What is "std.beta" in the output of report()? about report

[DominiqueMakowski]

@mooninnap good question ☺️

Currently, in report, in the case of GLMs, we used "beta" as a synonym for coefficient. This comes from the fact that betas are the symbols traditionally associated with the coefficients (see here for instance) in the mathematical formulas for regressions. "Std. beta" refers to the standardized coefficients as computed by the effectsize package.

However, I am aware that in some textbooks people have used the term "beta" to refer to the standardized coefficient. However, to my knowledge, this is an arbitrary convention which seems statistically rather confusing. I'd be interested to know more about the reasons for that denomination!

That being said, it's true that that beta might not be the optimal name, and I wonder if the APA7 finally came up with some clear guidelines regarding it. In the early stages of this package, we decided not to use "coefficients" as this doesn't apply to all cases (for instance, the beta of the intercept is not, technically, a coefficient).

[chaohanch]

@DominiqueMakowski Thank you for your response! I compared the output by standardize_parameters() in the "effectsize" package to the output by report(); I found the standardized coefficients by the "effectsize" package are the same as "beta" but not "std.beta" by report(). That's why I'm confused. Could you explain it a bit more?

Here is the output by standardize_parameters() in the "effectsize" package:

> standardize_parameters(mod.full)
          Parameter Std_Coefficient
1       (Intercept)      0.69407022
2    grammarUnGramm     -1.54836958
3    proficiencymid      0.16824769
4   proficiencyhigh      0.05134071
5 proficiencynative      0.05731125

Here is part of the output by report():

  - The effect of grammarUnGramm is negative and can be considered as very large and significant (beta = -1.55, SE = 0.14, 95% CI [-1.83, -1.27], std. beta = -1.55, p < .001).
  - The effect of proficiencymid is positive and can be considered as tiny and not significant (beta = 0.17, SE = 0.09, 95% CI [0.00, 0.34], std. beta = 0.05, p = 0.051).
  - The effect of proficiencyhigh is positive and can be considered as very small and not significant (beta = 0.05, SE = 0.10, 95% CI [-0.15, 0.25], std. beta = 0.17, p = 0.618).
  - The effect of proficiencynative is positive and can be considered as tiny and not significant (beta = 0.06, SE = 0.08, 95% CI [-0.10, 0.22], std. beta = 0.06, p = 0.480).

[strengejacke]

One convention I'm aware of is to refer to unstandardized a "B" (or "b"), and standardized are "Beta".

[DominiqueMakowski]

One convention I'm aware of is to refer to unstandardized a "B" (or "b"), and standardized are "Beta".

I've heard about this one, but that's confusing at best IMO. Unless made very explicit, referring as "something" and "std. something" seems better.*

Here is the output by standardize_parameters() in the "effectsize" package:

Mmh, intriguing, it seems like there might be an alignment issue in when the outputs of parameters and effectsize are merged 😕

[strengejacke]

Mmh, intriguing, it seems like there might be an alignment issue in when the outputs of parameters and effectsize are merged 😕

I always tend to use an extra row-ID, and always check whether sort should be set to FALSE in merge. We had these alignment issues quite for some time in model_parameters()...

[DominiqueMakowski]

report/R/model_table_regression.R

Lines 50 to 53 in 081adc5

	effsize <- effectsize::standardize_parameters(model, method = "refit", robust = FALSE, two_sd = FALSE, centrality = centrality, ...)
	effsize <- effsize[order(parameters$Parameter), ]
	effsize$Parameter <- NULL
	parameters <- cbind(parameters, effsize)

And yet we order by parameter 🤔 🤔

[strengejacke]

well, you retrieve the alphabetical order of parameter names from effsize. However, the order of parameter names is the same for both objects, effsize and parameters. So you actually shuffle the order of effsize, so it no longer matches the correct order.

[strengejacke]

Current behaviour:

library(easystats)
#> # Attaching packages
#> <U+2714> insight     0.8.1.1   <U+2714> bayestestR  0.5.2  
#> <U+2714> performance 0.4.4.1   <U+2714> parameters  0.5.0.1
#> <U+2714> see         0.4.0.1   <U+2714> effectsize  0.2.0  
#> <U+2714> modelbased  0.1.1     <U+2714> correlation 0.1.0  
#> <U+2714> report      0.1.0     
#> Warnings or errors in CRAN checks for package(s) 'bayestestR', 'parameters'.

data(iris)
iris$Proficiency <- as.factor(sample(1:4, nrow(iris), replace = TRUE))
levels(iris$Proficiency) <- c("low", "mid", "high", "native")

m <- lm(Sepal.Width ~ Sepal.Length + Proficiency, data= iris)

model_parameters(m)
#> Parameter            | Coefficient |   SE |        95% CI |     t |  df |      p
#> --------------------------------------------------------------------------------
#> (Intercept)          |        3.30 | 0.25 | [ 2.80, 3.80] | 13.02 | 145 | < .001
#> Sepal.Length         |       -0.06 | 0.04 | [-0.15, 0.02] | -1.52 | 145 | 0.130 
#> Proficiency [mid]    |        0.25 | 0.10 | [ 0.06, 0.43] |  2.58 | 145 | 0.011 
#> Proficiency [high]   |        0.19 | 0.10 | [-0.01, 0.38] |  1.89 | 145 | 0.061 
#> Proficiency [native] |        0.13 | 0.10 | [-0.06, 0.32] |  1.35 | 145 | 0.180

standardize_parameters(m)
#>           Parameter Std_Coefficient
#> 1       (Intercept)      -0.3057256
#> 2      Sepal.Length      -0.1233045
#> 3    Proficiencymid       0.5647469
#> 4   Proficiencyhigh       0.4249323
#> 5 Proficiencynative       0.3004431

report(m)
#> We fitted a linear model (estimated using OLS) to predict Sepal.Width with Sepal.Length and Proficiency (formula = Sepal.Width ~ Sepal.Length + Proficiency). Standardized parameters were obtained by fitting the model on a standardized version of the dataset. Effect sizes were labelled following Funder's (2019) recommendations.
#> 
#> The model explains a not significant and weak proportion of variance (R2 = 0.06, F(4, 145) = 2.37, p = 0.055, adj. R2 = 0.04). The model's intercept, corresponding to Sepal.Width = 0, Sepal.Length = 0 and Proficiency = low, is at 3.30 (SE = 0.25, 95% CI [2.80, 3.80], p < .001). Within this model:
#> 
#>   - The effect of Sepal.Length is negative and can be considered as medium and not significant (beta = -0.12, SE = 0.08, 95% CI [-0.28, 0.04], std. beta = 0.42, p = 0.130).
#>   - The effect of Proficiencymid is positive and can be considered as medium and significant (beta = 0.56, SE = 0.22, 95% CI [0.13, 1.00], std. beta = 0.56, p < .05).
#>   - The effect of Proficiencyhigh is positive and can be considered as small and not significant (beta = 0.42, SE = 0.22, 95% CI [-0.02, 0.87], std. beta = 0.30, p = 0.061).
#>   - The effect of Proficiencynative is positive and can be considered as very small and not significant (beta = 0.30, SE = 0.22, 95% CI [-0.14, 0.74], std. beta = -0.12, p = 0.180).

^{Created on 2020-02-25 by the reprex package (v0.3.0)}

[strengejacke]

should work now:

library(easystats)
#> # Attaching packages
#> <U+2714> insight     0.8.1.1   <U+2714> bayestestR  0.5.2  
#> <U+2714> performance 0.4.4.1   <U+2714> parameters  0.5.0.1
#> <U+2714> see         0.4.0.1   <U+2714> effectsize  0.2.0  
#> <U+2714> modelbased  0.1.1     <U+2714> correlation 0.1.0  
#> <U+2714> report      0.1.0     
#> Warnings or errors in CRAN checks for package(s) 'bayestestR', 'parameters'.

data(iris)
iris$Proficiency <- as.factor(sample(1:4, nrow(iris), replace = TRUE))
levels(iris$Proficiency) <- c("low", "mid", "high", "native")

m <- lm(Sepal.Width ~ Sepal.Length + Proficiency, data= iris)

model_parameters(m)
#> Parameter            | Coefficient |   SE |        95% CI |     t |  df |      p
#> --------------------------------------------------------------------------------
#> (Intercept)          |        3.23 | 0.25 | [ 2.73, 3.72] | 12.82 | 145 | < .001
#> Sepal.Length         |       -0.05 | 0.04 | [-0.14, 0.03] | -1.32 | 145 | 0.190 
#> Proficiency [mid]    |        0.37 | 0.10 | [ 0.19, 0.56] |  3.92 | 145 | < .001
#> Proficiency [high]   |        0.24 | 0.11 | [ 0.03, 0.45] |  2.23 | 145 | 0.027 
#> Proficiency [native] |        0.13 | 0.08 | [-0.04, 0.30] |  1.54 | 145 | 0.125

standardize_parameters(m)
#>           Parameter Std_Coefficient
#> 1       (Intercept)      -0.3428907
#> 2      Sepal.Length      -0.1035683
#> 3    Proficiencymid       0.8590401
#> 4   Proficiencyhigh       0.5420976
#> 5 Proficiencynative       0.2986141

report(m)
#> We fitted a linear model (estimated using OLS) to predict Sepal.Width with Sepal.Length and Proficiency (formula = Sepal.Width ~ Sepal.Length + Proficiency). Standardized parameters were obtained by fitting the model on a standardized version of the dataset. Effect sizes were labelled following Funder's (2019) recommendations.
#> 
#> The model explains a significant and weak proportion of variance (R2 = 0.11, F(4, 145) = 4.70, p = 0.001, adj. R2 = 0.09). The model's intercept, corresponding to Sepal.Width = 0, Sepal.Length = 0 and Proficiency = low, is at 3.23 (SE = 0.25, 95% CI [2.73, 3.72], p < .001). Within this model:
#> 
#>   - The effect of Sepal.Length is negative and can be considered as very small and not significant (beta = -0.05, SE = 0.04, 95% CI [-0.14, 0.03], std. beta = -0.10, p = 0.190).
#>   - The effect of Proficiencymid is positive and can be considered as large and significant (beta = 0.37, SE = 0.10, 95% CI [0.19, 0.56], std. beta = 0.86, p < .001).
#>   - The effect of Proficiencyhigh is positive and can be considered as medium and significant (beta = 0.24, SE = 0.11, 95% CI [0.03, 0.45], std. beta = 0.54, p < .05).
#>   - The effect of Proficiencynative is positive and can be considered as small and not significant (beta = 0.13, SE = 0.08, 95% CI [-0.04, 0.30], std. beta = 0.30, p = 0.125).

^{Created on 2020-02-25 by the reprex package (v0.3.0)}

[chaohanch]

I updated the package but "beta" by report() still matches "std.beta" by standardize_parameters(). Here is what I got:

library(easystats)
# # Attaching packages
# ✔ insight     0.8.1.1   ✔ bayestestR  0.5.2  
# ✔ performance 0.4.4.1   ✔ parameters  0.5.0.1
# ✔ see         0.4.0.1   ✔ effectsize  0.2.0  
# ✔ modelbased  0.1.1     ✔ report      0.1.0  
# Warnings or errors in CRAN checks for package(s) 'bayestestR', 'parameters'.


data(iris)
iris$Proficiency <- as.factor(sample(1:4, nrow(iris), replace = TRUE))
levels(iris$Proficiency) <- c("low", "mid", "high", "native")

m <- lm(Sepal.Width ~ Sepal.Length + Proficiency, data= iris)

model_parameters(m)
# Parameter            | Coefficient |   SE |        95% CI |     t |  df |      p
# --------------------------------------------------------------------------------
#   (Intercept)          |        3.36 | 0.26 | [ 2.84, 3.88] | 12.78 | 145 | < .001
# Sepal.Length         |       -0.06 | 0.04 | [-0.15, 0.02] | -1.43 | 145 | 0.156 
# Proficiency [mid]    |        0.09 | 0.11 | [-0.12, 0.30] |  0.87 | 145 | 0.387 
# Proficiency [high]   |        0.08 | 0.10 | [-0.11, 0.28] |  0.82 | 145 | 0.411 
# Proficiency [native] |        0.07 | 0.10 | [-0.12, 0.26] |  0.76 | 145 | 0.446 

standardize_parameters(m)
# Parameter Std_Coefficient
# 1       (Intercept)      -0.1353245
# 2      Sepal.Length      -0.1174270
# 3    Proficiencymid       0.2091567
# 4   Proficiencyhigh       0.1863652
# 5 Proficiencynative       0.1693196

report(m)
# We fitted a linear model (estimated using OLS) to predict Sepal.Width with Sepal.Length and Proficiency (formula = Sepal.Width ~ Sepal.Length + Proficiency). Standardized parameters were obtained by fitting the model on a standardized version of the dataset. Effect sizes were labelled following Funder's (2019) recommendations.
# 
# The model explains a not significant and weak proportion of variance (R2 = 0.02, F(4, 145) = 0.78, p = 0.543, adj. R2 = -0.01). The model's intercept, corresponding to Sepal.Width = 0, Sepal.Length = 0 and Proficiency = low, is at 3.36 (SE = 0.26, 95% CI [2.84, 3.88], p < .001). Within this model:
#   
#   - The effect of Sepal.Length is negative and can be considered as very small and not significant (beta = -0.12, SE = 0.08, 95% CI [-0.28, 0.05], std. beta = 0.19, p = 0.156).
# - The effect of Proficiencymid is positive and can be considered as small and not significant (beta = 0.21, SE = 0.24, 95% CI [-0.27, 0.69], std. beta = 0.21, p = 0.387).
# - The effect of Proficiencyhigh is positive and can be considered as very small and not significant (beta = 0.19, SE = 0.23, 95% CI [-0.26, 0.63], std. beta = 0.17, p = 0.411).
# - The effect of Proficiencynative is positive and can be considered as very small and not significant (beta = 0.17, SE = 0.22, 95% CI [-0.27, 0.61], std. beta = -0.12, p = 0.446).