StatisticalDashboard.py 文件源码

def _power_and_ttest(self, control_vals, exp_vals):
    control_mean = statistics.mean(control_vals)
    control_std = statistics.stdev(control_vals)
    exp_mean = statistics.mean(exp_vals)
    exp_std = statistics.stdev(exp_vals)

    pooled_stddev = self._compute_pooled_stddev(
        control_std, exp_std, control_vals, exp_vals)

    power = 0
    percent_diff = None
    if control_mean != 0 and pooled_stddev != 0:
      percent_diff = (control_mean - exp_mean) / float(control_mean)
      effect_size = (abs(percent_diff) * float(control_mean)) / float(pooled_stddev)
      power = smp.TTestIndPower().solve_power(
          effect_size,
          nobs1=len(control_vals),
          ratio=len(exp_vals) / float(len(control_vals)),
          alpha=self.ALPHA_ERROR, alternative='two-sided')

    ttest_result = stats.ttest_ind(control_vals, exp_vals, equal_var=False)
    p_val = ""
    if len(ttest_result) >= 2 and not math.isnan(ttest_result[1]):
      p_val = ttest_result[1]

    mean_diff = exp_mean - control_mean

    if p_val <= self.ALPHA_ERROR and mean_diff < 0:
      significance = "Negative"
    elif p_val <= self.ALPHA_ERROR and mean_diff > 0:
      significance = "Positive"
    else:
      significance = "Neutral"

    return {
        "power": power,
        "p_val": p_val,
        "control_mean": control_mean,
        "mean_diff": mean_diff,
        "percent_diff": 0 if percent_diff is None else percent_diff * -100,
        "significance": significance,
    }