1 * ratio`` ``ratio`` can be set to zero in order to get the power for a one sample test. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. power : float in interval (0,1) power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. ratio : float ratio of the number of observations in sample 2 relative to sample 1. see description of nobs1 The default for ratio is 1; to solve for ration given the other arguments it has to be explicitly set to None. alternative : str, 'two-sided' (default), 'larger', 'smaller' extra argument to choose whether the power is calculated for a two-sided (default) or one sided test. The one-sided test can be either 'larger', 'smaller'. Returns ------- value : float The value of the parameter that was set to None in the call. The value solves the power equation given the remaining parameters. Notes ----- The function uses scipy.optimize for finding the value that satisfies the power equation. It first uses ``brentq`` with a prior search for bounds. If this fails to find a root, ``fsolve`` is used. If ``fsolve`` also fails, then, for ``alpha``, ``power`` and ``effect_size``, ``brentq`` with fixed bounds is used. However, there can still be cases where this fails. rrrs rrNormalIndPower.solve_powers0pw"{^\rSrSrSrSSjrSU4SjjrSrU=r$) FTestPoweriaBStatistical Power calculations for generic F-test of a constraint This class is not recommended, use `FTestPowerF2` with corrected interface. This is based on Cohen's f as effect size measure. Warning: Methods in this class have the names df_num and df_denom reversed. See Also -------- FTestPowerF2 : Class with Cohen's f-squared as effect size, corrected keyword names. Examples -------- Sample size and power for multiple regression base on R-squared Compute effect size from R-squared >>> r2 = 0.1 >>> f2 = r2 / (1 - r2) >>> f = np.sqrt(f2) >>> r2, f2, f (0.1, 0.11111111111111112, 0.33333333333333337) Find sample size by solving for denominator df, wrongly named ``df_num`` >>> df1 = 1 # number of constraints in hypothesis test >>> df2 = FTestPower().solve_power(effect_size=f, alpha=0.1, power=0.9, df_denom=df1) >>> ncc = 1 # default >>> nobs = df2 + df1 + ncc >>> df2, nobs (76.46459758305376, 78.46459758305376) verify power at df2 >>> FTestPower().power(effect_size=f, alpha=0.1, df_denom=df1, df_num=df2) 0.8999999972109698 c[XX4US9nU$)aCalculate the power of a F-test. The effect size is Cohen's ``f``, square root of ``f2``. The sample size is given by ``nobs = df_denom + df_num + ncc`` Warning: The meaning of df_num and df_denom is reversed. Parameters ---------- effect_size : float Standardized effect size. The effect size is here Cohen's ``f``, square root of ``f2``. df_num : int or float Warning incorrect name denominator degrees of freedom, This corresponds to the number of constraints in Wald tests. df_denom : int or float Warning incorrect name numerator degrees of freedom. This corresponds to the df_resid in Wald tests. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. ncc : int degrees of freedom correction for non-centrality parameter. see Notes Returns ------- power : float Power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. Notes ----- sample size is given implicitly by df_num set ncc=0 to match t-test, or f-test in LikelihoodModelResults. ncc=1 matches the non-centrality parameter in R::pwr::pwr.f2.test ftest_power with ncc=0 should also be correct for f_test in regression models, with df_num and d_denom as defined there. (not verified yet) r`)raryr7rZr[r9r`r=s rrRFTestPower.powers`;SI rc >U(a+SU;a[R"S5 O[SU35e[TU]UUUUUUS9$)aMsolve for any one parameter of the power of a F-test for the one sample F-test the keywords are: effect_size, df_num, df_denom, alpha, power Exactly one needs to be ``None``, all others need numeric values. The effect size is Cohen's ``f``, square root of ``f2``. The sample size is given by ``nobs = df_denom + df_num + ncc``. Warning: The meaning of df_num and df_denom is reversed. Parameters ---------- effect_size : float Standardized effect size. The effect size is here Cohen's ``f``, square root of ``f2``. df_num : int or float Warning incorrect name denominator degrees of freedom, This corresponds to the number of constraints in Wald tests. Sample size is given by ``nobs = df_denom + df_num + ncc`` df_denom : int or float Warning incorrect name numerator degrees of freedom. This corresponds to the df_resid in Wald tests. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. power : float in interval (0,1) power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. ncc : int degrees of freedom correction for non-centrality parameter. see Notes kwargs : empty ``kwargs`` are not used and included for backwards compatibility. If ``nobs`` is used as keyword, then a warning is issued. All other keywords in ``kwargs`` raise a ValueError. Returns ------- value : float The value of the parameter that was set to None in the call. The value solves the power equation given the remaining parameters. Notes ----- The method uses scipy.optimize for finding the value that satisfies the power equation. It first uses ``brentq`` with a prior search for bounds. If this fails to find a root, ``fsolve`` is used. If ``fsolve`` also fails, then, for ``alpha``, ``power`` and ``effect_size``, ``brentq`` with fixed bounds is used. However, there can still be cases where this fails. r8znobs is not usedzincorrect keyword(s) r7rZr[r9rRr`)rrr.rr) ryr7rZr[r9rRr`kwargsrs rrFTestPower.solve_power'sYz  01 #8!ABBw"{=C?G^\rSrSrSrSSjrSU4SjjrSrU=r$) FTestPowerF2ira:Statistical Power calculations for generic F-test of a constraint This is based on Cohen's f^2 as effect size measure. Examples -------- Sample size and power for multiple regression base on R-squared Compute effect size from R-squared >>> r2 = 0.1 >>> f2 = r2 / (1 - r2) >>> f = np.sqrt(f2) >>> r2, f2, f (0.1, 0.11111111111111112, 0.33333333333333337) Find sample size by solving for denominator degrees of freedom. >>> df1 = 1 # number of constraints in hypothesis test >>> df2 = FTestPowerF2().solve_power(effect_size=f2, alpha=0.1, power=0.9, df_num=df1) >>> ncc = 1 # default >>> nobs = df2 + df1 + ncc >>> df2, nobs (76.46459758305376, 78.46459758305376) verify power at df2 >>> FTestPowerF2().power(effect_size=f, alpha=0.1, df_num=df1, df_denom=df2) 0.8999999972109698 c[XX4US9nU$)aCCalculate the power of a F-test. The effect size is Cohen's ``f^2``. The sample size is given by ``nobs = df_denom + df_num + ncc`` Parameters ---------- effect_size : float The effect size is here Cohen's ``f2``. This is equal to the noncentrality of an F-test divided by nobs. df_num : int or float Numerator degrees of freedom, This corresponds to the number of constraints in Wald tests. df_denom : int or float Denominator degrees of freedom. This corresponds to the df_resid in Wald tests. alpha : float in interval (0,1) Significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. ncc : int Degrees of freedom correction for non-centrality parameter. see Notes Returns ------- power : float Power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. Notes ----- The sample size is given implicitly by df_denom set ncc=0 to match t-test, or f-test in LikelihoodModelResults. ncc=1 matches the non-centrality parameter in R::pwr::pwr.f2.test ftest_power with ncc=0 should also be correct for f_test in regression models, with df_num and d_denom as defined there. (not verified yet) r )rcr s rrRFTestPowerF2.powersVk8L rc (>[TU]UUUUUUS9$)aSolve for any one parameter of the power of a F-test for the one sample F-test the keywords are: effect_size, df_num, df_denom, alpha, power Exactly one needs to be ``None``, all others need numeric values. The effect size is Cohen's ``f2``. The sample size is given by ``nobs = df_denom + df_num + ncc``, and can be found by solving for df_denom. Parameters ---------- effect_size : float The effect size is here Cohen's ``f2``. This is equal to the noncentrality of an F-test divided by nobs. df_num : int or float Numerator degrees of freedom, This corresponds to the number of constraints in Wald tests. df_denom : int or float Denominator degrees of freedom. This corresponds to the df_resid in Wald tests. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. power : float in interval (0,1) power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. ncc : int degrees of freedom correction for non-centrality parameter. see Notes Returns ------- value : float The value of the parameter that was set to None in the call. The value solves the power equation given the remaining parameters. Notes ----- The function uses scipy.optimize for finding the value that satisfies the power equation. It first uses ``brentq`` with a prior search for bounds. If this fails to find a root, ``fsolve`` is used. If ``fsolve`` also fails, then, for ``alpha``, ``power`` and ``effect_size``, ``brentq`` with fixed bounds is used. However, there can still be cases where this fails. r r)ryr7rZr[r9rRr`rs rrFTestPowerF2.solve_powers.lw"{=C?GUb/US-URS'[US-US-S9URS'UcURUUUUUS9$[TU]UUUUUS9$)aTsolve for any one parameter of the power of a F-test for the one sample F-test the keywords are: effect_size, nobs, alpha, power Exactly one needs to be ``None``, all others need numeric values. Parameters ---------- effect_size : float standardized effect size, mean divided by the standard deviation. effect size has to be positive. nobs : int or float sample size, number of observations. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. power : float in interval (0,1) power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. Returns ------- value : float The value of the parameter that was set to None in the call. The value solves the power equation given the remaining parameters. Notes ----- The function uses scipy.optimize for finding the value that satisfies the power equation. It first uses ``brentq`` with a prior search for bounds. If this fails to find a root, ``fsolve`` is used. If ``fsolve`` also fails, then, for ``alpha``, ``power`` and ``effect_size``, ``brentq`` with fixed bounds is used. However, there can still be cases where this fails. rgr8rPrlr7r8r9rYrR)rvrurx_solve_effect_sizerr)ryr7r8r9rRrYrs rrFTestAnovaPower.solve_power'sV  %-]DNN6 "'+1 6>m(ED  V $  **{04164<16 +8 8 w"{;?UnTRUTTTTS9$)Nr)r)r r7r9rYr8rRrys rr0FTestAnovaPower._solve_effect_size..funchs.K''K/3053;05 (7 7rrprTr)rrrprint) ryr7r8r9rRrYrrrs ` ```` rr"FTestAnovaPower._solve_effect_sizeds8 7 7tVF{{ !H rr)rPNNNNrP) rrrrrrRrrrrrs@rrrs3N6JN;CzEI01rrc>^\rSrSrSrSSjrSU4SjjrSrU=r$)GofChisquarePowerivzBStatistical Power calculations for one sample chisquare test cSSKJn U"XXCSS9$)aCalculate the power of a chisquare test for one sample Only two-sided alternative is implemented Parameters ---------- effect_size : float standardized effect size, according to Cohen's definition. see :func:`statsmodels.stats.gof.chisquare_effectsize` nobs : int or float sample size, number of observations. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. n_bins : int number of bins or cells in the distribution. Returns ------- power : float Power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. rchisquare_powerrstatsmodels.stats.gofr*)ryr7r8r9n_binsrr*s rrRGofChisquarePower.power{s4 :{&aHHrc&>[TU]UUUUUS9$)asolve for any one parameter of the power of a one sample chisquare-test for the one sample chisquare-test the keywords are: effect_size, nobs, alpha, power Exactly one needs to be ``None``, all others need numeric values. n_bins needs to be defined, a default=2 is used. Parameters ---------- effect_size : float standardized effect size, according to Cohen's definition. see :func:`statsmodels.stats.gof.chisquare_effectsize` nobs : int or float sample size, number of observations. alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. power : float in interval (0,1) power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. n_bins : int number of bins or cells in the distribution Returns ------- value : float The value of the parameter that was set to None in the call. The value solves the power equation given the remaining parameters. Notes ----- The function uses scipy.optimize for finding the value that satisfies the power equation. It first uses ``brentq`` with a prior search for bounds. If this fails to find a root, ``fsolve`` is used. If ``fsolve`` also fails, then, for ``alpha``, ``power`` and ``effect_size``, ``brentq`` with fixed bounds is used. However, there can still be cases where this fails. )r7r8r-r9rRr)ryr7r8r9rRr-rs rrGofChisquarePower.solve_powers-\w"{;?=CB'(2C2Crr'cB^\rSrSrSrSSjrSU4SjjrSrU=r$)_GofChisquareIndPowerizStatistical Power calculations for chisquare goodness-of-fit test TODO: this is not working yet for 2sample case need two nobs in function no one-sided chisquare test, is there one? use normal distribution? -> drop one-sided options? cDSSKJn X$-nSSU- SU- -- nU"XU5$)aCalculate the power of a chisquare for two independent sample Parameters ---------- effect_size : float standardize effect size, difference between the two means divided by the standard deviation. effect size has to be positive. nobs1 : int or float number of observations of sample 1. The number of observations of sample two is ratio times the size of sample 1, i.e. ``nobs2 = nobs1 * ratio`` alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. ratio : float ratio of the number of observations in sample 2 relative to sample 1. see description of nobs1 The default for ratio is 1; to solve for ration given the other arguments it has to be explicitely set to None. alternative : str, 'two-sided' (default) or 'one-sided' extra argument to choose whether the power is calculated for a two-sided (default) or one sided test. 'one-sided' assumes we are in the relevant tail. Returns ------- power : float Power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. rr)ror+) ryr7rir9rjr:r*rr8s rrR_GofChisquareIndPower.powers3F : BJe+,{%88rc (>[TU]UUUUUUS9$)asolve for any one parameter of the power of a two sample z-test for z-test the keywords are: effect_size, nobs1, alpha, power, ratio exactly one needs to be ``None``, all others need numeric values Parameters ---------- effect_size : float standardize effect size, difference between the two means divided by the standard deviation. nobs1 : int or float number of observations of sample 1. The number of observations of sample two is ratio times the size of sample 1, i.e. ``nobs2 = nobs1 * ratio`` alpha : float in interval (0,1) significance level, e.g. 0.05, is the probability of a type I error, that is wrong rejections if the Null Hypothesis is true. power : float in interval (0,1) power of the test, e.g. 0.8, is one minus the probability of a type II error. Power is the probability that the test correctly rejects the Null Hypothesis if the Alternative Hypothesis is true. ratio : float ratio of the number of observations in sample 2 relative to sample 1. see description of nobs1 The default for ratio is 1; to solve for ration given the other arguments it has to be explicitely set to None. alternative : str, 'two-sided' (default) or 'one-sided' extra argument to choose whether the power is calculated for a two-sided (default) or one sided test. 'one-sided' assumes we are in the relevant tail. Returns ------- value : float The value of the parameter that was set to None in the call. The value solves the power equation given the remaining parameters. Notes ----- The function uses scipy.optimize for finding the value that satisfies the power equation. It first uses ``brentq`` with a prior search for bounds. If this fails to find a root, ``fsolve`` is used. If ``fsolve`` also fails, then, for ``alpha``, ``power`` and ``effect_size``, ``brentq`` with fixed bounds is used. However, there can still be cases where this fails. rrrs rr!_GofChisquareIndPower.solve_powerrrrrrrrs@rr2r2s,67%'9TKO*59O9Orr2r)r%ro)roNr%)roN)rPNr)#rrnumpyr1scipyrrrstatsmodels.tools.rootfindingrrrrrr"r@rHrNrUr\rarcrerrrrrrr'r2rtt_solve_powertt_ind_solve_powerzt_ind_solve_powerrrrr=s+<**:%)+/,#L'TFJ(6r>@040f 2j:@sslhOhOTqOEqOfwOUwOtg?g?VL?5L?^rerjUCUCnmOEmO`))"_00#%11r