Package 'effectsize'

Title: Indices of Effect Size
Description: Provide utilities to work with indices of effect size for a wide variety of models and hypothesis tests (see list of supported models using the function 'insight::supported_models()'), allowing computation of and conversion between indices such as Cohen's d, r, odds, etc. References: Ben-Shachar et al. (2020) <doi:10.21105/joss.02815>.
Authors: Mattan S. Ben-Shachar [aut, cre] (<https://orcid.org/0000-0002-4287-4801>, @mattansb), Dominique Makowski [aut] (<https://orcid.org/0000-0001-5375-9967>, @Dom_Makowski), Daniel Lüdecke [aut] (<https://orcid.org/0000-0002-8895-3206>, @strengejacke), Indrajeet Patil [aut] (<https://orcid.org/0000-0003-1995-6531>, @patilindrajeets), Brenton M. Wiernik [aut] (<https://orcid.org/0000-0001-9560-6336>, @bmwiernik), Rémi Thériault [aut] (<https://orcid.org/0000-0003-4315-6788>, @rempsyc), Ken Kelley [ctb], David Stanley [ctb], Aaron Caldwell [ctb] , Jessica Burnett [rev] , Johannes Karreth [rev] , Philip Waggoner [aut, ctb]
Maintainer: Mattan S. Ben-Shachar <[email protected]>
License: MIT + file LICENSE
Version: 0.8.9
Built: 2024-11-03 09:26:48 UTC
Source: https://github.com/easystats/effectsize

Help Index


Convert χ2\chi^2 to ϕ\phi and Other Correlation-like Effect Sizes

Description

Convert between χ2\chi^2 (chi-square), ϕ\phi (phi), Cramer's VV, Tschuprow's TT, Cohen's ww, פ (Fei) and Pearson's CC for contingency tables or goodness of fit.

Usage

chisq_to_phi(
  chisq,
  n,
  nrow = 2,
  ncol = 2,
  adjust = TRUE,
  ci = 0.95,
  alternative = "greater",
  ...
)

chisq_to_cohens_w(
  chisq,
  n,
  nrow,
  ncol,
  p,
  ci = 0.95,
  alternative = "greater",
  ...
)

chisq_to_cramers_v(
  chisq,
  n,
  nrow,
  ncol,
  adjust = TRUE,
  ci = 0.95,
  alternative = "greater",
  ...
)

chisq_to_tschuprows_t(
  chisq,
  n,
  nrow,
  ncol,
  adjust = TRUE,
  ci = 0.95,
  alternative = "greater",
  ...
)

chisq_to_fei(chisq, n, nrow, ncol, p, ci = 0.95, alternative = "greater", ...)

chisq_to_pearsons_c(
  chisq,
  n,
  nrow,
  ncol,
  ci = 0.95,
  alternative = "greater",
  ...
)

phi_to_chisq(phi, n, ...)

Arguments

chisq

The χ2\chi^2 (chi-square) statistic.

n

Total sample size.

nrow, ncol

The number of rows/columns in the contingency table.

adjust

Should the effect size be corrected for small-sample bias? Defaults to TRUE; Advisable for small samples and large tables.

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "greater" (default) or "less" (one-sided CI), or "two.sided" (two-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

...

Arguments passed to or from other methods.

p

Vector of expected values. See stats::chisq.test().

phi

The ϕ\phi (phi) statistic.

Details

These functions use the following formulas:

ϕ=w=χ2/n\phi = w = \sqrt{\chi^2 / n}

Cramer’s V=ϕ/min(nrow,ncol)1\textrm{Cramer's } V = \phi / \sqrt{\min(\textit{nrow}, \textit{ncol}) - 1}

Tschuprow’s T=ϕ/(nrow1)×(ncol1)4\textrm{Tschuprow's } T = \phi / \sqrt[4]{(\textit{nrow} - 1) \times (\textit{ncol} - 1)}

פ=ϕ/[1/min(pE)]1פ = \phi / \sqrt{[1 / \min(p_E)] - 1}

Where pEp_E are the expected probabilities.

Pearson’s C=χ2/(χ2+n)\textrm{Pearson's } C = \sqrt{\chi^2 / (\chi^2 + n)}

For versions adjusted for small-sample bias of ϕ\phi, VV, and TT, see Bergsma, 2013.

Value

A data frame with the effect size(s), and confidence interval(s). See cramers_v().

Confidence (Compatibility) Intervals (CIs)

Unless stated otherwise, confidence (compatibility) intervals (CIs) are estimated using the noncentrality parameter method (also called the "pivot method"). This method finds the noncentrality parameter ("ncp") of a noncentral t, F, or χ2\chi^2 distribution that places the observed t, F, or χ2\chi^2 test statistic at the desired probability point of the distribution. For example, if the observed t statistic is 2.0, with 50 degrees of freedom, for which cumulative noncentral t distribution is t = 2.0 the .025 quantile (answer: the noncentral t distribution with ncp = .04)? After estimating these confidence bounds on the ncp, they are converted into the effect size metric to obtain a confidence interval for the effect size (Steiger, 2004).

For additional details on estimation and troubleshooting, see effectsize_CIs.

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

References

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

  • Cumming, G., & Finch, S. (2001). A primer on the understanding, use, and calculation of confidence intervals that are based on central and noncentral distributions. Educational and Psychological Measurement, 61(4), 532-574.

  • Ben-Shachar, M.S., Patil, I., Thériault, R., Wiernik, B.M., Lüdecke, D. (2023). Phi, Fei, Fo, Fum: Effect Sizes for Categorical Data That Use the Chi‑Squared Statistic. Mathematics, 11, 1982. doi:10.3390/math11091982

  • Bergsma, W. (2013). A bias-correction for Cramer's V and Tschuprow's T. Journal of the Korean Statistical Society, 42(3), 323-328.

  • Johnston, J. E., Berry, K. J., & Mielke Jr, P. W. (2006). Measures of effect size for chi-squared and likelihood-ratio goodness-of-fit tests. Perceptual and motor skills, 103(2), 412-414.

  • Rosenberg, M. S. (2010). A generalized formula for converting chi-square tests to effect sizes for meta-analysis. PloS one, 5(4), e10059.

See Also

phi() for more details.

Other effect size from test statistic: F_to_eta2(), t_to_d()

Examples

data("Music_preferences")

# chisq.test(Music_preferences)
#>
#> 	Pearson's Chi-squared test
#>
#> data:  Music_preferences
#> X-squared = 95.508, df = 6, p-value < 2.2e-16
#>

chisq_to_cohens_w(95.508,
  n = sum(Music_preferences),
  nrow = nrow(Music_preferences),
  ncol = ncol(Music_preferences)
)




data("Smoking_FASD")

# chisq.test(Smoking_FASD, p = c(0.015, 0.010, 0.975))
#>
#> 	Chi-squared test for given probabilities
#>
#> data:  Smoking_FASD
#> X-squared = 7.8521, df = 2, p-value = 0.01972

chisq_to_fei(
  7.8521,
  n = sum(Smoking_FASD),
  nrow = 1,
  ncol = 3,
  p = c(0.015, 0.010, 0.975)
)

Cohen's d and Other Standardized Differences

Description

Compute effect size indices for standardized differences: Cohen's d, Hedges' g and Glass’s delta (Δ\Delta). (This function returns the population estimate.) Pair with any reported stats::t.test().

Both Cohen's d and Hedges' g are the estimated the standardized difference between the means of two populations. Hedges' g provides a correction for small-sample bias (using the exact method) to Cohen's d. For sample sizes > 20, the results for both statistics are roughly equivalent. Glass’s delta is appropriate when the standard deviations are significantly different between the populations, as it uses only the second group's standard deviation.

Usage

cohens_d(
  x,
  y = NULL,
  data = NULL,
  pooled_sd = TRUE,
  mu = 0,
  paired = FALSE,
  adjust = FALSE,
  ci = 0.95,
  alternative = "two.sided",
  verbose = TRUE,
  ...
)

hedges_g(
  x,
  y = NULL,
  data = NULL,
  pooled_sd = TRUE,
  mu = 0,
  paired = FALSE,
  ci = 0.95,
  alternative = "two.sided",
  verbose = TRUE,
  ...
)

glass_delta(
  x,
  y = NULL,
  data = NULL,
  mu = 0,
  adjust = TRUE,
  ci = 0.95,
  alternative = "two.sided",
  verbose = TRUE,
  ...
)

Arguments

x, y

A numeric vector, or a character name of one in data. Any missing values (NAs) are dropped from the resulting vector. x can also be a formula (see stats::t.test()), in which case y is ignored.

data

An optional data frame containing the variables.

pooled_sd

If TRUE (default), a sd_pooled() is used (assuming equal variance). Else the mean SD from both groups is used instead.

mu

a number indicating the true value of the mean (or difference in means if you are performing a two sample test).

paired

If TRUE, the values of x and y are considered as paired. This produces an effect size that is equivalent to the one-sample effect size on x - y. See also repeated_measures_d() for more options.

adjust

Should the effect size be adjusted for small-sample bias using Hedges' method? Note that hedges_g() is an alias for cohens_d(adjust = TRUE).

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "two.sided" (default, two-sided CI), "greater" or "less" (one-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

verbose

Toggle warnings and messages on or off.

...

Arguments passed to or from other methods. When x is a formula, these can be subset and na.action.

Details

Set pooled_sd = FALSE for effect sizes that are to accompany a Welch's t-test (Delacre et al, 2021).

Value

A data frame with the effect size ( Cohens_d, Hedges_g, Glass_delta) and their CIs (CI_low and CI_high).

Confidence (Compatibility) Intervals (CIs)

Unless stated otherwise, confidence (compatibility) intervals (CIs) are estimated using the noncentrality parameter method (also called the "pivot method"). This method finds the noncentrality parameter ("ncp") of a noncentral t, F, or χ2\chi^2 distribution that places the observed t, F, or χ2\chi^2 test statistic at the desired probability point of the distribution. For example, if the observed t statistic is 2.0, with 50 degrees of freedom, for which cumulative noncentral t distribution is t = 2.0 the .025 quantile (answer: the noncentral t distribution with ncp = .04)? After estimating these confidence bounds on the ncp, they are converted into the effect size metric to obtain a confidence interval for the effect size (Steiger, 2004).

For additional details on estimation and troubleshooting, see effectsize_CIs.

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

Note

The indices here give the population estimated standardized difference. Some statistical packages give the sample estimate instead (without applying Bessel's correction).

References

  • Algina, J., Keselman, H. J., & Penfield, R. D. (2006). Confidence intervals for an effect size when variances are not equal. Journal of Modern Applied Statistical Methods, 5(1), 2.

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

  • Delacre, M., Lakens, D., Ley, C., Liu, L., & Leys, C. (2021, May 7). Why Hedges’ g*s based on the non-pooled standard deviation should be reported with Welch's t-test. doi:10.31234/osf.io/tu6mp

  • Hedges, L. V. & Olkin, I. (1985). Statistical methods for meta-analysis. Orlando, FL: Academic Press.

  • Hunter, J. E., & Schmidt, F. L. (2004). Methods of meta-analysis: Correcting error and bias in research findings. Sage.

See Also

rm_d(), sd_pooled(), t_to_d(), r_to_d()

Other standardized differences: mahalanobis_d(), means_ratio(), p_superiority(), rank_biserial(), repeated_measures_d()

Examples

data(mtcars)
mtcars$am <- factor(mtcars$am)

# Two Independent Samples ----------

(d <- cohens_d(mpg ~ am, data = mtcars))
# Same as:
# cohens_d("mpg", "am", data = mtcars)
# cohens_d(mtcars$mpg[mtcars$am=="0"], mtcars$mpg[mtcars$am=="1"])

# More options:
cohens_d(mpg ~ am, data = mtcars, pooled_sd = FALSE)
cohens_d(mpg ~ am, data = mtcars, mu = -5)
cohens_d(mpg ~ am, data = mtcars, alternative = "less")
hedges_g(mpg ~ am, data = mtcars)
glass_delta(mpg ~ am, data = mtcars)


# One Sample ----------

cohens_d(wt ~ 1, data = mtcars)

# same as:
# cohens_d("wt", data = mtcars)
# cohens_d(mtcars$wt)

# More options:
cohens_d(wt ~ 1, data = mtcars, mu = 3)
hedges_g(wt ~ 1, data = mtcars, mu = 3)


# Paired Samples ----------

data(sleep)

cohens_d(Pair(extra[group == 1], extra[group == 2]) ~ 1, data = sleep)

# same as:
# cohens_d(sleep$extra[sleep$group == 1], sleep$extra[sleep$group == 2], paired = TRUE)
# cohens_d(sleep$extra[sleep$group == 1] - sleep$extra[sleep$group == 2])
# rm_d(sleep$extra[sleep$group == 1], sleep$extra[sleep$group == 2], method = "z", adjust = FALSE)

# More options:
cohens_d(Pair(extra[group == 1], extra[group == 2]) ~ 1, data = sleep, mu = -1, verbose = FALSE)
hedges_g(Pair(extra[group == 1], extra[group == 2]) ~ 1, data = sleep, verbose = FALSE)


# Interpretation -----------------------
interpret_cohens_d(-1.48, rules = "cohen1988")
interpret_hedges_g(-1.48, rules = "sawilowsky2009")
interpret_glass_delta(-1.48, rules = "gignac2016")
# Or:
interpret(d, rules = "sawilowsky2009")

# Common Language Effect Sizes
d_to_u3(1.48)
# Or:
print(d, append_CLES = TRUE)

Effect Size for Paired Contingency Tables

Description

Cohen's g is an effect size of asymmetry (or marginal heterogeneity) for dependent (paired) contingency tables ranging between 0 (perfect symmetry) and 0.5 (perfect asymmetry) (see stats::mcnemar.test()). (Note this is not not a measure of (dis)agreement between the pairs, but of (a)symmetry.)

Usage

cohens_g(x, y = NULL, ci = 0.95, alternative = "two.sided", ...)

Arguments

x

a numeric vector or matrix. x and y can also both be factors.

y

a numeric vector; ignored if x is a matrix. If x is a factor, y should be a factor of the same length.

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "two.sided" (default, two-sided CI), "greater" or "less" (one-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

...

Ignored

Value

A data frame with the effect size (Cohens_g, Risk_ratio (possibly with the prefix log_), Cohens_h) and its CIs (CI_low and CI_high).

Confidence (Compatibility) Intervals (CIs)

Confidence intervals are based on the proportion (P=g+0.5P = g + 0.5) confidence intervals returned by stats::prop.test() (minus 0.5), which give a good close approximation.

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

References

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

See Also

Other effect sizes for contingency table: oddsratio(), phi()

Examples

data("screening_test")

phi(screening_test$Diagnosis, screening_test$Test1)

phi(screening_test$Diagnosis, screening_test$Test2)

# Both tests seem comparable - but are the tests actually different?

(tests <- table(Test1 = screening_test$Test1, Test2 = screening_test$Test2))

mcnemar.test(tests)

cohens_g(tests)

# Test 2 gives a negative result more than test 1!

Convert Between d, r, and Odds Ratio

Description

Enables a conversion between different indices of effect size, such as standardized difference (Cohen's d), (point-biserial) correlation r or (log) odds ratios.

Usage

d_to_r(d, n1, n2, ...)

r_to_d(r, n1, n2, ...)

oddsratio_to_d(OR, log = FALSE, ...)

logoddsratio_to_d(logOR, log = TRUE, ...)

d_to_oddsratio(d, log = FALSE, ...)

d_to_logoddsratio(d, log = TRUE, ...)

oddsratio_to_r(OR, n1, n2, log = FALSE, ...)

logoddsratio_to_r(logOR, log = TRUE, ...)

r_to_oddsratio(r, n1, n2, log = FALSE, ...)

r_to_logoddsratio(r, n1, n2, log = TRUE, ...)

Arguments

d, r, OR, logOR

Standardized difference value (Cohen's d), correlation coefficient (r), Odds ratio, or logged Odds ratio.

n1, n2

Group sample sizes. If either is missing, groups are assumed to be of equal size.

...

Arguments passed to or from other methods.

log

Take in or output the log of the ratio (such as in logistic models), e.g. when the desired input or output are log odds ratios instead odds ratios.

Details

Conversions between d and OR is done through these formulae:

  • d=log(OR)×3πd = \frac{\log(OR)\times\sqrt{3}}{\pi}

  • log(OR)=dπ(3)log(OR) = d * \frac{\pi}{\sqrt(3)}

Converting between d and r is done through these formulae:

  • d=hr1r2d = \frac{\sqrt{h} * r}{\sqrt{1 - r^2}}

  • r=dd2+hr = \frac{d}{\sqrt{d^2 + h}}

Where h=n1+n22n1+n1+n22n2h = \frac{n_1 + n_2 - 2}{n_1} + \frac{n_1 + n_2 - 2}{n_2}. When groups are of equal size, h reduces to approximately 4. The resulting r is also called the binomial effect size display (BESD; Rosenthal et al., 1982).

Value

Converted index.

References

  • Borenstein, M., Hedges, L. V., Higgins, J. P. T., & Rothstein, H. R. (2009). Converting among effect sizes. Introduction to meta-analysis, 45-49.

  • Jacobs, P., & Viechtbauer, W. (2017). Estimation of the biserial correlation and its sampling variance for use in meta-analysis. Research synthesis methods, 8(2), 161-180. doi:10.1002/jrsm.1218

  • Rosenthal, R., & Rubin, D. B. (1982). A simple, general purpose display of magnitude of experimental effect. Journal of educational psychology, 74(2), 166.

  • Sánchez-Meca, J., Marín-Martínez, F., & Chacón-Moscoso, S. (2003). Effect-size indices for dichotomized outcomes in meta-analysis. Psychological methods, 8(4), 448.

See Also

cohens_d()

Other convert between effect sizes: diff_to_cles, eta2_to_f2(), odds_to_probs(), oddsratio_to_riskratio(), w_to_fei()

Examples

r_to_d(0.5)
d_to_oddsratio(1.154701)
oddsratio_to_r(8.120534)

d_to_r(1)
r_to_oddsratio(0.4472136, log = TRUE)
oddsratio_to_d(1.813799, log = TRUE)

Convert Standardized Differences to Common Language Effect Sizes

Description

Convert Standardized Differences to Common Language Effect Sizes

Usage

d_to_p_superiority(d)

rb_to_p_superiority(rb)

rb_to_vda(rb)

d_to_u2(d)

d_to_u1(d)

d_to_u3(d)

d_to_overlap(d)

rb_to_wmw_odds(rb)

Arguments

d, rb

A numeric vector of Cohen's d / rank-biserial correlation or the output from cohens_d() / rank_biserial().

Details

This function use the following formulae for Cohen's d:

Pr(superiority)=Φ(d/2)Pr(superiority) = \Phi(d/\sqrt{2})


Cohen’s U3=Φ(d)\textrm{Cohen's } U_3 = \Phi(d)


Cohen’s U2=Φ(d/2)\textrm{Cohen's } U_2 = \Phi(|d|/2)


Cohen’s U1=(2×U21)/U2\textrm{Cohen's } U_1 = (2\times U_2 - 1)/U_2


Overlap=2×Φ(d/2)Overlap = 2 \times \Phi(-|d|/2)


And the following for the rank-biserial correlation:

Pr(superiority)=(rrb+1)/2Pr(superiority) = (r_{rb} + 1)/2


WMWOdds=Pr(superiority)/(1Pr(superiority))WMW_{Odds} = Pr(superiority) / (1 - Pr(superiority))

Value

A list of ⁠Cohen's U3⁠, Overlap, Pr(superiority), a numeric vector of Pr(superiority), or a data frame, depending on the input.

Note

For d, these calculations assume that the populations have equal variance and are normally distributed.

Vargha and Delaney's A is an alias for the non-parametric probability of superiority.

References

  • Cohen, J. (1977). Statistical power analysis for the behavioral sciences. New York: Routledge.

  • Reiser, B., & Faraggi, D. (1999). Confidence intervals for the overlapping coefficient: the normal equal variance case. Journal of the Royal Statistical Society, 48(3), 413-418.

  • Ruscio, J. (2008). A probability-based measure of effect size: robustness to base rates and other factors. Psychological methods, 13(1), 19–30.

See Also

cohens_u3() for descriptions of the effect sizes (also, cohens_d(), rank_biserial()).

Other convert between effect sizes: d_to_r(), eta2_to_f2(), odds_to_probs(), oddsratio_to_riskratio(), w_to_fei()


effectsize API

Description

Read the Support functions for model extensions vignette.

Usage

.es_aov_simple(
  aov_table,
  type = c("eta", "omega", "epsilon"),
  partial = TRUE,
  generalized = FALSE,
  include_intercept = FALSE,
  ci = 0.95,
  alternative = "greater",
  verbose = TRUE
)

.es_aov_strata(
  aov_table,
  DV_names,
  type = c("eta", "omega", "epsilon"),
  partial = TRUE,
  generalized = FALSE,
  include_intercept = FALSE,
  ci = 0.95,
  alternative = "greater",
  verbose = TRUE
)

.es_aov_table(
  aov_table,
  type = c("eta", "omega", "epsilon"),
  partial = TRUE,
  generalized = FALSE,
  include_intercept = FALSE,
  ci = 0.95,
  alternative = "greater",
  verbose = TRUE
)

Arguments

aov_table

Input data frame

type

Which effect size to compute?

partial, generalized, ci, alternative, verbose

See eta_squared().

include_intercept

Should the intercept ((Intercept)) be included?

DV_names

A character vector with the names of all the predictors, including the grouping variable (e.g., "Subject").


Confidence (Compatibility) Intervals

Description

More information regarding Confidence (Compatibiity) Intervals and how they are computed in effectsize.

Confidence (Compatibility) Intervals (CIs)

Unless stated otherwise, confidence (compatibility) intervals (CIs) are estimated using the noncentrality parameter method (also called the "pivot method"). This method finds the noncentrality parameter ("ncp") of a noncentral t, F, or χ2\chi^2 distribution that places the observed t, F, or χ2\chi^2 test statistic at the desired probability point of the distribution. For example, if the observed t statistic is 2.0, with 50 degrees of freedom, for which cumulative noncentral t distribution is t = 2.0 the .025 quantile (answer: the noncentral t distribution with ncp = .04)? After estimating these confidence bounds on the ncp, they are converted into the effect size metric to obtain a confidence interval for the effect size (Steiger, 2004).

For additional details on estimation and troubleshooting, see effectsize_CIs.

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Bootstrapped CIs

Some effect sizes are directionless–they do have a minimum value that would be interpreted as "no effect", but they cannot cross it. For example, a null value of Kendall's W is 0, indicating no difference between groups, but it can never have a negative value. Same goes for U2 and Overlap: the null value of U2U_2 is 0.5, but it can never be smaller than 0.5; am Overlap of 1 means "full overlap" (no difference), but it cannot be larger than 1.

When bootstrapping CIs for such effect sizes, the bounds of the CIs will never cross (and often will never cover) the null. Therefore, these CIs should not be used for statistical inference.

One-Sided CIs

Typically, CIs are constructed as two-tailed intervals, with an equal proportion of the cumulative probability distribution above and below the interval. CIs can also be constructed as one-sided intervals, giving only a lower bound or upper bound. This is analogous to computing a 1-tailed p value or conducting a 1-tailed hypothesis test.

Significance tests conducted using CIs (whether a value is inside the interval) and using p values (whether p < alpha for that value) are only guaranteed to agree when both are constructed using the same number of sides/tails.

Most effect sizes are not bounded by zero (e.g., r, d, g), and as such are generally tested using 2-tailed tests and 2-sided CIs.

Some effect sizes are strictly positive–they do have a minimum value, of 0. For example, R2R^2, η2\eta^2, sr2sr^2, and other variance-accounted-for effect sizes, as well as Cramer's V and multiple R, range from 0 to 1. These typically involve F- or χ2\chi^2-statistics and are generally tested using 1-tailed tests which test whether the estimated effect size is larger than the hypothesized null value (e.g., 0). In order for a CI to yield the same significance decision it must then by a 1-sided CI, estimating only a lower bound. This is the default CI computed by effectsize for these effect sizes, where alternative = "greater" is set.

This lower bound interval indicates the smallest effect size that is not significantly different from the observed effect size. That is, it is the minimum effect size compatible with the observed data, background model assumptions, and α\alpha level. This type of interval does not indicate a maximum effect size value; anything up to the maximum possible value of the effect size (e.g., 1) is in the interval.

One-sided CIs can also be used to test against a maximum effect size value (e.g., is R2R^2 significantly smaller than a perfect correlation of 1.0?) by setting alternative = "less". This estimates a CI with only an upper bound; anything from the minimum possible value of the effect size (e.g., 0) up to this upper bound is in the interval.

We can also obtain a 2-sided interval by setting alternative = "two.sided". These intervals can be interpreted in the same way as other 2-sided intervals, such as those for r, d, or g.

An alternative approach to aligning significance tests using CIs and 1-tailed p values that can often be found in the literature is to construct a 2-sided CI at a lower confidence level (e.g., 100(1-2α\alpha)% = 100 - 2*5% = 90%. This estimates the lower bound and upper bound for the above 1-sided intervals simultaneously. These intervals are commonly reported when conducting equivalence tests. For example, a 90% 2-sided interval gives the bounds for an equivalence test with α\alpha = .05. However, be aware that this interval does not give 95% coverage for the underlying effect size parameter value. For that, construct a 95% 2-sided CI.

data("hardlyworking")
fit <- lm(salary ~ n_comps, data = hardlyworking)
eta_squared(fit) # default, ci = 0.95, alternative = "greater"
#> For one-way between subjects designs, partial eta squared is equivalent
#>   to eta squared. Returning eta squared.
#> # Effect Size for ANOVA
#> 
#> Parameter | Eta2 |       95% CI
#> -------------------------------
#> n_comps   | 0.19 | [0.14, 1.00]
#> 
#> - One-sided CIs: upper bound fixed at [1.00].
eta_squared(fit, alternative = "less") # Test is eta is smaller than some value
#> For one-way between subjects designs, partial eta squared is equivalent
#>   to eta squared. Returning eta squared.
#> # Effect Size for ANOVA
#> 
#> Parameter | Eta2 |       95% CI
#> -------------------------------
#> n_comps   | 0.19 | [0.00, 0.24]
#> 
#> - One-sided CIs: lower bound fixed at [0.00].
eta_squared(fit, alternative = "two.sided") # 2-sided bounds for alpha = .05
#> For one-way between subjects designs, partial eta squared is equivalent
#>   to eta squared. Returning eta squared.
#> # Effect Size for ANOVA
#> 
#> Parameter | Eta2 |       95% CI
#> -------------------------------
#> n_comps   | 0.19 | [0.14, 0.25]
eta_squared(fit, ci = 0.9, alternative = "two.sided") # both 1-sided bounds for alpha = .05
#> For one-way between subjects designs, partial eta squared is equivalent
#>   to eta squared. Returning eta squared.
#> # Effect Size for ANOVA
#> 
#> Parameter | Eta2 |       90% CI
#> -------------------------------
#> n_comps   | 0.19 | [0.14, 0.24]

CI Does Not Contain the Estimate

For very large sample sizes or effect sizes, the width of the CI can be smaller than the tolerance of the optimizer, resulting in CIs of width 0. This can also result in the estimated CIs excluding the point estimate.

In these cases, consider an alternative method for computing CIs, such as the bootstrap.

References

Bauer, P., & Kieser, M. (1996). A unifying approach for confidence intervals and testing of equivalence and difference. Biometrika, 83(4), 934-–937. doi:10.1093/biomet/83.4.934

Rafi, Z., & Greenland, S. (2020). Semantic and cognitive tools to aid statistical science: Replace confidence and significance by compatibility and surprise. BMC Medical Research Methodology, 20(1), Article 244. doi:10.1186/s12874-020-01105-9

Schweder, T., & Hjort, N. L. (2016). Confidence, likelihood, probability: Statistical inference with confidence distributions. Cambridge University Press. doi:10.1017/CBO9781139046671

Steiger, J. H. (2004). Beyond the F test: Effect size confidence intervals and tests of close fit in the analysis of variance and contrast analysis. Psychological Methods, 9(2), 164–182. doi:10.1037/1082-989x.9.2.164

Xie, M., & Singh, K. (2013). Confidence distribution, the frequentist distribution estimator of a parameter: A review. International Statistical Review, 81(1), 3–-39. doi:10.1111/insr.12000


Deprecated / Defunct Functions

Description

Deprecated / Defunct Functions

Usage

convert_odds_to_probs(...)

convert_probs_to_odds(...)

convert_d_to_r(...)

convert_r_to_d(...)

convert_oddsratio_to_d(...)

convert_d_to_oddsratio(...)

convert_oddsratio_to_r(...)

convert_r_to_oddsratio(...)

interpret_d(...)

interpret_g(...)

interpret_delta(...)

interpret_parameters(...)

normalized_chi(...)

chisq_to_normalized(...)

convert_d_to_common_language(...)

d_to_common_language(...)

convert_rb_to_common_language(...)

rb_to_common_language(...)

common_language(...)

Arguments

...

Arguments to the deprecated function.


effectsize options

Description

Currently, the following global options are supported:

  • es.use_symbols logical: Should proper symbols be printed (TRUE) instead of transliterated effect size names (FALSE; default).


Effect Sizes

Description

This function tries to return the best effect-size measure for the provided input model. See details.

Usage

## S3 method for class 'BFBayesFactor'
effectsize(model, type = NULL, ci = 0.95, test = NULL, verbose = TRUE, ...)

effectsize(model, ...)

## S3 method for class 'aov'
effectsize(model, type = NULL, ...)

## S3 method for class 'htest'
effectsize(model, type = NULL, verbose = TRUE, ...)

Arguments

model

An object of class htest, or a statistical model. See details.

type

The effect size of interest. See details.

ci

Value or vector of probability of the CI (between 0 and 1) to be estimated. Default to 0.95 (⁠95%⁠).

test

The indices of effect existence to compute. Character (vector) or list with one or more of these options: "p_direction" (or "pd"), "rope", "p_map", "equivalence_test" (or "equitest"), "bayesfactor" (or "bf") or "all" to compute all tests. For each "test", the corresponding bayestestR function is called (e.g. rope() or p_direction()) and its results included in the summary output.

verbose

Toggle off warnings.

...

Arguments passed to or from other methods. See details.

Details

  • For an object of class htest, data is extracted via insight::get_data(), and passed to the relevant function according to:

    • A t-test depending on type: "cohens_d" (default), "hedges_g", or one of "p_superiority", "u1", "u2", "u3", "overlap".

      • For a Paired t-test: depending on type: "rm_rm", "rm_av", "rm_b", "rm_d", "rm_z".

    • A Chi-squared tests of independence or Fisher's Exact Test, depending on type: "cramers_v" (default), "tschuprows_t", "phi", "cohens_w", "pearsons_c", "cohens_h", "oddsratio", "riskratio", "arr", or "nnt".

    • A Chi-squared tests of goodness-of-fit, depending on type: "fei" (default) "cohens_w", "pearsons_c"

    • A One-way ANOVA test, depending on type: "eta" (default), "omega" or "epsilon" -squared, "f", or "f2".

    • A McNemar test returns Cohen's g.

    • A Wilcoxon test depending on type: returns "rank_biserial" correlation (default) or one of "p_superiority", "vda", "u2", "u3", "overlap".

    • A Kruskal-Wallis test depending on type: "epsilon" (default) or "eta".

    • A Friedman test returns Kendall's W. (Where applicable, ci and alternative are taken from the htest if not otherwise provided.)

  • For an object of class BFBayesFactor, using bayestestR::describe_posterior(),

    • A t-test depending on type: "cohens_d" (default) or one of "p_superiority", "u1", "u2", "u3", "overlap".

    • A correlation test returns r.

    • A contingency table test, depending on type: "cramers_v" (default), "phi", "tschuprows_t", "cohens_w", "pearsons_c", "cohens_h", "oddsratio", or "riskratio", "arr", or "nnt".

    • A proportion test returns p.

  • Objects of class anova, aov, aovlist or afex_aov, depending on type: "eta" (default), "omega" or "epsilon" -squared, "f", or "f2".

  • Other objects are passed to parameters::standardize_parameters().

For statistical models it is recommended to directly use the listed functions, for the full range of options they provide.

Value

A data frame with the effect size (depending on input) and and its CIs (CI_low and CI_high).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

See Also

vignette(package = "effectsize")

Examples

## Hypothesis Testing
## ------------------
data("Music_preferences")
Xsq <- chisq.test(Music_preferences)
effectsize(Xsq)
effectsize(Xsq, type = "cohens_w")

Tt <- t.test(1:10, y = c(7:20), alternative = "less")
effectsize(Tt)

Tt <- t.test(
  x = c(1.83, 0.50, 1.62, 2.48, 1.68, 1.88, 1.55, 3.06, 1.30),
  y = c(0.878, 0.647, 0.598, 2.05, 1.06, 1.29, 1.06, 3.14, 1.29),
  paired = TRUE
)
effectsize(Tt, type = "rm_b")

Aov <- oneway.test(extra ~ group, data = sleep, var.equal = TRUE)
effectsize(Aov)
effectsize(Aov, type = "omega")

Wt <- wilcox.test(1:10, 7:20, mu = -3, alternative = "less", exact = FALSE)
effectsize(Wt)
effectsize(Wt, type = "u2")

## Models and Anova Tables
## -----------------------
fit <- lm(mpg ~ factor(cyl) * wt + hp, data = mtcars)
effectsize(fit, method = "basic")

anova_table <- anova(fit)
effectsize(anova_table)
effectsize(anova_table, type = "epsilon")


## Bayesian Hypothesis Testing
## ---------------------------
bf_prop <- BayesFactor::proportionBF(3, 7, p = 0.3)
effectsize(bf_prop)

bf_corr <- BayesFactor::correlationBF(attitude$rating, attitude$complaints)
effectsize(bf_corr)

data(RCT_table)
bf_xtab <- BayesFactor::contingencyTableBF(RCT_table, sampleType = "poisson", fixedMargin = "cols")
effectsize(bf_xtab)
effectsize(bf_xtab, type = "oddsratio")
effectsize(bf_xtab, type = "arr")

bf_ttest <- BayesFactor::ttestBF(sleep$extra[sleep$group == 1],
  sleep$extra[sleep$group == 2],
  paired = TRUE, mu = -1
)
effectsize(bf_ttest)

Test Effect Size for Practical Equivalence to the Null

Description

Perform a Test for Practical Equivalence for indices of effect size.

Usage

## S3 method for class 'effectsize_table'
equivalence_test(
  x,
  range = "default",
  rule = c("classic", "cet", "bayes"),
  ...
)

Arguments

x

An effect size table, such as returned by cohens_d(), eta_squared(), F_to_r(), etc.

range

The range of practical equivalence of an effect. For one-sides CIs, a single value can be proved for the lower / upper bound to test against (but see more details below). For two-sided CIs, a single value is duplicated to c(-range, range). If "default", will be set to ⁠[-.1, .1]⁠.

rule

How should acceptance and rejection be decided? See details.

...

Arguments passed to or from other methods.

Details

The CIs used in the equivalence test are the ones in the provided effect size table. For results equivalent (ha!) to those that can be obtained using the TOST approach (e.g., Lakens, 2017), appropriate CIs should be extracted using the function used to make the effect size table (cohens_d, eta_squared, F_to_r, etc), with alternative = "two.sided". See examples.

The Different Rules

  • "classic" - the classic method:

    • If the CI is completely within the ROPE - Accept H0

    • Else, if the CI does not contain 0 - Reject H0

    • Else - Undecided

  • "cet" - conditional equivalence testing:

    • If the CI does not contain 0 - Reject H0

    • Else, If the CI is completely within the ROPE - Accept H0

    • Else - Undecided

  • "bayes" - The Bayesian approach, as put forth by Kruschke:

    • If the CI does is completely outside the ROPE - Reject H0

    • Else, If the CI is completely within the ROPE - Accept H0

    • Else - Undecided

Value

A data frame with the results of the equivalence test.

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

References

  • Campbell, H., & Gustafson, P. (2018). Conditional equivalence testing: An alternative remedy for publication bias. PLOS ONE, 13(4), e0195145. doi:10.1371/journal.pone.0195145

  • Kruschke, J. K. (2014). Doing Bayesian data analysis: A tutorial with R, JAGS, and Stan. Academic Press

  • Kruschke, J. K. (2018). Rejecting or accepting parameter values in Bayesian estimation. Advances in Methods and Practices in Psychological Science, 1(2), 270-280. doi:10.1177/2515245918771304

  • Lakens, D. (2017). Equivalence Tests: A Practical Primer for t Tests, Correlations, and Meta-Analyses. Social Psychological and Personality Science, 8(4), 355–362. doi:10.1177/1948550617697177

See Also

For more details, see bayestestR::equivalence_test().

Examples

data("hardlyworking")
model <- aov(salary ~ age + factor(n_comps) * cut(seniority, 3), data = hardlyworking)
es <- eta_squared(model, ci = 0.9, alternative = "two.sided")
equivalence_test(es, range = c(0, 0.15)) # TOST

data("RCT_table")
OR <- oddsratio(RCT_table, alternative = "greater")
equivalence_test(OR, range = c(0, 1))

ds <- t_to_d(
  t = c(0.45, -0.65, 7, -2.2, 2.25),
  df_error = c(675, 525, 2000, 900, 1875),
  ci = 0.9, alternative = "two.sided" # TOST
)
# Can also plot
if (require(see)) plot(equivalence_test(ds, range = 0.2))
if (require(see)) plot(equivalence_test(ds, range = 0.2, rule = "cet"))
if (require(see)) plot(equivalence_test(ds, range = 0.2, rule = "bayes"))

η2\eta^2 and Other Effect Size for ANOVA

Description

Functions to compute effect size measures for ANOVAs, such as Eta- (η\eta), Omega- (ω\omega) and Epsilon- (ϵ\epsilon) squared, and Cohen's f (or their partialled versions) for ANOVA tables. These indices represent an estimate of how much variance in the response variables is accounted for by the explanatory variable(s).

When passing models, effect sizes are computed using the sums of squares obtained from anova(model) which might not always be appropriate. See details.

Usage

eta_squared(
  model,
  partial = TRUE,
  generalized = FALSE,
  ci = 0.95,
  alternative = "greater",
  verbose = TRUE,
  ...
)

omega_squared(
  model,
  partial = TRUE,
  ci = 0.95,
  alternative = "greater",
  verbose = TRUE,
  ...
)

epsilon_squared(
  model,
  partial = TRUE,
  ci = 0.95,
  alternative = "greater",
  verbose = TRUE,
  ...
)

cohens_f(
  model,
  partial = TRUE,
  generalized = FALSE,
  squared = FALSE,
  method = c("eta", "omega", "epsilon"),
  model2 = NULL,
  ci = 0.95,
  alternative = "greater",
  verbose = TRUE,
  ...
)

cohens_f_squared(
  model,
  partial = TRUE,
  generalized = FALSE,
  squared = TRUE,
  method = c("eta", "omega", "epsilon"),
  model2 = NULL,
  ci = 0.95,
  alternative = "greater",
  verbose = TRUE,
  ...
)

eta_squared_posterior(
  model,
  partial = TRUE,
  generalized = FALSE,
  ss_function = stats::anova,
  draws = 500,
  verbose = TRUE,
  ...
)

Arguments

model

An ANOVA table (or an ANOVA-like table, e.g., outputs from parameters::model_parameters), or a statistical model for which such a table can be extracted. See details.

partial

If TRUE, return partial indices.

generalized

A character vector of observed (non-manipulated) variables to be used in the estimation of a generalized Eta Squared. Can also be TRUE, in which case generalized Eta Squared is estimated assuming none of the variables are observed (all are manipulated). (For afex_aov models, when TRUE, the observed variables are extracted automatically from the fitted model, if they were provided during fitting.

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "greater" (default) or "less" (one-sided CI), or "two.sided" (two-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

verbose

Toggle warnings and messages on or off.

...

Arguments passed to or from other methods.

  • Can be include_intercept = TRUE to include the effect size for the intercept (when it is included in the ANOVA table).

  • For Bayesian models, arguments passed to ss_function.

squared

Return Cohen's f or Cohen's f-squared?

method

What effect size should be used as the basis for Cohen's f?

model2

Optional second model for Cohen's f (/squared). If specified, returns the effect size for R-squared-change between the two models.

ss_function

For Bayesian models, the function used to extract sum-of-squares. Uses anova() by default, but can also be car::Anova() for simple linear models.

draws

For Bayesian models, an integer indicating the number of draws from the posterior predictive distribution to return. Larger numbers take longer to run, but provide estimates that are more stable.

Details

For aov (or lm), aovlist and afex_aov models, and for anova objects that provide Sums-of-Squares, the effect sizes are computed directly using Sums-of-Squares. (For maov (or mlm) models, effect sizes are computed for each response separately.)

For other ANOVA tables and models (converted to ANOVA-like tables via anova() methods), effect sizes are approximated via test statistic conversion of the omnibus F statistic provided by the (see F_to_eta2() for more details.)

Type of Sums of Squares

When model is a statistical model, the sums of squares (or F statistics) used for the computation of the effect sizes are based on those returned by anova(model). Different models have different default output type. For example, for aov and aovlist these are type-1 sums of squares, but for lmerMod (and lmerModLmerTest) these are type-3 sums of squares. Make sure these are the sums of squares you are interested in. You might want to convert your model to an ANOVA(-like) table yourself and then pass the result to eta_squared(). See examples below for use of car::Anova() and the afex package.

For type 3 sum of squares, it is generally recommended to fit models with orthogonal factor weights (e.g., contr.sum) and centered covariates, for sensible results. See examples and the afex package.

Un-Biased Estimate of Eta

Both Omega and Epsilon are unbiased estimators of the population's Eta, which is especially important is small samples. But which to choose?

Though Omega is the more popular choice (Albers and Lakens, 2018), Epsilon is analogous to adjusted R2 (Allen, 2017, p. 382), and has been found to be less biased (Carroll & Nordholm, 1975).

Cohen's f

Cohen's f can take on values between zero, when the population means are all equal, and an indefinitely large number as standard deviation of means increases relative to the average standard deviation within each group.

When comparing two models in a sequential regression analysis, Cohen's f for R-square change is the ratio between the increase in R-square and the percent of unexplained variance.

Cohen has suggested that the values of 0.10, 0.25, and 0.40 represent small, medium, and large effect sizes, respectively.

Eta Squared from Posterior Predictive Distribution

For Bayesian models (fit with brms or rstanarm), eta_squared_posterior() simulates data from the posterior predictive distribution (ppd) and for each simulation the Eta Squared is computed for the model's fixed effects. This means that the returned values are the population level effect size as implied by the posterior model (and not the effect size in the sample data). See rstantools::posterior_predict() for more info.

Value

A data frame with the effect size(s) between 0-1 (Eta2, Epsilon2, Omega2, Cohens_f or Cohens_f2, possibly with the partial or generalized suffix), and their CIs (CI_low and CI_high).

For eta_squared_posterior(), a data frame containing the ppd of the Eta squared for each fixed effect, which can then be passed to bayestestR::describe_posterior() for summary stats.

A data frame containing the effect size values and their confidence intervals.

Confidence (Compatibility) Intervals (CIs)

Unless stated otherwise, confidence (compatibility) intervals (CIs) are estimated using the noncentrality parameter method (also called the "pivot method"). This method finds the noncentrality parameter ("ncp") of a noncentral t, F, or χ2\chi^2 distribution that places the observed t, F, or χ2\chi^2 test statistic at the desired probability point of the distribution. For example, if the observed t statistic is 2.0, with 50 degrees of freedom, for which cumulative noncentral t distribution is t = 2.0 the .025 quantile (answer: the noncentral t distribution with ncp = .04)? After estimating these confidence bounds on the ncp, they are converted into the effect size metric to obtain a confidence interval for the effect size (Steiger, 2004).

For additional details on estimation and troubleshooting, see effectsize_CIs.

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

References

  • Albers, C., and Lakens, D. (2018). When power analyses based on pilot data are biased: Inaccurate effect size estimators and follow-up bias. Journal of experimental social psychology, 74, 187-195.

  • Allen, R. (2017). Statistics and Experimental Design for Psychologists: A Model Comparison Approach. World Scientific Publishing Company.

  • Carroll, R. M., & Nordholm, L. A. (1975). Sampling Characteristics of Kelley's epsilon and Hays' omega. Educational and Psychological Measurement, 35(3), 541-554.

  • Kelley, T. (1935) An unbiased correlation ratio measure. Proceedings of the National Academy of Sciences. 21(9). 554-559.

  • Olejnik, S., & Algina, J. (2003). Generalized eta and omega squared statistics: measures of effect size for some common research designs. Psychological methods, 8(4), 434.

  • Steiger, J. H. (2004). Beyond the F test: Effect size confidence intervals and tests of close fit in the analysis of variance and contrast analysis. Psychological Methods, 9, 164-182.

See Also

F_to_eta2()

Other effect sizes for ANOVAs: rank_epsilon_squared()

Examples

data(mtcars)
mtcars$am_f <- factor(mtcars$am)
mtcars$cyl_f <- factor(mtcars$cyl)

model <- aov(mpg ~ am_f * cyl_f, data = mtcars)

(eta2 <- eta_squared(model))

# More types:
eta_squared(model, partial = FALSE)
eta_squared(model, generalized = "cyl_f")
omega_squared(model)
epsilon_squared(model)
cohens_f(model)

model0 <- aov(mpg ~ am_f + cyl_f, data = mtcars) # no interaction
cohens_f_squared(model0, model2 = model)

## Interpretation of effect sizes
## ------------------------------

interpret_omega_squared(0.10, rules = "field2013")
interpret_eta_squared(0.10, rules = "cohen1992")
interpret_epsilon_squared(0.10, rules = "cohen1992")

interpret(eta2, rules = "cohen1992")


plot(eta2) # Requires the {see} package


# Recommended: Type-2 or -3 effect sizes + effects coding
# -------------------------------------------------------
contrasts(mtcars$am_f) <- contr.sum
contrasts(mtcars$cyl_f) <- contr.sum

model <- aov(mpg ~ am_f * cyl_f, data = mtcars)
model_anova <- car::Anova(model, type = 3)

epsilon_squared(model_anova)


# afex takes care of both type-3 effects and effects coding:
data(obk.long, package = "afex")
model <- afex::aov_car(value ~ gender + Error(id / (phase * hour)),
  data = obk.long, observed = "gender"
)

omega_squared(model)
eta_squared(model, generalized = TRUE) # observed vars are pulled from the afex model.


## Approx. effect sizes for mixed models
## -------------------------------------
model <- lme4::lmer(mpg ~ am_f * cyl_f + (1 | vs), data = mtcars)
omega_squared(model)


## Bayesian Models (PPD)
## ---------------------
fit_bayes <- rstanarm::stan_glm(
  mpg ~ factor(cyl) * wt + qsec,
  data = mtcars, family = gaussian(),
  refresh = 0
)

es <- eta_squared_posterior(fit_bayes,
  verbose = FALSE,
  ss_function = car::Anova, type = 3
)
bayestestR::describe_posterior(es, test = NULL)


# compare to:
fit_freq <- lm(mpg ~ factor(cyl) * wt + qsec,
  data = mtcars
)
aov_table <- car::Anova(fit_freq, type = 3)
eta_squared(aov_table)

Convert Between ANOVA Effect Sizes

Description

Convert Between ANOVA Effect Sizes

Usage

eta2_to_f2(es)

eta2_to_f(es)

f2_to_eta2(f2)

f_to_eta2(f)

Arguments

es

Any measure of variance explained such as Eta-, Epsilon-, Omega-, or R-Squared, partial or otherwise. See details.

f, f2

Cohen's f or f-squared.

Details

Any measure of variance explained can be converted to a corresponding Cohen's f via:

f2=η21η2f^2 = \frac{\eta^2}{1 - \eta^2}



η2=f21+f2\eta^2 = \frac{f^2}{1 + f^2}



If a partial Eta-Squared is used, the resulting Cohen's f is a partial-Cohen's f; If a less biased estimate of variance explained is used (such as Epsilon- or Omega-Squared), the resulting Cohen's f is likewise a less biased estimate of Cohen's f.

References

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

  • Steiger, J. H. (2004). Beyond the F test: Effect size confidence intervals and tests of close fit in the analysis of variance and contrast analysis. Psychological Methods, 9, 164-182.

See Also

eta_squared() for more details.

Other convert between effect sizes: d_to_r(), diff_to_cles, odds_to_probs(), oddsratio_to_riskratio(), w_to_fei()


Convert F and t Statistics to partial-η2\eta^2 and Other ANOVA Effect Sizes

Description

These functions are convenience functions to convert F and t test statistics to partial Eta- (η\eta), Omega- (ω\omega) Epsilon- (ϵ\epsilon) squared (an alias for the adjusted Eta squared) and Cohen's f. These are useful in cases where the various Sum of Squares and Mean Squares are not easily available or their computation is not straightforward (e.g., in liner mixed models, contrasts, etc.). For test statistics derived from lm and aov models, these functions give exact results. For all other cases, they return close approximations.
See Effect Size from Test Statistics vignette.

Usage

F_to_eta2(f, df, df_error, ci = 0.95, alternative = "greater", ...)

t_to_eta2(t, df_error, ci = 0.95, alternative = "greater", ...)

F_to_epsilon2(f, df, df_error, ci = 0.95, alternative = "greater", ...)

t_to_epsilon2(t, df_error, ci = 0.95, alternative = "greater", ...)

F_to_eta2_adj(f, df, df_error, ci = 0.95, alternative = "greater", ...)

t_to_eta2_adj(t, df_error, ci = 0.95, alternative = "greater", ...)

F_to_omega2(f, df, df_error, ci = 0.95, alternative = "greater", ...)

t_to_omega2(t, df_error, ci = 0.95, alternative = "greater", ...)

F_to_f(
  f,
  df,
  df_error,
  squared = FALSE,
  ci = 0.95,
  alternative = "greater",
  ...
)

t_to_f(t, df_error, squared = FALSE, ci = 0.95, alternative = "greater", ...)

F_to_f2(
  f,
  df,
  df_error,
  squared = TRUE,
  ci = 0.95,
  alternative = "greater",
  ...
)

t_to_f2(t, df_error, squared = TRUE, ci = 0.95, alternative = "greater", ...)

Arguments

df, df_error

Degrees of freedom of numerator or of the error estimate (i.e., the residuals).

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "greater" (default) or "less" (one-sided CI), or "two.sided" (two-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

...

Arguments passed to or from other methods.

t, f

The t or the F statistics.

squared

Return Cohen's f or Cohen's f-squared?

Details

These functions use the following formulae:

ηp2=F×dfnumF×dfnum+dfden\eta_p^2 = \frac{F \times df_{num}}{F \times df_{num} + df_{den}}


ϵp2=(F1)×dfnumF×dfnum+dfden\epsilon_p^2 = \frac{(F - 1) \times df_{num}}{F \times df_{num} + df_{den}}


ωp2=(F1)×dfnumF×dfnum+dfden+1\omega_p^2 = \frac{(F - 1) \times df_{num}}{F \times df_{num} + df_{den} + 1}


fp=ηp21ηp2f_p = \sqrt{\frac{\eta_p^2}{1-\eta_p^2}}



For t, the conversion is based on the equality of t2=Ft^2 = F when dfnum=1df_{num}=1.

Choosing an Un-Biased Estimate

Both Omega and Epsilon are unbiased estimators of the population Eta. But which to choose? Though Omega is the more popular choice, it should be noted that:

  1. The formula given above for Omega is only an approximation for complex designs.

  2. Epsilon has been found to be less biased (Carroll & Nordholm, 1975).

Value

A data frame with the effect size(s) between 0-1 (Eta2_partial, Epsilon2_partial, Omega2_partial, Cohens_f_partial or Cohens_f2_partial), and their CIs (CI_low and CI_high).

Confidence (Compatibility) Intervals (CIs)

Unless stated otherwise, confidence (compatibility) intervals (CIs) are estimated using the noncentrality parameter method (also called the "pivot method"). This method finds the noncentrality parameter ("ncp") of a noncentral t, F, or χ2\chi^2 distribution that places the observed t, F, or χ2\chi^2 test statistic at the desired probability point of the distribution. For example, if the observed t statistic is 2.0, with 50 degrees of freedom, for which cumulative noncentral t distribution is t = 2.0 the .025 quantile (answer: the noncentral t distribution with ncp = .04)? After estimating these confidence bounds on the ncp, they are converted into the effect size metric to obtain a confidence interval for the effect size (Steiger, 2004).

For additional details on estimation and troubleshooting, see effectsize_CIs.

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

Note

Adjusted (partial) Eta-squared is an alias for (partial) Epsilon-squared.

References

  • Albers, C., & Lakens, D. (2018). When power analyses based on pilot data are biased: Inaccurate effect size estimators and follow-up bias. Journal of experimental social psychology, 74, 187-195. doi:10.31234/osf.io/b7z4q

  • Carroll, R. M., & Nordholm, L. A. (1975). Sampling Characteristics of Kelley's epsilon and Hays' omega. Educational and Psychological Measurement, 35(3), 541-554.

  • Cumming, G., & Finch, S. (2001). A primer on the understanding, use, and calculation of confidence intervals that are based on central and noncentral distributions. Educational and Psychological Measurement, 61(4), 532-574.

  • Friedman, H. (1982). Simplified determinations of statistical power, magnitude of effect and research sample sizes. Educational and Psychological Measurement, 42(2), 521-526. doi:10.1177/001316448204200214

  • Mordkoff, J. T. (2019). A Simple Method for Removing Bias From a Popular Measure of Standardized Effect Size: Adjusted Partial Eta Squared. Advances in Methods and Practices in Psychological Science, 2(3), 228-232. doi:10.1177/2515245919855053

  • Morey, R. D., Hoekstra, R., Rouder, J. N., Lee, M. D., & Wagenmakers, E. J. (2016). The fallacy of placing confidence in confidence intervals. Psychonomic bulletin & review, 23(1), 103-123.

  • Steiger, J. H. (2004). Beyond the F test: Effect size confidence intervals and tests of close fit in the analysis of variance and contrast analysis. Psychological Methods, 9, 164-182.

See Also

eta_squared() for more details.

Other effect size from test statistic: chisq_to_phi(), t_to_d()

Examples

mod <- aov(mpg ~ factor(cyl) * factor(am), mtcars)
anova(mod)
(etas <- F_to_eta2(
  f = c(44.85, 3.99, 1.38),
  df = c(2, 1, 2),
  df_error = 26
))

if (require(see)) plot(etas)

# Compare to:
eta_squared(mod)


fit <- lmerTest::lmer(extra ~ group + (1 | ID), sleep)
# anova(fit)
# #> Type III Analysis of Variance Table with Satterthwaite's method
# #>       Sum Sq Mean Sq NumDF DenDF F value   Pr(>F)
# #> group 12.482  12.482     1     9  16.501 0.002833 **
# #> ---
# #> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

F_to_eta2(16.501, 1, 9)
F_to_omega2(16.501, 1, 9)
F_to_epsilon2(16.501, 1, 9)
F_to_f(16.501, 1, 9)


## Use with emmeans based contrasts
## --------------------------------
warp.lm <- lm(breaks ~ wool * tension, data = warpbreaks)

jt <- emmeans::joint_tests(warp.lm, by = "wool")
F_to_eta2(jt$F.ratio, jt$df1, jt$df2)

Classification of Foods

Description

Fictional data.

Format

A 2-by-3 table.

data("food_class")
food_class
#>           Soy Milk Meat
#> Vegan      47    0    0
#> Not-Vegan   0   12   21

See Also

Other effect size datasets: Music_preferences, Music_preferences2, RCT_table, Smoking_FASD, hardlyworking, rouder2016, screening_test


Format a Standardized Vector

Description

Transform a standardized vector into character, e.g., c("-1 SD", "Mean", "+1 SD").

Usage

format_standardize(
  x,
  reference = x,
  robust = FALSE,
  digits = 1,
  protect_integers = TRUE,
  ...
)

Arguments

x

A standardized numeric vector.

reference

The reference vector from which to compute the mean and SD.

robust

Logical, if TRUE, centering is done by subtracting the median from the variables and dividing it by the median absolute deviation (MAD). If FALSE, variables are standardized by subtracting the mean and dividing it by the standard deviation (SD).

digits

Number of digits for rounding or significant figures. May also be "signif" to return significant figures or "scientific" to return scientific notation. Control the number of digits by adding the value as suffix, e.g. digits = "scientific4" to have scientific notation with 4 decimal places, or digits = "signif5" for 5 significant figures (see also signif()).

protect_integers

Should integers be kept as integers (i.e., without decimals)?

...

Other arguments to pass to insight::format_value() such as digits, etc.

Examples

format_standardize(c(-1, 0, 1))
format_standardize(c(-1, 0, 1, 2), reference = rnorm(1000))
format_standardize(c(-1, 0, 1, 2), reference = rnorm(1000), robust = TRUE)

format_standardize(standardize(mtcars$wt), digits = 1)
format_standardize(standardize(mtcars$wt, robust = TRUE), digits = 1)

Workers' Salary and Other Information

Description

A sample (simulated) dataset, used in tests and some examples.

Format

A data frame with 500 rows and 5 variables:

salary

Salary, in Shmekels

xtra_hours

Number of overtime hours (on average, per week)

n_comps

Number of compliments given to the boss (observed over the last week)

age

Age in years

seniority

How many years with the company

is_senior

Has this person been working here for more than 4 years?

data("hardlyworking")
head(hardlyworking, n = 5)
#>     salary xtra_hours n_comps age seniority is_senior
#> 1 19744.65       4.16       1  32         3     FALSE
#> 2 11301.95       1.62       0  34         3     FALSE
#> 3 20635.62       1.19       3  33         5      TRUE
#> 4 23047.16       7.19       1  35         3     FALSE
#> 5 27342.15      11.26       0  33         4     FALSE

See Also

Other effect size datasets: Music_preferences, Music_preferences2, RCT_table, Smoking_FASD, food_class, rouder2016, screening_test


Generic Function for Interpretation

Description

Interpret a value based on a set of rules. See rules().

Usage

interpret(x, ...)

## S3 method for class 'numeric'
interpret(x, rules, name = attr(rules, "rule_name"), transform = NULL, ...)

## S3 method for class 'effectsize_table'
interpret(x, rules, transform = NULL, ...)

Arguments

x

Vector of value break points (edges defining categories), or a data frame of class effectsize_table.

...

Currently not used.

rules

Set of rules(). When x is a data frame, can be a name of an established set of rules.

name

Name of the set of rules (will be printed).

transform

a function (or name of a function) to apply to x before interpreting. See examples.

Value

  • For numeric input: A character vector of interpretations.

  • For data frames: the x input with an additional Interpretation column.

See Also

rules()

Examples

rules_grid <- rules(c(0.01, 0.05), c("very significant", "significant", "not significant"))
interpret(0.001, rules_grid)
interpret(0.021, rules_grid)
interpret(0.08, rules_grid)
interpret(c(0.01, 0.005, 0.08), rules_grid)

interpret(c(0.35, 0.15), c("small" = 0.2, "large" = 0.4), name = "Cohen's Rules")
interpret(c(0.35, 0.15), rules(c(0.2, 0.4), c("small", "medium", "large")))

bigness <- rules(c(1, 10), c("small", "medium", "big"))
interpret(abs(-5), bigness)
interpret(-5, bigness, transform = abs)

# ----------
d <- cohens_d(mpg ~ am, data = mtcars)
interpret(d, rules = "cohen1988")

d <- glass_delta(mpg ~ am, data = mtcars)
interpret(d, rules = "gignac2016")

interpret(d, rules = rules(1, c("tiny", "yeah okay")))

m <- lm(formula = wt ~ am * cyl, data = mtcars)
eta2 <- eta_squared(m)
interpret(eta2, rules = "field2013")

X <- chisq.test(mtcars$am, mtcars$cyl == 8)
interpret(oddsratio(X), rules = "chen2010")
interpret(cramers_v(X), "lovakov2021")

Interpret Bayes Factor (BF)

Description

Interpret Bayes Factor (BF)

Usage

interpret_bf(
  bf,
  rules = "jeffreys1961",
  log = FALSE,
  include_value = FALSE,
  protect_ratio = TRUE,
  exact = TRUE
)

Arguments

bf

Value or vector of Bayes factor (BF) values.

rules

Can be "jeffreys1961" (default), "raftery1995" or custom set of rules() (for the absolute magnitude of evidence).

log

Is the bf value log(bf)?

include_value

Include the value in the output.

protect_ratio

Should values smaller than 1 be represented as ratios?

exact

Should very large or very small values be reported with a scientific format (e.g., 4.24e5), or as truncated values (as "> 1000" and "< 1/1000").

Details

Argument names can be partially matched.

Rules

Rules apply to BF as ratios, so BF of 10 is as extreme as a BF of 0.1 (1/10).

  • Jeffreys (1961) ("jeffreys1961"; default)

    • BF = 1 - No evidence

    • 1 < BF <= 3 - Anecdotal

    • 3 < BF <= 10 - Moderate

    • 10 < BF <= 30 - Strong

    • 30 < BF <= 100 - Very strong

    • BF > 100 - Extreme.

  • Raftery (1995) ("raftery1995")

    • BF = 1 - No evidence

    • 1 < BF <= 3 - Weak

    • 3 < BF <= 20 - Positive

    • 20 < BF <= 150 - Strong

    • BF > 150 - Very strong

References

  • Jeffreys, H. (1961), Theory of Probability, 3rd ed., Oxford University Press, Oxford.

  • Raftery, A. E. (1995). Bayesian model selection in social research. Sociological methodology, 25, 111-164.

  • Jarosz, A. F., & Wiley, J. (2014). What are the odds? A practical guide to computing and reporting Bayes factors. The Journal of Problem Solving, 7(1), 2.

Examples

interpret_bf(1)
interpret_bf(c(5, 2, 0.01))

Interpret Standardized Differences

Description

Interpretation of standardized differences using different sets of rules of thumb.

Usage

interpret_cohens_d(d, rules = "cohen1988", ...)

interpret_hedges_g(g, rules = "cohen1988")

interpret_glass_delta(delta, rules = "cohen1988")

Arguments

d, g, delta

Value or vector of effect size values.

rules

Can be "cohen1988" (default), "gignac2016", "sawilowsky2009", "lovakov2021" or a custom set of rules().

...

Not directly used.

Rules

Rules apply to equally to positive and negative d (i.e., they are given as absolute values).

  • Cohen (1988) ("cohen1988"; default)

    • d < 0.2 - Very small

    • 0.2 <= d < 0.5 - Small

    • 0.5 <= d < 0.8 - Medium

    • d >= 0.8 - Large

  • Sawilowsky (2009) ("sawilowsky2009")

    • d < 0.1 - Tiny

    • 0.1 <= d < 0.2 - Very small

    • 0.2 <= d < 0.5 - Small

    • 0.5 <= d < 0.8 - Medium

    • 0.8 <= d < 1.2 - Large

    • 1.2 <= d < 2 - Very large

    • d >= 2 - Huge

  • Lovakov & Agadullina (2021) ("lovakov2021")

    • d < 0.15 - Very small

    • 0.15 <= d < 0.36 - Small

    • 0.36 <= d < 0.65 - Medium

    • d >= 0.65 - Large

  • Gignac & Szodorai (2016) ("gignac2016", based on the d_to_r() conversion, see interpret_r())

    • d < 0.2 - Very small

    • 0.2 <= d < 0.41 - Small

    • 0.41 <= d < 0.63 - Moderate

    • d >= 0.63 - Large

References

  • Lovakov, A., & Agadullina, E. R. (2021). Empirically Derived Guidelines for Effect Size Interpretation in Social Psychology. European Journal of Social Psychology.

  • Gignac, G. E., & Szodorai, E. T. (2016). Effect size guidelines for individual differences researchers. Personality and individual differences, 102, 74-78.

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

  • Sawilowsky, S. S. (2009). New effect size rules of thumb.

Examples

interpret_cohens_d(.02)
interpret_cohens_d(c(.5, .02))
interpret_cohens_d(.3, rules = "lovakov2021")

Interpret Cohen's g

Description

Interpret Cohen's g

Usage

interpret_cohens_g(g, rules = "cohen1988", ...)

Arguments

g

Value or vector of effect size values.

rules

Can be "cohen1988" (default) or a custom set of rules().

...

Not directly used.

Rules

Rules apply to equally to positive and negative g (i.e., they are given as absolute values).

  • Cohen (1988) ("cohen1988"; default)

    • d < 0.05 - Very small

    • 0.05 <= d < 0.15 - Small

    • 0.15 <= d < 0.25 - Medium

    • d >= 0.25 - Large

Note

"Since g is so transparently clear a unit, it is expected that workers in any given substantive area of the behavioral sciences will very frequently be able to set relevant [effect size] values without the proposed conventions, or set up conventions of their own which are suited to their area of inquiry." - Cohen, 1988, page 147.

References

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

Examples

interpret_cohens_g(.02)
interpret_cohens_g(c(.3, .15))

Interpret Direction

Description

Interpret Direction

Usage

interpret_direction(x)

Arguments

x

Numeric value.

Examples

interpret_direction(.02)
interpret_direction(c(.5, -.02))

Interpret Bayesian Diagnostic Indices

Description

Interpretation of Bayesian diagnostic indices, such as Effective Sample Size (ESS) and Rhat.

Usage

interpret_ess(ess, rules = "burkner2017")

interpret_rhat(rhat, rules = "vehtari2019")

Arguments

ess

Value or vector of Effective Sample Size (ESS) values.

rules

A character string (see Rules) or a custom set of rules().

rhat

Value or vector of Rhat values.

Rules

ESS

  • Bürkner, P. C. (2017) ("burkner2017"; default)

    • ESS < 1000 - Insufficient

    • ESS >= 1000 - Sufficient

Rhat

  • Vehtari et al. (2019) ("vehtari2019"; default)

    • Rhat < 1.01 - Converged

    • Rhat >= 1.01 - Failed

  • Gelman & Rubin (1992) ("gelman1992")

    • Rhat < 1.1 - Converged

    • Rhat >= 1.1 - Failed

References

  • Bürkner, P. C. (2017). brms: An R package for Bayesian multilevel models using Stan. Journal of Statistical Software, 80(1), 1-28.

  • Gelman, A., & Rubin, D. B. (1992). Inference from iterative simulation using multiple sequences. Statistical science, 7(4), 457-472.

  • Vehtari, A., Gelman, A., Simpson, D., Carpenter, B., & Bürkner, P. C. (2019). Rank-normalization, folding, and localization: An improved Rhat for assessing convergence of MCMC. arXiv preprint arXiv:1903.08008.

Examples

interpret_ess(1001)
interpret_ess(c(852, 1200))

interpret_rhat(1.00)
interpret_rhat(c(1.5, 0.9))

Interpret of CFA / SEM Indices of Goodness of Fit

Description

Interpretation of indices of fit found in confirmatory analysis or structural equation modelling, such as RMSEA, CFI, NFI, IFI, etc.

Usage

interpret_gfi(x, rules = "byrne1994")

interpret_agfi(x, rules = "byrne1994")

interpret_nfi(x, rules = "byrne1994")

interpret_nnfi(x, rules = "byrne1994")

interpret_cfi(x, rules = "byrne1994")

interpret_rfi(x, rules = "default")

interpret_ifi(x, rules = "default")

interpret_pnfi(x, rules = "default")

interpret_rmsea(x, rules = "byrne1994")

interpret_srmr(x, rules = "byrne1994")

## S3 method for class 'lavaan'
interpret(x, ...)

## S3 method for class 'performance_lavaan'
interpret(x, ...)

Arguments

x

vector of values, or an object of class lavaan.

rules

Can be the name of a set of rules (see below) or custom set of rules().

...

Currently not used.

Details

Indices of fit

  • Chisq: The model Chi-squared assesses overall fit and the discrepancy between the sample and fitted covariance matrices. Its p-value should be > .05 (i.e., the hypothesis of a perfect fit cannot be rejected). However, it is quite sensitive to sample size.

  • GFI/AGFI: The (Adjusted) Goodness of Fit is the proportion of variance accounted for by the estimated population covariance. Analogous to R2. The GFI and the AGFI should be > .95 and > .90, respectively (Byrne, 1994; "byrne1994").

  • NFI/NNFI/TLI: The (Non) Normed Fit Index. An NFI of 0.95, indicates the model of interest improves the fit by 95\ NNFI (also called the Tucker Lewis index; TLI) is preferable for smaller samples. They should be > .90 (Byrne, 1994; "byrne1994") or > .95 (Schumacker & Lomax, 2004; "schumacker2004").

  • CFI: The Comparative Fit Index is a revised form of NFI. Not very sensitive to sample size (Fan, Thompson, & Wang, 1999). Compares the fit of a target model to the fit of an independent, or null, model. It should be > .96 (Hu & Bentler, 1999; "hu&bentler1999") or .90 (Byrne, 1994; "byrne1994").

  • RFI: the Relative Fit Index, also known as RHO1, is not guaranteed to vary from 0 to 1. However, RFI close to 1 indicates a good fit.

  • IFI: the Incremental Fit Index (IFI) adjusts the Normed Fit Index (NFI) for sample size and degrees of freedom (Bollen's, 1989). Over 0.90 is a good fit, but the index can exceed 1.

  • PNFI: the Parsimony-Adjusted Measures Index. There is no commonly agreed-upon cutoff value for an acceptable model for this index. Should be > 0.50.

  • RMSEA: The Root Mean Square Error of Approximation is a parsimony-adjusted index. Values closer to 0 represent a good fit. It should be < .08 (Awang, 2012; "awang2012") or < .05 (Byrne, 1994; "byrne1994"). The p-value printed with it tests the hypothesis that RMSEA is less than or equal to .05 (a cutoff sometimes used for good fit), and thus should be not significant.

  • RMR/SRMR: the (Standardized) Root Mean Square Residual represents the square-root of the difference between the residuals of the sample covariance matrix and the hypothesized model. As the RMR can be sometimes hard to interpret, better to use SRMR. Should be < .08 (Byrne, 1994; "byrne1994").

See the documentation for fitmeasures().

What to report

For structural equation models (SEM), Kline (2015) suggests that at a minimum the following indices should be reported: The model chi-square, the RMSEA, the CFI and the SRMR.

Note

When possible, it is recommended to report dynamic cutoffs of fit indices. See https://dynamicfit.app/cfa/.

References

  • Awang, Z. (2012). A handbook on SEM. Structural equation modeling.

  • Byrne, B. M. (1994). Structural equation modeling with EQS and EQS/Windows. Thousand Oaks, CA: Sage Publications.

  • Fan, X., B. Thompson, and L. Wang (1999). Effects of sample size, estimation method, and model specification on structural equation modeling fit indexes. Structural Equation Modeling, 6, 56-83.

  • Hu, L. T., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural equation modeling: a multidisciplinary journal, 6(1), 1-55.

  • Kline, R. B. (2015). Principles and practice of structural equation modeling. Guilford publications.

  • Schumacker, R. E., and Lomax, R. G. (2004). A beginner's guide to structural equation modeling, Second edition. Mahwah, NJ: Lawrence Erlbaum Associates.

  • Tucker, L. R., and Lewis, C. (1973). The reliability coefficient for maximum likelihood factor analysis. Psychometrika, 38, 1-10.

Examples

interpret_gfi(c(.5, .99))
interpret_agfi(c(.5, .99))
interpret_nfi(c(.5, .99))
interpret_nnfi(c(.5, .99))
interpret_cfi(c(.5, .99))
interpret_rmsea(c(.07, .04))
interpret_srmr(c(.5, .99))
interpret_rfi(c(.5, .99))
interpret_ifi(c(.5, .99))
interpret_pnfi(c(.5, .99))


# Structural Equation Models (SEM)
structure <- " ind60 =~ x1 + x2 + x3
               dem60 =~ y1 + y2 + y3
               dem60 ~ ind60 "

model <- lavaan::sem(structure, data = lavaan::PoliticalDemocracy)

interpret(model)

Interpret Intraclass Correlation Coefficient (ICC)

Description

The value of an ICC lies between 0 to 1, with 0 indicating no reliability among raters and 1 indicating perfect reliability.

Usage

interpret_icc(icc, rules = "koo2016", ...)

Arguments

icc

Value or vector of Intraclass Correlation Coefficient (ICC) values.

rules

Can be "koo2016" (default) or custom set of rules().

...

Not used for now.

Rules

  • Koo (2016) ("koo2016"; default)

    • ICC < 0.50 - Poor reliability

    • 0.5 <= ICC < 0.75 - Moderate reliability

    • 0.75 <= ICC < 0.9 - Good reliability

    • **ICC >= 0.9 ** - Excellent reliability

References

  • Koo, T. K., and Li, M. Y. (2016). A guideline of selecting and reporting intraclass correlation coefficients for reliability research. Journal of chiropractic medicine, 15(2), 155-163.

Examples

interpret_icc(0.6)
interpret_icc(c(0.4, 0.8))

Interpret Kendall's Coefficient of Concordance W

Description

Interpret Kendall's Coefficient of Concordance W

Usage

interpret_kendalls_w(w, rules = "landis1977")

Arguments

w

Value or vector of Kendall's coefficient of concordance.

rules

Can be "landis1977" (default) or a custom set of rules().

Rules

  • Landis & Koch (1977) ("landis1977"; default)

    • 0.00 <= w < 0.20 - Slight agreement

    • 0.20 <= w < 0.40 - Fair agreement

    • 0.40 <= w < 0.60 - Moderate agreement

    • 0.60 <= w < 0.80 - Substantial agreement

    • w >= 0.80 - Almost perfect agreement

References

  • Landis, J. R., & Koch G. G. (1977). The measurement of observer agreement for categorical data. Biometrics, 33:159-74.


Interpret Odds Ratio

Description

Interpret Odds Ratio

Usage

interpret_oddsratio(OR, rules = "chen2010", log = FALSE, ...)

Arguments

OR

Value or vector of (log) odds ratio values.

rules

Can be "⁠chen2010"⁠ (default), "cohen1988" (through transformation to standardized difference, see oddsratio_to_d()) or custom set of rules().

log

Are the provided values log odds ratio.

...

Currently not used.

Rules

Rules apply to OR as ratios, so OR of 10 is as extreme as a OR of 0.1 (1/10).

  • Chen et al. (2010) ("chen2010"; default)

    • OR < 1.68 - Very small

    • 1.68 <= OR < 3.47 - Small

    • 3.47 <= OR < 6.71 - Medium

    • **OR >= 6.71 ** - Large

  • Cohen (1988) ("cohen1988", based on the oddsratio_to_d() conversion, see interpret_cohens_d())

    • OR < 1.44 - Very small

    • 1.44 <= OR < 2.48 - Small

    • 2.48 <= OR < 4.27 - Medium

    • **OR >= 4.27 ** - Large

References

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

  • Chen, H., Cohen, P., & Chen, S. (2010). How big is a big odds ratio? Interpreting the magnitudes of odds ratios in epidemiological studies. Communications in Statistics-Simulation and Computation, 39(4), 860-864.

  • Sánchez-Meca, J., Marín-Martínez, F., & Chacón-Moscoso, S. (2003). Effect-size indices for dichotomized outcomes in meta-analysis. Psychological methods, 8(4), 448.

Examples

interpret_oddsratio(1)
interpret_oddsratio(c(5, 2))

Interpret ANOVA Effect Sizes

Description

Interpret ANOVA Effect Sizes

Usage

interpret_omega_squared(es, rules = "field2013", ...)

interpret_eta_squared(es, rules = "field2013", ...)

interpret_epsilon_squared(es, rules = "field2013", ...)

interpret_r2_semipartial(es, rules = "field2013", ...)

Arguments

es

Value or vector of (partial) eta / omega / epsilon squared or semipartial r squared values.

rules

Can be "field2013" (default), "cohen1992" or custom set of rules().

...

Not used for now.

Rules

  • Field (2013) ("field2013"; default)

    • ES < 0.01 - Very small

    • 0.01 <= ES < 0.06 - Small

    • 0.06 <= ES < 0.14 - Medium

    • **ES >= 0.14 ** - Large

  • Cohen (1992) ("cohen1992") applicable to one-way anova, or to partial eta / omega / epsilon squared in multi-way anova.

    • ES < 0.02 - Very small

    • 0.02 <= ES < 0.13 - Small

    • 0.13 <= ES < 0.26 - Medium

    • ES >= 0.26 - Large

References

  • Field, A (2013) Discovering statistics using IBM SPSS Statistics. Fourth Edition. Sage:London.

  • Cohen, J. (1992). A power primer. Psychological bulletin, 112(1), 155.

See Also

https://imaging.mrc-cbu.cam.ac.uk/statswiki/FAQ/effectSize/

Examples

interpret_eta_squared(.02)
interpret_eta_squared(c(.5, .02), rules = "cohen1992")

Interpret p-Values

Description

Interpret p-Values

Usage

interpret_p(p, rules = "default")

Arguments

p

Value or vector of p-values.

rules

Can be "default", "rss" (for Redefine statistical significance rules) or custom set of rules().

Rules

  • Default

    • p >= 0.05 - Not significant

    • p < 0.05 - Significant

  • Benjamin et al. (2018) ("rss")

    • p >= 0.05 - Not significant

    • 0.005 <= p < 0.05 - Suggestive

    • p < 0.005 - Significant

References

  • Benjamin, D. J., Berger, J. O., Johannesson, M., Nosek, B. A., Wagenmakers, E. J., Berk, R., ... & Cesarini, D. (2018). Redefine statistical significance. Nature Human Behaviour, 2(1), 6-10.

Examples

interpret_p(c(.5, .02, 0.001))
interpret_p(c(.5, .02, 0.001), rules = "rss")

stars <- rules(c(0.001, 0.01, 0.05, 0.1), c("***", "**", "*", "+", ""),
  right = FALSE, name = "stars"
)
interpret_p(c(.5, .02, 0.001), rules = stars)

Interpret Probability of Direction (pd)

Description

Interpret Probability of Direction (pd)

Usage

interpret_pd(pd, rules = "default", ...)

Arguments

pd

Value or vector of probabilities of direction.

rules

Can be "default", "makowski2019" or a custom set of rules().

...

Not directly used.

Rules

  • Default (i.e., equivalent to p-values)

    • pd <= 0.975 - not significant

    • pd > 0.975 - significant

  • Makowski et al. (2019) ("makowski2019")

    • pd <= 0.95 - uncertain

    • pd > 0.95 - possibly existing

    • pd > 0.97 - likely existing

    • pd > 0.99 - probably existing

    • pd > 0.999 - certainly existing

References

  • Makowski, D., Ben-Shachar, M. S., Chen, S. H., and Lüdecke, D. (2019). Indices of effect existence and significance in the Bayesian framework. Frontiers in psychology, 10, 2767.

Examples

interpret_pd(.98)
interpret_pd(c(.96, .99), rules = "makowski2019")

Interpret Correlation Coefficient

Description

Interpret Correlation Coefficient

Usage

interpret_r(r, rules = "funder2019", ...)

interpret_phi(r, rules = "funder2019", ...)

interpret_cramers_v(r, rules = "funder2019", ...)

interpret_rank_biserial(r, rules = "funder2019", ...)

interpret_fei(r, rules = "funder2019", ...)

Arguments

r

Value or vector of correlation coefficient.

rules

Can be "funder2019" (default), "gignac2016", "cohen1988", "evans1996", "lovakov2021" or a custom set of rules().

...

Not directly used.

Details

Since Cohen's w does not have a fixed upper bound, for all by the most simple of cases (2-by-2 or 1-by-2 tables), interpreting Cohen's w as a correlation coefficient is inappropriate (Ben-Shachar, et al., 2024; Cohen, 1988, p. 222). Please us cramers_v() of the like instead.

Rules

Rules apply to positive and negative r alike.

  • Funder & Ozer (2019) ("funder2019"; default)

    • r < 0.05 - Tiny

    • 0.05 <= r < 0.1 - Very small

    • 0.1 <= r < 0.2 - Small

    • 0.2 <= r < 0.3 - Medium

    • 0.3 <= r < 0.4 - Large

    • r >= 0.4 - Very large

  • Gignac & Szodorai (2016) ("gignac2016")

    • r < 0.1 - Very small

    • 0.1 <= r < 0.2 - Small

    • 0.2 <= r < 0.3 - Moderate

    • r >= 0.3 - Large

  • Cohen (1988) ("cohen1988")

    • r < 0.1 - Very small

    • 0.1 <= r < 0.3 - Small

    • 0.3 <= r < 0.5 - Moderate

    • r >= 0.5 - Large

  • Lovakov & Agadullina (2021) ("lovakov2021")

    • r < 0.12 - Very small

    • 0.12 <= r < 0.24 - Small

    • 0.24 <= r < 0.41 - Moderate

    • r >= 0.41 - Large

  • Evans (1996) ("evans1996")

    • r < 0.2 - Very weak

    • 0.2 <= r < 0.4 - Weak

    • 0.4 <= r < 0.6 - Moderate

    • 0.6 <= r < 0.8 - Strong

    • r >= 0.8 - Very strong

Note

As ϕ\phi can be larger than 1 - it is recommended to compute and interpret Cramer's V instead.

References

  • Lovakov, A., & Agadullina, E. R. (2021). Empirically Derived Guidelines for Effect Size Interpretation in Social Psychology. European Journal of Social Psychology.

  • Funder, D. C., & Ozer, D. J. (2019). Evaluating effect size in psychological research: sense and nonsense. Advances in Methods and Practices in Psychological Science.

  • Gignac, G. E., & Szodorai, E. T. (2016). Effect size guidelines for individual differences researchers. Personality and individual differences, 102, 74-78.

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

  • Evans, J. D. (1996). Straightforward statistics for the behavioral sciences. Thomson Brooks/Cole Publishing Co.

  • Ben-Shachar, M.S., Patil, I., Thériault, R., Wiernik, B.M., Lüdecke, D. (2023). Phi, Fei, Fo, Fum: Effect Sizes for Categorical Data That Use the Chi‑Squared Statistic. Mathematics, 11, 1982. doi:10.3390/math11091982

See Also

Page 88 of APA's 6th Edition.

Examples

interpret_r(.015)
interpret_r(c(.5, -.02))
interpret_r(.3, rules = "lovakov2021")

Interpret Coefficient of Determination (R2R^2)

Description

Interpret Coefficient of Determination (R2R^2)

Usage

interpret_r2(r2, rules = "cohen1988")

Arguments

r2

Value or vector of R2R^2 values.

rules

Can be "cohen1988" (default), "falk1992", "chin1998", "hair2011", or custom set of rules()].

Rules

For Linear Regression

  • Cohen (1988) ("cohen1988"; default)

    • R2 < 0.02 - Very weak

    • 0.02 <= R2 < 0.13 - Weak

    • 0.13 <= R2 < 0.26 - Moderate

    • R2 >= 0.26 - Substantial

  • Falk & Miller (1992) ("falk1992")

    • R2 < 0.1 - Negligible

    • R2 >= 0.1 - Adequate

For PLS / SEM R-Squared of latent variables

  • Chin, W. W. (1998) ("chin1998")

    • R2 < 0.19 - Very weak

    • 0.19 <= R2 < 0.33 - Weak

    • 0.33 <= R2 < 0.67 - Moderate

    • R2 >= 0.67 - Substantial

  • Hair et al. (2011) ("hair2011")

    • R2 < 0.25 - Very weak

    • 0.25 <= R2 < 0.50 - Weak

    • 0.50 <= R2 < 0.75 - Moderate

    • R2 >= 0.75 - Substantial

References

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

  • Falk, R. F., & Miller, N. B. (1992). A primer for soft modeling. University of Akron Press.

  • Chin, W. W. (1998). The partial least squares approach to structural equation modeling. Modern methods for business research, 295(2), 295-336.

  • Hair, J. F., Ringle, C. M., & Sarstedt, M. (2011). PLS-SEM: Indeed a silver bullet. Journal of Marketing theory and Practice, 19(2), 139-152.

Examples

interpret_r2(.02)
interpret_r2(c(.5, .02))

Interpret Bayesian Posterior Percentage in ROPE.

Description

Interpretation of

Usage

interpret_rope(rope, ci = 0.9, rules = "default")

Arguments

rope

Value or vector of percentages in ROPE.

ci

The Credible Interval (CI) probability, corresponding to the proportion of HDI, that was used. Can be 1 in the case of "full ROPE".

rules

A character string (see details) or a custom set of rules().

Rules

  • Default

    • For CI < 1

      • Rope = 0 - Significant

      • 0 < Rope < 1 - Undecided

      • Rope = 1 - Negligible

    • For CI = 1

      • Rope < 0.01 - Significant

      • 0.01 < Rope < 0.025 - Probably significant

      • 0.025 < Rope < 0.975 - Undecided

      • 0.975 < Rope < 0.99 - Probably negligible

      • Rope > 0.99 - Negligible

References

BayestestR's reporting guidelines

Examples

interpret_rope(0, ci = 0.9)
interpret_rope(c(0.005, 0.99), ci = 1)

Interpret the Variance Inflation Factor (VIF)

Description

Interpret VIF index of multicollinearity.

Usage

interpret_vif(vif, rules = "default")

Arguments

vif

Value or vector of VIFs.

rules

Can be "default" or a custom set of rules().

Rules

  • Default

    • VIF < 5 - Low

    • 5 <= VIF < 10 - Moderate

    • VIF >= 10 - High

Examples

interpret_vif(c(1.4, 30.4))

Checks for a Valid Effect Size Name

Description

For use by other functions and packages.

Usage

is_effectsize_name(x, ignore_case = TRUE)

get_effectsize_name(x, ignore_case = TRUE)

get_effectsize_label(
  x,
  ignore_case = TRUE,
  use_symbols = getOption("es.use_symbols", FALSE)
)

Arguments

x

A character, or a vector.

ignore_case

Should case of input be ignored?

use_symbols

Should proper symbols be printed (TRUE) instead of transliterated effect size names (FALSE). See effectsize_options.


Mahalanobis' D (a multivariate Cohen's d)

Description

Compute effect size indices for standardized difference between two normal multivariate distributions or between one multivariate distribution and a defined point. This is the standardized effect size for Hotelling's T2T^2 test (e.g., DescTools::HotellingsT2Test()). D is computed as:

D=(Xˉ1Xˉ2μ)TΣp1(Xˉ1Xˉ2μ)D = \sqrt{(\bar{X}_1-\bar{X}_2-\mu)^T \Sigma_p^{-1} (\bar{X}_1-\bar{X}_2-\mu)}



Where Xˉi\bar{X}_i are the column means, Σp\Sigma_p is the pooled covariance matrix, and μ\mu is a vector of the null differences for each variable. When there is only one variate, this formula reduces to Cohen's d.

Usage

mahalanobis_d(
  x,
  y = NULL,
  data = NULL,
  pooled_cov = TRUE,
  mu = 0,
  ci = 0.95,
  alternative = "greater",
  verbose = TRUE,
  ...
)

Arguments

x, y

A data frame or matrix. Any incomplete observations (with NA values) are dropped. x can also be a formula (see details) in which case y is ignored.

data

An optional data frame containing the variables.

pooled_cov

Should equal covariance be assumed? Currently only pooled_cov = TRUE is supported.

mu

A named list/vector of the true difference in means for each variable. Can also be a vector of length 1, which will be recycled.

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "two.sided" (default, two-sided CI), "greater" or "less" (one-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

verbose

Toggle warnings and messages on or off.

...

Not used.

Details

To specify a x as a formula:

  • Two sample case: DV1 + DV2 ~ group or cbind(DV1, DV2) ~ group

  • One sample case: DV1 + DV2 ~ 1 or cbind(DV1, DV2) ~ 1

Value

A data frame with the Mahalanobis_D and potentially its CI (CI_low and CI_high).

Confidence (Compatibility) Intervals (CIs)

Unless stated otherwise, confidence (compatibility) intervals (CIs) are estimated using the noncentrality parameter method (also called the "pivot method"). This method finds the noncentrality parameter ("ncp") of a noncentral t, F, or χ2\chi^2 distribution that places the observed t, F, or χ2\chi^2 test statistic at the desired probability point of the distribution. For example, if the observed t statistic is 2.0, with 50 degrees of freedom, for which cumulative noncentral t distribution is t = 2.0 the .025 quantile (answer: the noncentral t distribution with ncp = .04)? After estimating these confidence bounds on the ncp, they are converted into the effect size metric to obtain a confidence interval for the effect size (Steiger, 2004).

For additional details on estimation and troubleshooting, see effectsize_CIs.

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

References

  • Del Giudice, M. (2017). Heterogeneity coefficients for Mahalanobis' D as a multivariate effect size. Multivariate Behavioral Research, 52(2), 216-221.

  • Mahalanobis, P. C. (1936). On the generalized distance in statistics. National Institute of Science of India.

  • Reiser, B. (2001). Confidence intervals for the Mahalanobis distance. Communications in Statistics-Simulation and Computation, 30(1), 37-45.

See Also

stats::mahalanobis(), cov_pooled()

Other standardized differences: cohens_d(), means_ratio(), p_superiority(), rank_biserial(), repeated_measures_d()

Examples

## Two samples --------------
mtcars_am0 <- subset(mtcars, am == 0,
  select = c(mpg, hp, cyl)
)
mtcars_am1 <- subset(mtcars, am == 1,
  select = c(mpg, hp, cyl)
)

mahalanobis_d(mtcars_am0, mtcars_am1)

# Or
mahalanobis_d(mpg + hp + cyl ~ am, data = mtcars)

mahalanobis_d(mpg + hp + cyl ~ am, data = mtcars, alternative = "two.sided")

# Different mu:
mahalanobis_d(mpg + hp + cyl ~ am,
  data = mtcars,
  mu = c(mpg = -4, hp = 15, cyl = 0)
)


# D is a multivariate d, so when only 1 variate is provided:
mahalanobis_d(hp ~ am, data = mtcars)

cohens_d(hp ~ am, data = mtcars)


# One sample ---------------------------
mahalanobis_d(mtcars[, c("mpg", "hp", "cyl")])

# Or
mahalanobis_d(mpg + hp + cyl ~ 1,
  data = mtcars,
  mu = c(mpg = 15, hp = 5, cyl = 3)
)

Ratio of Means

Description

Computes the ratio of two means (also known as the "response ratio"; RR) of variables on a ratio scale (with an absolute 0). Pair with any reported stats::t.test().

Usage

means_ratio(
  x,
  y = NULL,
  data = NULL,
  paired = FALSE,
  adjust = TRUE,
  log = FALSE,
  ci = 0.95,
  alternative = "two.sided",
  verbose = TRUE,
  ...
)

Arguments

x, y

A numeric vector, or a character name of one in data. Any missing values (NAs) are dropped from the resulting vector. x can also be a formula (see stats::t.test()), in which case y is ignored.

data

An optional data frame containing the variables.

paired

If TRUE, the values of x and y are considered as paired. The correlation between these variables will affect the CIs.

adjust

Should the effect size be adjusted for small-sample bias? Defaults to TRUE; Advisable for small samples.

log

Should the log-ratio be returned? Defaults to FALSE. Normally distributed and useful for meta-analysis.

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "two.sided" (default, two-sided CI), "greater" or "less" (one-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

verbose

Toggle warnings and messages on or off.

...

Arguments passed to or from other methods. When x is a formula, these can be subset and na.action.

Details

The Means Ratio ranges from 0 to \infty, with values smaller than 1 indicating that the second mean is larger than the first, values larger than 1 indicating that the second mean is smaller than the first, and values of 1 indicating that the means are equal.

Value

A data frame with the effect size (Means_ratio or Means_ratio_adjusted) and their CIs (CI_low and CI_high).

Confidence (Compatibility) Intervals (CIs)

Confidence intervals are estimated as described by Lajeunesse (2011 & 2015) using the log-ratio standard error assuming a normal distribution. By this method, the log is taken of the ratio of means, which makes this outcome measure symmetric around 0 and yields a corresponding sampling distribution that is closer to normality.

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

Note

The small-sample bias corrected response ratio reported from this function is derived from Lajeunesse (2015).

References

Lajeunesse, M. J. (2011). On the meta-analysis of response ratios for studies with correlated and multi-group designs. Ecology, 92(11), 2049-2055. doi:10.1890/11-0423.1

Lajeunesse, M. J. (2015). Bias and correction for the log response ratio in ecological meta-analysis. Ecology, 96(8), 2056-2063. doi:10.1890/14-2402.1

Hedges, L. V., Gurevitch, J., & Curtis, P. S. (1999). The meta-analysis of response ratios in experimental ecology. Ecology, 80(4), 1150–1156. doi:10.1890/0012-9658(1999)080[1150:TMAORR]2.0.CO;2

See Also

Other standardized differences: cohens_d(), mahalanobis_d(), p_superiority(), rank_biserial(), repeated_measures_d()

Examples

x <- c(1.83, 0.50, 1.62, 2.48, 1.68, 1.88, 1.55, 3.06, 1.30)
y <- c(0.878, 0.647, 0.598, 2.05, 1.06, 1.29, 1.06, 3.14, 1.29)
means_ratio(x, y)
means_ratio(x, y, adjust = FALSE)

means_ratio(x, y, log = TRUE)


# The ratio is scale invariant, making it a standardized effect size
means_ratio(3 * x, 3 * y)

Music Preference by College Major

Description

Fictional data.

Format

A 4-by-3 table, with a column for each major and a row for each type of music.

data("Music_preferences")
Music_preferences
#>       Pop Rock Jazz Classic
#> Psych 150  100  165     130
#> Econ   50   65   35      10
#> Law     2   55   40      25

See Also

Other effect size datasets: Music_preferences2, RCT_table, Smoking_FASD, food_class, hardlyworking, rouder2016, screening_test


Music Preference by College Major

Description

Fictional data, with more extreme preferences than Music_preferences

Format

A 4-by-3 table, with a column for each major and a row for each type of music.

data("Music_preferences2")
Music_preferences2
#>       Pop Rock Jazz Classic
#> Psych 151  130   12       7
#> Econ   77    6  111       4
#> Law     0    4    2     165

See Also

Other effect size datasets: Music_preferences, RCT_table, Smoking_FASD, food_class, hardlyworking, rouder2016, screening_test


Convert Between Odds and Probabilities

Description

Convert Between Odds and Probabilities

Usage

odds_to_probs(odds, log = FALSE, ...)

## S3 method for class 'data.frame'
odds_to_probs(odds, log = FALSE, select = NULL, exclude = NULL, ...)

probs_to_odds(probs, log = FALSE, ...)

## S3 method for class 'data.frame'
probs_to_odds(probs, log = FALSE, select = NULL, exclude = NULL, ...)

Arguments

odds

The Odds (or log(odds) when log = TRUE) to convert.

log

Take in or output log odds (such as in logistic models).

...

Arguments passed to or from other methods.

select

When a data frame is passed, character or list of of column names to be transformed.

exclude

When a data frame is passed, character or list of column names to be excluded from transformation.

probs

Probability values to convert.

Value

Converted index.

See Also

stats::plogis()

Other convert between effect sizes: d_to_r(), diff_to_cles, eta2_to_f2(), oddsratio_to_riskratio(), w_to_fei()

Examples

odds_to_probs(3)
odds_to_probs(1.09, log = TRUE)

probs_to_odds(0.95)
probs_to_odds(0.95, log = TRUE)

Odds Ratios, Risk Ratios and Other Effect Sizes for 2-by-2 Contingency Tables

Description

Compute Odds Ratios, Risk Ratios, Cohen's h, Absolute Risk Reduction or Number Needed to Treat. Report with any stats::chisq.test() or stats::fisher.test().

Note that these are computed with each column representing the different groups, and the first column representing the treatment group and the second column baseline (or control). Effects are given as treatment / control. If you wish you use rows as groups you must pass a transposed table, or switch the x and y arguments.

Usage

oddsratio(x, y = NULL, ci = 0.95, alternative = "two.sided", log = FALSE, ...)

riskratio(x, y = NULL, ci = 0.95, alternative = "two.sided", log = FALSE, ...)

cohens_h(x, y = NULL, ci = 0.95, alternative = "two.sided", ...)

arr(x, y = NULL, ci = 0.95, alternative = "two.sided", ...)

nnt(x, y = NULL, ci = 0.95, alternative = "two.sided", ...)

Arguments

x

a numeric vector or matrix. x and y can also both be factors.

y

a numeric vector; ignored if x is a matrix. If x is a factor, y should be a factor of the same length.

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "two.sided" (default, two-sided CI), "greater" or "less" (one-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

log

Take in or output the log of the ratio (such as in logistic models), e.g. when the desired input or output are log odds ratios instead odds ratios.

...

Ignored

Value

A data frame with the effect size (Odds_ratio, Risk_ratio (possibly with the prefix log_), Cohens_h, ARR, NNT) and its CIs (CI_low and CI_high).

Confidence (Compatibility) Intervals (CIs)

Confidence intervals are estimated using the standard normal parametric method (see Katz et al., 1978; Szumilas, 2010).

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

References

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

  • Katz, D. J. S. M., Baptista, J., Azen, S. P., & Pike, M. C. (1978). Obtaining confidence intervals for the risk ratio in cohort studies. Biometrics, 469-474.

  • Szumilas, M. (2010). Explaining odds ratios. Journal of the Canadian academy of child and adolescent psychiatry, 19(3), 227.

See Also

Other effect sizes for contingency table: cohens_g(), phi()

Examples

data("RCT_table")
RCT_table # note groups are COLUMNS

oddsratio(RCT_table)
oddsratio(RCT_table, alternative = "greater")

riskratio(RCT_table)

cohens_h(RCT_table)

arr(RCT_table)

nnt(RCT_table)

Convert Between Odds Ratios, Risk Ratios and Other Metrics of Change in Probabilities

Description

Convert Between Odds Ratios, Risk Ratios and Other Metrics of Change in Probabilities

Usage

oddsratio_to_riskratio(OR, p0, log = FALSE, verbose = TRUE, ...)

oddsratio_to_arr(OR, p0, log = FALSE, verbose = TRUE, ...)

oddsratio_to_nnt(OR, p0, log = FALSE, verbose = TRUE, ...)

logoddsratio_to_riskratio(logOR, p0, log = TRUE, verbose = TRUE, ...)

logoddsratio_to_arr(logOR, p0, log = TRUE, verbose = TRUE, ...)

logoddsratio_to_nnt(logOR, p0, log = TRUE, verbose = TRUE, ...)

riskratio_to_oddsratio(RR, p0, log = FALSE, verbose = TRUE, ...)

riskratio_to_arr(RR, p0, verbose = TRUE, ...)

riskratio_to_logoddsratio(RR, p0, log = TRUE, verbose = TRUE, ...)

riskratio_to_nnt(RR, p0, verbose = TRUE, ...)

arr_to_riskratio(ARR, p0, verbose = TRUE, ...)

arr_to_oddsratio(ARR, p0, log = FALSE, verbose = TRUE, ...)

arr_to_logoddsratio(ARR, p0, log = TRUE, verbose = TRUE, ...)

arr_to_nnt(ARR, ...)

nnt_to_oddsratio(NNT, p0, log = FALSE, verbose = TRUE, ...)

nnt_to_logoddsratio(NNT, p0, log = TRUE, verbose = TRUE, ...)

nnt_to_riskratio(NNT, p0, verbose = TRUE, ...)

nnt_to_arr(NNT, ...)

Arguments

OR, logOR, RR, ARR, NNT

Odds-ratio of odds(p1)/odds(p0), log-Odds-ratio of log(odds(p1)/odds(p0)), Risk ratio of p1/p0, Absolute Risk Reduction of p1 - p0, or Number-needed-to-treat of 1/(p1 - p0). OR and logOR can also be a logistic regression model.

p0

Baseline risk

log

If:

  • TRUE:

    • In ⁠oddsratio_to_*()⁠, OR input is treated as log(OR).

    • In ⁠*_to_oddsratio()⁠, returned value is log(OR).

  • FALSE:

    • In ⁠logoddsratio_to_*()⁠, logOR input is treated as OR.

    • In ⁠*_to_logoddsratio()⁠, returned value is OR.

verbose

Toggle warnings and messages on or off.

...

Arguments passed to and from other methods.

Value

Converted index, or if OR/logOR is a logistic regression model, a parameter table with the converted indices.

References

Grant, R. L. (2014). Converting an odds ratio to a range of plausible relative risks for better communication of research findings. Bmj, 348, f7450.

See Also

oddsratio(), riskratio(), arr(), and nnt().

Other convert between effect sizes: d_to_r(), diff_to_cles, eta2_to_f2(), odds_to_probs(), w_to_fei()

Examples

p0 <- 0.4
p1 <- 0.7

(OR <- probs_to_odds(p1) / probs_to_odds(p0))
(RR <- p1 / p0)
(ARR <- p1 - p0)
(NNT <- arr_to_nnt(ARR))

riskratio_to_oddsratio(RR, p0 = p0)
oddsratio_to_riskratio(OR, p0 = p0)
riskratio_to_arr(RR, p0 = p0)
arr_to_oddsratio(nnt_to_arr(NNT), p0 = p0)

m <- glm(am ~ factor(cyl),
  data = mtcars,
  family = binomial()
)
oddsratio_to_riskratio(m, verbose = FALSE) # RR is relative to the intercept if p0 not provided

Cohen's Us and Other Common Language Effect Sizes (CLES)

Description

Cohen's U1U_1, U2U_2, and U3U_3, probability of superiority, proportion of overlap, Wilcoxon-Mann-Whitney odds, and Vargha and Delaney's A are CLESs. These are effect sizes that represent differences between two (independent) distributions in probabilistic terms (See details). Pair with any reported stats::t.test() or stats::wilcox.test().

Usage

p_superiority(
  x,
  y = NULL,
  data = NULL,
  mu = 0,
  paired = FALSE,
  parametric = TRUE,
  ci = 0.95,
  alternative = "two.sided",
  verbose = TRUE,
  ...
)

cohens_u1(
  x,
  y = NULL,
  data = NULL,
  mu = 0,
  parametric = TRUE,
  ci = 0.95,
  alternative = "two.sided",
  iterations = 200,
  verbose = TRUE,
  ...
)

cohens_u2(
  x,
  y = NULL,
  data = NULL,
  mu = 0,
  parametric = TRUE,
  ci = 0.95,
  alternative = "two.sided",
  iterations = 200,
  verbose = TRUE,
  ...
)

cohens_u3(
  x,
  y = NULL,
  data = NULL,
  mu = 0,
  parametric = TRUE,
  ci = 0.95,
  alternative = "two.sided",
  iterations = 200,
  verbose = TRUE,
  ...
)

p_overlap(
  x,
  y = NULL,
  data = NULL,
  mu = 0,
  parametric = TRUE,
  ci = 0.95,
  alternative = "two.sided",
  iterations = 200,
  verbose = TRUE,
  ...
)

vd_a(
  x,
  y = NULL,
  data = NULL,
  mu = 0,
  ci = 0.95,
  alternative = "two.sided",
  verbose = TRUE,
  ...
)

wmw_odds(
  x,
  y = NULL,
  data = NULL,
  mu = 0,
  paired = FALSE,
  ci = 0.95,
  alternative = "two.sided",
  verbose = TRUE,
  ...
)

Arguments

x, y

A numeric vector, or a character name of one in data. Any missing values (NAs) are dropped from the resulting vector. x can also be a formula (see stats::t.test()), in which case y is ignored.

data

An optional data frame containing the variables.

mu

a number indicating the true value of the mean (or difference in means if you are performing a two sample test).

paired

If TRUE, the values of x and y are considered as paired. This produces an effect size that is equivalent to the one-sample effect size on x - y.

parametric

Use parametric estimation (see cohens_d()) or non-parametric estimation (see rank_biserial()). See details.

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "two.sided" (default, two-sided CI), "greater" or "less" (one-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

verbose

Toggle warnings and messages on or off.

...

Arguments passed to or from other methods. When x is a formula, these can be subset and na.action.

iterations

The number of bootstrap replicates for computing confidence intervals. Only applies when ci is not NULL and parametric = FALSE.

Details

These measures of effect size present group differences in probabilistic terms:

  • Probability of superiority is the probability that, when sampling an observation from each of the groups at random, that the observation from the second group will be larger than the sample from the first group. For the one-sample (or paired) case, it is the probability that the sample (or difference) is larger than mu. (Vargha and Delaney's A is an alias for the non-parametric probability of superiority.)

  • Cohen's U1U_1 is the proportion of the total of both distributions that does not overlap.

  • Cohen's U2U_2 is the proportion of one of the groups that exceeds the same proportion in the other group.

  • Cohen's U3U_3 is the proportion of the second group that is smaller than the median of the first group.

  • Overlap (OVL) is the proportional overlap between the distributions. (When parametric = FALSE, bayestestR::overlap() is used.)

Wilcoxon-Mann-Whitney odds are the odds of non-parametric superiority (via probs_to_odds()), that is the odds that, when sampling an observation from each of the groups at random, that the observation from the second group will be larger than the sample from the first group.

Where U1U_1, U2U_2, and Overlap are agnostic to the direction of the difference between the groups, U3U_3 and probability of superiority are not.

The parametric version of these effects assumes normality of both populations and homoscedasticity. If those are not met, the non parametric versions should be used.

Value

A data frame containing the common language effect sizes (and optionally their CIs).

Confidence (Compatibility) Intervals (CIs)

For parametric CLES, the CIs are transformed CIs for Cohen's d (see d_to_u3()). For non-parametric (parametric = FALSE) CLES, the CI of Pr(superiority) is a transformed CI of the rank-biserial correlation (rb_to_p_superiority()), while for all others, confidence intervals are estimated using the bootstrap method (using the {boot} package).

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Bootstrapped CIs

Some effect sizes are directionless–they do have a minimum value that would be interpreted as "no effect", but they cannot cross it. For example, a null value of Kendall's W is 0, indicating no difference between groups, but it can never have a negative value. Same goes for U2 and Overlap: the null value of U2U_2 is 0.5, but it can never be smaller than 0.5; am Overlap of 1 means "full overlap" (no difference), but it cannot be larger than 1.

When bootstrapping CIs for such effect sizes, the bounds of the CIs will never cross (and often will never cover) the null. Therefore, these CIs should not be used for statistical inference.

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

Note

If mu is not 0, the effect size represents the difference between the first shifted sample (by mu) and the second sample.

References

  • Cohen, J. (1977). Statistical power analysis for the behavioral sciences. New York: Routledge.

  • Reiser, B., & Faraggi, D. (1999). Confidence intervals for the overlapping coefficient: the normal equal variance case. Journal of the Royal Statistical Society, 48(3), 413-418.

  • Ruscio, J. (2008). A probability-based measure of effect size: robustness to base rates and other factors. Psychological methods, 13(1), 19–30.

  • Vargha, A., & Delaney, H. D. (2000). A critique and improvement of the CL common language effect size statistics of McGraw and Wong. Journal of Educational and Behavioral Statistics, 25(2), 101-132.

  • O’Brien, R. G., & Castelloe, J. (2006, March). Exploiting the link between the Wilcoxon-Mann-Whitney test and a simple odds statistic. In Proceedings of the Thirty-first Annual SAS Users Group International Conference (pp. 209-31). Cary, NC: SAS Institute.

  • Agresti, A. (1980). Generalized odds ratios for ordinal data. Biometrics, 59-67.

See Also

sd_pooled()

Other standardized differences: cohens_d(), mahalanobis_d(), means_ratio(), rank_biserial(), repeated_measures_d()

Other rank-based effect sizes: rank_biserial(), rank_epsilon_squared()

Examples

cohens_u2(mpg ~ am, data = mtcars)

p_superiority(mpg ~ am, data = mtcars, parametric = FALSE)

wmw_odds(mpg ~ am, data = mtcars)

x <- c(1.83, 0.5, 1.62, 2.48, 1.68, 1.88, 1.55, 3.06, 1.3)
y <- c(0.878, 0.647, 0.598, 2.05, 1.06, 1.29, 1.06, 3.14, 1.29)

p_overlap(x, y)
p_overlap(y, x) # direction of effect does not matter

cohens_u3(x, y)
cohens_u3(y, x) # direction of effect does matter

ϕ\phi and Other Contingency Tables Correlations

Description

Compute phi (ϕ\phi), Cramer's V, Tschuprow's T, Cohen's w, פ (Fei), Pearson's contingency coefficient for contingency tables or goodness-of-fit. Pair with any reported stats::chisq.test().

Usage

phi(x, y = NULL, adjust = TRUE, ci = 0.95, alternative = "greater", ...)

cramers_v(x, y = NULL, adjust = TRUE, ci = 0.95, alternative = "greater", ...)

tschuprows_t(
  x,
  y = NULL,
  adjust = TRUE,
  ci = 0.95,
  alternative = "greater",
  ...
)

cohens_w(
  x,
  y = NULL,
  p = rep(1, length(x)),
  ci = 0.95,
  alternative = "greater",
  ...
)

fei(x, p = rep(1, length(x)), ci = 0.95, alternative = "greater", ...)

pearsons_c(
  x,
  y = NULL,
  p = rep(1, length(x)),
  ci = 0.95,
  alternative = "greater",
  ...
)

Arguments

x

a numeric vector or matrix. x and y can also both be factors.

y

a numeric vector; ignored if x is a matrix. If x is a factor, y should be a factor of the same length.

adjust

Should the effect size be corrected for small-sample bias? Defaults to TRUE; Advisable for small samples and large tables.

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "greater" (default) or "less" (one-sided CI), or "two.sided" (two-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

...

Ignored.

p

a vector of probabilities of the same length as x. An error is given if any entry of p is negative.

Details

phi (ϕ\phi), Cramer's V, Tschuprow's T, Cohen's w, and Pearson's C are effect sizes for tests of independence in 2D contingency tables. For 2-by-2 tables, phi, Cramer's V, Tschuprow's T, and Cohen's w are identical, and are equal to the simple correlation between two dichotomous variables, ranging between 0 (no dependence) and 1 (perfect dependence).

For larger tables, Cramer's V, Tschuprow's T or Pearson's C should be used, as they are bounded between 0-1. (Cohen's w can also be used, but since it is not bounded at 1 (can be larger) its interpretation is more difficult.) For square table, Cramer's V and Tschuprow's T give the same results, but for non-square tables Tschuprow's T is more conservative: while V will be 1 if either columns are fully dependent on rows (for each column, there is only one non-0 cell) or rows are fully dependent on columns, T will only be 1 if both are true.

For goodness-of-fit in 1D tables Cohen's W, פ (Fei) or Pearson's C can be used. Cohen's w has no upper bound (can be arbitrarily large, depending on the expected distribution). Fei is an adjusted Cohen's w, accounting for the expected distribution, making it bounded between 0-1 (Ben-Shachar et al, 2023). Pearson's C is also bounded between 0-1.

To summarize, for correlation-like effect sizes, we recommend:

  • For a 2x2 table, use phi()

  • For larger tables, use cramers_v()

  • For goodness-of-fit, use fei()

Value

A data frame with the effect size (Cramers_v, phi (possibly with the suffix ⁠_adjusted⁠), Cohens_w, Fei) and its CIs (CI_low and CI_high).

Confidence (Compatibility) Intervals (CIs)

Unless stated otherwise, confidence (compatibility) intervals (CIs) are estimated using the noncentrality parameter method (also called the "pivot method"). This method finds the noncentrality parameter ("ncp") of a noncentral t, F, or χ2\chi^2 distribution that places the observed t, F, or χ2\chi^2 test statistic at the desired probability point of the distribution. For example, if the observed t statistic is 2.0, with 50 degrees of freedom, for which cumulative noncentral t distribution is t = 2.0 the .025 quantile (answer: the noncentral t distribution with ncp = .04)? After estimating these confidence bounds on the ncp, they are converted into the effect size metric to obtain a confidence interval for the effect size (Steiger, 2004).

For additional details on estimation and troubleshooting, see effectsize_CIs.

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

References

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

  • Ben-Shachar, M.S., Patil, I., Thériault, R., Wiernik, B.M., Lüdecke, D. (2023). Phi, Fei, Fo, Fum: Effect Sizes for Categorical Data That Use the Chi‑Squared Statistic. Mathematics, 11, 1982. doi:10.3390/math11091982

  • Johnston, J. E., Berry, K. J., & Mielke Jr, P. W. (2006). Measures of effect size for chi-squared and likelihood-ratio goodness-of-fit tests. Perceptual and motor skills, 103(2), 412-414.

  • Rosenberg, M. S. (2010). A generalized formula for converting chi-square tests to effect sizes for meta-analysis. PloS one, 5(4), e10059.

See Also

chisq_to_phi() for details regarding estimation and CIs.

Other effect sizes for contingency table: cohens_g(), oddsratio()

Examples

## 2-by-2 tables
## -------------
data("RCT_table")
RCT_table # note groups are COLUMNS

phi(RCT_table)
pearsons_c(RCT_table)



## Larger tables
## -------------
data("Music_preferences")
Music_preferences

cramers_v(Music_preferences)

cohens_w(Music_preferences)

pearsons_c(Music_preferences)



## Goodness of fit
## ---------------
data("Smoking_FASD")
Smoking_FASD

fei(Smoking_FASD)

cohens_w(Smoking_FASD)

pearsons_c(Smoking_FASD)

# Use custom expected values:
fei(Smoking_FASD, p = c(0.015, 0.010, 0.975))

cohens_w(Smoking_FASD, p = c(0.015, 0.010, 0.975))

pearsons_c(Smoking_FASD, p = c(0.015, 0.010, 0.975))

Methods for {effectsize} Tables

Description

Printing, formatting and plotting methods for effectsize tables.

Usage

## S3 method for class 'effectsize_table'
plot(x, ...)

## S3 method for class 'effectsize_table'
print(x, digits = 2, use_symbols = getOption("es.use_symbols", FALSE), ...)

## S3 method for class 'effectsize_table'
print_md(x, digits = 2, use_symbols = getOption("es.use_symbols", FALSE), ...)

## S3 method for class 'effectsize_table'
print_html(
  x,
  digits = 2,
  use_symbols = getOption("es.use_symbols", FALSE),
  ...
)

## S3 method for class 'effectsize_table'
format(
  x,
  digits = 2,
  output = c("text", "markdown", "html"),
  use_symbols = getOption("es.use_symbols", FALSE),
  ...
)

## S3 method for class 'effectsize_difference'
print(x, digits = 2, append_CLES = NULL, ...)

Arguments

x

Object to print.

...

Arguments passed to or from other functions.

digits

Number of digits for rounding or significant figures. May also be "signif" to return significant figures or "scientific" to return scientific notation. Control the number of digits by adding the value as suffix, e.g. digits = "scientific4" to have scientific notation with 4 decimal places, or digits = "signif5" for 5 significant figures (see also signif()).

use_symbols

Should proper symbols be printed (TRUE) instead of transliterated effect size names (FALSE). See effectsize_options.

output

Which output is the formatting intended for? Affects how title and footers are formatted.

append_CLES

Which Common Language Effect Sizes should be printed as well? Only applicable to Cohen's d, Hedges' g for independent samples of equal variance (pooled sd) or for the rank-biserial correlation for independent samples (See d_to_cles).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

See Also

insight::display()


Semi-Partial (Part) Correlation Squared (ΔR2\Delta R^2)

Description

Compute the semi-partial (part) correlation squared (also known as ΔR2\Delta R^2). Currently, only lm() models are supported.

Usage

r2_semipartial(
  model,
  type = c("terms", "parameters"),
  ci = 0.95,
  alternative = "greater",
  ...
)

Arguments

model

An lm model.

type

Type, either "terms", or "parameters".

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "greater" (default) or "less" (one-sided CI), or "two.sided" (two-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

...

Arguments passed to or from other methods.

Details

This is similar to the last column of the "Conditional Dominance Statistics" section of the parameters::dominance_analysis() output. For each term, the model is refit without the columns on the model matrix that correspond to that term. The R2R^2 of this sub-model is then subtracted from the R2R^2 of the full model to yield the ΔR2\Delta R^2. (For type = "parameters", this is done for each column in the model matrix.)

Note that this is unlike parameters::dominance_analysis(), where term deletion is done via the formula interface, and therefore may lead to different results.

For other, non-lm() models, as well as more verbose information and options, please see the documentation for parameters::dominance_analysis().

Value

A data frame with the effect size.

Confidence (Compatibility) Intervals (CIs)

Confidence intervals are based on the normal approximation as provided by Alf and Graf (1999). An adjustment to the lower bound of the CI is used, to improve the coverage properties of the CIs, according to Algina et al (2008): If the F test associated with the sr2sr^2 is significant (at 1-ci level), but the lower bound of the CI is 0, it is set to a small value (arbitrarily to a 10th of the estimated sr2sr^2); if the F test is not significant, the lower bound is set to 0. (Additionally, lower and upper bound are "fixed" so that they cannot be smaller than 0 or larger than 1.)

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

References

  • Alf Jr, E. F., & Graf, R. G. (1999). Asymptotic confidence limits for the difference between two squared multiple correlations: A simplified approach. Psychological Methods, 4(1), 70-75. doi:10.1037/1082-989X.4.1.70

  • Algina, J., Keselman, H. J., & Penfield, R. D. (2008). Confidence intervals for the squared multiple semipartial correlation coefficient. Journal of Modern Applied Statistical Methods, 7(1), 2-10. doi:10.22237/jmasm/1209614460

See Also

eta_squared(), cohens_f() for comparing two models, parameters::dominance_analysis() and parameters::standardize_parameters().

Examples

data("hardlyworking")

m <- lm(salary ~ factor(n_comps) + xtra_hours * seniority, data = hardlyworking)

r2_semipartial(m)

r2_semipartial(m, type = "parameters")



# Compare to `eta_squared()`
# --------------------------
npk.aov <- lm(yield ~ N + P + K, npk)

# When predictors are orthogonal,
# eta_squared(partial = FALSE) gives the same effect size:
performance::check_collinearity(npk.aov)

eta_squared(npk.aov, partial = FALSE)

r2_semipartial(npk.aov)


# Compare to `dominance_analysis()`
# ---------------------------------
m_full <- lm(salary ~ ., data = hardlyworking)

r2_semipartial(m_full)

# Compare to last column of "Conditional Dominance Statistics":
parameters::dominance_analysis(m_full)

Dominance Effect Sizes for Rank Based Differences

Description

Compute the rank-biserial correlation (rrbr_{rb}) and Cliff's delta (δ\delta) effect sizes for non-parametric (rank sum) differences. These effect sizes of dominance are closely related to the Common Language Effect Sizes. Pair with any reported stats::wilcox.test().

Usage

rank_biserial(
  x,
  y = NULL,
  data = NULL,
  mu = 0,
  paired = FALSE,
  ci = 0.95,
  alternative = "two.sided",
  verbose = TRUE,
  ...
)

cliffs_delta(
  x,
  y = NULL,
  data = NULL,
  mu = 0,
  ci = 0.95,
  alternative = "two.sided",
  verbose = TRUE,
  ...
)

Arguments

x, y

A numeric or ordered vector, or a character name of one in data. Any missing values (NAs) are dropped from the resulting vector. x can also be a formula (see stats::wilcox.test()), in which case y is ignored.

data

An optional data frame containing the variables.

mu

a number indicating the value around which (a-)symmetry (for one-sample or paired samples) or shift (for independent samples) is to be estimated. See stats::wilcox.test.

paired

If TRUE, the values of x and y are considered as paired. This produces an effect size that is equivalent to the one-sample effect size on x - y.

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "two.sided" (default, two-sided CI), "greater" or "less" (one-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

verbose

Toggle warnings and messages on or off.

...

Arguments passed to or from other methods. When x is a formula, these can be subset and na.action.

Details

The rank-biserial correlation is appropriate for non-parametric tests of differences - both for the one sample or paired samples case, that would normally be tested with Wilcoxon's Signed Rank Test (giving the matched-pairs rank-biserial correlation) and for two independent samples case, that would normally be tested with Mann-Whitney's U Test (giving Glass' rank-biserial correlation). See stats::wilcox.test. In both cases, the correlation represents the difference between the proportion of favorable and unfavorable pairs / signed ranks (Kerby, 2014). Values range from -1 complete dominance of the second sample (all values of the second sample are larger than all the values of the first sample) to +1 complete dominance of the fist sample (all values of the second sample are smaller than all the values of the first sample).

Cliff's delta is an alias to the rank-biserial correlation in the two sample case.

Value

A data frame with the effect size r_rank_biserial and its CI (CI_low and CI_high).

Ties

When tied values occur, they are each given the average of the ranks that would have been given had no ties occurred. This results in an effect size of reduced magnitude. A correction has been applied for Kendall's W.

Confidence (Compatibility) Intervals (CIs)

Confidence intervals for the rank-biserial correlation (and Cliff's delta) are estimated using the normal approximation (via Fisher's transformation).

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

References

  • Cureton, E. E. (1956). Rank-biserial correlation. Psychometrika, 21(3), 287-290.

  • Glass, G. V. (1965). A ranking variable analogue of biserial correlation: Implications for short-cut item analysis. Journal of Educational Measurement, 2(1), 91-95.

  • Kerby, D. S. (2014). The simple difference formula: An approach to teaching nonparametric correlation. Comprehensive Psychology, 3, 11-IT.

  • King, B. M., & Minium, E. W. (2008). Statistical reasoning in the behavioral sciences. John Wiley & Sons Inc.

  • Cliff, N. (1993). Dominance statistics: Ordinal analyses to answer ordinal questions. Psychological bulletin, 114(3), 494.

  • Tomczak, M., & Tomczak, E. (2014). The need to report effect size estimates revisited. An overview of some recommended measures of effect size.

See Also

Other standardized differences: cohens_d(), mahalanobis_d(), means_ratio(), p_superiority(), repeated_measures_d()

Other rank-based effect sizes: p_superiority(), rank_epsilon_squared()

Examples

data(mtcars)
mtcars$am <- factor(mtcars$am)
mtcars$cyl <- factor(mtcars$cyl)

# Two Independent Samples ----------
(rb <- rank_biserial(mpg ~ am, data = mtcars))
# Same as:
# rank_biserial("mpg", "am", data = mtcars)
# rank_biserial(mtcars$mpg[mtcars$am=="0"], mtcars$mpg[mtcars$am=="1"])
# cliffs_delta(mpg ~ am, data = mtcars)

# More options:
rank_biserial(mpg ~ am, data = mtcars, mu = -5)
print(rb, append_CLES = TRUE)


# One Sample ----------
# from help("wilcox.test")
x <- c(1.83, 0.50, 1.62, 2.48, 1.68, 1.88, 1.55, 3.06, 1.30)
y <- c(0.878, 0.647, 0.598, 2.05, 1.06, 1.29, 1.06, 3.14, 1.29)
depression <- data.frame(first = x, second = y, change = y - x)

rank_biserial(change ~ 1, data = depression)

# same as:
# rank_biserial("change", data = depression)
# rank_biserial(mtcars$wt)

# More options:
rank_biserial(change ~ 1, data = depression, mu = -0.5)


# Paired Samples ----------
(rb <- rank_biserial(Pair(first, second) ~ 1, data = depression))

# same as:
# rank_biserial(depression$first, depression$second, paired = TRUE)

interpret_rank_biserial(0.78)
interpret(rb, rules = "funder2019")

Effect Size for Rank Based ANOVA

Description

Compute rank epsilon squared (ER2E^2_R) or rank eta squared (ηH2\eta^2_H) (to accompany stats::kruskal.test()), and Kendall's W (to accompany stats::friedman.test()) effect sizes for non-parametric (rank sum) one-way ANOVAs.

Usage

rank_epsilon_squared(
  x,
  groups,
  data = NULL,
  ci = 0.95,
  alternative = "greater",
  iterations = 200,
  verbose = TRUE,
  ...
)

rank_eta_squared(
  x,
  groups,
  data = NULL,
  ci = 0.95,
  alternative = "greater",
  iterations = 200,
  verbose = TRUE,
  ...
)

kendalls_w(
  x,
  groups,
  blocks,
  data = NULL,
  blocks_on_rows = TRUE,
  ci = 0.95,
  alternative = "greater",
  iterations = 200,
  verbose = TRUE,
  ...
)

Arguments

x

Can be one of:

  • A numeric or ordered vector, or a character name of one in data.

  • A list of vectors (for rank_eta/epsilon_squared()).

  • A matrix of ⁠blocks x groups⁠ (for kendalls_w()) (or ⁠groups x blocks⁠ if blocks_on_rows = FALSE). See details for the blocks and groups terminology used here.

  • A formula in the form of:

    • DV ~ groups for rank_eta/epsilon_squared().

    • DV ~ groups | blocks for kendalls_w() (See details for the blocks and groups terminology used here).

groups, blocks

A factor vector giving the group / block for the corresponding elements of x, or a character name of one in data. Ignored if x is not a vector.

data

An optional data frame containing the variables.

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "two.sided" (default, two-sided CI), "greater" or "less" (one-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

iterations

The number of bootstrap replicates for computing confidence intervals. Only applies when ci is not NULL.

verbose

Toggle warnings and messages on or off.

...

Arguments passed to or from other methods. When x is a formula, these can be subset and na.action.

blocks_on_rows

Are blocks on rows (TRUE) or columns (FALSE).

Details

The rank epsilon squared and rank eta squared are appropriate for non-parametric tests of differences between 2 or more samples (a rank based ANOVA). See stats::kruskal.test. Values range from 0 to 1, with larger values indicating larger differences between groups.

Kendall's W is appropriate for non-parametric tests of differences between 2 or more dependent samples (a rank based rmANOVA), where each group (e.g., experimental condition) was measured for each block (e.g., subject). This measure is also common as a measure of reliability of the rankings of the groups between raters (blocks). See stats::friedman.test. Values range from 0 to 1, with larger values indicating larger differences between groups / higher agreement between raters.

Value

A data frame with the effect size and its CI.

Confidence (Compatibility) Intervals (CIs)

Confidence intervals for ER2E^2_R, ηH2\eta^2_H, and Kendall's W are estimated using the bootstrap method (using the {boot} package).

Ties

When tied values occur, they are each given the average of the ranks that would have been given had no ties occurred. This results in an effect size of reduced magnitude. A correction has been applied for Kendall's W.

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Bootstrapped CIs

Some effect sizes are directionless–they do have a minimum value that would be interpreted as "no effect", but they cannot cross it. For example, a null value of Kendall's W is 0, indicating no difference between groups, but it can never have a negative value. Same goes for U2 and Overlap: the null value of U2U_2 is 0.5, but it can never be smaller than 0.5; am Overlap of 1 means "full overlap" (no difference), but it cannot be larger than 1.

When bootstrapping CIs for such effect sizes, the bounds of the CIs will never cross (and often will never cover) the null. Therefore, these CIs should not be used for statistical inference.

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

References

  • Kendall, M.G. (1948) Rank correlation methods. London: Griffin.

  • Tomczak, M., & Tomczak, E. (2014). The need to report effect size estimates revisited. An overview of some recommended measures of effect size. Trends in sport sciences, 1(21), 19-25.

See Also

Other rank-based effect sizes: p_superiority(), rank_biserial()

Other effect sizes for ANOVAs: eta_squared()

Examples

# Rank Eta/Epsilon Squared
# ========================

rank_eta_squared(mpg ~ cyl, data = mtcars)

rank_epsilon_squared(mpg ~ cyl, data = mtcars)



# Kendall's W
# ===========
dat <- data.frame(
  cond = c("A", "B", "A", "B", "A", "B"),
  ID = c("L", "L", "M", "M", "H", "H"),
  y = c(44.56, 28.22, 24, 28.78, 24.56, 18.78)
)
(W <- kendalls_w(y ~ cond | ID, data = dat, verbose = FALSE))

interpret_kendalls_w(0.11)
interpret(W, rules = "landis1977")

Fictional Results from a Workers' Randomized Control Trial

Description

Fictional Results from a Workers' Randomized Control Trial

Format

A 2-by-2 table, with a column for each group and a row for the diagnosis.

data("RCT_table")
RCT_table
#>            Group
#> Diagnosis   Treatment Control
#>   Sick             71      30
#>   Recovered        50     100

See Also

Other effect size datasets: Music_preferences, Music_preferences2, Smoking_FASD, food_class, hardlyworking, rouder2016, screening_test


Standardized Mean Differences for Repeated Measures

Description

Compute effect size indices for standardized mean differences in repeated measures data. Pair with any reported stats::t.test(paired = TRUE).

In a repeated-measures design, the same subjects are measured in multiple conditions or time points. Unlike the case of independent groups, there are multiple sources of variation that can be used to standardized the differences between the means of the conditions / times.

Usage

repeated_measures_d(
  x,
  y,
  data = NULL,
  mu = 0,
  method = c("rm", "av", "z", "b", "d", "r"),
  adjust = TRUE,
  ci = 0.95,
  alternative = "two.sided",
  verbose = TRUE,
  ...
)

rm_d(
  x,
  y,
  data = NULL,
  mu = 0,
  method = c("rm", "av", "z", "b", "d", "r"),
  adjust = TRUE,
  ci = 0.95,
  alternative = "two.sided",
  verbose = TRUE,
  ...
)

Arguments

x, y

Paired numeric vectors, or names of ones in data. x can also be a formula:

  • Pair(x,y) ~ 1 for wide data.

  • y ~ condition | id for long data, possibly with repetitions.

data

An optional data frame containing the variables.

mu

a number indicating the true value of the mean (or difference in means if you are performing a two sample test).

method

Method of repeated measures standardized differences. See details.

adjust

Apply Hedges' small-sample bias correction? See hedges_g().

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "two.sided" (default, two-sided CI), "greater" or "less" (one-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

verbose

Toggle warnings and messages on or off.

...

Arguments passed to or from other methods. When x is a formula, these can be subset and na.action.

Value

A data frame with the effect size and their CIs (CI_low and CI_high).

Standardized Mean Differences for Repeated Measures

Unlike Cohen's d for independent groups, where standardization naturally is done by the (pooled) population standard deviation (cf. Glass’s Δ\Delta), when measured across two conditions are dependent, there are many more options for what error term to standardize by. Additionally, some options allow for data to be replicated (many measurements per condition per individual), others require a single observation per condition per individual (aka, paired data; so replications are aggregated).

(It should be noted that all of these have awful and confusing notations.)

Standardize by...

  • Difference Score Variance: dzd_{z} (Requires paired data) - This is akin to computing difference scores for each individual and then computing a one-sample Cohen's d (Cohen, 1988, pp. 48; see examples).

  • Within-Subject Variance: drmd_{rm} (Requires paired data) - Cohen suggested adjusting dzd_{z} to estimate the "standard" between-subjects d by a factor of 2(1r)\sqrt{2(1-r)}, where r is the Pearson correlation between the paired measures (Cohen, 1988, pp. 48).

  • Control Variance: dbd_{b} (aka Becker's d) (Requires paired data) - Standardized by the variance of the control condition (or in a pre- post-treatment setting, the pre-treatment condition). This is akin to Glass' delta (glass_delta()) (Becker, 1988). Note that this is taken here as the second condition (y).

  • Average Variance: davd_{av} (Requires paired data) - Instead of standardizing by the variance in the of the control (or pre) condition, Cumming suggests standardizing by the average variance of the two paired conditions (Cumming, 2013, pp. 291).

  • All Variance: Just dd - This is the same as computing a standard independent-groups Cohen's d (Cohen, 1988). Note that CIs do account for the dependence, and so are typically more narrow (see examples).

  • Residual Variance: drd_{r} (Requires data with replications) - Divide by the pooled variance after all individual differences have been partialled out (i.e., the residual/level-1 variance in an ANOVA or MLM setting). In between-subjects designs where each subject contributes a single response, this is equivalent to classical Cohen’s d. Priors in the BayesFactor package are defined on this scale (Rouder et al., 2012).

    Note that for paired data, when the two conditions have equal variance, drmd_{rm}, davd_{av}, dbd_{b} are equal to dd.

Confidence (Compatibility) Intervals (CIs)

Confidence intervals are estimated using the standard normal parametric method (see Algina & Keselman, 2003; Becker, 1988; Cooper et al., 2009; Hedges & Olkin, 1985; Pustejovsky et al., 2014).

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

Note

rm_d() is an alias for repeated_measures_d().

References

  • Algina, J., & Keselman, H. J. (2003). Approximate confidence intervals for effect sizes. Educational and Psychological Measurement, 63(4), 537-553.

  • Becker, B. J. (1988). Synthesizing standardized mean‐change measures. British Journal of Mathematical and Statistical Psychology, 41(2), 257-278.

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

  • Cooper, H., Hedges, L., & Valentine, J. (2009). Handbook of research synthesis and meta-analysis. Russell Sage Foundation, New York.

  • Cumming, G. (2013). Understanding the new statistics: Effect sizes, confidence intervals, and meta-analysis. Routledge.

  • Hedges, L. V. & Olkin, I. (1985). Statistical methods for meta-analysis. Orlando, FL: Academic Press.

  • Pustejovsky, J. E., Hedges, L. V., & Shadish, W. R. (2014). Design-comparable effect sizes in multiple baseline designs: A general modeling framework. Journal of Educational and Behavioral Statistics, 39(5), 368-393.

  • Rouder, J. N., Morey, R. D., Speckman, P. L., & Province, J. M. (2012). Default Bayes factors for ANOVA designs. Journal of mathematical psychology, 56(5), 356-374.

See Also

cohens_d(), and lmeInfo::g_mlm() and emmeans::effsize() for more flexible methods.

Other standardized differences: cohens_d(), mahalanobis_d(), means_ratio(), p_superiority(), rank_biserial()

Examples

# Paired data -------

data("sleep")
sleep2 <- reshape(sleep,
  direction = "wide",
  idvar = "ID", timevar = "group"
)

repeated_measures_d(Pair(extra.1, extra.2) ~ 1, data = sleep2)

# Same as:
# repeated_measures_d(sleep$extra[sleep$group==1],
#                     sleep$extra[sleep$group==2])
# repeated_measures_d(extra ~ group | ID, data = sleep)


# More options:
repeated_measures_d(Pair(extra.1, extra.2) ~ 1, data = sleep2, mu = -1)
repeated_measures_d(Pair(extra.1, extra.2) ~ 1, data = sleep2, alternative = "less")

# Other methods
repeated_measures_d(Pair(extra.1, extra.2) ~ 1, data = sleep2, method = "av")
repeated_measures_d(Pair(extra.1, extra.2) ~ 1, data = sleep2, method = "b")
repeated_measures_d(Pair(extra.1, extra.2) ~ 1, data = sleep2, method = "d")
repeated_measures_d(Pair(extra.1, extra.2) ~ 1, data = sleep2, method = "z", adjust = FALSE)

# d_z is the same as Cohen's d for one sample (of individual difference):
cohens_d(extra.1 - extra.2 ~ 1, data = sleep2)



# Repetition data -----------

data("rouder2016")

# For rm, ad, z, b, data is aggregated
repeated_measures_d(rt ~ cond | id, data = rouder2016)

# same as:
rouder2016_wide <- tapply(rouder2016[["rt"]], rouder2016[1:2], mean)
repeated_measures_d(rouder2016_wide[, 1], rouder2016_wide[, 2])

# For r or d, data is not aggragated:
repeated_measures_d(rt ~ cond | id, data = rouder2016, method = "r")
repeated_measures_d(rt ~ cond | id, data = rouder2016, method = "d", adjust = FALSE)

# d is the same as Cohen's d for two independent groups:
cohens_d(rt ~ cond, data = rouder2016, ci = NULL)

Jeff Rouder's Example Dataset for Repeated Measures

Description

A dataset "with 25 people each observing 50 trials in 2 conditions", published as effectSizePuzzler.txt by Jeff Rouder on March 24, 2016 (http://jeffrouder.blogspot.com/2016/03/the-effect-size-puzzler.html).

The data is used in examples and tests of rm_d().

Format

A data frame with 2500 rows and 3 variables:

id

participant: 1...25

cond

condition: 1,2

rt

response time in seconds

data("rouder2016")
head(rouder2016, n = 5)
#>   id cond    rt
#> 1  1    1 0.560
#> 2  1    1 0.930
#> 3  1    1 0.795
#> 4  1    1 0.615
#> 5  1    1 1.028

See Also

Other effect size datasets: Music_preferences, Music_preferences2, RCT_table, Smoking_FASD, food_class, hardlyworking, screening_test


Create an Interpretation Grid

Description

Create a container for interpretation rules of thumb. Usually used in conjunction with interpret.

Usage

rules(values, labels = NULL, name = NULL, right = TRUE)

is.rules(x)

Arguments

values

Vector of reference values (edges defining categories or critical values).

labels

Labels associated with each category. If NULL, will try to infer it from values (if it is a named vector or a list), otherwise, will return the breakpoints.

name

Name of the set of rules (will be printed).

right

logical, for threshold-type rules, indicating if the thresholds themselves should be included in the interval to the right (lower values) or in the interval to the left (higher values).

x

An arbitrary R object.

See Also

interpret()

Examples

rules(c(0.05), c("significant", "not significant"), right = FALSE)
rules(c(0.2, 0.5, 0.8), c("small", "medium", "large"))
rules(c("small" = 0.2, "medium" = 0.5), name = "Cohen's Rules")

Results from 2 Screening Tests

Description

A sample (simulated) dataset, used in tests and some examples.

Format

A data frame with 1600 rows and 3 variables:

Diagnosis

Ground truth

Test1

Results given by the 1st test

Test2

Results given by the 2nd test

data("screening_test")
head(screening_test, n = 5)
#>   Diagnosis Test1 Test2
#> 1       Neg "Neg" "Neg"
#> 2       Neg "Neg" "Neg"
#> 3       Neg "Neg" "Neg"
#> 4       Neg "Neg" "Neg"
#> 5       Neg "Neg" "Neg"

See Also

Other effect size datasets: Music_preferences, Music_preferences2, RCT_table, Smoking_FASD, food_class, hardlyworking, rouder2016


Pooled Indices of (Co)Deviation

Description

The Pooled Standard Deviation is a weighted average of standard deviations for two or more groups, assumed to have equal variance. It represents the common deviation among the groups, around each of their respective means.

Usage

sd_pooled(x, y = NULL, data = NULL, verbose = TRUE, ...)

mad_pooled(x, y = NULL, data = NULL, constant = 1.4826, verbose = TRUE, ...)

cov_pooled(x, y = NULL, data = NULL, verbose = TRUE, ...)

Arguments

x, y

A numeric vector, or a character name of one in data. Any missing values (NAs) are dropped from the resulting vector. x can also be a formula (see stats::t.test()), in which case y is ignored.

data

An optional data frame containing the variables.

verbose

Toggle warnings and messages on or off.

...

Arguments passed to or from other methods. When x is a formula, these can be subset and na.action.

constant

scale factor.

Details

The standard version is calculated as:

(xixˉ)2n1+n22\sqrt{\frac{\sum (x_i - \bar{x})^2}{n_1 + n_2 - 2}}

The robust version is calculated as:

1.4826×Median({xMedianx,yMediany})1.4826 \times Median(|\left\{x - Median_x,\,y - Median_y\right\}|)

Value

Numeric, the pooled standard deviation. For cov_pooled() a matrix.

See Also

cohens_d(), mahalanobis_d()

Examples

sd_pooled(mpg ~ am, data = mtcars)
mad_pooled(mtcars$mpg, factor(mtcars$am))

cov_pooled(mpg + hp + cyl ~ am, data = mtcars)

Frequency of FASD for Smoking Mothers

Description

Fictional data.

Format

A 1-by-3 table, with a column for each diagnosis.

data("Smoking_FASD")
Smoking_FASD
#>  FAS PFAS   TD 
#>   17   11  640

See Also

Other effect size datasets: Music_preferences, Music_preferences2, RCT_table, food_class, hardlyworking, rouder2016, screening_test


Convert t, z, and F to Cohen's d or partial-r

Description

These functions are convenience functions to convert t, z and F test statistics to Cohen's d and partial r. These are useful in cases where the data required to compute these are not easily available or their computation is not straightforward (e.g., in liner mixed models, contrasts, etc.).
See Effect Size from Test Statistics vignette.

Usage

t_to_d(t, df_error, paired = FALSE, ci = 0.95, alternative = "two.sided", ...)

z_to_d(z, n, paired = FALSE, ci = 0.95, alternative = "two.sided", ...)

F_to_d(
  f,
  df,
  df_error,
  paired = FALSE,
  ci = 0.95,
  alternative = "two.sided",
  ...
)

t_to_r(t, df_error, ci = 0.95, alternative = "two.sided", ...)

z_to_r(z, n, ci = 0.95, alternative = "two.sided", ...)

F_to_r(f, df, df_error, ci = 0.95, alternative = "two.sided", ...)

Arguments

t, f, z

The t, the F or the z statistics.

paired

Should the estimate account for the t-value being testing the difference between dependent means?

ci

Confidence Interval (CI) level

alternative

a character string specifying the alternative hypothesis; Controls the type of CI returned: "two.sided" (default, two-sided CI), "greater" or "less" (one-sided CI). Partial matching is allowed (e.g., "g", "l", "two"...). See One-Sided CIs in effectsize_CIs.

...

Arguments passed to or from other methods.

n

The number of observations (the sample size).

df, df_error

Degrees of freedom of numerator or of the error estimate (i.e., the residuals).

Details

These functions use the following formulae to approximate r and d:

rpartial=t/t2+dferrorr_{partial} = t / \sqrt{t^2 + df_{error}}



rpartial=z/z2+Nr_{partial} = z / \sqrt{z^2 + N}



d=2t/dferrord = 2 * t / \sqrt{df_{error}}



dz=t/dferrord_z = t / \sqrt{df_{error}}



d=2z/Nd = 2 * z / \sqrt{N}

The resulting d effect size is an approximation to Cohen's d, and assumes two equal group sizes. When possible, it is advised to directly estimate Cohen's d, with cohens_d(), emmeans::eff_size(), or similar functions.

Value

A data frame with the effect size(s)(r or d), and their CIs (CI_low and CI_high).

Confidence (Compatibility) Intervals (CIs)

Unless stated otherwise, confidence (compatibility) intervals (CIs) are estimated using the noncentrality parameter method (also called the "pivot method"). This method finds the noncentrality parameter ("ncp") of a noncentral t, F, or χ2\chi^2 distribution that places the observed t, F, or χ2\chi^2 test statistic at the desired probability point of the distribution. For example, if the observed t statistic is 2.0, with 50 degrees of freedom, for which cumulative noncentral t distribution is t = 2.0 the .025 quantile (answer: the noncentral t distribution with ncp = .04)? After estimating these confidence bounds on the ncp, they are converted into the effect size metric to obtain a confidence interval for the effect size (Steiger, 2004).

For additional details on estimation and troubleshooting, see effectsize_CIs.

CIs and Significance Tests

"Confidence intervals on measures of effect size convey all the information in a hypothesis test, and more." (Steiger, 2004). Confidence (compatibility) intervals and p values are complementary summaries of parameter uncertainty given the observed data. A dichotomous hypothesis test could be performed with either a CI or a p value. The 100 (1 - α\alpha)% confidence interval contains all of the parameter values for which p > α\alpha for the current data and model. For example, a 95% confidence interval contains all of the values for which p > .05.

Note that a confidence interval including 0 does not indicate that the null (no effect) is true. Rather, it suggests that the observed data together with the model and its assumptions combined do not provided clear evidence against a parameter value of 0 (same as with any other value in the interval), with the level of this evidence defined by the chosen α\alpha level (Rafi & Greenland, 2020; Schweder & Hjort, 2016; Xie & Singh, 2013). To infer no effect, additional judgments about what parameter values are "close enough" to 0 to be negligible are needed ("equivalence testing"; Bauer & Kiesser, 1996).

Plotting with see

The see package contains relevant plotting functions. See the plotting vignette in the see package.

References

  • Friedman, H. (1982). Simplified determinations of statistical power, magnitude of effect and research sample sizes. Educational and Psychological Measurement, 42(2), 521-526. doi:10.1177/001316448204200214

  • Wolf, F. M. (1986). Meta-analysis: Quantitative methods for research synthesis (Vol. 59). Sage.

  • Rosenthal, R. (1994) Parametric measures of effect size. In H. Cooper and L.V. Hedges (Eds.). The handbook of research synthesis. New York: Russell Sage Foundation.

  • Steiger, J. H. (2004). Beyond the F test: Effect size confidence intervals and tests of close fit in the analysis of variance and contrast analysis. Psychological Methods, 9, 164-182.

  • Cumming, G., & Finch, S. (2001). A primer on the understanding, use, and calculation of confidence intervals that are based on central and noncentral distributions. Educational and Psychological Measurement, 61(4), 532-574.

See Also

cohens_d()

Other effect size from test statistic: F_to_eta2(), chisq_to_phi()

Examples

## t Tests
res <- t.test(1:10, y = c(7:20), var.equal = TRUE)
t_to_d(t = res$statistic, res$parameter)
t_to_r(t = res$statistic, res$parameter)
t_to_r(t = res$statistic, res$parameter, alternative = "less")

res <- with(sleep, t.test(extra[group == 1], extra[group == 2], paired = TRUE))
t_to_d(t = res$statistic, res$parameter, paired = TRUE)
t_to_r(t = res$statistic, res$parameter)
t_to_r(t = res$statistic, res$parameter, alternative = "greater")


## Linear Regression
model <- lm(rating ~ complaints + critical, data = attitude)
(param_tab <- parameters::model_parameters(model))

(rs <- t_to_r(param_tab$t[2:3], param_tab$df_error[2:3]))

# How does this compare to actual partial correlations?
correlation::correlation(attitude,
  select = "rating",
  select2 = c("complaints", "critical"),
  partial = TRUE
)

Convert Between Effect Sizes for Contingency Tables Correlations

Description

Enables a conversion between different indices of effect size, such as Cohen's w to פ (Fei), and Cramer's V to Tschuprow's T.

Usage

w_to_fei(w, p)

w_to_v(w, nrow, ncol)

w_to_t(w, nrow, ncol)

w_to_c(w)

fei_to_w(fei, p)

v_to_w(v, nrow, ncol)

t_to_w(t, nrow, ncol)

c_to_w(c)

v_to_t(v, nrow, ncol)

t_to_v(t, nrow, ncol)

Arguments

w, c, v, t, fei

Effect size to be converted

p

Vector of expected values. See stats::chisq.test().

nrow, ncol

The number of rows/columns in the contingency table.

References

  • Ben-Shachar, M.S., Patil, I., Thériault, R., Wiernik, B.M., Lüdecke, D. (2023). Phi, Fei, Fo, Fum: Effect Sizes for Categorical Data That Use the Chi‑Squared Statistic. Mathematics, 11, 1982. doi:10.3390/math11091982

  • Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd Ed.). New York: Routledge.

See Also

cramers_v() chisq_to_fei()

Other convert between effect sizes: d_to_r(), diff_to_cles, eta2_to_f2(), odds_to_probs(), oddsratio_to_riskratio()

Examples

library(effectsize)

## 2D tables
## ---------
data("Music_preferences2")
Music_preferences2

cramers_v(Music_preferences2, adjust = FALSE)

v_to_t(0.80, 3, 4)

tschuprows_t(Music_preferences2)



## Goodness of fit
## ---------------
data("Smoking_FASD")
Smoking_FASD

cohens_w(Smoking_FASD, p = c(0.015, 0.010, 0.975))

w_to_fei(0.11, p = c(0.015, 0.010, 0.975))

fei(Smoking_FASD, p = c(0.015, 0.010, 0.975))


## Power analysis
## --------------
# See https://osf.io/cg64s/

p0 <- c(0.35, 0.65)
Fei <- 0.3

pwr::pwr.chisq.test(
  w = fei_to_w(Fei, p = p0),
  df = length(p0) - 1,
  sig.level = 0.01,
  power = 0.85
)