statsmodels.stats.proportion.proportion_confint

statsmodels.stats.proportion.proportion_confint(count, nobs, alpha=0.05, method='normal')[source]

Confidence interval for a binomial proportion

Parameters:

  • count ({int or float, array_like}) – number of successes, can be pandas Series or DataFrame. Arrays must contain integer values if method is “binom_test”.

  • nobs ({int or float, array_like}) – total number of trials. Arrays must contain integer values if method is “binom_test”.

  • alpha (float) – Significance level, default 0.05. Must be in (0, 1)

  • method ({"normal", "agresti_coull", "beta", "wilson", "binom_test"}) –

    default: “normal” method to use for confidence interval. Supported methods:

    • normal : asymptotic normal approximation

    • agresti_coull : Agresti-Coull interval

    • beta : Clopper-Pearson interval based on Beta distribution

    • wilson : Wilson Score interval

    • jeffreys : Jeffreys Bayesian Interval

    • binom_test : Numerical inversion of binom_test

Returns:

ci_low, ci_upp – lower and upper confidence level with coverage (approximately) 1-alpha. When a pandas object is returned, then the index is taken from count.

Return type:

{float, ndarray, Series DataFrame}

Notes

Beta, the Clopper-Pearson exact interval has coverage at least 1-alpha, but is in general conservative. Most of the other methods have average coverage equal to 1-alpha, but will have smaller coverage in some cases.

The “beta” and “jeffreys” interval are central, they use alpha/2 in each tail, and alpha is not adjusted at the boundaries. In the extreme case when count is zero or equal to nobs, then the coverage will be only 1 - alpha/2 in the case of “beta”.

The confidence intervals are clipped to be in the [0, 1] interval in the case of “normal” and “agresti_coull”.

Method “binom_test” directly inverts the binomial test in scipy.stats. which has discrete steps.

TODO: binom_test intervals raise an exception in small samples if one

interval bound is close to zero or one.

References