Fix bug in effsize_object, remove unnecessary numba function

Author: Jacobluke-

dabest/_effsize_objects.py

@@ -762,7 +762,7 @@ def __pre_calc(self):
             for ix, tname in enumerate(current_tuple[1:]):
                 if self.__is_paired == "sequential":
                     cname = current_tuple[ix]
-                    control = grouped_data[tname]
+                    control = grouped_data[cname]
                 test = grouped_data[tname]
 
                 result = TwoGroupsEffectSize(

dabest/_stats_tools/confint_1group.py

@@ -135,4 +135,4 @@ def summary_ci_1group(
     }
 
     del B
-    return out
+    return out

dabest/_stats_tools/confint_2group_diff.py

@@ -120,7 +120,7 @@ def _calc_accel(jack_dist):
         return numer / denom
 
 
-@njit(cache=True, parallel=True)
+@njit(cache=True) # parallelization must be turned off for random number generation
 def bootstrap_indices(is_paired, x0_len, x1_len, resamples, random_seed):
     np.random.seed(random_seed)
     indices = np.empty((resamples, x0_len if is_paired else x0_len + x1_len), dtype=np.int64)
@@ -157,7 +157,7 @@ def compute_bootstrapped_diff(
 
     return out
 
-@njit(cache=True, parallel=True)
+@njit(cache=True) # parallelization must be turned off for random number generation
 def delta2_bootstrap_loop(x1, x2, x3, x4, resamples, pooled_sd, rng_seed, is_paired):
     np.random.seed(rng_seed)
     out_delta_g = np.empty(resamples)
@@ -268,7 +268,7 @@ def _compute_alpha_from_ci(ci):
     return (100.0 - ci) / 100.0
 
 
-@njit(cache=True)
+# @njit(cache=True)
 def _compute_quantile(z, bias, acceleration):
     numer = bias + z
     denom = 1 - (acceleration * numer)

dabest/_stats_tools/effsize.py

@@ -8,6 +8,7 @@
 from scipy.special import gamma
 from scipy.stats import mannwhitneyu
 
+
 # %% auto 0
 __all__ = ['two_group_difference', 'func_difference', 'cohens_d', 'cohens_h', 'hedges_g', 'cliffs_delta', 'weighted_delta']
 
@@ -208,6 +209,7 @@ def cohens_d(control:list|tuple|np.ndarray,
     return M / divisor
 
 # %% ../../nbs/API/effsize.ipynb 8
+# @njit(cache=True) # It uses np.seterr which is not supported by Numba
 def cohens_h(control:list|tuple|np.ndarray, 
              test:list|tuple|np.ndarray
             )->float:
@@ -382,4 +384,4 @@ def weighted_delta(difference, group_var):
     '''
 
     weight = np.true_divide(1, group_var)
-    return np.sum(difference*weight)/np.sum(weight)
+    return np.sum(difference*weight)/np.sum(weight)

nbs/API/confint_2group_diff.ipynb

@@ -176,7 +176,7 @@
     "        return numer / denom\n",
     "\n",
     "\n",
-    "@njit(cache=True, parallel=True)\n",
+    "@njit(cache=True) # parallelization must be turned off for random number generation\n",
     "def bootstrap_indices(is_paired, x0_len, x1_len, resamples, random_seed):\n",
     "    np.random.seed(random_seed)\n",
     "    indices = np.empty((resamples, x0_len if is_paired else x0_len + x1_len), dtype=np.int64)\n",
@@ -213,7 +213,7 @@
     "\n",
     "    return out\n",
     "\n",
-    "@njit(cache=True, parallel=True)\n",
+    "@njit(cache=True) # parallelization must be turned off for random number generation\n",
     "def delta2_bootstrap_loop(x1, x2, x3, x4, resamples, pooled_sd, rng_seed, is_paired):\n",
     "    np.random.seed(rng_seed)\n",
     "    out_delta_g = np.empty(resamples)\n",
@@ -324,7 +324,7 @@
     "    return (100.0 - ci) / 100.0\n",
     "\n",
     "\n",
-    "@njit(cache=True)\n",
+    "# @njit(cache=True)\n",
     "def _compute_quantile(z, bias, acceleration):\n",
     "    numer = bias + z\n",
     "    denom = 1 - (acceleration * numer)\n",
@@ -365,37 +365,26 @@
     "    return control_var / control_N + test_var / test_N\n",
     "\n",
     "\n",
-    "@njit(cache=True)\n",
     "def calculate_weighted_delta(group_var, differences):\n",
     "    \"\"\"\n",
     "    Compute the weighted deltas.\n",
     "    \"\"\"\n",
     "\n",
     "    weight = 1 / group_var\n",
     "    denom = np.sum(weight)\n",
-    "    num = np.sum(weight[i] * differences[i] for i in range(0, len(weight)))\n",
+    "    num = 0.0\n",
+    "    for i in range(len(weight)):\n",
+    "        num += weight[i] * differences[i]\n",
     "\n",
     "    return num / denom"
    ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "87e0c164",
-   "metadata": {},
-   "outputs": [],
-   "source": []
   }
  ],
  "metadata": {
   "kernelspec": {
    "display_name": "python3",
    "language": "python",
    "name": "python3"
-  },
-  "language_info": {
-   "name": "python",
-   "version": "3.11.8"
   }
  },
  "nbformat": 4,

nbs/API/effsize.ipynb

@@ -288,7 +288,7 @@
    "outputs": [],
    "source": [
     "#| export\n",
-    "@njit(cache=True)\n",
+    "# @njit(cache=True) # It uses np.seterr which is not supported by Numba\n",
     "def cohens_h(control:list|tuple|np.ndarray, \n",
     "             test:list|tuple|np.ndarray\n",
     "            )->float:\n",
@@ -383,7 +383,7 @@
     "    U, _ = mannwhitneyu(t, c, alternative='two-sided') # Not supported by numba.\n",
     "    cliffs_delta = ((2 * U) / (control_n * test_n)) - 1\n",
     "\n",
-    "    return cliffs_delta"
+    "    return cliffs_delta\n"
    ]
   },
   {
@@ -510,10 +510,6 @@
    "display_name": "python3",
    "language": "python",
    "name": "python3"
-  },
-  "language_info": {
-   "name": "python",
-   "version": "3.11.8"
   }
  },
  "nbformat": 4,

nbs/API/effsize_objects.ipynb

@@ -923,7 +923,7 @@
     "            for ix, tname in enumerate(current_tuple[1:]):\n",
     "                if self.__is_paired == \"sequential\":\n",
     "                    cname = current_tuple[ix]\n",
-    "                    control = grouped_data[tname]\n",
+    "                    control = grouped_data[cname]\n",
     "                test = grouped_data[tname]\n",
     "\n",
     "                result = TwoGroupsEffectSize(\n",