mariposa (version 0.7.3)

partial_cor: Partial Correlation

Description

partial_cor() computes partial correlations between variables while controlling for one or more other variables — the SPSS PARTIAL CORR procedure (Stata: pcorr). It answers the question "how strongly are x and y related once the influence of the control variables is removed?"

Usage

partial_cor(data, ..., controls, weights = NULL)

Value

An object of class "partial_cor" whose $correlations tibble holds one row per variable pair (and group combination) with:

partial_r

Partial correlation controlling for the controls

zero_order_r

The ordinary (zero-order) Pearson correlation of the pair, for comparison — SPSS /STATISTICS=CORR

df

Degrees of freedom, n - 2 - k with k control variables (non-integer for weighted data, Charter §5.1)

t_stat, p_value

t-test of the partial correlation (two-tailed, the SPSS default)

n

Listwise-complete sample size (rounded weighted N when weighted)

For three or more analysis variables, $matrices additionally holds the full partial-correlation matrix per group.

Arguments

data

Your survey data (a data frame or tibble). If grouped (via dplyr::group_by()), separate partial correlations are computed per group (SPSS SPLIT FILE).

...

Variables to correlate (unquoted, supports tidyselect). At least two.

controls

Control variable(s) to partial out (unquoted, supports tidyselect; SPSS BY list). At least one, must not overlap with the analysis variables.

weights

Optional survey weights (unquoted variable name), treated as frequency weights matching SPSS WEIGHT BY.

Details

Understanding the Output

Comparing partial_r against zero_order_r tells you what the controls contribute:

• Partial clearly smaller than zero-order: much of the original association runs through the control variables (confounding or mediation).

• Partial similar to zero-order: the association is largely independent of the controls.

• Partial larger than zero-order: a suppressor situation — the controls masked part of the association.

When to Use This

• Check whether a bivariate correlation survives controlling for demographics (age, education, ...)

• Separate the direct association of two attitudes from what a shared cause explains

For full multivariate control with several predictors, use linear_regression instead.

Technical Details

Cases are deleted listwise across the analysis and control variables (SPSS /MISSING=LISTWISE, the PARTIAL CORR default). The partial correlation is computed from the Pearson correlation matrix of all variables via the inverse of the control-variable block; the same Pearson formula as pearson_cor is used throughout, so weighted results follow the SPSS frequency-weight convention with unrounded sum(w) in df and test statistics. Significance is two-tailed (the SPSS default).

An SPSS v29 PARTIAL CORR reference run is pending; until it lands the statistics are verified against the independent residual-of-regressions characterization (see the SPSS compatibility vignette).

See Also

pearson_cor for zero-order correlations.

linear_regression for multivariate control.

summary.partial_cor for detailed output.

Other correlation: kendall_tau(), pearson_cor(), spearman_rho()

Examples

library(dplyr)
data(survey_data)

# Does the satisfaction-income correlation survive controlling for age?
partial_cor(survey_data, life_satisfaction, income, controls = age)

# Several variables, several controls
partial_cor(survey_data, trust_government, trust_media, trust_science,
            controls = c(age, political_orientation))

# Weighted (SPSS WEIGHT BY)
partial_cor(survey_data, life_satisfaction, income,
            controls = age, weights = sampling_weight)

# Grouped (SPSS SPLIT FILE)
survey_data %>%
  group_by(gender) %>%
  partial_cor(life_satisfaction, income, controls = age)

# --- Three-layer output ---
result <- partial_cor(survey_data, life_satisfaction, income, controls = age)
result              # compact overview
summary(result)     # full detailed output