Last update: 2024-04-16
This note describes the contents at https://github.com/rnj0nes/HCAP23. This repo is an update for the rj0nes/HCAP22 repo. The update revises the 2022 code addressing the following issues:
- Code is written to work with public versions of HRS/HCAP data (the rjones/HCAP22 repo was written to work with interim working files of HCAP data that included variables not included in public versions)
- Errors in variable handling are corrected
- Embedded LaTeX code in Stata do-files is removed
- Analyses are integrated into a single workflow (confirmatory factor analysis, normalization and standardization, algorithmic classification)
The code provides a record of data handling and analyses in two manuscripts (Manly et al 2022, Jones et al 2024). The code, as posted in this repo, cannot be used without access to non-public sensitive data. In a closing section I discuss approaches to modify this code to run without access to public data.
This repo contains (mostly) Stata code used in analyses of the Health and Retirement Study (HRS) Harmonized Cognitive Assessment Protocol (HCAP). Using these command files will require:
- Stata (files were prepared using version 18)
- R (version 4.1+)
- Mplus (version 7.1+)
The workhorse software is Stata.
The Stata code is organized inspired by the principles outlined in JS Long's Workflow of Data Analysis Using Stata (2008, Stata Corp). As such there is a "master" file that calls the individual Stata do files that organize the data and perform the analyses. It is intended that a single code fine does one major activity. To run the code, open the 000-master.do command file and run all of the lines. These lines include all of the sub-command files (e.g., 001-Environment-settings.do). The workflow organization is illustrated in Figure 1.
This figure illustrates the flow of data -- with input data shown in blue and green boxes on the left -- through Stata code programs, and the output generated. Output generated include other Stata data files (dta): also the do files generate other do files (discussed below) and other kinds of output, but this figure does not illustrate those outputs.
With this figure, you can see that one of the non-public data files (it is green: hcapnormexcldsummary_20210426.dta) is called by the Stata do file 017-call-data-nonpublic.do, which generates an output data file (w017.dta) which is in turn used by the Stata do file 041-Normalization-Standardization.do.
Note the command file Compile_Public_Data is not included in the repo.
Ideally, each component do file (e.g., 031-factor-score-estimation.do) should be able to be executed on it's own (that is, without having to run all of the previous do files) with the exception of 001-Environment-settings.do. So if a user wanted to only run the factor score estimation program, a small do file
include 001-Environment-settings.do
include 031-Factor-score-esitmation.do
should be sufficient, so long as the input dta files are available in the working folder. The figure shows that the input data for 031-factor-score-estimation.do is w021.dta.
I make use of personal ado files. I think the list of these is:
lowercase.ado(changes all variables to lower case)runmplus.ado(run Mplus)vlabel.ado(apply value labels)z1.ado(minmax normalization)scoreit.ado(create additive sums)itemsummary.ado(called by scoreit.ado)rdoc.ado(hack of Jann's texdoc.ado)
These are all included in this repo, except for runmplus.ado . Only one of these includes a sthlp file (scoreit.sthlp). Others may have helpful information available if you use Stata's viewsource command (e.g., viewsource z1.ado) after you have installed the ado.
runmplus.ado and associated files can be installed running the following from the Stata command line:
net install runmplus , from(https://s3.amazonaws.com/runmplusbucket) force
Note you must have Mplus (https://statmodel.com) to use the runmplus.ado.
I also make use of many ado files created by other Stata users, and hopefully most of these are available at SSC. Unfortunately I do not keep track of which of the commands I call are custom ados. The user will discover this, frustratingly, by encountering the Stata error message: command XYZ is unrecognized (where XYZ is the name of the command you tried to run). When this happens, the first thing to try is running ssc install XYZ from the Stata command line. If that does not work, try findit XYZ from the Stata command line. If neither of these work, email me (rich_jones@brown.edu) and I will add it to the repo.
Final note, a very easy way to avoid having to download and install my custom ados, and custom SSC ados, is to add a folder holding my custom ados to your adopath. The way to do this is to run
adopath +"https://quantsci.s3.amazonaws.com/Jones-ADO/"
from the Stata command line before trying to execute the command files. Now, it is possible you might run these programs and still get a command XYZ is unrecognized error, but in this case the missing command is a command available from SSC (or elsewhere). A way to protect yourself from this error is to also add a copy of my plus commands to your adopath, with:
adopath +"https://quantsci.s3.amazonaws.com/Jones-ADO-plus/"
Or, you could encounter these errors and add the commands to your own "plus" as they come up. That way is more painful and time consuming, but then you have these ados in your workflow more permanently.
includes the other do files.
Sets the working folder path, notes custom ados used.
Simply uses the public HCAP respondent data file as provided by HRS.
Simply uses the public HCAP informant data file as provided by HRS.
Calls (uses) the HRS Core data file from the 2016 wave, and the tracker file. The file that is called is the public access Stata data file, but the downloaded CSV has been processed into a dta file. Some variable handling for demographic variables is also included in this file.
This do file calls and minimally processes non-public data relevant to identifying persons in the HCAP sample who likely have dementia and should be excluded from the normative reference sample. See Manly et al (2022).
This small program generates a variable hcapage. This variable was present in early, working copies of the HCAP data, but is not included in the public data release. In order to make my code developed in these early working files perform correctly when calling the public data, I generate the hcapage variable.
This do file transforms public data cognitive performance scores into analytic variables used in the factor analysis. This is where missing data handling is contained, min/max normalization for selected variables, collapsing categories for sparsely observed categories, etc. Extensive annotations within the code justifying the variable handling and noting differences in variable handling from the rnj0nes/hcap22 repo.
This do file uses the runmplus.ado to call the Mplus software and conduct the confirmatory factor analysis models. Most of these models are described in Jones et al (2024). Factor score estimates are also generated. Note we generate both Bayesian plausible values and expected a posteriori (EAP) factor score estimates. They have different properties and useful in different situations. Standard errors for EAP estimates are also provided, when they are available. A list of scores generated is presented below.
This do file does the normalization and standardization of factor score estimates, including the adjustment for the effects of age, sex, education, and race/ethnicity.
This program cannot run without the non-public data. However, one of the main bits of output of this do file are the adjustment equations. The user will notice that this repo contains many more do files than are listed in Figure 1. For example, we have
do file |
Function |
|---|---|
gen_mem_blom_from_mem.do |
Generates Pmem_blom -- a Blom-transformed version of the mem variable (a factor score estimate from 031-factor-score-estimation.do). The Blom transformation is accomplished using observed values in the HRS/HCAP sample and hard coded into this do file using restricted cubic splines. |
gen_spage_from_age.do |
Generates cubic spline functions for age (spage1-sapge3) from hcapage. |
gen_Emem_blom.do |
Generates Emem_blom from spage1-sapge3, female, black, hisp, and schlyrs. This generated do file includes the parameter estimates from the adjustment regression equations. Capital E in Emem_blom is meant to indicate this is an expected score. The _blom indicates this is based on the Blom-transformed version of the factor score estimate, and the mem refers to the domain that is being scored. |
genTmem.do |
Generates Tmem from Emem_blom and Pmem_blom, and the HRS/HCAP observed standard deviation, and adjustment model r-squared. Tmem should have a mean of 50 and standard deviation of 10. |
There is one version of gen_spage_from_age.do, but there are versions of the gen_#_blom_from_#.do, gen_E#_blom.do, and genT#.do for many different kinds of factor score estimates (# is mem, exf, lfl, gcp, etc.). The table below summarizes the kinds of factor score estimates:
| Name | Domain | Kind of model | Kind of score |
|---|---|---|---|
| gmemm1 | Memory | Second order | PV |
| memm1 | Single factor | PV | |
| gmem | Second order | EAP | |
| mem | Single factor | EAP | |
| gexfm1 | Executive function | Second order | PV |
| exfm1 | Single factor | PV | |
| gexf | Second order | EAP | |
| exf | Single factor | EAP | |
| glflm1 | Language, fluency | Second order | PV |
| lflm1 | Single factor | PV | |
| glfl | Second order | EAP | |
| lfl | Single factor | EAP | |
| vdori1 | Orientation | Sum of correct | Not applicable |
| vdvis1 | Visuospatial | CERAD constructional praxis | Not applicable |
| gcpm1 | Global cognition | Second order | PV |
| gcp | Second order | EAP | |
| h1rmsetotal | MMSE score | Total score | Not applicable |
| Note: EAP, Expected a posteriori; PV, Bayesian plausible value |
This program imputes missing informant data and respondent cognitive performance data. It uses Mplus to do the imputation. It does make use of non-public data. A single imputation is drawn, each missing observation being replaced with a random draw from a posterior distribution of plausible values given the other variables in the data.
Assigns persons to normal vs MCI vs Dementia categories as described in Manly et al (2022). Note that in the manuscript, and in this code, cognitive impairment indicators use Bayesian plausible values: these have a stochastic component.
If the user would like to approximate the algorithm without relying upon the non-public data, it is possible to come close to the results that do use the non-public data. I will discuss with reference to Figure 2, which is the workflow diagram with some modifications.
Without sensitive data: the data files that are not available are hashed out. Their resulting output dta files are also hashed out. Without sensitive data, the do file 017-Call-data-nonpublic.do is not needed. The analysis files 041-Normalization-standardization.do and 051-Imputation.do will have to be modified, because as written these do files call processed non-public data.
The original 041-Normalization-standardization.do does two things (I should have made this two files, each doing one thing: sorry). It both generates and executes do files that construct Blom-transformed versions of factor scores estimated from 031-factor-score-estimation.do, generated predicted values for the cognitive performance scores given demographic variables, and standardizes these predicted values to a T distribution (mean 50, SD 10) taking into account the HRS/HCAP variability and r-squared of the adjustment model. These steps are described above in a little more detail. The edit that would be necessary to 041-Normalization-standardization.do to make it work without the non-public data is to strip out the code that generates the do files, and only retains the code that executes the generated do files. All of the generated do files should be in this repo. If anything is missing, let me know. Let's call this modified version of file 041: 041-normalization-standardization-public-version.do. The output of this program (I suggest calling/labeling/naming the output data file 041-output-public.dta) should be identical to that generated using the sensitive data.
The original 051-imputation.do file will also have to be modified. It will have to have variables from the sensitive data removed from the imputation model. I would call the revised program 051-imputation-public-version.do and the resulting output w051-postimputation-public.dta. The output from this file will not be the same as the original w051 dta file, because the input data will have changed as used in the imputation model. Even if the sensitive data used in the imputation model were not very important predictors of missingness or outcomes among those for whom data were not missing, because the imputation values are single plausible values from a posterior distribution, the imputed values will be different. Of course, values for persons without missing data will be identical.
Manly, J. J., Jones, R. N., Langa, K. M., Ryan, L. H., Levine, D. A., McCammon, R., Heeringa, S. G., & Weir, D. (2022). Estimating the Prevalence of Dementia and Mild Cognitive Impairment in the US: The 2016 Health and Retirement Study Harmonized Cognitive Assessment Protocol Project. JAMA Neurol, 79(12), 1242-1249. https://doi.org/10.1001/jamaneurol.2022.3543
Jones, R. N., Manly, J. J., Langa, K. M., Ryan, L. H., Levine, D. A., McCammon, R., & Weir, D. (2024). Factor structure of the Harmonized Cognitive Assessment Protocol neuropsychological battery in the Health and Retirement Study. Journal of the International Neuropsychological Society, 30(1), 47-55. https://doi.org/10.1017/S135561772300019X
Please feel free to contact me with questions, missing commands or files, or any other challenges you have using these resources.
Rich Jones