"""Sandwich covariance estimators Created on Sun Nov 27 14:10:57 2011 Author: Josef Perktold Author: Skipper Seabold for HCxxx in linear_model.RegressionResults License: BSD-3 Notes ----- for calculating it, we have two versions version 1: use pinv pinv(x) scale pinv(x) used currently in linear_model, with scale is 1d (or diagonal matrix) (x'x)^(-1) x' scale x (x'x)^(-1), scale in general is (nobs, nobs) so pretty large general formulas for scale in cluster case are in [4], which can be found (as of 2017-05-20) at http://www.tandfonline.com/doi/abs/10.1198/jbes.2010.07136 This paper also has the second version. version 2: (x'x)^(-1) S (x'x)^(-1) with S = x' scale x, S is (kvar,kvars), (x'x)^(-1) is available as normalized_covparams. S = sum (x*u) dot (x*u)' = sum x*u*u'*x' where sum here can aggregate over observations or groups. u is regression residual. x is (nobs, k_var) u is (nobs, 1) x*u is (nobs, k_var) For cluster robust standard errors, we first sum (x*w) over other groups (including time) and then take the dot product (sum of outer products) S = sum_g(x*u)' dot sum_g(x*u) For HAC by clusters, we first sum over groups for each time period, and then use HAC on the group sums of (x*w). If we have several groups, we have to sum first over all relevant groups, and then take the outer product sum. This can be done by summing using indicator functions or matrices or with explicit loops. Alternatively we calculate separate covariance matrices for each group, sum them and subtract the duplicate counted intersection. Not checked in details yet: degrees of freedom or small sample correction factors, see (two) references (?) This is the general case for MLE and GMM also in MLE hessian H, outerproduct of jacobian S, cov_hjjh = HJJH, which reduces to the above in the linear case, but can be used generally, e.g. in discrete, and is misnomed in GenericLikelihoodModel in GMM it's similar but I would have to look up the details, (it comes out in sandwich form by default, it's in the sandbox), standard Newey West or similar are on the covariance matrix of the moment conditions quasi-MLE: MLE with mis-specified model where parameter estimates are fine (consistent ?) but cov_params needs to be adjusted similar or same as in sandwiches. (I did not go through any details yet.) TODO ---- * small sample correction factors, Done for cluster, not yet for HAC * automatic lag-length selection for Newey-West HAC, -> added: nlag = floor[4(T/100)^(2/9)] Reference: xtscc paper, Newey-West note this will not be optimal in the panel context, see Peterson * HAC should maybe return the chosen nlags * get consistent notation, varies by paper, S, scale, sigma? * replace diag(hat_matrix) calculations in cov_hc2, cov_hc3 References ---------- [1] John C. Driscoll and Aart C. Kraay, “Consistent Covariance Matrix Estimation with Spatially Dependent Panel Data,” Review of Economics and Statistics 80, no. 4 (1998): 549-560. [2] Daniel Hoechle, "Robust Standard Errors for Panel Regressions with Cross-Sectional Dependence", The Stata Journal [3] Mitchell A. Petersen, “Estimating Standard Errors in Finance Panel Data Sets: Comparing Approaches,” Review of Financial Studies 22, no. 1 (January 1, 2009): 435 -480. [4] A. Colin Cameron, Jonah B. Gelbach, and Douglas L. Miller, “Robust Inference With Multiway Clustering,” Journal of Business and Economic Statistics 29 (April 2011): 238-249. not used yet: A.C. Cameron, J.B. Gelbach, and D.L. Miller, “Bootstrap-based improvements for inference with clustered errors,” The Review of Economics and Statistics 90, no. 3 (2008): 414–427. """ import numpy as np from statsmodels.tools.grouputils import combine_indices, group_sums from statsmodels.stats.moment_helpers import se_cov __all__ = ['cov_cluster', 'cov_cluster_2groups', 'cov_hac', 'cov_nw_panel', 'cov_white_simple', 'cov_hc0', 'cov_hc1', 'cov_hc2', 'cov_hc3', 'se_cov', 'weights_bartlett', 'weights_uniform'] #----------- from linear_model.RegressionResults ''' HC0_se White's (1980) heteroskedasticity robust standard errors. Defined as sqrt(diag(X.T X)^(-1)X.T diag(e_i^(2)) X(X.T X)^(-1) where e_i = resid[i] HC0_se is a property. It is not evaluated until it is called. When it is called the RegressionResults instance will then have another attribute cov_HC0, which is the full heteroskedasticity consistent covariance matrix and also `het_scale`, which is in this case just resid**2. HCCM matrices are only appropriate for OLS. HC1_se MacKinnon and White's (1985) alternative heteroskedasticity robust standard errors. Defined as sqrt(diag(n/(n-p)*HC_0) HC1_se is a property. It is not evaluated until it is called. When it is called the RegressionResults instance will then have another attribute cov_HC1, which is the full HCCM and also `het_scale`, which is in this case n/(n-p)*resid**2. HCCM matrices are only appropriate for OLS. HC2_se MacKinnon and White's (1985) alternative heteroskedasticity robust standard errors. Defined as (X.T X)^(-1)X.T diag(e_i^(2)/(1-h_ii)) X(X.T X)^(-1) where h_ii = x_i(X.T X)^(-1)x_i.T HC2_se is a property. It is not evaluated until it is called. When it is called the RegressionResults instance will then have another attribute cov_HC2, which is the full HCCM and also `het_scale`, which is in this case is resid^(2)/(1-h_ii). HCCM matrices are only appropriate for OLS. HC3_se MacKinnon and White's (1985) alternative heteroskedasticity robust standard errors. Defined as (X.T X)^(-1)X.T diag(e_i^(2)/(1-h_ii)^(2)) X(X.T X)^(-1) where h_ii = x_i(X.T X)^(-1)x_i.T HC3_se is a property. It is not evaluated until it is called. When it is called the RegressionResults instance will then have another attribute cov_HC3, which is the full HCCM and also `het_scale`, which is in this case is resid^(2)/(1-h_ii)^(2). HCCM matrices are only appropriate for OLS. ''' # Note: HCCM stands for Heteroskedasticity Consistent Covariance Matrix def _HCCM(results, scale): ''' sandwich with pinv(x) * diag(scale) * pinv(x).T where pinv(x) = (X'X)^(-1) X and scale is (nobs,) ''' H = np.dot(results.model.pinv_wexog, scale[:,None]*results.model.pinv_wexog.T) return H def cov_hc0(results): """ See statsmodels.RegressionResults """ het_scale = results.resid**2 # or whitened residuals? only OLS? cov_hc0 = _HCCM(results, het_scale) return cov_hc0 def cov_hc1(results): """ See statsmodels.RegressionResults """ het_scale = results.nobs/(results.df_resid)*(results.resid**2) cov_hc1 = _HCCM(results, het_scale) return cov_hc1 def cov_hc2(results): """ See statsmodels.RegressionResults """ # probably could be optimized h = np.diag(np.dot(results.model.exog, np.dot(results.normalized_cov_params, results.model.exog.T))) het_scale = results.resid**2/(1-h) cov_hc2_ = _HCCM(results, het_scale) return cov_hc2_ def cov_hc3(results): """ See statsmodels.RegressionResults """ # above probably could be optimized to only calc the diag h = np.diag(np.dot(results.model.exog, np.dot(results.normalized_cov_params, results.model.exog.T))) het_scale=(results.resid/(1-h))**2 cov_hc3_ = _HCCM(results, het_scale) return cov_hc3_ #--------------------------------------- def _get_sandwich_arrays(results, cov_type=''): """Helper function to get scores from results Parameters """ if isinstance(results, tuple): # assume we have jac and hessian_inv jac, hessian_inv = results xu = jac = np.asarray(jac) hessian_inv = np.asarray(hessian_inv) elif hasattr(results, 'model'): if hasattr(results, '_results'): # remove wrapper results = results._results # assume we have a results instance if hasattr(results.model, 'jac'): xu = results.model.jac(results.params) hessian_inv = np.linalg.inv(results.model.hessian(results.params)) elif hasattr(results.model, 'score_obs'): xu = results.model.score_obs(results.params) hessian_inv = np.linalg.inv(results.model.hessian(results.params)) else: xu = results.model.wexog * results.wresid[:, None] hessian_inv = np.asarray(results.normalized_cov_params) # experimental support for freq_weights if hasattr(results.model, 'freq_weights') and not cov_type == 'clu': # we do not want to square the weights in the covariance calculations # assumes that freq_weights are incorporated in score_obs or equivalent # assumes xu/score_obs is 2D # temporary asarray xu /= np.sqrt(np.asarray(results.model.freq_weights)[:, None]) else: raise ValueError('need either tuple of (jac, hessian_inv) or results' + 'instance') return xu, hessian_inv def _HCCM1(results, scale): ''' sandwich with pinv(x) * scale * pinv(x).T where pinv(x) = (X'X)^(-1) X and scale is (nobs, nobs), or (nobs,) with diagonal matrix diag(scale) Parameters ---------- results : result instance need to contain regression results, uses results.model.pinv_wexog scale : ndarray (nobs,) or (nobs, nobs) scale matrix, treated as diagonal matrix if scale is one-dimensional Returns ------- H : ndarray (k_vars, k_vars) robust covariance matrix for the parameter estimates ''' if scale.ndim == 1: H = np.dot(results.model.pinv_wexog, scale[:,None]*results.model.pinv_wexog.T) else: H = np.dot(results.model.pinv_wexog, np.dot(scale, results.model.pinv_wexog.T)) return H def _HCCM2(hessian_inv, scale): ''' sandwich with (X'X)^(-1) * scale * (X'X)^(-1) scale is (kvars, kvars) this uses results.normalized_cov_params for (X'X)^(-1) Parameters ---------- results : result instance need to contain regression results, uses results.normalized_cov_params scale : ndarray (k_vars, k_vars) scale matrix Returns ------- H : ndarray (k_vars, k_vars) robust covariance matrix for the parameter estimates ''' if scale.ndim == 1: scale = scale[:,None] xxi = hessian_inv H = np.dot(np.dot(xxi, scale), xxi.T) return H #TODO: other kernels, move ? def weights_bartlett(nlags): '''Bartlett weights for HAC this will be moved to another module Parameters ---------- nlags : int highest lag in the kernel window, this does not include the zero lag Returns ------- kernel : ndarray, (nlags+1,) weights for Bartlett kernel ''' #with lag zero return 1 - np.arange(nlags+1)/(nlags+1.) def weights_uniform(nlags): '''uniform weights for HAC this will be moved to another module Parameters ---------- nlags : int highest lag in the kernel window, this does not include the zero lag Returns ------- kernel : ndarray, (nlags+1,) weights for uniform kernel ''' #with lag zero return np.ones(nlags+1) kernel_dict = {'bartlett': weights_bartlett, 'uniform': weights_uniform} def S_hac_simple(x, nlags=None, weights_func=weights_bartlett): '''inner covariance matrix for HAC (Newey, West) sandwich assumes we have a single time series with zero axis consecutive, equal spaced time periods Parameters ---------- x : ndarray (nobs,) or (nobs, k_var) data, for HAC this is array of x_i * u_i nlags : int or None highest lag to include in kernel window. If None, then nlags = floor(4(T/100)^(2/9)) is used. weights_func : callable weights_func is called with nlags as argument to get the kernel weights. default are Bartlett weights Returns ------- S : ndarray, (k_vars, k_vars) inner covariance matrix for sandwich Notes ----- used by cov_hac_simple options might change when other kernels besides Bartlett are available. ''' if x.ndim == 1: x = x[:,None] n_periods = x.shape[0] if nlags is None: nlags = int(np.floor(4 * (n_periods / 100.)**(2./9.))) weights = weights_func(nlags) S = weights[0] * np.dot(x.T, x) #weights[0] just for completeness, is 1 for lag in range(1, nlags+1): s = np.dot(x[lag:].T, x[:-lag]) S += weights[lag] * (s + s.T) return S def S_white_simple(x): '''inner covariance matrix for White heteroscedastistity sandwich Parameters ---------- x : ndarray (nobs,) or (nobs, k_var) data, for HAC this is array of x_i * u_i Returns ------- S : ndarray, (k_vars, k_vars) inner covariance matrix for sandwich Notes ----- this is just dot(X.T, X) ''' if x.ndim == 1: x = x[:,None] return np.dot(x.T, x) def S_hac_groupsum(x, time, nlags=None, weights_func=weights_bartlett): '''inner covariance matrix for HAC over group sums sandwich This assumes we have complete equal spaced time periods. The number of time periods per group need not be the same, but we need at least one observation for each time period For a single categorical group only, or a everything else but time dimension. This first aggregates x over groups for each time period, then applies HAC on the sum per period. Parameters ---------- x : ndarray (nobs,) or (nobs, k_var) data, for HAC this is array of x_i * u_i time : ndarray, (nobs,) timeindes, assumed to be integers range(n_periods) nlags : int or None highest lag to include in kernel window. If None, then nlags = floor[4(T/100)^(2/9)] is used. weights_func : callable weights_func is called with nlags as argument to get the kernel weights. default are Bartlett weights Returns ------- S : ndarray, (k_vars, k_vars) inner covariance matrix for sandwich References ---------- Daniel Hoechle, xtscc paper Driscoll and Kraay ''' #needs groupsums x_group_sums = group_sums(x, time).T #TODO: transpose return in grou_sum return S_hac_simple(x_group_sums, nlags=nlags, weights_func=weights_func) def S_crosssection(x, group): '''inner covariance matrix for White on group sums sandwich I guess for a single categorical group only, categorical group, can also be the product/intersection of groups This is used by cov_cluster and indirectly verified ''' x_group_sums = group_sums(x, group).T #TODO: why transposed return S_white_simple(x_group_sums) def cov_crosssection_0(results, group): '''this one is still wrong, use cov_cluster instead''' #TODO: currently used version of groupsums requires 2d resid scale = S_crosssection(results.resid[:,None], group) scale = np.squeeze(scale) cov = _HCCM1(results, scale) return cov def cov_cluster(results, group, use_correction=True): '''cluster robust covariance matrix Calculates sandwich covariance matrix for a single cluster, i.e. grouped variables. Parameters ---------- results : result instance result of a regression, uses results.model.exog and results.resid TODO: this should use wexog instead use_correction : bool If true (default), then the small sample correction factor is used. Returns ------- cov : ndarray, (k_vars, k_vars) cluster robust covariance matrix for parameter estimates Notes ----- same result as Stata in UCLA example and same as Peterson ''' #TODO: currently used version of groupsums requires 2d resid xu, hessian_inv = _get_sandwich_arrays(results, cov_type='clu') if not hasattr(group, 'dtype') or group.dtype != np.dtype('int'): clusters, group = np.unique(group, return_inverse=True) else: clusters = np.unique(group) scale = S_crosssection(xu, group) nobs, k_params = xu.shape n_groups = len(clusters) #replace with stored group attributes if available cov_c = _HCCM2(hessian_inv, scale) if use_correction: cov_c *= (n_groups / (n_groups - 1.) * ((nobs-1.) / float(nobs - k_params))) return cov_c def cov_cluster_2groups(results, group, group2=None, use_correction=True): '''cluster robust covariance matrix for two groups/clusters Parameters ---------- results : result instance result of a regression, uses results.model.exog and results.resid TODO: this should use wexog instead use_correction : bool If true (default), then the small sample correction factor is used. Returns ------- cov_both : ndarray, (k_vars, k_vars) cluster robust covariance matrix for parameter estimates, for both clusters cov_0 : ndarray, (k_vars, k_vars) cluster robust covariance matrix for parameter estimates for first cluster cov_1 : ndarray, (k_vars, k_vars) cluster robust covariance matrix for parameter estimates for second cluster Notes ----- verified against Peterson's table, (4 decimal print precision) ''' if group2 is None: if group.ndim !=2 or group.shape[1] != 2: raise ValueError('if group2 is not given, then groups needs to be ' + 'an array with two columns') group0 = group[:, 0] group1 = group[:, 1] else: group0 = group group1 = group2 group = (group0, group1) cov0 = cov_cluster(results, group0, use_correction=use_correction) #[0] because we get still also returns bse cov1 = cov_cluster(results, group1, use_correction=use_correction) # cov of cluster formed by intersection of two groups cov01 = cov_cluster(results, combine_indices(group)[0], use_correction=use_correction) #robust cov matrix for union of groups cov_both = cov0 + cov1 - cov01 #return all three (for now?) return cov_both, cov0, cov1 def cov_white_simple(results, use_correction=True): ''' heteroscedasticity robust covariance matrix (White) Parameters ---------- results : result instance result of a regression, uses results.model.exog and results.resid TODO: this should use wexog instead Returns ------- cov : ndarray, (k_vars, k_vars) heteroscedasticity robust covariance matrix for parameter estimates Notes ----- This produces the same result as cov_hc0, and does not include any small sample correction. verified (against LinearRegressionResults and Peterson) See Also -------- cov_hc1, cov_hc2, cov_hc3 : heteroscedasticity robust covariance matrices with small sample corrections ''' xu, hessian_inv = _get_sandwich_arrays(results) sigma = S_white_simple(xu) cov_w = _HCCM2(hessian_inv, sigma) #add bread to sandwich if use_correction: nobs, k_params = xu.shape cov_w *= nobs / float(nobs - k_params) return cov_w def cov_hac_simple(results, nlags=None, weights_func=weights_bartlett, use_correction=True): ''' heteroscedasticity and autocorrelation robust covariance matrix (Newey-West) Assumes we have a single time series with zero axis consecutive, equal spaced time periods Parameters ---------- results : result instance result of a regression, uses results.model.exog and results.resid TODO: this should use wexog instead nlags : int or None highest lag to include in kernel window. If None, then nlags = floor[4(T/100)^(2/9)] is used. weights_func : callable weights_func is called with nlags as argument to get the kernel weights. default are Bartlett weights Returns ------- cov : ndarray, (k_vars, k_vars) HAC robust covariance matrix for parameter estimates Notes ----- verified only for nlags=0, which is just White just guessing on correction factor, need reference options might change when other kernels besides Bartlett are available. ''' xu, hessian_inv = _get_sandwich_arrays(results) sigma = S_hac_simple(xu, nlags=nlags, weights_func=weights_func) cov_hac = _HCCM2(hessian_inv, sigma) if use_correction: nobs, k_params = xu.shape cov_hac *= nobs / float(nobs - k_params) return cov_hac cov_hac = cov_hac_simple #alias for users #---------------------- use time lags corrected for groups #the following were copied from a different experimental script, #groupidx is tuple, observations assumed to be stacked by group member and #sorted by time, equal number of periods is not required, but equal spacing is. #I think this is pure within group HAC: apply HAC to each group member #separately def lagged_groups(x, lag, groupidx): ''' assumes sorted by time, groupidx is tuple of start and end values not optimized, just to get a working version, loop over groups ''' out0 = [] out_lagged = [] for l,u in groupidx: if l+lag < u: #group is longer than lag out0.append(x[l+lag:u]) out_lagged.append(x[l:u-lag]) if out0 == []: raise ValueError('all groups are empty taking lags') #return out0, out_lagged return np.vstack(out0), np.vstack(out_lagged) def S_nw_panel(xw, weights, groupidx): '''inner covariance matrix for HAC for panel data no denominator nobs used no reference for this, just accounting for time indices ''' nlags = len(weights)-1 S = weights[0] * np.dot(xw.T, xw) #weights just for completeness for lag in range(1, nlags+1): xw0, xwlag = lagged_groups(xw, lag, groupidx) s = np.dot(xw0.T, xwlag) S += weights[lag] * (s + s.T) return S def cov_nw_panel(results, nlags, groupidx, weights_func=weights_bartlett, use_correction='hac'): '''Panel HAC robust covariance matrix Assumes we have a panel of time series with consecutive, equal spaced time periods. Data is assumed to be in long format with time series of each individual stacked into one array. Panel can be unbalanced. Parameters ---------- results : result instance result of a regression, uses results.model.exog and results.resid TODO: this should use wexog instead nlags : int or None Highest lag to include in kernel window. Currently, no default because the optimal length will depend on the number of observations per cross-sectional unit. groupidx : list of tuple each tuple should contain the start and end index for an individual. (groupidx might change in future). weights_func : callable weights_func is called with nlags as argument to get the kernel weights. default are Bartlett weights use_correction : 'cluster' or 'hac' or False If False, then no small sample correction is used. If 'cluster' (default), then the same correction as in cov_cluster is used. If 'hac', then the same correction as in single time series, cov_hac is used. Returns ------- cov : ndarray, (k_vars, k_vars) HAC robust covariance matrix for parameter estimates Notes ----- For nlags=0, this is just White covariance, cov_white. If kernel is uniform, `weights_uniform`, with nlags equal to the number of observations per unit in a balance panel, then cov_cluster and cov_hac_panel are identical. Tested against STATA `newey` command with same defaults. Options might change when other kernels besides Bartlett and uniform are available. ''' if nlags == 0: #so we can reproduce HC0 White weights = [1, 0] #to avoid the scalar check in hac_nw else: weights = weights_func(nlags) xu, hessian_inv = _get_sandwich_arrays(results) S_hac = S_nw_panel(xu, weights, groupidx) cov_hac = _HCCM2(hessian_inv, S_hac) if use_correction: nobs, k_params = xu.shape if use_correction == 'hac': cov_hac *= nobs / float(nobs - k_params) elif use_correction in ['c', 'clu', 'cluster']: n_groups = len(groupidx) cov_hac *= n_groups / (n_groups - 1.) cov_hac *= ((nobs-1.) / float(nobs - k_params)) return cov_hac def cov_nw_groupsum(results, nlags, time, weights_func=weights_bartlett, use_correction=0): '''Driscoll and Kraay Panel robust covariance matrix Robust covariance matrix for panel data of Driscoll and Kraay. Assumes we have a panel of time series where the time index is available. The time index is assumed to represent equal spaced periods. At least one observation per period is required. Parameters ---------- results : result instance result of a regression, uses results.model.exog and results.resid TODO: this should use wexog instead nlags : int or None Highest lag to include in kernel window. Currently, no default because the optimal length will depend on the number of observations per cross-sectional unit. time : ndarray of int this should contain the coding for the time period of each observation. time periods should be integers in range(maxT) where maxT is obs of i weights_func : callable weights_func is called with nlags as argument to get the kernel weights. default are Bartlett weights use_correction : 'cluster' or 'hac' or False If False, then no small sample correction is used. If 'hac' (default), then the same correction as in single time series, cov_hac is used. If 'cluster', then the same correction as in cov_cluster is used. Returns ------- cov : ndarray, (k_vars, k_vars) HAC robust covariance matrix for parameter estimates Notes ----- Tested against STATA xtscc package, which uses no small sample correction This first averages relevant variables for each time period over all individuals/groups, and then applies the same kernel weighted averaging over time as in HAC. Warning: In the example with a short panel (few time periods and many individuals) with mainly across individual variation this estimator did not produce reasonable results. Options might change when other kernels besides Bartlett and uniform are available. References ---------- Daniel Hoechle, xtscc paper Driscoll and Kraay ''' xu, hessian_inv = _get_sandwich_arrays(results) #S_hac = S_nw_panel(xw, weights, groupidx) S_hac = S_hac_groupsum(xu, time, nlags=nlags, weights_func=weights_func) cov_hac = _HCCM2(hessian_inv, S_hac) if use_correction: nobs, k_params = xu.shape if use_correction == 'hac': cov_hac *= nobs / float(nobs - k_params) elif use_correction in ['c', 'cluster']: n_groups = len(np.unique(time)) cov_hac *= n_groups / (n_groups - 1.) cov_hac *= ((nobs-1.) / float(nobs - k_params)) return cov_hac