Returning to our worked example, we will use indicator variable “R” to denote the missingness of BMI at age 7 years (for example, R=1 if BMI at age 7 years is observed, and 0 otherwise). In this specific example, @@ -196,7 +196,7 @@
You can draw your own DAG using a Shiny app, by editing the text in the box below.
- +As a final check of our mDAG, we will use the midoc
function exploreDAG to explore whether relationships in the
dataset are consistent with the proposed mDAG, specifying both our mDAG
@@ -254,7 +254,7 @@
Tips for specifying a “missingness” DAG
First specify the DAG for the analysis model, as it would be if @@ -368,13 +368,11 @@
FALSE). Further note that if you do not upload a
file, the default bmi dataset will be
analysed.
-
+
Tips for imputation model variable selection
plot is set to FALSE). Further note that if
you do not upload a file, the default bmi dataset will be
analysed.
-
+
Note Given multiple partially observed variables, we
can specify a list of imputation models (one for each partially observed
variable) in proposeMI (although note that the Shiny app
@@ -827,7 +825,7 @@
bmi dataset will be analysed.
-
+
Check complete records analysis is valid under the proposed analysis model and directed acyclic graph (DAG). Validity means that complete records -analysis will allow unbiased estimation of regression parameters, -associations, and causal effects.
+analysis will allow unbiased estimation of the estimand(s) of interest, +including regression parameters, associations, and causal effects.checkCRA(y="bmi7", covs="matage", r_cra="r",
+ # Example DAG for which complete records analysis is not valid
+checkCRA(y="bmi7", covs="matage", r_cra="r",
mdag="matage -> bmi7 mated -> matage mated -> bmi7 sep_unmeas ->
mated sep_unmeas -> r")
#> Based on the proposed directed acyclic graph (DAG), the analysis model
@@ -118,6 +119,8 @@ Examples#>
#> { mated }
#> { mated, sep_unmeas }
+ ## For the DAG in the example above, complete records analysis is valid
+ ## if a different set of covariates is used
checkCRA(y="bmi7", covs="matage mated", r_cra="r",
mdag="matage -> bmi7 mated -> matage mated -> bmi7 sep_unmeas ->
mated sep_unmeas -> r")
@@ -125,6 +128,8 @@ Examples#> outcome and complete record indicator are independent given analysis
#> model covariates. Hence, complete records analysis is valid.
#>
+
+# Example DAG for which complete records analysis is not valid
checkCRA(y="bmi7", covs="matage mated", r_cra="r",
mdag="matage -> bmi7 mated -> matage mated -> bmi7 sep_unmeas ->
mated sep_unmeas -> r bmi7 -> r")
@@ -140,11 +145,12 @@ Examples#>
#> There are no other variables which could be added to the model to make
#> the analysis model outcome and complete record indicator conditionally
-#> independent, without changing the estimand of interest.
-#>
-#> Consider using a different strategy e.g. multiple imputation.
+#> independent. Consider using a different strategy e.g. multiple imputation.
#>
- checkCRA(y="bmi7", covs="matage mated", r_cra="r",
+
+# Example DAG for which complete records is not valid, but could be valid
+ ## for a different estimand
+checkCRA(y="bmi7", covs="matage mated", r_cra="r",
mdag="matage -> bmi7 mated -> matage mated -> bmi7 sep_unmeas ->
mated sep_unmeas -> r matage -> bmi3 mated -> bmi3 bmi3 -> bmi7
bmi3 -> r")
@@ -163,11 +169,9 @@ Examples#> independent, without changing the estimand of interest.
#>
#> Consider using a different strategy e.g. multiple imputation.
-#>
-#> Consider whether an estimand based on a different set of variables
-#> could be of interest. For example, the analysis model outcome and
-#> complete record indicator are independent given each of the following
-#> sets of variables:
+#> Alternatively, consider whether a different estimand could be of
+#> interest. For example, the analysis model outcome and complete record
+#> indicator are independent given each of the following sets of variables:
#>
#> { bmi3, mated }
#> { bmi3, matage, mated }
diff --git a/reference/checkmi.html b/reference/checkmi.html
index 2e30895..1477f43 100644
--- a/reference/checkmi.html
+++ b/reference/checkmi.html
@@ -98,13 +98,16 @@ Details
Examples
- checkMI(dep="bmi7", preds="matage mated pregsize", r_dep="r",
+ # Example DAG for which multiple imputation is valid
+checkMI(dep="bmi7", preds="matage mated pregsize", r_dep="r",
mdag="matage -> bmi7 mated -> matage mated -> bmi7 sep_unmeas ->
mated sep_unmeas -> r pregsize -> bmi7 pregsize -> bwt sep_unmeas -> bwt")
#> Based on the proposed directed acyclic graph (DAG), the incomplete
#> variable and its missingness indicator are independent given imputation
#> model predictors. Hence, multiple imputation methods which assume data
#> are missing at random are valid in principle.
+
+# Example DAG for which multiple imputation is not valid, due to a collider
checkMI(dep="bmi7", preds="matage mated bwt", r_dep="r",
mdag="matage -> bmi7 mated -> matage mated -> bmi7 sep_unmeas ->
mated sep_unmeas -> r pregsize -> bmi7 pregsize -> bwt sep_unmeas -> bwt")
@@ -124,13 +127,6 @@ Examples#>
#> { pregsize }
#> { pregsize, sep_unmeas }
-checkMI(dep="matage", preds="bmi7 mated", r_dep="r",
- mdag="matage -> bmi7 mated -> matage mated -> bmi7 sep_unmeas ->
- mated sep_unmeas -> r pregsize -> bmi7 pregsize -> bwt sep_unmeas -> bwt")
-#> Based on the proposed directed acyclic graph (DAG), the incomplete
-#> variable and its missingness indicator are independent given imputation
-#> model predictors. Hence, multiple imputation methods which assume data
-#> are missing at random are valid in principle.