M/PhRdZddlZddlZddlmZGddZdZdZ dZ dS) a The RegressionFDR class implements the 'Knockoff' approach for controlling false discovery rates (FDR) in regression analysis. The knockoff approach does not require standard errors. Thus one application is to provide inference for parameter estimates that are not smooth functions of the data. For example, the knockoff approach can be used to do inference for parameter estimates obtained from the LASSO, of from stepwise variable selection. The knockoff approach controls FDR for parameter estimates that may be dependent, such as coefficient estimates in a multiple regression model. The knockoff approach is applicable whenever the test statistic can be computed entirely from x'y and x'x, where x is the design matrix and y is the vector of responses. Reference --------- Rina Foygel Barber, Emmanuel Candes (2015). Controlling the False Discovery Rate via Knockoffs. Annals of Statistics 43:5. https://candes.su.domains/publications/downloads/FDR_regression.pdf N)summary2c&eZdZdZddZdZdZdS) RegressionFDRa Control FDR in a regression procedure. Parameters ---------- endog : array_like The dependent variable of the regression exog : array_like The independent variables of the regression regeffects : RegressionEffects instance An instance of a RegressionEffects class that can compute effect sizes for the regression coefficients. method : str The approach used to assess and control FDR, currently must be 'knockoff'. Returns ------- Returns an instance of the RegressionFDR class. The `fdr` attribute holds the estimated false discovery rates. Notes ----- This class Implements the knockoff method of Barber and Candes. This is an approach for controlling the FDR of a variety of regression estimation procedures, including correlation coefficients, OLS regression, OLS with forward selection, and LASSO regression. For other approaches to FDR control in regression, see the statsmodels.stats.multitest module. Methods provided in that module use Z-scores or p-values, and therefore require standard errors for the coefficient estimates to be available. The default method for constructing the augmented design matrix is the 'equivariant' approach, set `design_method='sdp'` to use an alternative approach involving semidefinite programming. See Barber and Candes for more information about both approaches. The sdp approach requires that the cvxopt package be installed. knockoffc :t|dr |j|_n)dt|jdD|_t j|}t j|}d|vrd|d<|j\}}|ddkrt|\}} } n|ddkrt|\}} } |t j |z }||_ t j || fd|_ ||_ | |_|||_t j|jdd \} } } t j| }t%|j|z | z}d||dk<t j| | d dz }||}d ||d k<d|z|z }||z }|| |_|| |_||_| |_t1j|j }|j|d<|j|d<|j|d<||_dS)Ncolumnscg|]}d|zS)zx%d).0js [/var/www/html/test/jupyter/venv/lib/python3.11/site-packages/statsmodels/stats/_knockoff.py z*RegressionFDR.__init__..OsCCC519CCC design_methodequisdpaxisT)return_inverse return_countsright)sider)indexStatzFDR+FDR)hasattrrxnamesrangeshapenpasarray_design_knockoff_equi_design_knockoff_sdpmeanendog concatenateexogexog1exog2statsuniquecumsumlen searchsortedfdrfdrp_ufdr_unqpd DataFramefdr_df)selfr&r( regeffectsmethodkwargsnobsnvarr)r*_unqinvcntccdenomiinumerr1r0dfs r __init__zRegressionFDR.__init__Is. 4 # # D,DKKCCeDJqM.B.BCCCDKz$ 5!! & ( (&,F? #Z d / "f , ,3D99OE5!! O $ - -2488OE5!& NE5>:::   %%d++  $*T04666 S#Ys^^DJ"$s*eai_S3$W 5 5 5 92b1f E U"ems8I   \ , , ,Z6 Y6 H5  rctj|j|kr|j|j|kdStjS)zI Returns the threshold statistic for a given target FDR. r)r!minr2r3inf)r7tfdrs r thresholdzRegressionFDR.thresholds; 6$*   % %9TZ4/03 36Mrctj}|d||j|S)NzRegression FDR results)rSummary add_titleadd_dfr6)r7summs r summaryzRegressionFDR.summarys<!! /000 DK    rN)r)__name__ __module__ __qualname____doc__rFrKrQr rr rrsR''R9999vrrc ddlm}m}n#t$rt dwxYw|j\}}t j|dzd}t j|}||z }t j |j |}|t j | }t j t j |t j |f}||}t j t j| t j|fd} || } d|z} || } t j|\} } t j ||z|f} d| | |z| z| f<|| } d|jd<||| || g| g}t j|d }t j |j |}t)|||}|||fS) zw Use semidefinite programming to construct a knockoff design matrix. Requires cvxopt to be installed. r)solversmatrixz3SDP knockoff designs require installation of cvxoptrrF show_progressx)cvxoptrWrX ImportError ValueErrorr r!sumsqrtdotTonesr'zeroseye diag_indicesoptionsrr"ravel _get_knmat)r(rWrXr;r<xnmSigmach0G0h1ir G1solslxcovexogns r r$r$sP********* PPPNOOOPJD$ &q!  C '#,,C #:D F464 E ~A $7 8 8B B "&,,t 5A > > >B B UB B ?4 DAq 49d# $ $BBqvz1} B',GOO$ ++aR"t , ,C CH   # # % %B 6$&$  D tT2 & &E ?s %c|j\}}|d|zkrd}t|tj|dzd}tj|}||z }tj|j|}tj|\}}tj |}td|zd} | tj |z} t||| } || | fS)a Construct an equivariant design matrix for knockoff analysis. Follows the 'equi-correlated knockoff approach of equation 2.4 in Barber and Candes. Constructs a pair of design matrices exogs, exogn such that exogs is a scaled/centered version of the input matrix exog, exogn is another matrix of the same shape with cov(exogn) = cov(exogs), and the covariances between corresponding columns of exogn and exogs are as small as possible. rYz6The equivariant knockoff can ony be used when n >= 2*prr) r r^r!r_r`rarblinalgeigrHrcri) r(r;r<msgrjrtevr=evminrsrus r r#r#sJD$ af}}Foo &q!  C '#,,C #:D 6$&$  D IMM$  EB F2JJE QuWaB bgdmm B tT2 & &E ?rc6|j\}}tj|}|tj|| z}tj|\}}|||fxxd|zz cc<tj||f}tj|\} } |tj | tj | j |z}tj|\}} tj |d\} } } | tj | z} | j } |dddftj ||j zj } || z tj || z}|S)NrY)sizer)r r!rwr?outerrfrandomnormalqrrarbsvdr`solve)r(rtrsr;r<ashrpr umatur=ashrxcexrus r ririsXJD$ )--  CBHR   C ?4 DAq1IIIRIII 9  $  . .D 9<<  DAqBF1bfQS$'' ( ((Dill4  GD!)--Q''KD"aBGBKKD 6D QQQW+ df55 5 8B 2ItT** *E Lr) rUnumpyr!pandasr4statsmodels.iolibrrr$r#rir rr rs2&&&&&&uuuuuuuup+++\"""Jr