M/PhgdZddlZddlZddlmZddlmZGddZ dZ dd Z dd Z ddZ dZddZdS)zM Created on Thu Apr 2 14:34:25 2020 Author: Josef Perktold License: BSD-3 N)stats) HolderTuplecNeZdZdZdZ d dZd dZdZd dZd d Z dd Z dS)CombineResultszResults from combined estimate of means or effect sizes This currently includes intermediate results that might be removed c ||j|t||_|jdz |_tj|j |_ tj|j |_ |j |jdz z |_dd|jz z |_i|_dS)N)__dict__updatelistkeys _ini_keyskdf_residnpsqrt var_hksj_fesd_eff_w_fe_hksj var_hksj_resd_eff_w_re_hksjqh2i2cache_ci)selfkwdss _/var/www/html/test/jupyter/venv/lib/python3.11/site-packages/statsmodels/stats/meta_analysis.py__init__zCombineResults.__init__s T"""diikk**  !#(8 9 9 "(8 9 9&DFQJ'a$'k/ 皙?Nc4||f|jvr|j||fS||j}||d|ini}|d d|i|}d|_n|dur*tj|dz }d|_nv|0|dz }tj|dz |}d |_nDd } d dl} | | tj|dz }d|_|j ||j zz } |j ||j zz} | | f}||j||f<|S) aconfidence intervals for the effect size estimate of samples Additional information needs to be provided for confidence intervals that are not based on normal distribution using available variance. This is likely to change in future. Parameters ---------- alpha : float in (0, 1) Significance level for confidence interval. Nominal coverage is ``1 - alpha``. use_t : None or bool If use_t is None, then the attribute `use_t` determines whether normal or t-distribution is used for confidence intervals. Specifying use_t overrides the attribute. If use_t is false, then confidence intervals are based on the normal distribution. If it is true, then the t-distribution is used. nobs : None or float Number of observations used for degrees of freedom computation. Only used if use_t is true. ci_func : None or callable User provided function to compute confidence intervals. This is not used yet and will allow using non-standard confidence intervals. Returns ------- ci_eff : tuple of ndarrays Tuple (ci_low, ci_upp) with confidence interval computed for each sample. Notes ----- CombineResults currently only has information from the combine_effects function, which does not provide details about individual samples. Nuse_talphaci_funcFnormalrtz`use_t=True` requires `nobs` for each sample or `ci_func`. Using normal distribution for confidence interval of individual samples.r) rr!ci_sample_distrrnormisfr&warningswarneffsd_eff) rr"r!nobsr#rci_effcritrmsgr+ci_lowci_upps rconf_int_sampleszCombineResults.conf_int_samples's]T 5>T] * *=%0 0 =JE  ','8GU##bDW1151D11F#,D ~~z~~eai00'/$$##axH 7;;uqy(;;D+.D((HC$OOOMM#&&& :>>%!)44D+3D(Xt{ 22FXt{ 22Ff%F)/ uen% rc||j}|dur#tj|dz }n(tj|dz |j}t jddg}|j}|j }|||z|j zz}|||z|j zz}|||zt j |j zz} |||zt j |jzz} ||| | fS)aconfidence interval for the overall mean estimate Parameters ---------- alpha : float in (0, 1) Significance level for confidence interval. Nominal coverage is ``1 - alpha``. use_t : None or bool If use_t is None, then the attribute `use_t` determines whether normal or t-distribution is used for confidence intervals. Specifying use_t overrides the attribute. If use_t is false, then confidence intervals are based on the normal distribution. If it is true, then the t-distribution is used. Returns ------- ci_eff_fe : tuple of floats Confidence interval for mean effects size based on fixed effects model with scale=1. ci_eff_re : tuple of floats Confidence interval for mean effects size based on random effects model with scale=1 ci_eff_fe_wls : tuple of floats Confidence interval for mean effects size based on fixed effects model with estimated scale corresponding to WLS, ie. HKSJ. ci_eff_re_wls : tuple of floats Confidence interval for mean effects size based on random effects model with estimated scale corresponding to WLS, ie. HKSJ. If random effects method is fully iterated, i.e. Paule-Mandel, then the estimated scale is 1. NFr$r)r!rr)r*r&rrasarraymean_effect_femean_effect_re sd_eff_w_fe sd_eff_w_rerrr) rr"r!r1sgnm_fem_re ci_eff_fe ci_eff_re ci_eff_fe_wls ci_eff_re_wlss rconf_intzCombineResults.conf_intwsD =JE E>>:>>%!),,DD7;;uqy$-88Dj"a!!""3:(888 3:(888 sTzBGD4D,E,EEE sTzBGD4D,E,EEE )]MAArctj|j|jdz }t |j||jdz d}|S)aTest whether the means of all samples are the same currently no options, test uses chisquare distribution default might change depending on `use_t` Returns ------- res : HolderTuple instance The results include the following attributes: - statistic : float Test statistic, ``q`` in meta-analysis, this is the pearson_chi2 statistic for the fixed effects model. - pvalue : float P-value based on chisquare distribution. - df : float Degrees of freedom, equal to number of studies or samples minus 1. rchi2) statisticpvaluedfdistr)rrFsfrrr)rrHress rtest_homogeneityzCombineResults.test_homogeneitysO(tvtvz22DF!'!VaZ &((( rc|||\}}tj|j|j|||j|jg}|||}|j|j |dd|dddtj gg}|j |j |dd|ddtj dgg}|j|j |dd|dddtj gg} |j |j|dd|ddtj dgg} tj|||| | gd}gd} || fS)aCreate array with sample statistics and mean estimates Parameters ---------- alpha : float in (0, 1) Significance level for confidence interval. Nominal coverage is ``1 - alpha``. use_t : None or bool If use_t is None, then the attribute `use_t` determines whether normal or t-distribution is used for confidence intervals. Specifying use_t overrides the attribute. If use_t is false, then confidence intervals are based on the normal distribution. If it is true, then the t-distribution is used. Returns ------- res : ndarray Array with columns ['eff', "sd_eff", "ci_low", "ci_upp", "w_fe","w_re"]. Rows include statistics for samples and estimates of overall mean. column_names : list of str The names for the columns, used when creating summary DataFrame. r"r!rrr$)axis)r-r.r3r4w_few_re)r5r column_stackr-r.weights_rel_feweights_rel_rerDr9r;nanr:r<rr concatenate) rr"r!r3r4rLcires_feres_re res_fe_wls res_re_wls column_namess r summary_arrayzCombineResults.summary_arraysn4..U%.HHotx%v#2D4GIJJ]]e] 4 4&(8a58RU1Xq"&23&(8a58RU1Xrvq23*D,A!uQxAq1bf67 *D,A!uQxAq26167 nc66:zJ"#%%%LLL L  rc||j}t|jgdz}|||\}}t j|||}|S)a|Create DataFrame with sample statistics and mean estimates Parameters ---------- alpha : float in (0, 1) Significance level for confidence interval. Nominal coverage is ``1 - alpha``. use_t : None or bool If use_t is None, then the attribute `use_t` determines whether normal or t-distribution is used for confidence intervals. Specifying use_t overrides the attribute. If use_t is false, then confidence intervals are based on the normal distribution. If it is true, then the t-distribution is used. Returns ------- res : DataFrame pandas DataFrame instance with columns ['eff', "sd_eff", "ci_low", "ci_upp", "w_fe","w_re"]. Rows include statistics for samples and estimates of overall mean. N)z fixed effectz random effectzfixed effect wlszrandom effect wlsrO)indexcolumns)r!r row_namesr_pd DataFrame)rr"r!labelsrL col_namesresultss r summary_framezCombineResults.summary_framesl0 =JEt~&&<<<<++%u+EEY,s&)DDDrFc ddlm}|||}|rtj|gd}tj|ddg|dgjz }|d |d||j|jd|} | S) aUForest plot with means and confidence intervals Parameters ---------- ax : None or matplotlib axis instance If ax is provided, then the plot will be added to it. alpha : float in (0, 1) Significance level for confidence interval. Nominal coverage is ``1 - alpha``. use_t : None or bool If use_t is None, then the attribute `use_t` determines whether normal or t-distribution is used for confidence intervals. Specifying use_t overrides the attribute. If use_t is false, then confidence intervals are based on the normal distribution. If it is true, then the t-distribution is used. use_exp : bool If `use_exp` is True, then the effect size and confidence limits will be exponentiated. This transform log-odds-ration into odds-ratio, and similarly for risk-ratio. ax : AxesSubplot, optional If given, this axes is used to plot in instead of a new figure being created. kwds : optional keyword arguments Keywords are forwarded to the dot_plot function that creates the plot. Returns ------- fig : Matplotlib figure instance See Also -------- dot_plot r)dot_plotrO)r-r3r4r3r4r-)points intervalslines line_orderr')statsmodels.graphics.dotplotsrkrirexpabsvaluesra) rr"r!use_expaxrrkres_dfhwfigs r plot_forestzCombineResults.plot_forestsL ;:::::##%u#==  AVF#>#>#>?@@F VFHh/065'?3II J JhLfUmr#\flLLFJLL r)rNNN)rN)rNFN) __name__ __module__ __qualname____doc__rr5rDrMr_riryr'rrrrs "=A!%NNNN`3B3B3B3Bj6,!,!,!,!\B;@------rrc|dz|dz z|dz|dz zz||zdz z }tj|}||z}ddd|zdz z z } ||z |z } | | z} ||z |z | dzdz |dz z z} | | fS)aeffect sizes for mean difference for use in meta-analysis mean1, sd1, nobs1 are for treatment mean2, sd2, nobs2 are for control Effect sizes are computed for the mean difference ``mean1 - mean2`` standardized by an estimate of the within variance. This does not have option yet. It uses standardized mean difference with bias correction as effect size. This currently does not use np.asarray, all computations are possible in pandas. Parameters ---------- mean1 : array mean of second sample, treatment groups sd1 : array standard deviation of residuals in treatment groups, within nobs1 : array number of observations in treatment groups mean2, sd2, nobs2 : arrays mean, standard deviation and number of observations of control groups Returns ------- smd_bc : array bias corrected estimate of standardized mean difference var_smdbc : array estimate of variance of smd_bc Notes ----- Status: API will still change. This is currently intended for support of meta-analysis. References ---------- Borenstein, Michael. 2009. Introduction to Meta-Analysis. Chichester: Wiley. Chen, Ding-Geng, and Karl E. Peace. 2013. Applied Meta-Analysis with R. Chapman & Hall/CRC Biostatistics Series. Boca Raton: CRC Press/Taylor & Francis Group. r$rrP gQ@)rr) mean1sd1nobs1mean2sd2nobs2var_diffsd_diffr/bias_correctionsmdsmd_bc var_smdbcs reffectsize_smdrFslQ%!)$Q%!)$%).):g !?z+zero_correction not recognized or supported)rrdr)z risk-ratiorr)z odds-ratioor)arcsinearcsinasrz;statistic not recognized, use one of "rd", "rr", "or", "as")getNotImplementedErrorr logical_orrlogrr)count1rcount2rrGzero_correction zero_kwdscc1cc2nobs_trr2 zero_mask1 zero_mask2zmaskn1n2p1p2rrd_varr-var_efflog_rr log_rr_varlog_or log_or_varas_as_vars reffectsize_2proportionsrst cc E ! !fnfn F " "mmM3CDD  cc '##cc;!#&&&A+&E/2JA+&E/2J M*j 1 1E #)u$ $B #)u$ $B 3,2 B 3,2 B&  WR &+ & & & WR &+ & & &N"" "Wq2v#bAFmb&88 * * *bfRjj("f]R'1r6R-"*<<  * * *bfQVnn,rvbzz9BF1r6NNJ"B-",-R1r6]R5G0HH  1 1 1i $$ry'='==b&1r6/Q&K!#&&& <riteratedFrc t|}|tt|}tj|dz }|} |} t j| } d| z } | d} | | z }|| z}|}d| z }t j|}| | dzzd}|| | zdz| z z}|dz }| dvrt| | fi|\}}nK| dvr&| | dz| z z }||z |z }ntdd| |zz }|d}||dz }|| z}|}d|z }t j|}|| |z dzz|z }| | |z dzz|z } || |z dzz|z }!|| |z dzz|z }"tdit}#|#S) aJ combining effect sizes for effect sizes using meta-analysis This currently does not use np.asarray, all computations are possible in pandas. Parameters ---------- effect : array mean of effect size measure for all samples variance : array variance of mean or effect size measure for all samples method_re : {"iterated", "chi2"} method that is use to compute the between random effects variance "iterated" or "pm" uses Paule and Mandel method to iteratively estimate the random effects variance. Options for the iteration can be provided in the ``kwds`` "chi2" or "dl" uses DerSimonian and Laird one-step estimator. row_names : list of strings (optional) names for samples or studies, will be included in results summary and table. alpha : float in (0, 1) significance level, default is 0.05, for the confidence intervals Returns ------- results : CombineResults Contains estimation results and intermediate statistics, and includes a method to return a summary table. Statistics from intermediate calculations might be removed at a later time. Notes ----- Status: Basic functionality is verified, mainly compared to R metafor package. However, API might still change. This computes both fixed effects and random effects estimates. The random effects results depend on the method to estimate the RE variance. Scale estimate In fixed effects models and in random effects models without fully iterated random effects variance, the model will in general not account for all residual variance. Traditional meta-analysis uses a fixed scale equal to 1, that might not produce test statistics and confidence intervals with the correct size. Estimating the scale to account for residual variance often improves the small sample properties of inference and confidence intervals. This adjustment to the standard errors is often referred to as HKSJ method based attributed to Hartung and Knapp and Sidik and Jonkman. However, this is equivalent to estimating the scale in WLS. The results instance includes both, fixed scale and estimated scale versions of standard errors and confidence intervals. References ---------- Borenstein, Michael. 2009. Introduction to Meta-Analysis. Chichester: Wiley. Chen, Ding-Geng, and Karl E. Peace. 2013. Applied Meta-Analysis with R. Chapman & Hall/CRC Biostatistics Series. Boca Raton: CRC Press/Taylor & Francis Group. Nr$rr)rpm)rFdlz(method_re should be "iterated" or "chi2"r')lenr rangerr)r*rrsumlower_fit_tau_iterative ValueErrorrlocals)$effectvariance method_rercr!r"rrr1r-rr. weights_fe w_total_ferUeff_w_fer9 var_eff_w_fer;rrItau2_c weights_re w_total_rerVeff_w_rer: var_eff_w_rer< scale_hksj_re scale_hksj_ferrrLs$ rcombine_effectsrsD F AqNN :>>%!) $ $D CG WW  FWJ""J*,N#H\\^^Nz>L',''K c1f !!!$$A*s    ! !1 $z 11A QB...$S'::T::aa   n , , *a-,,..; ;B!|CDDDgn%J""J*.."3"33N#H\\^^Nz>L',''K 3#7!";;@@BBRGM3#7!";;@@BBRGM!S>%9A$==BBDDrIK!S>%9A$==BBDDrIK  $ $688 $ $C Jrh㈵>2c|}|jd}d}t|D]}d||zz } | || dz } || z dz} | | } | |dz z } | dkrd}d}nrs......ssssssssl >>>BEK<@qqqqhGK'+vvvvr0000fD******r