Abridged example structure of a wide dataset formatted as required
for the devMSMs. Column A denotes the ID variable, column B (green)
denotes a time in-variant confounder (e.g., race, birth information),
diff --git a/articles/Specify_Core_Inputs.html b/articles/Specify_Core_Inputs.html
index 9c9c63c9..a0455951 100644
--- a/articles/Specify_Core_Inputs.html
+++ b/articles/Specify_Core_Inputs.html
@@ -38,6 +38,7 @@
Any developmental construct measured at one or more time points including at a final outcome time point
+
Exposure
+
Exposure or experience that constitutes the causal event of interest and is measured at at least two time points, with at least one time point occurring prior to the outcome.
-
Exposure
-
Exposure or experience that constitutes the causal event of interest and is measured at at least two time points, with at least one time point occurring prior to the outcome
+
Outcome
+
Any developmental construct measured at least once at a final outcome time point upon which the exposure is theorized to have causal effects.
Exposure Time Points
-
Time points in development when the exposure was measured at which balancing formulas will be created
+
Time points in development when the exposure was measured, at which balancing formulas will be created.
Exposure Epochs
-(optional) Further delineation of exposure time points into meaningful units of developmental time, each of which could encompass multiple exposure time points, that together constitute exposure main effects in the outcome model and exposure histories
+(optional) Further delineation of exposure time points into meaningful units of developmental time, each of which could encompass multiple exposure time points, that together constitute exposure main effects in the outcome model and exposure histories.
Exposure Histories
-
Sequences of relatively high ('h') or low ('p') levels of exposure at each exposure time point or exposure epoch
+
Sequences of relatively high ('h') or low ('l') levels of exposure at each exposure time point or exposure epoch.
Exposure Dosage
-
Total cumulative exposure epochs during which an individual experienced high (or low) levels of exposure across an entire exposure history
-
-
-
Backdoor Path
-
Any alternative, non-causal way (other than the direct causal effects from exposure to outcome) that the exposure at a given time point associates with the outcome within a causal diagram, often via confounders
+
Total cumulative exposure epochs/time points during which an individual experienced high (or low) levels of exposure, across an entire exposure history.
Confounder
-
Pre-exposure variable that represents a common cause of exposure at a given time point and outcome; adjusting for all of which successfully blocks all backdoor paths
+
Pre-exposure variable that represents a common cause of exposure at a given time point and outcome; adjusting for all of which successfully blocks all backdoor paths.
Time-varying confounder
-
A confounder measured at one or more time points prior to the exposure time point which reflects a construct that varies in time.
+
A confounder that often changes over time (even if it is not measured at every time point), and is affected by prior exposure, either directly or indirectly.
Time invariant confounder
-
A confounder measured at a single time point prior to the exposure time point which reflects a construct that is stable over time.
+
A confounder that occurs only at a single time point, prior to the exposure and remains stable and/or is not possibly affected by exposure.
Collider
-
A variable that represents a common effect of exposure at a given time point and outcome; adjusting for which opens a backdoor path
+
A variable that represents a common effect of exposure at a given time point and outcome; adjusting for which introduces bias.
@@ -182,7 +178,7 @@
Terminology
-
Balancing Formula
+
Balancing Formula
Balancing formula are formula created at each exposure time point, regressing exposure on confounders, for the purpose of creating and/or assessing balancing weights. The general form is given by:
Summary of the formula terminology used in the devMSMs
package for creating balancing weights and assessing their performance
in devMSMs (Workflows vignettes Steps 4-6). The different kinds
@@ -285,20 +281,20 @@
Terminology
-
6 core functions of devMSMs
+
7 core functions of devMSMs
-
There are 6 key functions that are used in the devMSMs workflow.
+
There are 7 key functions that are used in the devMSMs workflow.
-
+
Summary of the 7 core functions of the devMSMs package.
This vignette guides a user through the process of using devMSMs to fit marginal structural models (MSMs) with a a continuously distributed exposure variable. The users should first view the Terminology, Data Requirements, and Specifying Core Inputs vignettes.
+
The code contained in this vignette is also available, integrated code from the other vignettes, in the ExampleWorkflow.rmd file. This workflow is designed to complement the conceptual and high-level practical details provided in the manuscript [add link here]. We strongly suggest users familiarize themselves with concepts of the MSM process outlined in the manuscript and the practical steps and functions put forth in the following sections before implementing this workflow.
+
Installation
+
+
Until devMSMs is available on CRAN, you will need to install it directly from Github (https://github.com/istallworthy/devMSMs), as shown below.
All devMSMs functions have an option to save out objects as ‘.rds’ files. Users can also save out content from print, summary, and plot methods, as illustrated in the sections below. To save, users must supply a path to a home directory (home_dir) when creating their initial MSM object. Users can save to the home directory using the default file labels (and .txt file type) using save.out = TRUE. When saving tables, users have the option to supply their own name and file type (e.g., save.out = “custom_name.png”). Allowable file types are: .png, .html, .pdf, .tex, and .md. All sub-folders referenced by each function are created automatically within the home directory. We recommend saving outputs for future use and provide commented out examples here. When an output is saved out, the function automatically provides a path file to aid the user in reading in that output in the future.
+
Some functions output tables. These are all from the tinytables package and can be further customized (e.g., dimensions, footnotes, captions, combined, etc.) according to the options provided by the package (https://vincentarelbundock.github.io/tinytable/vignettes/tinytable.html).
+
+
+save.out=FALSE
+
+
+
Phase 0: Preliminary Conceptual & Data Preparation
+
+
Please see the accompanying manuscript for steps P1 (creating hypotheses) and P2 (creating a DAG).
+
STEP P3. Specify Core Inputs
+
+
The first step is to create an initial MSM object by specifying the core variables and data for use with the package. Please see the Specifying Core Inputs vignette for more detail on the following core inputs.
+
Below, we specify data, exposure, time invariant and time-varying confounders, as well as exposure epochs.
+
There are several other optional fields that a user could specify in the MSM object.
+
The user also has the option to specify concur_conf, indicating, as a list of character strings, the names of any time-varying confounders (e.g., “variable.time”) they wish to be included concurrently in the weights formulas (overriding the default which is to only include lagged confounders). This should only be done if the user has strong evidence to differentiate confounders from mediators in their relation with the exposure contemporaneously.
+
P3.1 Recommended: Home Directory
+
+
We do not specify a home directory given the nature of this example, but we do recommend doing so to save out core function outputs.
We specify that the first two exposure time points (6 and 15 months) will be considered infancy, the second two (34 and 25 months) toddlerhood, and the final (58 months) childhood.
See the Specify Core Inputs vignette for more information.
+
Below, we specify low economic strain at all epochs (“l-l-l”) as our reference event in comparison to high levels at all epochs (“h-h-h”) to examine an example question comparing the causal effects of 0 vs 3 doses of exposure to economic strain on children’s behavior problems.
We specify no concurrent confounders as, given our data, we are unable to disentangle them from mediators or colliders.
+
+
STEP P4. Data Preparation & Inspection
+
+
P4.3b. Required: Read in Wide Data
+
+
We highly recommend first implementing the Data Requirements & Preparation Vignette https://istallworthy.github.io/devMSMs/articles/Data_Requirements.html before assigning to the variable, data, one of the following wide data formats (see Figure 1) for use in the package:
+
+
a single data frame of data in wide format with no missing data
+
a mids object (output from mice::mice()) of data imputed in wide format
+
a list of data imputed in wide format as data frames
+
+
See the Data Preparation vignette for more detail.
+
We first load in 2 imputed datasets as a mice object. These data are simulated based on data from the Family Life Project (FLP), a longitudinal study following 1,292 families representative of two geographic areas (three counties in North Carolina and three counties in Pennsylvania) with high rural child poverty (Vernon-Feagans et al., 2013; Burchinal et al., 2008). We take the example exposure of economic strain (“ESETA1”) measured at 6, 15, 24, 35, and 58 months in relation to the outcome of behavior problems (“StrDif_Tot”) measured at 58 months. (See Data Requirements & Preparation vignette for beginning with other data types, including missing data).
P4.5. Recommended: Inspect Exposure Histories and Data
+
+
For all users, we highly recommend use of the helper inspectData() function (with the a complete dataset in long or wide format or imputed data in the case of missingness) to summarize exposure, outcome, and confounders and inspect the sample distribution among exposure histories. Based on any user-specified exposure epochs and high and low quantile values (for continuous exposures), this function outputs a table showing the sample distribution across all histories.
+
We strongly suggest visually inspecting this table and revising the designation of epochs and/or high and low quantile values (for continuous exposures) until each history contains a reasonable number of participants. While there is no gold standard required number per history cell, users should guard against extrapolation beyond the scope of the data. For example, in our data, when using 75th and 25th percentile cutoffs, there were histories that represented less than two cases and thus we re-evaluated our cutoffs. Users may wish to revise any epoch designation and high and low cutoff values, where applicable. The function conducts summaries and history distribution inspection for each imputed dataset if imputed data are supplied.
+
The required inputs for inspectData() are: complete data (as a data frame in wide or long format, a list of imputed data frames in wide format, or a mids object), exposure (e.g., “variable”), and outcome (e.g., “variable.t”). If the exposure is continuously distributed, the user is required to supply to hi_lo_cut values demarcating high and low levels.
+
Optional inputs are a home directory (if save.out = TRUE), epochs, high/low cutoff values for continuous exposures, and specification of reference and comparison histories.
+
The helper inspectData() function outputs the following files into the home directory: a correlation plot of all variables in the dataset, tables of exposure and outcome descriptive statistics, and two summary tables of the confounders considered at each time point.
+
+
+inspectData(data =data,
+ obj =obj,
+ outcome =outcome,
+ hi_lo_cut =hi_lo_cut,
+ reference =reference,
+ comparison =comparison,
+ verbose =TRUE,
+ save.out =save.out)
+#> Imputation 1
+#> USER ALERT: Out of the total of 1292 individuals in the sample, below is the distribution of the 333 (26%) individuals that fall into 2 user-selected exposure histories (out of the 27 total) created from 30th and 60th percentile values for low and high levels of exposure-epoch Infancy, Toddlerhood, Childhood.
+#> USER ALERT: Please inspect the distribution of the sample across the following exposure histories and ensure there is sufficient spread to avoid extrapolation and low precision:
+#>
+#> +---------------+-----+
+#> | epoch_history | n |
+#> +===============+=====+
+#> | h-h-h | 214 |
+#> +---------------+-----+
+#> | l-l-l | 119 |
+#> +---------------+-----+
+#>
+#> Table: Summary of user-selected exposure histories based on exposure main effects Infancy, Toddlerhood, Childhood:
+#> USER ALERT: Below are the 55 variables spanning 32 unique domains that will be treated as confounding variables for the relation between ESETA1 and StrDif_Tot.58.
+#> Please inspect this list carefully. It should include all time-varying covariates, time invariant covariates, as well as lagged levels of exposure and outcome variables if they were collected at time points earlier than the outcome time point.
+#> [1] "B18Raw.15" "B18Raw.24" "B18Raw.6" "BioDadInHH2"
+#> [5] "caregiv_health" "CORTB.15" "CORTB.24" "CORTB.6"
+#> [9] "DrnkFreq" "EARS_TJo.24" "EARS_TJo.35" "fscore.35"
+#> [13] "gov_assist" "HOMEETA1.15" "HOMEETA1.24" "HOMEETA1.35"
+#> [17] "HOMEETA1.6" "HomeOwnd" "IBRAttn.15" "IBRAttn.24"
+#> [21] "IBRAttn.6" "InRatioCor.15" "InRatioCor.24" "InRatioCor.35"
+#> [25] "InRatioCor.6" "KFASTScr" "LESMnNeg.24" "LESMnNeg.35"
+#> [29] "LESMnPos.24" "LESMnPos.35" "MDI.15" "MDI.6"
+#> [33] "peri_health" "PmAge2" "PmBlac2" "PmEd2"
+#> [37] "PmMrSt2" "RHasSO.15" "RHasSO.24" "RHasSO.35"
+#> [41] "RHasSO.6" "RHealth" "RMomAgeU" "SAAmylase.15"
+#> [45] "SAAmylase.24" "SAAmylase.6" "SmokTotl" "state"
+#> [49] "StrDif_Tot.35" "SurpPreg" "SWghtLB" "TcBlac2"
+#> [53] "WndNbrhood.24" "WndNbrhood.35" "WndNbrhood.6"
+#>
+#> The following variables are designated as numeric:
+#> [1] "ALI_Le.35, CORTB.15, CORTB.24, CORTB.6, ESETA1.15, ESETA1.24, ESETA1.35, ESETA1.58, ESETA1.6, fscore.35, fscore.58, HOMEETA1.15, HOMEETA1.24, HOMEETA1.35, HOMEETA1.58, HOMEETA1.6, IBRAttn.15, IBRAttn.24, IBRAttn.6, ID, InRatioCor.15, InRatioCor.24, InRatioCor.35, InRatioCor.58, InRatioCor.6, LESMnNeg.24, LESMnNeg.35, LESMnPos.24, LESMnPos.35, PmAge2, SAAmylase.15, SAAmylase.24, SAAmylase.6, StrDif_Tot.35, StrDif_Tot.58, WndNbrhood.24, WndNbrhood.35, WndNbrhood.58, WndNbrhood.6, ESETA1.Childhood, ESETA1.Infancy, ESETA1.Toddlerhood"
+#>
+#> The following variables are designated as factors:
+#> [1] "BioDadInHH2, DrnkFreq, HomeOwnd, PmBlac2, PmMrSt2, RHasSO.15, RHasSO.24, RHasSO.35, RHasSO.58, RHasSO.6, RHealth, SmokTotl, state, SurpPreg, TcBlac2"
+#>
+#> Table: Other variable types
+#>
+#> | |variable |type |
+#> |:--------------|:--------------|:-------|
+#> |B18Raw.15 |B18Raw.15 |integer |
+#> |B18Raw.24 |B18Raw.24 |integer |
+#> |B18Raw.58 |B18Raw.58 |integer |
+#> |B18Raw.6 |B18Raw.6 |integer |
+#> |caregiv_health |caregiv_health |integer |
+#> |EARS_TJo.24 |EARS_TJo.24 |integer |
+#> |EARS_TJo.35 |EARS_TJo.35 |integer |
+#> |gov_assist |gov_assist |integer |
+#> |KFASTScr |KFASTScr |integer |
+#> |MDI.15 |MDI.15 |integer |
+#> |MDI.6 |MDI.6 |integer |
+#> |peri_health |peri_health |integer |
+#> |PmEd2 |PmEd2 |integer |
+#> |RMomAgeU |RMomAgeU |integer |
+#> |SWghtLB |SWghtLB |integer |
+#>
+#> <table class="kable_wrapper">
+#> <caption>Summary of ESETA1 Information</caption>
+#> <tbody>
+#> <tr>
+#> <td>
+#>
+#> | | ESETA1.6 | ESETA1.15 | ESETA1.24 | ESETA1.35 | ESETA1.58 |
+#> |:--|:---------------|:---------------|:---------------|:----------------|:----------------|
+#> | |Min. :-2.8090 |Min. :-3.1010 |Min. :-2.8580 |Min. :-3.24100 |Min. :-2.47800 |
+#> | |1st Qu.:-0.4193 |1st Qu.:-0.4153 |1st Qu.:-0.4665 |1st Qu.:-0.57050 |1st Qu.:-0.51100 |
+#> | |Median : 0.3030 |Median : 0.2165 |Median : 0.1445 |Median : 0.07300 |Median : 0.05650 |
+#> | |Mean : 0.2765 |Mean : 0.2115 |Mean : 0.1154 |Mean : 0.08565 |Mean : 0.07673 |
+#> | |3rd Qu.: 0.9000 |3rd Qu.: 0.8620 |3rd Qu.: 0.7258 |3rd Qu.: 0.74075 |3rd Qu.: 0.71425 |
+#> | |Max. : 4.0350 |Max. : 3.4740 |Max. : 3.2840 |Max. : 3.01400 |Max. : 3.17300 |
+#>
+#> </td>
+#> </tr>
+#> </tbody>
+#> </table>
+#>
+#> Table: Summary of StrDif_Tot.58 Information
+#>
+#> | Min.| X1st.Qu.| Median| Mean| X3rd.Qu.| Max.|
+#> |------:|--------:|------:|---------:|--------:|-----:|
+#> | -0.497| 0.297| 0.48| 0.4897314| 0.689| 1.448|
+#> Imputation 2
+#> USER ALERT: Out of the total of 1292 individuals in the sample, below is the distribution of the 333 (26%) individuals that fall into 2 user-selected exposure histories (out of the 27 total) created from 30th and 60th percentile values for low and high levels of exposure-epoch Infancy, Toddlerhood, Childhood.
+#> USER ALERT: Please inspect the distribution of the sample across the following exposure histories and ensure there is sufficient spread to avoid extrapolation and low precision:
+#>
+#> +---------------+-----+
+#> | epoch_history | n |
+#> +===============+=====+
+#> | h-h-h | 218 |
+#> +---------------+-----+
+#> | l-l-l | 115 |
+#> +---------------+-----+
+#>
+#> Table: Summary of user-selected exposure histories based on exposure main effects Infancy, Toddlerhood, Childhood:
+#> USER ALERT: Below are the 55 variables spanning 32 unique domains that will be treated as confounding variables for the relation between ESETA1 and StrDif_Tot.58.
+#> Please inspect this list carefully. It should include all time-varying covariates, time invariant covariates, as well as lagged levels of exposure and outcome variables if they were collected at time points earlier than the outcome time point.
+#> [1] "B18Raw.15" "B18Raw.24" "B18Raw.6" "BioDadInHH2"
+#> [5] "caregiv_health" "CORTB.15" "CORTB.24" "CORTB.6"
+#> [9] "DrnkFreq" "EARS_TJo.24" "EARS_TJo.35" "fscore.35"
+#> [13] "gov_assist" "HOMEETA1.15" "HOMEETA1.24" "HOMEETA1.35"
+#> [17] "HOMEETA1.6" "HomeOwnd" "IBRAttn.15" "IBRAttn.24"
+#> [21] "IBRAttn.6" "InRatioCor.15" "InRatioCor.24" "InRatioCor.35"
+#> [25] "InRatioCor.6" "KFASTScr" "LESMnNeg.24" "LESMnNeg.35"
+#> [29] "LESMnPos.24" "LESMnPos.35" "MDI.15" "MDI.6"
+#> [33] "peri_health" "PmAge2" "PmBlac2" "PmEd2"
+#> [37] "PmMrSt2" "RHasSO.15" "RHasSO.24" "RHasSO.35"
+#> [41] "RHasSO.6" "RHealth" "RMomAgeU" "SAAmylase.15"
+#> [45] "SAAmylase.24" "SAAmylase.6" "SmokTotl" "state"
+#> [49] "StrDif_Tot.35" "SurpPreg" "SWghtLB" "TcBlac2"
+#> [53] "WndNbrhood.24" "WndNbrhood.35" "WndNbrhood.6"
+#>
+#> The following variables are designated as numeric:
+#> [1] "ALI_Le.35, CORTB.15, CORTB.24, CORTB.6, ESETA1.15, ESETA1.24, ESETA1.35, ESETA1.58, ESETA1.6, fscore.35, fscore.58, HOMEETA1.15, HOMEETA1.24, HOMEETA1.35, HOMEETA1.58, HOMEETA1.6, IBRAttn.15, IBRAttn.24, IBRAttn.6, ID, InRatioCor.15, InRatioCor.24, InRatioCor.35, InRatioCor.58, InRatioCor.6, LESMnNeg.24, LESMnNeg.35, LESMnPos.24, LESMnPos.35, PmAge2, SAAmylase.15, SAAmylase.24, SAAmylase.6, StrDif_Tot.35, StrDif_Tot.58, WndNbrhood.24, WndNbrhood.35, WndNbrhood.58, WndNbrhood.6, ESETA1.Childhood, ESETA1.Infancy, ESETA1.Toddlerhood"
+#>
+#> The following variables are designated as factors:
+#> [1] "BioDadInHH2, DrnkFreq, HomeOwnd, PmBlac2, PmMrSt2, RHasSO.15, RHasSO.24, RHasSO.35, RHasSO.58, RHasSO.6, RHealth, SmokTotl, state, SurpPreg, TcBlac2"
+#>
+#> Table: Other variable types
+#>
+#> | |variable |type |
+#> |:--------------|:--------------|:-------|
+#> |B18Raw.15 |B18Raw.15 |integer |
+#> |B18Raw.24 |B18Raw.24 |integer |
+#> |B18Raw.58 |B18Raw.58 |integer |
+#> |B18Raw.6 |B18Raw.6 |integer |
+#> |caregiv_health |caregiv_health |integer |
+#> |EARS_TJo.24 |EARS_TJo.24 |integer |
+#> |EARS_TJo.35 |EARS_TJo.35 |integer |
+#> |gov_assist |gov_assist |integer |
+#> |KFASTScr |KFASTScr |integer |
+#> |MDI.15 |MDI.15 |integer |
+#> |MDI.6 |MDI.6 |integer |
+#> |peri_health |peri_health |integer |
+#> |PmEd2 |PmEd2 |integer |
+#> |RMomAgeU |RMomAgeU |integer |
+#> |SWghtLB |SWghtLB |integer |
+#>
+#> <table class="kable_wrapper">
+#> <caption>Summary of ESETA1 Information</caption>
+#> <tbody>
+#> <tr>
+#> <td>
+#>
+#> | | ESETA1.6 | ESETA1.15 | ESETA1.24 | ESETA1.35 | ESETA1.58 |
+#> |:--|:---------------|:---------------|:---------------|:----------------|:----------------|
+#> | |Min. :-2.8090 |Min. :-3.1010 |Min. :-2.8580 |Min. :-3.24100 |Min. :-2.47800 |
+#> | |1st Qu.:-0.3245 |1st Qu.:-0.4133 |1st Qu.:-0.5020 |1st Qu.:-0.56900 |1st Qu.:-0.53350 |
+#> | |Median : 0.3440 |Median : 0.2215 |Median : 0.1160 |Median : 0.07400 |Median : 0.08300 |
+#> | |Mean : 0.3201 |Mean : 0.2022 |Mean : 0.1229 |Mean : 0.08662 |Mean : 0.07949 |
+#> | |3rd Qu.: 0.9423 |3rd Qu.: 0.8482 |3rd Qu.: 0.7380 |3rd Qu.: 0.74625 |3rd Qu.: 0.70550 |
+#> | |Max. : 4.0350 |Max. : 3.4740 |Max. : 3.2840 |Max. : 3.01400 |Max. : 3.17300 |
+#>
+#> </td>
+#> </tr>
+#> </tbody>
+#> </table>
+#>
+#> Table: Summary of StrDif_Tot.58 Information
+#>
+#> | Min.| X1st.Qu.| Median| Mean| X3rd.Qu.| Max.|
+#> |------:|--------:|------:|--------:|--------:|-----:|
+#> | -0.497| 0.29675| 0.472| 0.489219| 0.69325| 1.448|
+#> [[1]]
+#> NULL
+#>
+#> [[2]]
+#> NULL
+
+
Here, we see summaries of the data types as well as reasonable cell counts in each of our specified histories, for each imputed dataset.
+
+
PHASE 1: Confounder Adjustment
+
+
The goal of this first phase is to minimize the associations between confounders and exposure using IPTW balancing weights. We strongly advise the user to carefully inspect each weights formula to ensure weights are created and evaluated appropriately at each step.
1a. Create Full Weights Formulas at each Exposure Time Point
+
+
We first create comprehensive, full weights formulas relating exposure to confounders at each time point using the createFormulas() function (type = “full”). This step creates full formulas containing all measured confounding variables at each exposure time point, including all time-invariant confounders and lagged time-varying confounders. The code automatically excludes time-varying confounders at the contemporaneous time point given that they cannot be decisively differentiated from mediators which should not be balanced on (Thoemmes & Ong, 2016), although this can be modified by the user if they have strong reason to believe a concurrent variable is truly a confounder (see below).
+
If the user wishes to specify any interactions between confounders in the weights formulas, they need to manually create them in the data before listing them here. Keep in mind that any interactions that include factor variables will be decomposed into interactions at each factor level.
+
The required input to create full weights formulas using the createFormulas() function are: MSM object (e.g., “obj”) and setting type = “full”.
+
Optional inputs to create full weights formulas using the createFormulas() function are as follows.
+
The user may specify a list of custom formulas by specifying to custom a list of formulas, one for each exposure time point (e.g., “exposure.time ~ variable.time + variable +…”) in formula format. We recommend first running the createFormulas() function without custom formulas (custom = NULL) and using the output as a model of the required format for custom formulas. The createFormulas() function will automatically check custom formulas to ensure that there is a correctly formatted formula for each exposure time point with exposure as the dependent variable. However, the user is responsible for ensuring the custom formulas contain the appropriate confounders for the formula type they are generating.
We chose not to create custom formulas and instead use createFormulas() to make them automatically in this example.
+
We first create full formulas.
+
+
+type<-"full"
+
+full_formulas<-createFormulas(obj =obj,
+ type =type,
+ save.out =save.out)
+
+
The function returns a list of formulas, one for each exposure time point. We inspect them below. Each full formula contains all time invariant confounders as well as all lagged time-varying confounders at each time point. This inspection is an important step, to verify that all appropriate confounders are present in each formula.
+
We inspect the formulas below.
+
+
+print(full_formulas)
+#> USER ALERT: Please manually inspect the full balancing formula below:
+#> At time point 6, the full formula for ESETA1.6 is:
+#> ESETA1.6 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist
+#> USER ALERT: Please manually inspect the full balancing formula below:
+#> At time point 15, the full formula for ESETA1.15 is:
+#> ESETA1.15 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist + SAAmylase.6 + MDI.6 + RHasSO.6 + WndNbrhood.6 + IBRAttn.6 + B18Raw.6 + HOMEETA1.6 + InRatioCor.6 + CORTB.6 + ESETA1.6
+#> USER ALERT: Please manually inspect the full balancing formula below:
+#> At time point 24, the full formula for ESETA1.24 is:
+#> ESETA1.24 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist + SAAmylase.6 + SAAmylase.15 + MDI.6 + MDI.15 + RHasSO.6 + RHasSO.15 + WndNbrhood.6 + IBRAttn.6 + IBRAttn.15 + B18Raw.6 + B18Raw.15 + HOMEETA1.6 + HOMEETA1.15 + InRatioCor.6 + InRatioCor.15 + CORTB.6 + CORTB.15 + ESETA1.6 + ESETA1.15
+#> USER ALERT: Please manually inspect the full balancing formula below:
+#> At time point 35, the full formula for ESETA1.35 is:
+#> ESETA1.35 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist + SAAmylase.6 + SAAmylase.15 + SAAmylase.24 + MDI.6 + MDI.15 + RHasSO.6 + RHasSO.15 + RHasSO.24 + WndNbrhood.6 + WndNbrhood.24 + IBRAttn.6 + IBRAttn.15 + IBRAttn.24 + B18Raw.6 + B18Raw.15 + B18Raw.24 + HOMEETA1.6 + HOMEETA1.15 + HOMEETA1.24 + InRatioCor.6 + InRatioCor.15 + InRatioCor.24 + CORTB.6 + CORTB.15 + CORTB.24 + EARS_TJo.24 + LESMnPos.24 + LESMnNeg.24 + ESETA1.6 + ESETA1.15 + ESETA1.24
+#> USER ALERT: Please manually inspect the full balancing formula below:
+#> At time point 58, the full formula for ESETA1.58 is:
+#> ESETA1.58 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist + SAAmylase.6 + SAAmylase.15 + SAAmylase.24 + MDI.6 + MDI.15 + RHasSO.6 + RHasSO.15 + RHasSO.24 + RHasSO.35 + WndNbrhood.6 + WndNbrhood.24 + WndNbrhood.35 + IBRAttn.6 + IBRAttn.15 + IBRAttn.24 + B18Raw.6 + B18Raw.15 + B18Raw.24 + HOMEETA1.6 + HOMEETA1.15 + HOMEETA1.24 + HOMEETA1.35 + InRatioCor.6 + InRatioCor.15 + InRatioCor.24 + InRatioCor.35 + CORTB.6 + CORTB.15 + CORTB.24 + EARS_TJo.24 + EARS_TJo.35 + LESMnPos.24 + LESMnPos.35 + LESMnNeg.24 + LESMnNeg.35 + StrDif_Tot.35 + fscore.35 + ESETA1.6 + ESETA1.15 + ESETA1.24 + ESETA1.35
+
+
+
1b. Conduct Exploratory Pre-Balance Assessment
+
+
The next step examines the initial imbalance, or how strongly exposure relates to each confounder at each time point, for all measured confounders prior to weighting using the assessBalance() function. This function draws on the calcBalStats() function (see the Assessing Balance for Time-Varying Exposure section in the accompanying manuscript).
+The assessBalance() function outputs balance statistics (correlations for continuous exposures and standardized mean differences for binary exposures) relating exposure at each time point to confounders in a table as well as in plots. This function also provides summary balance statistics averaging across all time points (and imputed datasets if they are supplied).
+
The required inputs for using the assessBalance() function to conduct pre-balance testing are: data (data frame, a mids object, or a list of imputed datasets as dataframes in wide format) and an MSM object (e.g., “obj”). Please see the Assessing Balance for Time-Varying Exposures vignette for more detail on how this function calculates balance.
+
The optional inputs are as follows.
+
The user may specify balance_thresh, or a threshold(s) for determining confounder balance, in one of two ways.
+* First, they can provide a single number value (0-1) for the absolute value of the standardized balance statistic (either the correlation for continuous exposures or standardized group mean difference for binary exposures) for exposure and confounders below which confounders are considered balanced, and above which they are considered imbalanced (default is 0.1; Stuart, 2010).
+
+
Second, users may make an a priori assertion that some confounders are more important than others based on theory and existing research. In this case, they can provide two numbers that represent the balance thresholds for important and less important confounders, respectively. If the user supplies two balance thresholds, they must also supply a list of important confounders (time-varying: “variable.t”, time invariant: “variable”) to the imp_conf field. The balance threshold specification should be kept consistent throughout the use of this workflow.
+
+
Below, as recommended, we provide two balancing thresholds and identify income and parent education as important confounders in the relation between economic strain and behavior problems.
The function returns a list (one entry per imputed dataset, when applicable), that contains a table for each exposure time point. The able contains all confounders for that time point, each with an associated standardized balance statistics relating confounder to exposure at that time point, user-supplied balance threshold, and a binary indicator of whether the confounder is balanced.
+
As shown below, we can print, summarize, and plot several versions of the balance statistics with the option to supply save.out to save viewed output to the home directory.
+
Each of these functions takes an optional t field to view balance statistics for any one of your exposure time points. t takes an integer value from 1 to the total number of time points. If it is not specified, the output is shown for all exposure time points.
+
With imputed data, each of these functions takes an option i field that can be used to view balance for any one imputed data set. If it is not specified, output is shown averaged across the absolute values of the balance statistics of the imputed datasets. It can be useful to average across imputed datasets to get an overall sense of balance. For non-imputed data, do not specify i.
+
We can view prebalance statistics for any single imputed dataset (e.g., first imputed dataset), using the i field. Note that we supply t as integers 1 through however number of time points at which exposure is measured. For example, my first time point measures ESETA1 at 6 months which corresponds to t = 1.
Or, we can view prebalance statistics averaged across imputed data sets at different time points by not specifying i. This can also be used to view balance statistics when the data are not imputed.
We can also summarize the assessBalance() output to view the average remaining relation between confounders and exposure as well as a summary table showing the total number of imbalanced confounders at each exposure time point. We can view this in one imputed dataset or averaged across them.
+
+
+summary(prebalance_stats,
+ i =1,
+ save.out =save.out)
+#> USER ALERT: For imputation 1: As shown below, 55 out of 241 (22.8%) covariates across time points remain imbalanced with a remaining median absolute correlation of 0.13 (max: 0.31):
+#> +-----------+-----------------------+----------------------------+
+#> | Exposure | Total # of covariates | # of imbalanced covariates |
+#> +===========+=======================+============================+
+#> | ESETA1.6 | 28 | 14 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.15 | 38 | 11 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.24 | 47 | 11 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.35 | 59 | 11 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.58 | 69 | 8 |
+#> +-----------+-----------------------+----------------------------+
+#>
+#> Table: Imbalanced Covariates for imputation 1
+
+summary(prebalance_stats,
+ save.out =save.out)
+#> USER ALERT: Averaging across imputated datasets: As shown below, 60 out of 241 (24.9%) covariates across time points remain imbalanced with a remaining median absolute correlation of 0.13 (max: 0.31):
+#> +-----------+-----------------------+----------------------------+
+#> | Exposure | Total # of covariates | # of imbalanced covariates |
+#> +===========+=======================+============================+
+#> | ESETA1.6 | 28 | 15 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.15 | 38 | 12 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.24 | 47 | 11 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.35 | 59 | 14 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.58 | 69 | 8 |
+#> +-----------+-----------------------+----------------------------+
+#>
+#> Table: Imbalanced Covariates Averaged Across Imputed Datasets
+
+
Averaging across imputed datasets, we see that xx confounders are imbalanced with respect to the economic strain exposure and their respective balance threshold.
+
Lastly, we can plot a balance summary at one or more time points, from one imputed dataset or averaged across them. The dotted red lines denote our balance thresholds and the points colored and labeled in red denote confounders that are imbalanced in relation to their respective balance thresholds.
+
+
+plot(prebalance_stats,
+ i =1,
+ t =1,
+ save.out =save.out)
+
+
+
+
+
+
+
+
+plot(prebalance_stats,
+ t =1,
+ save.out =save.out)
+
+
+
+
+
+
+
+
+The love plots depict the standardized associations between confounder and exposure at each exposure time point, with the vertical red dashed lines indicating balance thresholds. Imbalanced confounders are shown in red with variable name labels.
The goal of this second step is to create shortened, more parsimonious weights formulas for determining the optimal IPTW weighting method that most successfully reduces imbalance in the data.
+
2a. Create Simplified Weights Formulas
+
+
First, we create shorter, more parsimonious weights formulas relating exposure to confounders at each time point using the createFormulas() function (type = “short”). For each exposure time point, these formulas contain all time invariant confounders as well as time-varying confounders only at the t-1 lag. The logic here is that balancing on confounders at the most recent prior time point (t-1 only) may achieve balance on levels at more distal time points, given the stability of many confounders over time. Importantly, we will empirically assess and relax this assumption if needed at subsequent steps (Steps 3a-b).
+
The required input to create shortened weights formulas using the createFormulas() function are: a MSM object (e.g., ‘obj’) and setting type = “short”.
+
In addition to the optional input outlined in Step 1a, the user also has the option to specify in keep_conf, a list of any time-varying confounders (e.g., “variable.t”) to always retain as lagged confounders in these shortened formulas. The user may use this argument to retain specific time-varying confounders that would otherwise be excluded at this step if they occur at lags greater than t-1 for each formula.
+
We create short formulas below.
+
+
+type<-"short"
+
+short_formulas<-createFormulas(obj =obj,
+ type =type,
+ save.out =save.out)
+
+
We again get a list with entries containing a formula for each exposure time point.
+
And then inspect them to make sure they contain only time-varying covariates at a lag of one prior to the exposure time point.
+
+
+print(short_formulas)
+#> USER ALERT: Please manually inspect the short balancing formula below that includes time-varying confounders at t-1 only:
+#> At time point 6, the short formula for ESETA1.6 is:
+#> ESETA1.6 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist
+#> USER ALERT: Please manually inspect the short balancing formula below that includes time-varying confounders at t-1 only:
+#> At time point 15, the short formula for ESETA1.15 is:
+#> ESETA1.15 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist + SAAmylase.6 + MDI.6 + RHasSO.6 + WndNbrhood.6 + IBRAttn.6 + B18Raw.6 + HOMEETA1.6 + InRatioCor.6 + CORTB.6 + ESETA1.6
+#> USER ALERT: Please manually inspect the short balancing formula below that includes time-varying confounders at t-1 only:
+#> At time point 24, the short formula for ESETA1.24 is:
+#> ESETA1.24 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist + SAAmylase.15 + MDI.15 + RHasSO.15 + IBRAttn.15 + B18Raw.15 + HOMEETA1.15 + InRatioCor.15 + CORTB.15 + ESETA1.15
+#> USER ALERT: Please manually inspect the short balancing formula below that includes time-varying confounders at t-1 only:
+#> At time point 35, the short formula for ESETA1.35 is:
+#> ESETA1.35 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist + SAAmylase.24 + RHasSO.24 + WndNbrhood.24 + IBRAttn.24 + B18Raw.24 + HOMEETA1.24 + InRatioCor.24 + CORTB.24 + EARS_TJo.24 + LESMnPos.24 + LESMnNeg.24 + ESETA1.24
+#> USER ALERT: Please manually inspect the short balancing formula below that includes time-varying confounders at t-1 only:
+#> At time point 58, the short formula for ESETA1.58 is:
+#> ESETA1.58 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist + RHasSO.35 + WndNbrhood.35 + HOMEETA1.35 + InRatioCor.35 + EARS_TJo.35 + LESMnPos.35 + LESMnNeg.35 + StrDif_Tot.35 + fscore.35 + ESETA1.35
+
+
These formulas are considerably shorter than the full formulas. For instance, at the 58-month exposure time point, the formula contains all time invariant confounders and only time-varying confounders at the 35-month time point.
+
+
2b. Create IPTW Balancing Weights Using Multiple Weighting Methods
+
+
Having created shorter, simplified weights formulas, we now create the first round of IPTW balancing weights (Thoemmes & Ong, 2016) using the createWeights() function, the shortened weights formulas, and all available weighting methods. The function calls the weightitMSM() function from the WeightIt package (Greifer, 2023) that uses the time-specific formulas to create weights at each time point before automatically multiplying them together to create one weight per person. Weights are stabilized, as recommended (Cole & Hernan, 2008; Thoemmes & Ong, 2016), and distributions can be saved for inspection.
+
The required inputs for using the createWeights() function to create the initial around of IPTW balancing weights are: an MSM object (e.g, ‘obj’), complete data (data frame, mids object, or a list of imputed datasets as dataframes in wide format), and the short formulas (see Step 2a).
+
We specify the short formulas below.
+
+
+formulas<-short_formulas
+
+
The optional inputs are as follows.
+
For method, provide one of the following methods for calculating balancing weights using weightitMSM() from the methods that have been validated for longitudinal exposures: “cbps” (Covariate Balancing Propensity Score weighting), “gbm” (generalized boosted model), “glm” (generalized linear model; default), or “super” (SuperLearner via the SuperLearner package; Polley et al., 2013). More information can be found in the WeightIt documentation.
+
We begin with specifying CBPS as a weighting method.
+
+
+method<-"cbps"
+
+
The createWeights() function can also take any number of additional arguments that are accapted by the weightitMSM() function (e.g., ‘criterion’, distribution’, ‘SL.library’). The package defaults correspond to the weightIt defaults. If the user selects the SuperLearner (“super”) method, the default super learner libraries (‘SL.library’) are “SL.glm” and “SL.glm.interaction” but an alternative library can be entered as an input to the createWeights function. For binary exposures, the “cbps” method allows you to specify estimand as either ATE, ATT, or ATC. With “glm”, “super”, and “bart” you can specify ATE, ATT, ATC, ATO, ATM, or ATOS. With “gbm”, you can specify ATE, ATT, ATC, ATO, or ATM. The default estimand for binary exposures is ATE. We advise the interested user to review the WeightIt documentation for more information about the additional optional arguments available for each of the weighting methods. Users have the option to specify verbose = TRUE to view information about weights creation.
+
The function returns a list of weights objects each in the form of WeightItMSM output (with an entry for each imputed dataset when appropriate).
+
CBPS
+
+
Below, we create IPTW weights using the default CBPS method.
+
+
+weights.cbps<-createWeights(obj =obj,
+ data =data,
+ method =method,
+ formulas =formulas,
+ verbose =TRUE,
+ maxit =1, # just for testing to speed up --will remove
+ save.out =save.out)
+#> initial value 0.903706
+#> final value 0.463025
+#> stopped after 2 iterations
+#> Warning: The optimziation failed to converge; try again with a higher value of
+#> `maxit`.
+#> initial value 0.585184
+#> final value 0.381680
+#> stopped after 2 iterations
+#> Warning: The optimziation failed to converge; try again with a higher value of
+#> `maxit`.
+#>
+#> For imputation 1 and the `cbps` weighting method, the median weight value is 0.68 (SD = 1.01; range = 0-11).
+
+
These take a while to run. Note that if you save the weights (by supplying save.out = TRUE or a custom file name to createWeights() and a home directory to initMWM()), the function outputs the file path to use for reading in the weights for future use. This can be useful given that some weighting methods can take a long time to run, especially for imputed datasets.
+
If we had previously saved out CBPS weights, we could read them in instead of re-creating them.
Given that a separate set of weights is created for each imputed dataset, we conduct our inspections on each imputed dataset.
+
First, we view the basic statistics of the CBPS weights for a given imputed dataset. Here, we note a median weight value of 0.77 (SD= 1.18) but with a fairly extensive range of 0 - 9.
+
+
+print(weights.cbps,
+ i =1)
+#>
+#> For imputation 1 and the `cbps` weighting method, the median weight value is 0.68 (SD = 1.01; range = 0-11).
+
+
Next, we look inside the output to summarize the weighting process (drawing on the summary method from weightIt, for any given imputed dataset.
+
+
+summary(weights.cbps[[1]])
+#> Time 1
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 10.69
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 798 1145 737 1235 877
+#> all 7.7683 8.696 9.0943 9.1878 10.69
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 1.131 0.654 0.418 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 567.09
+#>
+#> Time 2
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 10.69
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 798 1145 737 1235 877
+#> all 7.7683 8.696 9.0943 9.1878 10.69
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 1.131 0.654 0.418 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 567.09
+#>
+#> Time 3
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 10.69
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 798 1145 737 1235 877
+#> all 7.7683 8.696 9.0943 9.1878 10.69
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 1.131 0.654 0.418 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 567.09
+#>
+#> Time 4
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 10.69
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 798 1145 737 1235 877
+#> all 7.7683 8.696 9.0943 9.1878 10.69
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 1.131 0.654 0.418 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 567.09
+#>
+#> Time 5
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 10.69
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 798 1145 737 1235 877
+#> all 7.7683 8.696 9.0943 9.1878 10.69
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 1.131 0.654 0.418 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 567.09
+
+
This summary also provides the effective sample size, or the sample size that results from applying the weights to the original sample size, for each time point. Weighting can often result is an effective or weighted sample size that is smaller than the orignal sample size and is something to keep in mind when evaluating weighting methods. In this example, we see that our original 1,292 sample is reduced to 605 upon weighting with the CBPS method.
+
We then view a distribution plot of the weights for any one imputed dataset. The user has the option to supply save.out to save plots to the home directory.
As shown above, the distribution has a heavy right tail (typical of real-world data). The right tail of the distribution represents individuals who experienced statistically unexpected levels of exposure given their levels of confounders.
+
We then create and inspect IPTW balancing weights using all other available methods in order to evaluate and compare their performance in subsequent steps. Here, we summarize and plot averaging across all imputed datasets in order to get a sense for their overall performance. Example inspections are for the first imputed dataset.
+
GLM
+
+
+
+method<-"glm"
+
+weights.glm<-createWeights(obj =obj,
+ data =data,
+ method =method,
+ formulas =formulas,
+ save.out =save.out)
+
+print(weights.glm,
+ i =1)
+#>
+#> For imputation 1 and the `glm` weighting method, the median weight value is 1.27 (SD = 10.03; range = 0-271).
+
+summary(weights.glm[[1]])
+#> Time 1
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 271.3462
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1030 475 697 1145 591
+#> all 52.1877 58.3054 63.4626 178.9983 271.3462
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 3.824 0.96 1.144 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 82.74
+#>
+#> Time 2
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 271.3462
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1030 475 697 1145 591
+#> all 52.1877 58.3054 63.4626 178.9983 271.3462
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 3.824 0.96 1.144 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 82.74
+#>
+#> Time 3
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 271.3462
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1030 475 697 1145 591
+#> all 52.1877 58.3054 63.4626 178.9983 271.3462
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 3.824 0.96 1.144 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 82.74
+#>
+#> Time 4
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 271.3462
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1030 475 697 1145 591
+#> all 52.1877 58.3054 63.4626 178.9983 271.3462
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 3.824 0.96 1.144 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 82.74
+#>
+#> Time 5
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 271.3462
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1030 475 697 1145 591
+#> all 52.1877 58.3054 63.4626 178.9983 271.3462
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 3.824 0.96 1.144 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 82.74
+
+plot(weights.glm,
+ i =1,
+ save.out =save.out)
+
+
+
+
+
+
+
+
As shown above, the GLM method produces a higher median of 1.27 and a much greater range of weights.
The GBM method produces a similar mean as GLM and a similarly large range (0-216).
+
+
Bart
+
+
+
+method<-"bart"
+
+weights.bart<-createWeights(obj =obj,
+ data =data,
+ method =method,
+ formulas =formulas,
+ save.out =save.out)
+
+print(weights.bart,
+ i =1)
+#>
+#> For imputation 1 and the `bart` weighting method, the median weight value is 1.24 (SD = 15; range = 0-499).
+
+summary(weights.bart[[1]])
+#> Time 1
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 499.4559
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 79 591 1235 475 1145
+#> all 51.3105 72.3414 90.2131 107.8639 499.4559
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 5.503 0.996 1.377 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 41.33
+#>
+#> Time 2
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 499.4559
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 79 591 1235 475 1145
+#> all 51.3105 72.3414 90.2131 107.8639 499.4559
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 5.503 0.996 1.377 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 41.33
+#>
+#> Time 3
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 499.4559
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 79 591 1235 475 1145
+#> all 51.3105 72.3414 90.2131 107.8639 499.4559
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 5.503 0.996 1.377 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 41.33
+#>
+#> Time 4
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 499.4559
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 79 591 1235 475 1145
+#> all 51.3105 72.3414 90.2131 107.8639 499.4559
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 5.503 0.996 1.377 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 41.33
+#>
+#> Time 5
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 499.4559
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 79 591 1235 475 1145
+#> all 51.3105 72.3414 90.2131 107.8639 499.4559
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 5.503 0.996 1.377 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 41.33
+
+plot(weights.bart,
+ i =1,
+ save.out =save.out)
+
+
+
+
+
+
+
+
The bart method has a similar median and an even larger range (0-945).
+
+
Super
+
+
+
+method<-"super"
+
+weights.super<-createWeights(obj =obj,
+ data =data,
+ method =method,
+ formulas =formulas,
+ save.out =save.out)
+#> Loading required package: nnls
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning: glm.fit: algorithm did not converge
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+#> Warning in predict.lm(object, newdata, se.fit, scale = 1, type = if (type == :
+#> prediction from rank-deficient fit; attr(*, "non-estim") has doubtful cases
+
+print(weights.super,
+ i =1)
+#>
+#> For imputation 1 and the `super` weighting method, the median weight value is 1.27 (SD = 9.99; range = 0-270).
+
+summary(weights.super[[1]])
+#> Time 1
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 270.0915
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1030 475 697 1145 591
+#> all 52.0024 58.0163 63.2837 178.6145 270.0915
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 3.816 0.959 1.141 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 83.08
+#>
+#> Time 2
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 270.0915
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1030 475 697 1145 591
+#> all 52.0024 58.0163 63.2837 178.6145 270.0915
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 3.816 0.959 1.141 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 83.08
+#>
+#> Time 3
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 270.0915
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1030 475 697 1145 591
+#> all 52.0024 58.0163 63.2837 178.6145 270.0915
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 3.816 0.959 1.141 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 83.08
+#>
+#> Time 4
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 270.0915
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1030 475 697 1145 591
+#> all 52.0024 58.0163 63.2837 178.6145 270.0915
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 3.816 0.959 1.141 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 83.08
+#>
+#> Time 5
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 270.0915
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1030 475 697 1145 591
+#> all 52.0024 58.0163 63.2837 178.6145 270.0915
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 3.816 0.959 1.141 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 83.08
+
+plot(weights.super,
+ i =1,
+ save.out =save.out)
+
+
+
+
+
+
+
+
The super method produces a similar median and a range of 0-270.
+
+
2c. Assess All Weighting Methods to Determine Optimal Method
+
+
Next, we evaluate how well the weights created using each of the different weighting methods reduced imbalance using the assessBalance() function. This function calls the calcBalStats() function using the short formulas and specifies that the balance statistics should be calculated using the IPTW weights supplied.
+
The required inputs for using the assessBalance() function to assess balance for the first round of IPTW weights are: complete data (data frame, mids object, or a list of imputed datasets as dataframes in wide format), a MSM object (e.g., ‘obj’), and the weights that were just created.
+
The optional inputs are described in Step 1b.
+
The assessBalance() function returns a data frame (or list) of balance statistics, balance thresholds, and binary balanced tag (1 = balanced, 0 = imbalanced) for each confounder relevant to each exposure time point. The function outputs balance statistics (correlations for continuous exposures and standardized mean differences for binary exposures) relating exposure at each time point to confounders in a table as well as in plots. This function also provides summary balance statistics averaging across all time points (and imputed datasets if they are supplied).
+
We retain the same optional important confounders (imp_conf) and balance threshold (balance_thresh) as we specified earlier.
+
CBPS
+
+
We first assess balance for the CBPS weighting method.
The function returns a list of balance statistics for each expsoure time point, with one entry per imputed dataset (where applicable).
+
For assessing balance using imputed data, we can average across imputed datasets to view an overall summary of the performance method. We can also examine each imputed dataset individually.
+
We summarize the CBPS balance statistics across imputed datasets and examine balance for the first imputed dataset. The user has the option to supply save.out to save printed output to the home directory.
+
+
+summary(balance_stats.cbps,
+ save.out =save.out)
+#> Averaging across imputed datasets using cbps weighting method.
+#> USER ALERT: Averaging across imputed datasets using `cbps` weighting method: As shown below, 0 out of 241 (0.0%) covariates across time points remain imbalanced:
+#> +-----------+-----------------------+----------------------------+
+#> | Exposure | Total # of covariates | # of imbalanced covariates |
+#> +===========+=======================+============================+
+#> | ESETA1.6 | 28 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.15 | 38 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.24 | 47 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.35 | 59 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.58 | 69 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#>
+#> Table: Imbalanced Covariates Averaging Across Imputed Datasetrs using `cbps`
+
+summary(balance_stats.cbps,
+ i =1,
+ save.out =save.out)
+#> USER ALERT: For imputation 1 using `cbps` weighting method: As shown below, 4 out of 241 (1.7%) covariates across time points remain imbalanced with a remaining median absolute correlation of 0.06 (max: 0.10):
+#> +-----------+-----------------------+----------------------------+
+#> | Exposure | Total # of covariates | # of imbalanced covariates |
+#> +===========+=======================+============================+
+#> | ESETA1.6 | 28 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.15 | 38 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.24 | 47 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.35 | 59 | 2 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.58 | 69 | 2 |
+#> +-----------+-----------------------+----------------------------+
+#>
+#> Table: Imbalanced Covariates for imputation 1 using `cbps`
+
+
Averaging across imputed datasets, we find only 2 imbalanced confounders remaining (both at the 35-month exposure time point) out of 241, with a median remaining correlation of 0.06 and a maximum of 0.07.
+
We can then visualize the imbalance with love plots, averaging across imputed datasets, for each exposure time point (with the first time point shown here). We can also plot balance for any given imputed datset.
+
+
+plot(balance_stats.cbps,
+ t =4,
+ save.out =save.out)
+
+
+
+
+
+
+
+
+plot(balance_stats.cbps,
+ t =4,
+ i =1,
+ save.out =save.out)
+
+
+
+
+
+
+
+
We can see that income at 6 and 15 months remains imbalanced at the 4th (35-month) time point.
+
We can then inspect the balance statistics for each confounder in relation to exposure at all time points or each one individually, either averaging across imputed datasets or for any one individually. The user has the option to supply save.out to save printed output to the home directory. Below, we show an example of balance statistics for the first time point, averaged across imputed datasets.
From these summaries, we identify the optimal weighting method for a dataset, or the one that yields the best confounder balance. To do this, we can consider several criteria. We note that there exist no gold-standard, hard and fast rules for identifying optimal balance (especially when using imputed data). However, we can draw from the following guidance:
+
+
Fewest imbalanced confounders remaining relative to the user-specified balance threshold(s) (from summary output);
+
Lowest median absolute balance statistic, across all confounders and time points, reflecting the best overall attenuation of confounding (from summary output);
+
Lowest maximum absolute balance statistic, across all confounders and time points (and imputed datasets, where applicable), indicating weakest remaining relation between exposure and confounder for the least balanced confounder (from summary output);
+
+
Reasonable effective sample size following weighting (for all imputed datasets, where applicable), indicating reasonable power to detect effects (from weightiIt summary output). For the first three, we examine summaries for each of the weighting methods.
+
+
+
+summary(balance_stats.cbps,
+ save.out =save.out)
+#> Averaging across imputed datasets using cbps weighting method.
+#> USER ALERT: Averaging across imputed datasets using `cbps` weighting method: As shown below, 0 out of 241 (0.0%) covariates across time points remain imbalanced:
+#> +-----------+-----------------------+----------------------------+
+#> | Exposure | Total # of covariates | # of imbalanced covariates |
+#> +===========+=======================+============================+
+#> | ESETA1.6 | 28 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.15 | 38 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.24 | 47 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.35 | 59 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.58 | 69 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#>
+#> Table: Imbalanced Covariates Averaging Across Imputed Datasetrs using `cbps`
+
+summary(balance_stats.glm,
+ save.out =save.out)
+#> USER ALERT: Averaging across imputed datasets using `glm` weighting method: As shown below, 31 out of 241 (12.9%) covariates across time points remain imbalanced with a remaining median absolute correlation of 0.13 (max: 0.29):
+#> +-----------+-----------------------+----------------------------+
+#> | Exposure | Total # of covariates | # of imbalanced covariates |
+#> +===========+=======================+============================+
+#> | ESETA1.6 | 28 | 7 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.15 | 38 | 18 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.24 | 47 | 2 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.35 | 59 | 2 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.58 | 69 | 2 |
+#> +-----------+-----------------------+----------------------------+
+#>
+#> Table: Imbalanced Covariates Averaging Across Imputed Datasetrs using `glm`
+
+# summary(balance_stats.gbm,
+# save.out = save.out)
+
+summary(balance_stats.bart,
+ save.out =save.out)
+#> USER ALERT: Averaging across imputed datasets using `bart` weighting method: As shown below, 16 out of 241 (6.6%) covariates across time points remain imbalanced with a remaining median absolute correlation of 0.12 (max: 0.18):
+#> +-----------+-----------------------+----------------------------+
+#> | Exposure | Total # of covariates | # of imbalanced covariates |
+#> +===========+=======================+============================+
+#> | ESETA1.6 | 28 | 3 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.15 | 38 | 10 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.24 | 47 | 2 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.35 | 59 | 1 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.58 | 69 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#>
+#> Table: Imbalanced Covariates Averaging Across Imputed Datasetrs using `bart`
+
+summary(balance_stats.super,
+ save.out =save.out)
+#> USER ALERT: Averaging across imputed datasets using `super` weighting method: As shown below, 31 out of 241 (12.9%) covariates across time points remain imbalanced with a remaining median absolute correlation of 0.13 (max: 0.29):
+#> +-----------+-----------------------+----------------------------+
+#> | Exposure | Total # of covariates | # of imbalanced covariates |
+#> +===========+=======================+============================+
+#> | ESETA1.6 | 28 | 7 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.15 | 38 | 18 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.24 | 47 | 2 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.35 | 59 | 2 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.58 | 69 | 2 |
+#> +-----------+-----------------------+----------------------------+
+#>
+#> Table: Imbalanced Covariates Averaging Across Imputed Datasetrs using `super`
+
+
From this, we find that the CBPS method has the fewest imbalanced confounders (with only 2), lowest median balance statistic, and lowest max balance statistic.
+
To examine the fourth criterion, we use the weightIt summary method to examine effective sample sizes (“Weighted”) compared to the orignal sample size (“Unweighted”) across weighting methods. We do this just for the first imputed dataset.
The goal of this next step is to more closely inspect the balance reults of the best-performing weights created by the shortened weights formulas, and add to the shortened formulas any time-varying confounders at lags > t-1 that were not successfully balanced, to create a final round of weights.
+
3a. Examine Balance of Optimal Weighting Method
+
+
We next inspect the balance produced by the weights created in the previous step with the best-performing weights method (i.e., using the SuperLearner method). Here, we revisit our assumption that balancing on the most proximal time-varying confounders (t-1) confers balance for those same confounders at more distal prior time points (t- 1+).
+
We more closely inspect the balance of the CBPS weights, averaged across imputed datasets.
With real-world data, it is often difficult to fully balance the many confounding variables, especially across time. If a user does find that no confounders remain imbalanced, they can skip to Step 3d. Given that we identified remaining imbalanced confounders, we proceed to Step 3b.
+
Step 3b. Update Simplified Formulas
+
+
Subsequently, we update the shortened formulas to include any time-varying confounders (t-1 +) that were not successfully balanced by the full formulas, as shown above. To do this, we create a final round of weights formulas using the createFormulas() function (setting type = “update” and providing the balance statistics of the bal_stats field). The createFormulas() function draws from user-provided balance statistics to automatically identify and add to the formulas at each exposure time point any time-varying confounders at lags greater than 1 that remain imbalanced after weighting. The function displays each weights formula in the console with a message to the user about any time-varying confounders that were added.
+
The required input to update the shortened weights formulas using the createFormulas() function are: an MSM object (e.g., “obj”), setting type = “update”, and providing to bal_stats the balance statistics that were just created in Step 3a.
+
The optional input are detailed in Step 1a.
+
The function returns a list of weights formulas labeled by type, exposure, outcome, and exposure time point.
+
Below, we update our short formulas using the balance statistics from the best-performing weights.
We then inspect these new formulas to make sure that the imbalanced covariates were added to the appropriate formulas. The user has the option to supply save.out to save printed output to the home directory.
+
+
+print(updated_formulas,
+ save.out =save.out)
+#> USER ALERT: Please manually inspect the updated balancing formula below that includes time-varying confounders at t-1 and those greater at further lags that remained imbalanced:
+#> At time point 6, the update formula for ESETA1.6 is:
+#> ESETA1.6 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist
+#> USER ALERT: Please manually inspect the updated balancing formula below that includes time-varying confounders at t-1 and those greater at further lags that remained imbalanced:
+#> At time point 15, the update formula for ESETA1.15 is:
+#> ESETA1.15 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist + SAAmylase.6 + MDI.6 + RHasSO.6 + WndNbrhood.6 + IBRAttn.6 + B18Raw.6 + HOMEETA1.6 + InRatioCor.6 + CORTB.6 + ESETA1.6
+#> USER ALERT: Please manually inspect the updated balancing formula below that includes time-varying confounders at t-1 and those greater at further lags that remained imbalanced:
+#> At time point 24, the update formula for ESETA1.24 is:
+#> ESETA1.24 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist + SAAmylase.15 + MDI.15 + RHasSO.15 + IBRAttn.15 + B18Raw.15 + HOMEETA1.15 + InRatioCor.15 + CORTB.15 + ESETA1.15
+#> USER ALERT: Please manually inspect the updated balancing formula below that includes time-varying confounders at t-1 and those greater at further lags that remained imbalanced:
+#> At time point 35, the update formula for ESETA1.35 is:
+#> ESETA1.35 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist + SAAmylase.24 + RHasSO.24 + WndNbrhood.24 + IBRAttn.24 + B18Raw.24 + HOMEETA1.24 + InRatioCor.24 + CORTB.24 + EARS_TJo.24 + LESMnPos.24 + LESMnNeg.24 + ESETA1.24
+#> USER ALERT: Please manually inspect the updated balancing formula below that includes time-varying confounders at t-1 and those greater at further lags that remained imbalanced:
+#> At time point 58, the update formula for ESETA1.58 is:
+#> ESETA1.58 ~ state + BioDadInHH2 + PmAge2 + PmBlac2 + TcBlac2 + PmMrSt2 + PmEd2 + KFASTScr + RMomAgeU + RHealth + HomeOwnd + SWghtLB + SurpPreg + SmokTotl + DrnkFreq + peri_health + caregiv_health + gov_assist + RHasSO.35 + WndNbrhood.35 + HOMEETA1.35 + InRatioCor.35 + EARS_TJo.35 + LESMnPos.35 + LESMnNeg.35 + StrDif_Tot.35 + fscore.35 + ESETA1.35
+
+
As shown above, income at 6 and 15 months (“InRatioCor.6” and “InRatioCor.15”) were added to the 35-month weights formula. These were not originally in that weights formula given that they are lags greater than t -1. Their remaining imbalance suggests that achieving balance for 24-month income did not successfully balance prior levels of income. We will then use these weights formulas to recreate CBPS weights in an effort to achieve the greatest reduction in balance.
+
+
Step 3c. Create Final Balancing Weights
+
+
Next, we create a final set of balancing weights using the optimal weighting method identified in Step 2c and the updated formulas from the previous step using the createWeights() function (method = “…’), with the SuperLearner method being the optimal weighting method identified in Step 2c. The function calls the weightitMSM() function from the WeightIt package (Greifer, 2023) that uses the time-specific formulas to create weights at each time point before automatically multiplying them together to create one weight per person. Weights are stabilized, as is recommended (Cole & Hernan, 2008; Thoemmes & Ong, 2016) and distributions are saved out in the home directory for inspection.
+
The required inputs for using the createWeights() function to create the final round of IPTW balancing weights using the updated short weights formulas are: complete data (data frame, mids object, or a list of imputed datasets as dataframes in wide format), an MSM object (e.g., “obj”), the best-performing weights method, and the updated formulas (see Step 3a).
+
The optional input for the createWeights() function are listed in Step 2b.
+
The function returns a list in the form of WeightItMSM output.
+
Below, we use the updated formulas and the CPBS weighting method to create a final round of IPTW balancing weights.
We then inspect these final weights by imputed dataset. The user has the option to supply save.out to save printed or plot output to the home directory.
+
+
+print(final_weights,
+ i =1)
+#>
+#> For imputation 1 and the `cbps` weighting method, the median weight value is 0.77 (SD = 1.18; range = 0-9).
+
+plot(final_weights,
+ i =1,
+ save.out =save.out)
+
+
+
+
+
+
+
+
As shown above, these weights have a median value of 0.77 and a range of 0-9, the same as before.
+
+
3d. Trim Final Balancing Weights
+
+
The next step is to trim (i.e., winsorize) this final set of weights to eliminate the heavy right tail of its distribution using the trimWeights() function. This function draws on the Weightit package (Griefer, 2023) and plots and summarizes trimmed weights. This function outputs a list of trimmed weights with either a single nested list (labeled “0” if data are in data frame format) or with nested lists for each imputed dataset (if data are imputed).
+
The required input for the trimWeights() function is the final weights we just created.
+
The optional input allows the user to specify a quantile value (0-1; default is 0.95) above which the weights will be replaced with the weight value of that quantile, to reduce the heavy right tail.
+
Below, we use the default 95th percentile for trimming weights.
+
+
+quantile<-0.95
+
+
We trim the final weights below.
+
+
+weights<-final_weights
+
+trim_weights<-trimWeights(weights =weights,
+ at =quantile,
+ save.out =save.out)
+
+
The function returns a list of weights objects, containing trimmed weights, each in the form of weightitMSM output ( with one entry per imputed dataset, where applicable). Each entry also specifies the quantile value at which the weights were trimmed.
+
We then inspect the trimmed weights from one of the imputed datasets below. The user has the option to supply save.out to save plots to the home directory.
+
+
+print(trim_weights,
+ i =1)
+#>
+#> For imputation 1 and the `cbps` weighting method, after trimming at 95th quantile, the median weight value is 0.77 (SD = 0.93; range = 0-4).
+
+plot(trim_weights,
+ i =1,
+ save.out =save.out)
+
+
+
+
+
+
+
+
As shown above, the weights still have a median value of 0.77 but a smaller standard deviation and a range that now only goes from 0-4.
+
+
Sensitvity Analyses
+
+
We then create trimmed weights using two other quantile values at + /- ~0.3 of the previously chosen quantile value, in order to conduct the recommended sensitivity analyses at subsequent steps.
+
We first create weights at the 92nd quantile value.
+
+
+quantile<-0.92
+
+trim_weights.s1<-trimWeights(weights =weights,
+ at =quantile,
+ save.out =save.out)
+
+print(trim_weights.s1,
+ i =1)
+#>
+#> For imputation 1 and the `cbps` weighting method, after trimming at 92th quantile, the median weight value is 0.77 (SD = 0.82; range = 0-3).
+
+plot(trim_weights.s1,
+ i =1,
+ save.out =save.out)
+
+
+
+
+
+
+
+
Trimming at the 92nd quantile preserves the same median of 0.77 but with an even smaller standard deviation and range.
+
And then at the 98th quantile value.
+
+
+quantile<-0.98
+
+trim_weights.s2<-trimWeights(weights =weights,
+ at =quantile,
+ save.out =save.out)
+
+print(trim_weights.s2,
+ i =1)
+#>
+#> For imputation 1 and the `cbps` weighting method, after trimming at 98th quantile, the median weight value is 0.77 (SD = 1.06; range = 0-5).
+
+plot(trim_weights.s2,
+ i =1,
+ save.out =save.out)
+
+
+
+
+
+
+
+
Trimming instead at the 98th quantile produces a larger standard deviation and range.
+
We find comparable descriptive statistics for all sets of weights, with the upper range value varying by quantile cutoff. We will assess the consequences of any differences (e.g., different ranges) in subsequent steps.
+
+
Step 4: Conduct Final Balance Assessment
+
+
Having created and trimmed the final set of IPTW balancing weights, the next step is to conduct a final evaluation of how well they reduce imbalance. We assess the performance of the final trimmed and untrimmed weights using assessBalance() function.
+
The required inputs for using the assessBalance() function to assess how the final, trimmed weights achieve balance for the full formulas are: complete data (data frame, mids object, or a list of imputed datasets as dataframes in wide format), an MSM object (e.g., “obj”), and weights.
+
The optional inputs for the assessBalance() function are detailed in Step 1b.
+summary(final_balance_stats_untrim,
+ save.out =save.out)
+#> USER ALERT: Averaging across imputed datasets using `cbps` weighting method: As shown below, 2 out of 241 (0.8%) covariates across time points remain imbalanced with a remaining median absolute correlation of 0.06 (max: 0.07):
+#> +-----------+-----------------------+----------------------------+
+#> | Exposure | Total # of covariates | # of imbalanced covariates |
+#> +===========+=======================+============================+
+#> | ESETA1.6 | 28 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.15 | 38 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.24 | 47 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.35 | 59 | 2 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.58 | 69 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#>
+#> Table: Imbalanced Covariates Averaging Across Imputed Datasetrs using `cbps`
+
+summary(final_weights[[1]])
+#> Time 1
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 8.5138
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1246 1145 294 697 919
+#> all 7.4581 7.6788 7.9902 8.2032 8.5138
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 1.066 0.725 0.443 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 605.04
+#>
+#> Time 2
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 8.5138
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1246 1145 294 697 919
+#> all 7.4581 7.6788 7.9902 8.2032 8.5138
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 1.066 0.725 0.443 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 605.04
+#>
+#> Time 3
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 8.5138
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1246 1145 294 697 919
+#> all 7.4581 7.6788 7.9902 8.2032 8.5138
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 1.066 0.725 0.443 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 605.04
+#>
+#> Time 4
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 8.5138
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1246 1145 294 697 919
+#> all 7.4581 7.6788 7.9902 8.2032 8.5138
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 1.066 0.725 0.443 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 605.04
+#>
+#> Time 5
+#> Summary of weights
+#>
+#> - Weight ranges:
+#>
+#> Min Max
+#> all 0 |---------------------------| 8.5138
+#>
+#> - Units with the 5 most extreme weights:
+#>
+#> 1246 1145 294 697 919
+#> all 7.4581 7.6788 7.9902 8.2032 8.5138
+#>
+#> - Weight statistics:
+#>
+#> Coef of Var MAD Entropy # Zeros
+#> all 1.066 0.725 0.443 0
+#>
+#> - Effective Sample Sizes:
+#>
+#> Total
+#> Unweighted 1292.
+#> Weighted 605.04
+
+
We see that the non-trimmed weights result in no imbalanced confounders and an effective sample size of 611.
+
In this scenario, given that, for the trimmed weights, the imbalanced confounder (income at 6 months) is only related to exposure at -0.051 and the trimmed weights yield a higher effective sample size, we proceed with the trimmed weights.
+
We then summarize the final balance statistics, averaging across imputed datasets. The user has the option to supply save.out to save plot output to the home directory.
+
+
+summary(final_balance_stats_trim,
+ save.out =save.out)
+#> USER ALERT: Averaging across imputed datasets using `cbps` weighting method: As shown below, 2 out of 241 (0.8%) covariates across time points remain imbalanced with a remaining median absolute correlation of 0.06 (max: 0.06):
+#> +-----------+-----------------------+----------------------------+
+#> | Exposure | Total # of covariates | # of imbalanced covariates |
+#> +===========+=======================+============================+
+#> | ESETA1.6 | 28 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.15 | 38 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.24 | 47 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.35 | 59 | 2 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.58 | 69 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#>
+#> Table: Imbalanced Covariates Averaging Across Imputed Datasetrs using `cbps`
+
+plot(final_balance_stats_trim,
+ t =2,
+ save.out =save.out)
+
+
+
+
+
+
+
+
In the outcome modeling step (Step 5), users have the option to include any remaining imbalanced confounders that are time invariant as covariates in the final outcome model. In this case, we would manually list out these imbalanced confounders that are time invariant and assign them to covariates.
+
+
+# covariates <- c("list_imbalanced_ti_conf")
+
+
+
Sensitvity Analyses
+
+
Subsequently, we also assess balance for the weights trimmed at the two additional quantile values to assess whether the final balance assessment is sensitive to the trim value.
+
We first assess balance for the weights trimmed at the 93rd quantile value.
From this, we similarly find that income at 6 months is imbalanced with respect to exposure at 15 months (albeit with a slighter stronger correlation than the main analyses).
+
+
We next assess balance for the weights trimmed at the 98th quantile value.
+
+
+weights<-trim_weights.s2
+
+final_balance_stats.s2<-assessBalance(data =data,
+ obj =obj,
+ weights =weights,
+ imp_conf =imp_conf,
+ balance_thresh =balance_thresh,
+ save.out =save.out)
+
+summary(final_balance_stats.s2,
+ save.out =save.out)
+#> USER ALERT: Averaging across imputed datasets using `cbps` weighting method: As shown below, 2 out of 241 (0.8%) covariates across time points remain imbalanced with a remaining median absolute correlation of 0.06 (max: 0.06):
+#> +-----------+-----------------------+----------------------------+
+#> | Exposure | Total # of covariates | # of imbalanced covariates |
+#> +===========+=======================+============================+
+#> | ESETA1.6 | 28 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.15 | 38 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.24 | 47 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.35 | 59 | 2 |
+#> +-----------+-----------------------+----------------------------+
+#> | ESETA1.58 | 69 | 0 |
+#> +-----------+-----------------------+----------------------------+
+#>
+#> Table: Imbalanced Covariates Averaging Across Imputed Datasetrs using `cbps`
+
+
From this, we find no remaining imbalanced confounders (similar to the untrimmed results).
+
+
PHASE 2: Assess Substantive Associations between Exposure and Outcome
+
+
Having created IPTW balancing weights that minimize associations between confounders and exposure at each time point, we can move to the substantive modeling phase.
+
Step 5: Fit Weighted Outcome Model & Summarize & Visualize Results
+
+
The goal of this final step is to fit a weighted model relating exposure at meaningful epochs of developmental time to the outcome, before summarizing and visualizing the results. In this step, the user models and compares various counterfactuals, or the effects of different developmental histories of exposure on the outcome, to test substantive hypotheses about dose and timing.
+
+
Step 5a. Select & Fit a Weighted Outcome Model
+
+
First, we use the fitModel() function to fit a weighted generalized linear model relating exposure to outcome. The function draws on the glm_weightit() function of the WeightIt package (Greifer, 2023). The exposure main effects in models reflect exposure levels at each exposure time point unless exposure epochs are specified. One of the benefits of creating balancing weights is that they can be used in a variety of different marginal outcome models and those encompassed in this function are only a subset of possible models. Note that these models can get complex and we do not advise interpreting the individual terms.
+
The required inputs for using the fitModel() function are: complete data (data frame, mids object, or a list of imputed datasets as dataframes in wide format), an MSM object (e.g., “obj”), an outcome, a list of trimmed weights, and a model from the list below (“m0”, “m1”, “m2”, or “m3”).
+
We first inspect the following list of models.
+
+
M0: Baseline model regressing the outcome on the main effects of exposure (e.g., infancy, toddlerhood, childhood).
+
M1: Covariate model regressing the outcome on the main effects of exposure as well as user-specified covariates (e.g., confounders measured at baseline or the first time point that remained imbalanced after weighting in Step 4).
+
M2: Interaction model regressing the outcome on the main effects of exposure as well as all interactions between exposure main effects (e.g., infancy:toddlerhood) of the user-specified interaction order
+
M3: Full model regressing the outcome on the main effects of exposure, user-specified covariates, as well as all exposure main effect interactions and interactions between expoure and covariates, of the user-specified interaction order
+
+
Below, we specify the M0 model.
+
+
+m<-"m0"
+
+
If the user selects a covariate model (“m1” or “m3”), they are required to supply a list to covariates that corresponds to covariates in the wide data (see Step 4).
+
The optional inputs to the fitModel() function are as follows.
+
If the user selects an interaction model (“m2” or “m3”), they are required to provide a interaction order integer in the int_order field that reflects the maximum interaction (e.g., 3) (that will automatically include all lower order interactions (e.g., 2-way)). The interaction order cannot exceed the total number of exposure main effects (3 in our case, as we specified 3 exposure epochs).
+
If we were to specify an interaction model (“m2” or “m3”), we would also need to specify the interaction order ( to be included with all lower-level interactoins) between exposure main effects to the fitModel() function below.
+
+
+#int_order <- 2
+
+
The user can also specify a family (as a function, not in quotations; e.g., gaussian) and link (in quotations, e.g., “identity”) functions for the generalized linear model (defaults are gaussian and “link”, respectively). The possible families are: binomial, gaussian, Gama, inverse.gaussian, poisson, quasi, quasibinomial, and quasipoisson. For binomial and Poisson families, set family to quasibinomial and quasipoisson, respectively, to avoid a warning about non-integer numbers of successes. The `quasi’ versions of the family objects give the same point estimates and standard errors and do not give the warning. The gaussian family accepts the links: “identity”, “log” and “inverse”; the binomial family the links “logit”, “probit”, “cauchit”, (corresponding to logistic, normal and Cauchy CDFs respectively) “log” and “cloglog” (complementary log-log); the Gamma family the links “inverse”, “identity”, and “log”; the poisson family the links “log”, “identity”, and “sqrt”; and the inverse.gaussian family the links 1/mu^2, inverse, identity and log. The quasi family accepts the links “logit”, “probit”, “cloglog”, “identity”, “inverse”, “log”, “1/mu^2”, and “sqrt”, and the function power can be used to create a power link function.
+
Below, we retain the default family and link functions.
+
+
+family<-gaussian
+
+link<-"identity"
+
+
This function returns a list of fitted model objects, each as glm output.
+
We fit the outcome model using trimmed weights below.
+
+
+weights<-trim_weights
+
+models<-fitModel(data =data,
+ obj =obj,
+ weights =weights,
+ outcome =outcome,
+ model =m,
+ family =family,
+ link =link,
+ save.out =save.out)
+
+
The function returns a list of models (one per imputed dataset, where applicable).
+
We then inspect the model output. The user has the option to supply save.out to save printed output to the home directory.
+
We inspect the model averaged across imputed datasets.
+
+
+print(models,
+ save.out =save.out)
+#> Please inspect the Wald test to determine if the exposures collectively predict significant variation in the outcome compared to a model without exposure terms.
+#> We strongly suggest only conducting history comparisons if the test is significant.
+#> test statistic df1 df2 dfcom p.value riv
+#> 1 ~~ 2 -1.181783 3 0.4300163 NA 1 2.021014
+#>
+#>
+#> The marginal model, m0, pooled across imputed datasets is summarized below:
+#> +--------------------+-------+-----------------+--------+
+#> | | (1) |
+#> +--------------------+-------+-----------------+--------+
+#> | | Est. | CI | p |
+#> +====================+=======+=================+========+
+#> | (Intercept) | 0.483 | [0.452, 0.513] | <0.001 |
+#> +--------------------+-------+-----------------+--------+
+#> | ESETA1.Childhood | 0.028 | [-0.112, 0.169] | 0.442 |
+#> +--------------------+-------+-----------------+--------+
+#> | ESETA1.Infancy | 0.003 | [-0.042, 0.048] | 0.891 |
+#> +--------------------+-------+-----------------+--------+
+#> | ESETA1.Toddlerhood | 0.012 | [-0.054, 0.079] | 0.650 |
+#> +--------------------+-------+-----------------+--------+
+#> | Num.Obs. | 1292 | | |
+#> +--------------------+-------+-----------------+--------+
+
+
Importantly, we first examine the results of the Wald test at the top of the output. In the case of imputed data, the likelihood ratio test is done by pooling parameters via Rubin’s Rules. This test compares the user-specified model to a nested version of that model that omits all exposure variables to test whether exposure predicts variation in the outcome. Models are pooled prior to conducting the likelihood ratio test for imputed data.)
+
+
Sensitvity Analyses
+
+
We then conduct sensitivity analyses, fitting the same model with weights that were trimmed at two different values.
+
Of note if save.out = TRUE using the default file naming, saving the model output will overwrite the output from the main model. To save out sensitivity analyses, we recomend the user supply a new name (e.g., save.out = “model_m1_s1.rds”).
+
We first fit the same model to the weights trimmed at the 92nd quantile.
+
+
+weights<-trim_weights.s1
+
+models.s1<-fitModel(data =data,
+ obj =obj,
+ weights =weights,
+ outcome =outcome,
+ model =m,
+ family =family,
+ link =link,
+ save.out =save.out)
+
+print(models.s1,
+ save.out =save.out)
+#> Please inspect the Wald test to determine if the exposures collectively predict significant variation in the outcome compared to a model without exposure terms.
+#> We strongly suggest only conducting history comparisons if the test is significant.
+#> test statistic df1 df2 dfcom p.value riv
+#> 1 ~~ 2 -0.42805 3 0.6223825 NA 1 1.252613
+#>
+#>
+#> The marginal model, m0, pooled across imputed datasets is summarized below:
+#> +--------------------+-------+-----------------+--------+
+#> | | (1) |
+#> +--------------------+-------+-----------------+--------+
+#> | | Est. | CI | p |
+#> +====================+=======+=================+========+
+#> | (Intercept) | 0.481 | [0.450, 0.513] | <0.001 |
+#> +--------------------+-------+-----------------+--------+
+#> | ESETA1.Childhood | 0.028 | [-0.085, 0.141] | 0.400 |
+#> +--------------------+-------+-----------------+--------+
+#> | ESETA1.Infancy | 0.006 | [-0.040, 0.051] | 0.784 |
+#> +--------------------+-------+-----------------+--------+
+#> | ESETA1.Toddlerhood | 0.015 | [-0.061, 0.091] | 0.601 |
+#> +--------------------+-------+-----------------+--------+
+#> | Num.Obs. | 1292 | | |
+#> +--------------------+-------+-----------------+--------+
+
+
We similarly find a significant likelihoood ratio test.
+
+
+
We then fit the same model with the weights trimmed at the 98th quantile.
+
+
+weights<-trim_weights.s2
+
+models.s2<-fitModel(data =data,
+ obj =obj,
+ weights =weights,
+ outcome =outcome,
+ model =m,
+ family =family,
+ link =link,
+ save.out =save.out)
+
+print(models.s2,
+ save.out =save.out)
+#> Please inspect the Wald test to determine if the exposures collectively predict significant variation in the outcome compared to a model without exposure terms.
+#> We strongly suggest only conducting history comparisons if the test is significant.
+#> test statistic df1 df2 dfcom p.value riv
+#> 1 ~~ 2 -1.653469 3 0.3504546 NA 1 2.86164
+#>
+#>
+#> The marginal model, m0, pooled across imputed datasets is summarized below:
+#> +--------------------+--------+-----------------+--------+
+#> | | (1) |
+#> +--------------------+--------+-----------------+--------+
+#> | | Est. | CI | p |
+#> +====================+========+=================+========+
+#> | (Intercept) | 0.485 | [0.454, 0.515] | <0.001 |
+#> +--------------------+--------+-----------------+--------+
+#> | ESETA1.Childhood | 0.029 | [-0.150, 0.209] | 0.487 |
+#> +--------------------+--------+-----------------+--------+
+#> | ESETA1.Infancy | -0.002 | [-0.055, 0.052] | 0.934 |
+#> +--------------------+--------+-----------------+--------+
+#> | ESETA1.Toddlerhood | 0.009 | [-0.061, 0.080] | 0.748 |
+#> +--------------------+--------+-----------------+--------+
+#> | Num.Obs. | 1292 | | |
+#> +--------------------+--------+-----------------+--------+
+
+
With a comparable result.
+
+
+
+
Step 5b. Estimate, Compare, and Visualize Model-Predicted Outcome as a Function of Exposure History
+
+
In this final step, we use the fitted model results to test substantive hypotheses about dose and timing. We estimate and then compare the average marginal estimates of the outcome for each user-specified exposure history (i.e., permutation of high (“h) and low (“l”) levels of exposure at each exposure epoch) using the compareHistories() function. This draws primarily from the avg_predictions() and hypotheses() functions in the marginaleffects package (Arel-Bundock, 2023).
+
First, the compareHistories() function creates average predictions of the outcome for each exposure history. For each of the n combinations of user-specified exposure histories, we set the value of those predictors in the full dataset to the values in that combination, leaving all other variables as they are. This gives us n datasets, each the same size as our original dataset used to fit the model. For the n datasets, we then compute the predicted values given the model before taking the average predicted value for each of the n datasets. These n averaged predicted values are the expected potential outcomes under each combination. (For imputed data, the function outputs pooled predicted values using Rubin’s Rules.)
+
Next, using the predicted values, the function conducts comparisons of the different histories (pooling across imputed datasets for imputed data using Rubin’s Rules). Lastly, the function implements correction for multiple comparisons (treating each run of the function as a family) before plotting the results. Box plots display outcome on the x-axis and exposure history on the y-axis and whiskers display standard errors.
+
The required inputs for using the compareHistories() function are: an MSM object (e.g., “obj”) and a fitted model output (from the previous steps).
+
The optional inputs are as follows.
+
To create histories of high and low values with continuous exposures, to hi_lo_cut the user can specify a list of two quantile values (0-1; default is median split +/- 0.001) demarcating high and low levels of exposure, respectively. (Imputed data are stacked to calculate cutoff values.) We suggest drawing on existing hypotheses and examining the variability in the exposure variable to determine high and low cutoffs. We recommend users begin by specifying meaningful high and low percentile cutoffs and examining how many individuals in the sample fall into each of the user-specified exposure histories created by those percentile cutoffs (see the Specify Core Inputs vignette). While there are no gold standard recommendations for sufficient cell numbers per history, users should ensure that there is is reasonable coverage in all of their histories to avoid extrapolation and maximize precision.
+
Below, we draw on the high and low cutoffs that we previously specified, that 60th and 30th percentile values to denote high and low levels of economic strain, respectively.
The user also has the option to estimate and compare only a custom subset of user-specified exposure histories (i.e., sequences of high and low levels of exposure at each epoch or time point) using the reference and comparison fields. To conduct these recommended customized comparisons, users must provide at least one unique valid history (e.g., “l-l-l”) as a reference by, in quotations, provide a string (or a list of strings) of lowercase l’s and h’s (each separated by -), each corresponding to each exposure epoch (or time point), that signify the sequence of exposure levels (“low” or “high”, respectively). If the user supplies a reference history, they are required to provide at least one unique and valid history for comparison by, in quotations, providing to comparison a string (or list of strings) of l’s and h’s (each separated by “-”), with each corresponding to each exposure epoch, that signify the sequence of exposure levels (“low” or “high”, respectively) that constitutes the comparison exposure history/histories to be compared to the reference. If the user supplies one or more comparisons, at least one reference must be specified. Each reference exposure history will be compared to each comparison history and all comparisons will be subject to multiple comparison correction.
+
If no reference or comparison is specified, all histories will be compared to each other. If there are more than 4 exposure main effects (either as epochs or exposure time points), the user is required to select a subset of history comparisons (Step 5b), given that the base code (see the hypotheses() function from the marginaleffects package; Arel-Bundock, 2023) cannot accommodate all pairwise history comparisons for more than 5 time points.
+
The user can also specify a multiple comparison method in mc_method by in quotations, providing the shorthand for the method (“holm”, “hochberg”,“hommel”, “bonferroni”, “BH” (default), “BY”, “fdr”, or “n” (see stats::p.adjust documentation; R Core Team) for multiple comparison correction to be applied to the final (pooled across imputed datasets when applicable) contrasts comparing effects of different exposure histories on outcome (default is Benjamini-Hochburg). Each code run is considered a family. If the user iterates through this function specifying different comparisons each time, we strongly recommend interpreting the outcome of the most inclusive set of comparisons to avoid false discovery.
+
Below, we retain the default Benjamini-Hochburg method for multiple comparison.
+
+
+mc_comp_method<-"BH"
+
+
Based on their substantive interests, the user also has the option to choose which level of dosage (“h” or “l”) is tallied in labels of dose counts in the tables and figures (dose_level; default is “h”). For example, if the exposure variable was coded in such a way that lower levels are conceptualized as the exposure (e.g., lower income), the user may wish to choose dosage level “l”.
+
Below, given our interest in histories of high economic strain, we specify that we wish to tally high doses of exposure.
+
+
+dose_level<-"h"
+
+
Lastly, the user can provide alternate plotting labels for exposure and outcome in the exp_lab and out_lab fields (defaults are variable names), as well as a list (equal to number of exposure main effects +1) of colors or a Brewer color palette (colors; default is “Dark2”). See RColorBrewer::display.brewer.all() or https://r-graph-gallery.com/38-rcolorbrewers-palettes.html).
The function returns a data frame of user-specified history comparisons containing contrast estimates, standard errors, statistics, p-values, low and high confidence intervals, and corrected p-values, labeled by history and dose.
The function outputs a list with two entries. The first contains predicted values for each history and the second contains the comparisons.
+
We then inspect the compared histories output (now averaged across all imputed datsets). The user has the option to supply save.out to save printed output to the home directory.
+
+
+print(results)
+#> Summary of Exposure Main Effects:
+#> USER ALERT: Out of the total of 1292 individuals in the sample, below is the distribution of the 333 (26%) individuals that fall into 2 user-selected exposure histories (out of the 27 total) created from 30th and 60th percentile values for low and high levels of exposure-epoch Infancy, Toddlerhood, Childhood.
+#> USER ALERT: Please inspect the distribution of the sample across the following exposure histories and ensure there is sufficient spread to avoid extrapolation and low precision:
+#>
+#> +---------------+-----+
+#> | epoch_history | n |
+#> +===============+=====+
+#> | h-h-h | 214 |
+#> +---------------+-----+
+#> | l-l-l | 119 |
+#> +---------------+-----+
+#>
+#> Table: Summary of user-selected exposure histories based on exposure main effects Infancy, Toddlerhood, Childhood:
+#>
+#> Below are the pooled average predictions by user-specified history:
+#> +-------+----------+-----------+----------+-----------+
+#> | term | estimate | std.error | conf.low | conf.high |
+#> +=======+==========+===========+==========+===========+
+#> | l-l-l | 0.47 | 0.014 | 0.44 | 0.49 |
+#> +-------+----------+-----------+----------+-----------+
+#> | h-h-h | 0.5 | 0.019 | 0.43 | 0.57 |
+#> +-------+----------+-----------+----------+-----------+
+#>
+#> Conducting multiple comparison correction for all pairings between comparison histories and each refernece history using the BH method.
+#>
+#>
+#> +-------------------+------------+------------+-----------+-------------+------------+--------------+
+#> | term | estimate | std.error | p.value | conf.low | conf.high | p.value_corr |
+#> +===================+============+============+===========+=============+============+==============+
+#> | (h-h-h) - (l-l-l) | 0.02991272 | 0.01979075 | 0.2172254 | -0.02903923 | 0.08886468 | 0.2172254 |
+#> +-------------------+------------+------------+-----------+-------------+------------+--------------+
+
+
We first confirm the distribution of our sample across our exposure histories is reasonable.
We then inspect the history comparison and conclude that there is no evidence that children who experienced different histories of exposure to economic strain over infancy, toddlerhood, and early childhood differ in their behavioral problems in early childhood.
+
Lastly, we plot the results. The user has the option to supply save.out to save plot output to the home directory.
We then conduct sensitivity analyses by assessing and comparing histories drawing from models that used weights trimmed at two different values. Of note if save.out = TRUE using the default file naming, saving the model output will overwrite the output from the main model. To save out sensitivity analyses, we recommend the user supply a new name (e.g., save.out = “history_comparisons_s1.rds”).
+
We first compare the same histories using the model fit with weights trimmed at the 92nd quantile value.
As shown above, results indicate a marginal but non-significant contrast between “l-l-l” and “h-h-h” histories of economic strain exposure in relation to behavior problems in early childhood.
+
+
We then compare the same histories using the model fit with weights trimmed at the 98th quantile value.
Burchinal, M., Howes, C., Pianta, R., Bryant, D., Early, D., Clifford, R., & Barbarin, O. (2008). Predicting Child Outcomes at the End of Kindergarten from the Quality of Pre-Kindergarten Teacher–Child Interactions and Instruction. Applied Developmental Science, 12(3), 140–153. https://doi.org/10.1080/10888690802199418
+
Cole, S. R., & Hernán, M. A. (2008). Constructing Inverse Probability Weights for Marginal Structural Models. American Journal of Epidemiology, 168(6), 656–664. https://doi.org/10.1093/aje/kwn164.
+
Greifer, Noah. 2023.WeightIt: Weighting for Covariate Balance in Observational Studies. https://CRAN.R-project.org/package=WeightIt.
+
Lumley, Thomas. 2023. “survey: Analysis of Complex Survey Samples.”
+
Polley, Eric, Erin LeDell, Chris Kennedy, and Mark van der Laan. 2023. SuperLearner: SuperLearner Prediction. https://CRAN.R-project.org/package=SuperLearner.
+
R Core Team (2013). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, URLhttp://www.R-project.org/.
+
Stuart, E. A. (2010). Matching methods for causal inference: A review and a look forward. Statistical Science: A Review Journal of the Institute of Mathematical Statistics, 25(1), 1–21. https://doi.org/10.1214/09-STS313.
+
Thoemmes, F., & Ong, A. D. (2016). A Primer on Inverse Probability of Treatment Weighting and Marginal Structural Models. https://doi.org/10.1177/2167696815621645.
+
Vernon-Feagans, L., Cox, M., Willoughby, M., Burchinal, M., Garrett-Peters, P., Mills-Koonce, R., Garrett-Peiers, P., Conger, R. D., & Bauer, P. J. (2013). The Family Life Project: An Epidemiological and Developmental Study of Young Children Living in Poor Rural Communities. Monographs of the Society for Research in Child Development, 78(5), i–150.
