Exploring the Relation between Contextual Social Determinants of Health and COVID-19 Occurrence and Hospitalization

Round 1

Reviewer 1 Report

This manuscript describes a secondary data analysis of EHR data from PCORnet to examine the association of COVID-19 diagnosis and hospitalization with various social determinants of health (SDOH) variables in Florida patients. The manuscript is generally well-written (with some minor English language and grammar issues) and addresses an important topic that is of increasing importance across health care. Utilizing EHR data to better understand the impact of SDOH on patient  outcomes is valuable. There are numerous areas for improvement which would enhance the overall impact of the manuscript.

Introduction

1) Lines 44-46: Provide more context for the survey cited in these lines. Is this a national survey?  A single state survey?

2) Line 50: Provide some key examples of social and built environment SDOH factors.

3) Line 56: Operationalize the term "contextual-level" SDOH. How is this different from just SDOH generally?

4) Line 65: Provide an example of correlated SDOH.

5) Lines 76-81: Unclear that these examples are necessary.  EHRs are generally understood to be a useful source of research data and this is an informatics-focused journal, so readers should be familiar.

Materials and Methods

6) Describe the completeness of the health records used (e.g. to what degree is there missing data that would impact the analyses or reveal another barrier in trying to replicate this type of work)?

7) Lines 112-114: Provide a justification for matching in a 1:4 ratio. Why only matching on age, sex, and index month? There is a growing body of literature looking at COVID co-morbidities leading to worsening outcomes.

8) Line 121: Why include death as a category under hospitalization? Can't death also occur in an outpatient setting if treated inappropriately or if a patient had other co-morbid conditions worsened by COVID?

9) Table 1 lists 10-factor domains. What is the 11th (as noted in the text)?

10) Line 165: Specify which version of ICD codes were used.

11) Is it appropriate to use all of the SDOH variables? Being comprehensive is important but so is making theoretical sense - the authors should essentially be able to hypothesize for each variable the directionality of the outcome on an a priori basis. Because the sample size is reasonably large, it can detect differences that are statistically significant but practically meaningless.

Discussion

12) Lines 296-304: Is there any literature describing the impact of religious beliefs on vaccination practices? This could impact spread if religiosity was associated with decreased vaccine uptake (unless the dataset did not address the time period when vaccines became available).

13) Lines 311-315: As noted in the limitations but the authors - confounding across SDOH variables was likely not fully eliminated. This is even more of a justification for a combined approach with theory-based variable selection.

14) Lines 324-326: Is there a relationship between vacant business properties and population density? If yes, perhaps vacant businesses indicate less population and thus decreased COVID diagnoses.

15) Lines 339-340: Findings may not be generalizable to other states (not just other U.S. regions). There is large variation in state policy and other factors that are associated with SDOH (e.g. drinking water contamination in Flint, Michigan).

Writing would benefit from a review by someone who has expertise in SDOH and is a native English speaker.

Author Response

First, I would like to express our sincere thanks to the time and effort of the reviewer. Your comments and suggestions helped us further improve our study. According to each point of your comments, we revised our manuscript.

Introduction

1) Lines 44-46: Provide more context for the survey cited in these lines. Is this a national survey?  A single state survey?

Thanks for point out this confusion. We rewrite this sentence to separate the state-wide survey and the systematic review of the impact of COVID-19 on the health system. The new description is below:

According to Massachusetts state survey, 20% of the respondents are missing either critical urgent care or essential routine care, mostly due to limited health care capacity, while a second review reported a median of 37% reduction in medical service worldwide.

2) Line 50: Provide some key examples of social and built environment SDOH factors.

We added the following examples to the social and built environment factors.

Social factors: religion influences, social class.

Built environment: transportation, opportunity for physical activity.

3) Line 56: Operationalize the term "contextual-level" SDOH. How is this different from just SDOH generally?

We add the following description to further distinguish contextual-level SDOH with individual level SDOH.

Contextual SDoH is the social and built factors within community or region that influence health outcomes while the individual SDoH are the factors measured at patient level.

4) Line 65: Provide an example of correlated SDOH.

We added the following sentence to example as correlated SDOH:

However, most prior studies examined a limited number of contextual SDoH and thus may suffer from unmeasured confounding by co-exposures to other unincluded contextual SDoH, especially considering that a number of SDoH are correlated, i.e., the occupation and the income level at the community.

5) Lines 76-81: Unclear that these examples are necessary.  EHRs are generally understood to be a useful source of research data and this is an informatics-focused journal, so readers should be familiar.

Thanks for the suggestion. We have removed this content from the manuscript and combined the remaining description with the following paragraph. The modified paragraph was included below:

A number of COVID-19 research projects are based on EHR data. The topics of these projects, to name a few, include COVID-19 surveillance (data infrastructure and fore-casting) [26–29], COVID-19 outcome prediction [30–41], and social determinants of COVID-19 analyses [21,42–44]. In summary, although a number of studies [16–25] have associated contextual SDoH with COVID-19 outcomes, either poor geographical represented data from limited study sites were used or a few contextual SDoH factors were analyzed in these literature. There is still a large knowledge gap in our understanding of the relationships between contextual SDoH and COVID-19 outcomes. In this study, multiple contextual SDoH factors with EHR data from a large clinical research network (CRN) in the National Patient-Centered Clinical Research Network (PCORnet) to conduct an exploratory study on the impact of contextual SDoH on COVID-19 occurrence and hospitalization in Florida. Our analysis explored the major contextual SDoH factors that were potentially related to COVID-19 outcomes. To our best knowledge, this was the first study that explore the relation between a wide range of contextual SDoH and COVID-19 outcomes using EHRs.

Materials and Methods

6) Describe the completeness of the health records used (e.g. to what degree is there missing data that would impact the analyses or reveal another barrier in trying to replicate this type of work)?

Thanks for raise this issue. The electronic health records did not contain missing variables. The imputation was conducted with the SDoH factors as the measurement could be missing at certain locations where the patients resided. The method used was described in Hu’s paper under citation [58].

7) Lines 112-114: Provide a justification for matching in a 1:4 ratio. Why only matching on age, sex, and index month? There is a growing body of literature looking at COVID co-morbidities leading to worsening outcomes.

Thanks for raise this issue. We adopted this ratio from our previous study among ADRD and breast cancer analysis [1]. For the co-morbidities, indeed the co-morbidities would affect the outcome among COVID-19 patients. However, using the co-morbidities as the matching criteria would greatly reduce the number of patients in the matching cohort and thus reduce the number of COVID-19 patients as they could not be matched. As a result, instead of using the comorbidities as matching criteria, we considered them as covariates in our data analysis where the effect of the comorbidities were measured and controlled. We added the paper as a justification for our matching ratio.

[1] Chen, A.; Li, Y.; Woodard, J.N.; Islam, J.Y.; Yang, S.; George, T.J.; Shenkman, E.A.; Bian, J.; Guo, Y. The Impact of Race–Ethnicity and Diagnosis of Alzheimer’s Disease and Related Dementias on Mammography Use. Cancers 2022, 14, 4726. https://doi.org/10.3390/cancers14194726

8) Line 121: Why include death as a category under hospitalization? Can't death also occur in an outpatient setting if treated inappropriately or if a patient had other co-morbid conditions worsened by COVID?

Thanks for mention this. This would be a limitation of the study due to the nature of EHRs. The EHR system currently would not be able to record the death information of outpatients as it is not connected to the electronic death register system from the government. Although there have been proposals to resolve this issue, the solution was not implemented in the oneFlorida+ EHR system. We added the following paragraph in the discussion section to discuss this limitation.

Third, there are also many known limitations to EHR. For example, inaccuracy and vague ICD coding is a known issue to the EHR that could affect lead to misidentification on patient condition. Also, the EHRs is currently not recording death cases outside inpatient cases due to not being linked with the government’s death register system.

9) Table 1 lists 10-factor domains. What is the 11th (as noted in the text)?

Thanks for point out this issue. We have revised Table 1 to separate the Green Space, CDC Social Vulnerability Index, Area Deprivation Index, and Social Capital. Currently the revised table better shows the 11 factor domains.

10) Line 165: Specify which version of ICD codes were used.

Thanks for point out this issue, we have clarified that we use both ICD-9 and ICD-10 in this process.

11) Is it appropriate to use all of the SDOH variables? Being comprehensive is important but so is making theoretical sense - the authors should essentially be able to hypothesize for each variable the directionality of the outcome on an a priori basis. Because the sample size is reasonably large, it can detect differences that are statistically significant but practically meaningless.

Thanks for point out this issue. Indeed, this leads to the tread-off made at the study design phase. Firstly, our study did not seek to draw any causal relationship between the SDoH and the COVID-19 outcomes. Second, as our study targeted at discovering the SDoH factors that potentially associated with the incidence and severity of the study, we decided to include a wide variety of factors as the first step of our studies.

Discussion

12) Lines 296-304: Is there any literature describing the impact of religious beliefs on vaccination practices? This could impact spread if religiosity was associated with decreased vaccine uptake (unless the dataset did not address the time period when vaccines became available).

Thanks for the suggestion. We performed a search through the literature and identify some literature discussing the effect of religion on COVID-19 vaccine uptake.

Janet W. Rich-Edwards, Carissa M. Rocheleau, Ming Ding, Jennifer A. Hankins, Laura M. Katuska, Xenia Kumph, Andrea L. Steege, James M. Boiano, and Christina C. Lawson, 2022:

COVID-19 Vaccine Uptake and Factors Affecting Hesitancy Among US Nurses, March–June 2021, American Journal of Public Health 112, 1620_1629, https://doi.org/10.2105/AJPH.2022.307050

Sides E, Jones LF, Kamal A, et al, Attitudes towards coronavirus (COVID-19) vaccine and sources of information across diverse ethnic groups in the UK: a qualitative study from June to October 2020, BMJ Open 2022;12:e060992. doi: 10.1136/bmjopen-2022-060992

Wirsiy FS, Nkfusai CN, Ako-Arrey DE, Dongmo EK, Manjong FT, Cumber SN. Acceptability of COVID-19 Vaccine in Africa. Int J MCH AIDS. 2021;10(1):134-138. doi: 10.21106/ijma.482. Epub 2021 Apr 8. PMID: 33868778; PMCID: PMC8039868.

We included them into the discussion and added the following discussion.

An explanation for this observation could be to the suboptimal uptake of COVID-19 vaccine caused by religion factors. A number of studies had reported that religion factors had negative impact on COVID-19 vaccination in the United States, the United Kingdom, and Africa.

13) Lines 311-315: As noted in the limitations but the authors - confounding across SDOH variables was likely not fully eliminated. This is even more of a justification for a combined approach with theory-based variable selection.

Thanks for raise this point. Indeed, the combined approach were the golden standard for analyzing known SDoH factors, especially with specific domain and theory, on the causal relations with the outcomes. However, our study was designed to explore and identify new SDoH factors associated with COVID-19 incidence and outcome. These factors may or may not had theory support and we did not seek to draw any conclusions on causal relations.

14) Lines 324-326: Is there a relationship between vacant business properties and population density? If yes, perhaps vacant businesses indicate less population and thus decreased COVID diagnoses.

Thanks for raise this question. After checking the rest SDOH variables, we found that the percentage of vacant resident property was highly correlated with the percent of vacant business properties. This indicated that the population density had been low in the area. We added the following section into our discussion section:

This could be the result of the low population density in the area as the percentage of vacant business properties was found highly correlated with the percentage of vacant resident properties.

15) Lines 339-340: Findings may not be generalizable to other states (not just other U.S. regions). There is large variation in state policy and other factors that are associated with SDOH (e.g. drinking water contamination in Flint, Michigan).

Thanks for point out this issue, I have modified the original description to the following:

Our findings on the contextual SDoH factors are not necessarily generalizable to the other US regions or states, where state policy and other factors associated to the SDoH differs from that in Florida.

Reviewer 2 Report

The study addressed the knowledge gap contextual social determinants of health  (SDoH) related to COVID-19 occurrence and outcomes.

The proposed study linked 199 contextual SDoH factors covering 11 domains of social and built environment.

Major concerns

Using EHR data there will be a percentage of missingness, for example: Zip codes. 

“Missing data for all environmental factors and covariates were imputed using the chained equation method”. Is this addresses all the covariates?

Authors will be interested to know some of the contextual SDoH factors such as ‘Green Space’ in table 1.

I think this needs to be mentioned in the abstract as well

“86 contextual SDoH variables were included in the analysis”

Most contextual SDoH found statistically significant with COVID-19 outcomes in our study are consistent with the literature.

It should be discussed that which of the covariates are newly discovered in Table 4 precisely, otherwise the study can be viewed as a repetition of previous works by just increasing the number of variables in the initial phase.

This could indicate that religious practices could be a confounder of COVID-19 outcomes

bed utilization might be a confounder

Authors needs to elaborate more on these confounders as these values are makes it to table 4.

It would be also be helpful for authors to know the handful of contextual SDoH

In the abstract, in short forms, if the full name makes abstract lengthy.

Minor Comments

For the sentence “For environmental factors measured at 146

spatial scales larger than that of ZIP code, i.e., country,”

Did the authors mean county instead of country?

“geolocation would affect the effect of SDoH”, check for additional space

Author Response

First, I would like to express our sincere thanks to the time and effort of the reviewer. Your comments and suggestions helped us further improve our study. According to each point of your comments, we revised our manuscript.

The study addressed the knowledge gap contextual social determinants of health (SDoH) related to COVID-19 occurrence and outcomes.

The proposed study linked 199 contextual SDoH factors covering 11 domains of social and built environment.

Major concerns

Using EHR data there will be a percentage of missingness, for example: Zip codes. 

“Missing data for all environmental factors and covariates were imputed using the chained equation method”. Is this addresses all the covariates?

Thanks for point out this confusion. Our study only included the patients with valid ZIP code. The imputation was applied where certain contextual SDoH value was missing at the patients’ residential address.

We updated the description of the data extraction to the following:

We extracted EHR data between January 1st, 2020 and May 20th, 2021 on patients with valid ZIP codes whose latest address was in Florida.

We updated the description of the imputation to the following to exclude the covariates which are not imputed:

Missing data for all environmental factors were imputed using the chained equation method.

Authors will be interested to know some of the contextual SDoH factors such as ‘Green Space’ in table 1.

Thanks for provide this suggestion. We added additional description to each of the category of SDoH factor. The updated description is included below:

Vacant land record the vacancy of resident and business address at census tract level and thus represented the prosperity & population density. The vacant land variables were obtained from the US Department of Housing and Urban Development [46]. The social vulnerability index was a measure of social vulnerability obtained from the Centers for Disease Control and Prevention [47]. The area deprivation index Area deprivation was measured with the Area Deprivation Index obtained from Neighborhood Atlas [48]. Social capital data were obtained from the US Census Bureau Business Patterns [49], with the types of establishments determined using the North American Industry Classification System (NACIS) codes [50]. Crime and safety variables were obtained from the Uniform Crime Reporting Program [51]. Hospital bed capacity variables were ob-tained from Definitive Healthcare [52]. Healthcare status variables were obtained from the Area Health Resources Files [53]. Walkability was measured using the well-validated walkability index [54]. The food access and food environment data were obtained from the US Department of Agriculture’s Food Access Research Atlas [55] and Food Environment Atlas [56]. The green space measured the ratio of green space at the geological location. It was measured using the normalized difference vegetation index (NDVI) from the National Aeronautics and Space Administration based on satellite imaging [57].

I think this needs to be mentioned in the abstract as well

“86 contextual SDoH variables were included in the analysis”

Thanks for the suggestion, we added an additional sentence for this in the abstract:

After removed the highly correlated SDoH variables, 86 contextual SDoH variables were included in the data analysis.

Most contextual SDoH found statistically significant with COVID-19 outcomes in our study are consistent with the literature.

It should be discussed that which of the covariates are newly discovered in Table 4 precisely, otherwise the study can be viewed as a repetition of previous works by just increasing the number of variables in the initial phase.

Thanks for point out this issue. We added additional description to point out the unique SDoH factors we found and explanation of the possible reason.

We identified some new SDoH factors that are associated with COVID-19 incidence and outcome. First, we found the percentage of vacant business properties was a reflection of the local economic status and was negatively correlated with the area’s prosperity [78]. This could be the result of the low population density in the area as the percentage of vacant business properties was found highly correlated with the percentage of vacant resident properties. Additionally, we also found higher percentage of farmers’ markets that report selling baked/prepared food products was associated with a lower likelihood of getting hospitalized in COVID-19 positive patients. This factor was not reported before as a factor for COVID-19 outcome. Upon investigation, we found this factor highly correlated with the percentage of farmers’ markets that report selling vegetables. This revels that healthy diet and lifestyle had great effect on COVID-19 outcome.

This could indicate that religious practices could be a confounder of COVID-19 outcomes

bed utilization might be a confounder

Authors needs to elaborate more on these confounders as these values are makes it to table 4.

Thanks for the suggestion. For religious practice, we add a discussion where we found the religious factor was reported to affect COVID-19 vaccine uptake. The added discussion section is included below:

Our findings on the contextual SDoH factors are not necessarily generalizable to the other US regions or states, where state policy and other factors associated to the SDoH differs from that in Florida.

For the bed utilization, we found it correlated with the variables from healthcare indicator. Indicating the available medical resources in the area. We added the discussion section below for the bed utilization.

The rate of hospital bed utilization rate was proposed as a measure of a hospital’s ability to function safely and effectively, with high bed utilization being associated with a greater risk of hospital-associated infection [77]. However, we found that higher bed utilization was associated with a lower likelihood of COVID-19 infection. Further analysis revealed high correlation between hospital bed utilization and the number of available beds suggesting that bed utilization might be a confounder for the available medical resources.

It would be also be helpful for authors to know the handful of contextual SDoH

In the abstract, in short forms, if the full name makes abstract lengthy.

Thanks for the suggestion, we added the short names of the SDoH to the abstract. The sentence describing the findings on SDoH factor is included below:

Adjusting for race, ethnicity and comobidities, we found six contextual SDoH variables (i.e., hospital available beds & utilization, percent of vacant property, # of golf course, and % of minor) related to the occurrence of COVID-19, and three variables (i.e., farmer market, low access, and religion) related to the hospitalization of COVID-19.

Minor Comments

For the sentence “For environmental factors measured at 146

spatial scales larger than that of ZIP code, i.e., country,”

Did the authors mean county instead of country?

Thanks for point out this error. We corrected it to county.

“geolocation would affect the effect of SDoH”, check for additional space

Thanks for point out this error. We have removed the additional space.