
Robustness of an inference: percent bias to invalidate, RIR and ITCV
Source:R/Konfound.R
Konfound.RdWith the threshold estimate delta# = t_crit * se, Frank et al give
the percent bias to invalidate an inference and the equivalent number
of cases that would have to be replaced (RIR). Frank (2000) gives
the impact threshold for a confounding variable: an omitted variable
correlated sqrt(ITCV) with both the predictor and the outcome
would exactly overturn the inference. Degrees of freedom follow the
reference implementation of these indices,
df = n - n_covariates - 3.
Value
List with estimate (percent bias), pct_bias,
rir, rir_cases, delta_threshold, t,
t_crit, r_xy, r_crit, itcv,
r_cv, impact, df, significant,
n, method.
Details
Formula: pct = 100 (1 - delta#/|est|); RIR = n (1 - delta#/|est|); ITCV = (|r_xy| - r#)/(1 - r#) with r# = t_crit/sqrt(t_crit^2 + df).
References
Frank, Maroulis, Duong and Kelcey (2013), What would it take to change an inference?, Educational Evaluation and Policy Analysis 35(4):437-460, doi:10.3102/0162373713493129 ; Frank (2000), Impact of a confounding variable on a regression coefficient, Sociological Methods and Research 29(2):147-194. doi:10.1177/0049124100029002001
Examples
Konfound(est = 5L, se = 5L, n = 5L, threshold = 0.5)
#> $estimate
#> [1] 18.35034
#>
#> $pct_bias
#> [1] 18.35034
#>
#> $rir
#> [1] 0.9175171
#>
#> $rir_cases
#> [1] 0
#>
#> $delta_threshold
#> [1] 4.082483
#>
#> $t
#> [1] 1
#>
#> $t_crit
#> [1] 0.8164966
#>
#> $r_xy
#> [1] 0.5773503
#>
#> $r_crit
#> [1] 0.5
#>
#> $itcv
#> [1] 0.1547005
#>
#> $r_cv
#> [1] 0.3933199
#>
#> $impact
#> [1] 0.1547005
#>
#> $df
#> [1] 2
#>
#> $significant
#> [1] 1
#>
#> $n
#> [1] 5
#>
#> $method
#> [1] "Frank et al (2013) percent bias / RIR and Frank (2000) ITCV"
#>
#> attr(,"class")
#> [1] "morie_rich_result" "list"