Development of Iron Status Measures during Youth: Associations with Sex, Neighborhood Socioeconomic Status, Cognitive Performance, and Brain Structure

Project overview

This page includes detailed output of all analyses executed for this project. It begins with an overview of the sample characteristics. The results the are organized into four parts:
1. The developmental analyses. These include effects of age, sex, menarche, and socioeconomic status on each hematological measure of iron status.
2. Cognitive analyses. These include the cross-sectional associations between measures of iron status and cognitive factor scores from the Penn Computerized Neurocognitive Battery. Additional analyses are broken down by sex (including a moderated mediation analysis) and individuals who do not meet the laboratory threshold for iron deficiency.
3. Principal component analysis. We use PCA to to isolate an iron status index from an inflammatory index. We then conduct demographic and cognitive analyses using the PCA-derived iron status index.
4. Cross-sectional associations between hemoglobin concentration and brain white matter integrity (from diffusion-weighted MRI).

The Rmarkdown code used to generate this file can be found at the github repository.

Sample

This sample was drawn from the Philadelphia Neurodevelopmental Cohort (PNC). As previously described in detail, the PNC consists of 9,498 participants aged 8-22 years that underwent cognitive assessment and a subset of 1,601 youths that also completed neuroimaging(Calkins et al., 2015; Satterthwaite et al., 2014). This study used a subset of the PNC sample for which electronic medical records (EMR) of blood draws that included at least one indicator of peripheral iron status (hemoglobin, transferrin, or ferritin) were available (n = 6,926, 92,720 total observations). Study participants were excluded from analyses if their EMR indicated a significant or severe physical medical condition (n = 1,954). Medical condition severity was determined based on prior work in this sample(Merikangas et al., 2015). Lab results with serum ferritin levels above the normal reference range were excluded from all analyses (described below; 307 total observations). For participants with repeated measures, all blood draws that met inclusion criteria were included. The final sample included 30,190 hemoglobin levels from 4,874 participants, 1,353 ferritin levels from 537 participants, and 1,335 transferrin levels from 526 participants (see eTable1 for demographic summary statistics for each measure).

eTable 1. Sample demographics by iron status metric.

Iron status metric	Participants (N)	Visits (N)	Male/Female	Age range (y)	Mean (SD)
Ferritin	537	1353	218/319	0.85-22.78	13.68 (4.84)
Hgb	4874	30190	2161/2713	0.00-22.98	13.18 (4.80)
Transferrin	526	1335	199/327	0.21-22.97	13.66 (4.82)

Overall sample demographics.

Participants (N)	Visits (N)	Male/Female	Age range (y)	Mean (SD)
4899	32878	2178/2721	0.00-22.98	13.22 (4.80)

---

Results

Part 1: Developmental effects

Age effects for individual blood iron measures

Fitting GAMMs to model the age and sex effects for Hgb, ferritin, and transferrin.

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## Hgb ~ s(Age, k = 5, fx = F) + oSex + s(Age, by = oSex, k = 5, 
##     fx = F)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept) 13.04404    0.01917  680.53   <2e-16 ***
## oSex.L      -0.68666    0.02710  -25.34   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                     edf Ref.df     F p-value    
## s(Age)            3.942  3.942 348.4  <2e-16 ***
## s(Age):oSexFemale 3.942  3.942 205.3  <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.0541   
## lmer.REML = 1.0033e+05  Scale est. = 1.2581    n = 30190

## 
## Sig change in Male:
## Increasing from 0.000 to 20.559
## 
## Sig change in Female:
## Increasing from 0.000 to 10.395
## Decreasing from 11.665 to 17.440

## 
## Family: Gamma 
## Link function: log 
## 
## Formula:
## Ferritin ~ s(Age, k = 5, fx = F) + oSex + s(Age, by = oSex, k = 5, 
##     fx = F)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  3.40642    0.04017  84.801  < 2e-16 ***
## oSex.L      -0.43179    0.05610  -7.697 2.68e-14 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                     edf Ref.df      F   p-value    
## s(Age)            2.869  2.869  7.898 0.0000317 ***
## s(Age):oSexFemale 1.321  1.321 10.914  0.000278 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.0527   
## glmer.ML = 381.06  Scale est. = 0.39681   n = 1353

## 
## Sig change in Male:
## Increasing from 0.851 to 22.779
## 
## Sig change in Female:
## Increasing from 0.851 to 7.904
## Decreasing from 10.548 to 16.829

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## Transferrin ~ s(Age, k = 5, fx = F) + oSex + s(Age, by = oSex, 
##     k = 5, fx = F)
## 
## Parametric coefficients:
##             Estimate Std. Error t value     Pr(>|t|)    
## (Intercept)  275.764      2.535 108.782      < 2e-16 ***
## oSex.L        20.422      3.581   5.703 0.0000000145 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                     edf Ref.df     F p-value  
## s(Age)            1.725  1.725 1.313  0.2312  
## s(Age):oSexFemale 2.485  2.485 3.772  0.0134 *
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.0496   
## lmer.REML =  14196  Scale est. = 1418.1    n = 1335

## 
## Sig change in Male:
## No change
## 
## Sig change in Female:
## Increasing from 10.159 to 15.422

eTable 2. Regression table for age-by-sex interaction.

Response var.	edf	F	R2partial	P	PFDR
Hgb	3.94	205.3	0.01	< .001	< .001
Ferritin	1.32	10.9	0.02	< .001	< .001
Transferrin	2.48	3.8	0.01	.013	.013
Note:
abbr. edf: Effective degrees of freedom; FDR: False discovery rate.

Puberty effects in females

We saw a divergence in the developmental effects around the onset of adolescence between males and females. A possible cause of this sex difference is the effect of menstruation on the body’s iron status. To test this hypothesis, we look at the same metrics in females relative to the self-reported onset of menarche.

If our hypothesis is correct, we should see nonlinear effects such that the slope of Hgb by age differs pre and post menarche. In the regression models below, age_menarche_centered is age in years recentered on the self-reported age of menarche in female participants.

Number of participants included: 1,809
Number of observations included: 10,471

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## Hgb ~ s(age_menarche_centered, k = 5, fx = F)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept) 12.57868    0.02968   423.8   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                            edf Ref.df    F p-value    
## s(age_menarche_centered) 3.791  3.791 23.4  <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.00283   
## lmer.REML =  33759  Scale est. = 1.101     n = 10471

## 
## Sig change in age_menarche_centered:
## Increasing from -9.755 to -0.326
## Decreasing from 1.163 to 5.531

## 
## Family: Gamma 
## Link function: log 
## 
## Formula:
## Ferritin ~ s(age_menarche_centered, k = 5, fx = F)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  2.98791    0.06519   45.84   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                            edf Ref.df     F p-value  
## s(age_menarche_centered) 2.237  2.237 3.611  0.0134 *
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  -0.0197   
## glmer.ML = 138.37  Scale est. = 0.42581   n = 484

## 
## Sig change in age_menarche_centered:
## Decreasing from 0.790 to 5.699

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## Transferrin ~ s(age_menarche_centered, k = 5, fx = F)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  295.042      4.163   70.88   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                            edf Ref.df     F p-value  
## s(age_menarche_centered) 2.788  2.788 3.478   0.023 *
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  -0.0139   
## lmer.REML = 5053.4  Scale est. = 1438.4    n = 471

## 
## Sig change in age_menarche_centered:
## Increasing from 0.512 to 4.125

eTable 3. Association between iron status measures and menarche timing.

Response var.	edf	F	R2partial	P	PFDR
Hgb	3.79	23.4	0.01	< .001	< .001
Ferritin	2.24	3.6	0.02	.013	.02
Transferrin	2.79	3.5	0.02	.023	.023
Note:
abbr. edf: Effective degrees of freedom; FDR: False discovery rate.

Socioeconomic status

Next, we investigate whether the neighborhood socioeconomic status (based on geocoded data) was associated with iron status.

Main effects

Does iron status vary by SES, irrespective of age and sex?

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## Hgb ~ oSex + s(Age, k = 5) + s(Age, by = oSex, k = 5) + s(envSES, 
##     k = 4)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept) 13.07024    0.01927  678.11   <2e-16 ***
## oSex.L      -0.66764    0.02699  -24.74   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                     edf Ref.df      F p-value    
## s(Age)            3.937  3.937 331.84  <2e-16 ***
## s(Age):oSexFemale 3.921  3.921 202.92  <2e-16 ***
## s(envSES)         2.446  2.446  34.05  <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.0679   
## lmer.REML = 1.0025e+05  Scale est. = 1.2568    n = 30190

## 
## Family: Gamma 
## Link function: log 
## 
## Formula:
## Ferritin ~ oSex + s(Age, k = 5) + s(Age, by = oSex, k = 5) + 
##     s(envSES, k = 4)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  3.41286    0.04010  85.105  < 2e-16 ***
## oSex.L      -0.44234    0.05661  -7.814 1.11e-14 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                     edf Ref.df     F p-value   
## s(Age)            2.660  2.660 4.434 0.00336 **
## s(Age):oSexFemale 2.117  2.117 6.479 0.00119 **
## s(envSES)         1.000  1.000 1.609 0.20479   
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.0547   
## glmer.ML = 379.63  Scale est. = 0.39447   n = 1353

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## Transferrin ~ oSex + s(Age, k = 5) + s(Age, by = oSex, k = 5) + 
##     s(envSES, k = 4)
## 
## Parametric coefficients:
##             Estimate Std. Error t value     Pr(>|t|)    
## (Intercept)  275.725      2.542 108.462      < 2e-16 ***
## oSex.L        20.283      3.620   5.602 0.0000000257 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                     edf Ref.df     F p-value  
## s(Age)            1.728  1.728 1.272  0.2386  
## s(Age):oSexFemale 2.490  2.490 3.694  0.0146 *
## s(envSES)         1.001  1.001 0.073  0.7881  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.0488   
## lmer.REML =  14193  Scale est. = 1418      n = 1335

term	edf	ref.df	statistic	p.value	fdr.p
s(envSES)	2.445758	2.445758	34.0499920	0.0000000	0.0000000
s(envSES)	1.000005	1.000005	1.6094202	0.2047934	0.3071901
s(envSES)	1.000863	1.000863	0.0727617	0.7881131	0.7881131

Interaction effects

We observed a significant effect of SES for Hgb. Now, we test whether SES moderates the developmental trajectory of iron status in males and females. Due to the complexity of a three-way interaction GAMM, males and females were modeled separately:
Hgb ~ ti(envSES) + ti(Age)+ ti(envSES,Age) for males and females separately.

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## Hgb ~ s(envSES, k = 4, fx = F) + s(Age, k = 5, fx = F) + ti(envSES, 
##     Age, k = 4, fx = F)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept) 12.59805    0.02392   526.8   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                  edf Ref.df      F  p-value    
## s(envSES)      1.000  1.000 47.238  < 2e-16 ***
## s(Age)         3.825  3.825 35.313  < 2e-16 ***
## ti(envSES,Age) 2.928  2.928  6.613 0.000329 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.026   
##   Scale est. = 1.1208    n = 16220

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## Hgb ~ s(envSES, k = 4, fx = F) + s(Age, k = 5, fx = F) + ti(envSES, 
##     Age, k = 4, fx = F)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept) 13.47342    0.03087   436.5   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                  edf Ref.df       F  p-value    
## s(envSES)      1.000  1.000  35.056  < 2e-16 ***
## s(Age)         3.945  3.945 276.551  < 2e-16 ***
## ti(envSES,Age) 3.474  3.474   5.256 0.000774 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.0731   
##   Scale est. = 1.4072    n = 13970

term	edf	ref.df	statistic	p.value	fdr.p	partialR2
ti(envSES,Age)	3.473973	3.473973	5.255564	0.0007737	0.0007737	0.0022714
ti(envSES,Age)	2.928267	2.928267	6.612679	0.0003288	0.0006575	0.0018424

Part 2: Cogntion

Test the association between cognitive data and the individual blood iron measures

This association was testing using GAMs that co-vary for age and sex. e.g.
gam(NAR_Overall_Accuracy ~ sex + s(COG_AGE) + s(Hgb), data = Hgb,method = "REML")

Demographics table for cognitive models

Metric	N	Male/Female	Age range	Mean (SD)
Ferritin	61	19/42	8.83-20.67	14.28 (3.17)
Hgb	1302	533/769	8.08-21.92	14.29 (3.28)
Transferrin	66	21/45	8.75-20.67	15.27 (2.91)
Total	1429	573/856	8.08-21.92	14.33 (3.26)

Participants who had at least one iron status metric within six months prior to participation in the PNC were also included in cognitive (\(n\) = 1,429; male/female = 573/856; ages 8.08-21.92y, \(M (SD)\) = 14.33 (3.26)).

Hemoglobin

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## NAR_Overall_Accuracy ~ sex + s(COG_AGE) + s(Hgb, k = 4)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)  
## (Intercept) -0.10581    0.07793  -1.358   0.1748  
## sex          0.08488    0.04704   1.805   0.0714 .
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##              edf Ref.df     F    p-value    
## s(COG_AGE) 4.844  5.935 95.60    < 2e-16 ***
## s(Hgb)     1.003  1.005 26.09 0.00000109 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.331   Deviance explained = 33.5%
## -REML = 1548.3  Scale est. = 0.61857   n = 1302

## Now testing cognitive subdomains of Executive, Social, and Memory.

eTable 4. Association between hemoglobin and cognitive subdomains.

Response var.	edf	F	R2partial	P	PFDR
Executive	1.00	29.7	0.02	< .001	< .001
Social	1.55	3.5	0.01	.023	.023
Memory	1.00	9.9	0.01	.002	.003
Note:
abbr. edf: Effective degrees of freedom; FDR: False discovery rate.

Reviewer response: Comparing timing thresholds.

[[1]]

Association between hemoglobin and cognitive subdomains (1 year).

Response var.	edf	F	R2partial	P	PFDR
Executive	1	29.1	0.01	< .001	< .001
Social	1	8.9	0.00	.003	.003
Memory	1	9.0	0.00	.003	.003
Overall	1	26.5	0.01	< .001	< .001
Note:
abbr. edf: Effective degrees of freedom; FDR: False discovery rate.

[[2]]

Association between hemoglobin and cognitive subdomains (3 month).

Response var.	edf	F	R2partial	P	PFDR
Executive	1.10	23.4	0.03	< .001	< .001
Social	1.35	1.6	0.00	.148	.148
Memory	1.00	9.3	0.01	.002	.003
Overall	1.02	21.6	0.02	< .001	< .001
Note:
abbr. edf: Effective degrees of freedom; FDR: False discovery rate.

[[3]]

Association between hemoglobin and cognitive subdomains (6 month).

Response var.	edf	F	R2partial	P	PFDR
Executive	1.00	29.7	0.02	< .001	< .001
Social	1.55	3.5	0.01	.023	.023
Memory	1.00	9.9	0.01	.002	.002
Overall	1.00	26.1	0.02	< .001	< .001
Note:
abbr. edf: Effective degrees of freedom; FDR: False discovery rate.

Ferritin

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## NAR_Overall_Accuracy ~ sex + s(COG_AGE, fx = F) + s(Ferritin, 
##     k = 4, fx = F)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)
## (Intercept)  0.05011    0.43265   0.116    0.908
## sex         -0.01364    0.24836  -0.055    0.956
## 
## Approximate significance of smooth terms:
##               edf Ref.df      F   p-value    
## s(COG_AGE)  2.252  2.827 10.787 0.0000201 ***
## s(Ferritin) 1.000  1.000  0.843     0.362    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.329   Deviance explained = 37.6%
## GCV = 0.75565  Scale est. = 0.69059   n = 61

## [1] 0.01408571

eTable 5. Association between ferritin and cognitive subdomains.

Response var.	edf	F	R2partial	P	PFDR
Executive	1	1.5	0.02	.223	.668
Social	1	0.0	0.00	.937	.937
Memory	1	0.1	0.00	.735	.937
Note:
abbr. edf: Effective degrees of freedom; FDR: False discovery rate.

Transferrin

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## NAR_Overall_Accuracy ~ s(COG_AGE, fx = F) + sex + s(Transferrin, 
##     k = 4, fx = F)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)
## (Intercept)   0.6375     0.4174   1.527    0.132
## sex          -0.2371     0.2398  -0.989    0.327
## 
## Approximate significance of smooth terms:
##                  edf Ref.df     F p-value  
## s(COG_AGE)     2.246   2.81 3.784  0.0198 *
## s(Transferrin) 1.000   1.00 0.667  0.4171  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.133   Deviance explained =   19%
## GCV = 0.82834  Scale est. = 0.76251   n = 66

## [1] 0.003921891

eTable 6. Association between transferrin and cognitive subdomains.

Response var.	edf	F	R2partial	P	PFDR
Executive	1	0.2	0.00	.657	.657
Social	1	1.8	0.01	.181	.543
Memory	1	0.2	0.00	.625	.657
Note:
abbr. edf: Effective degrees of freedom; FDR: False discovery rate.

Hemoglobin: a deeper dive into the cognitive effects.

Since only Hgb is significantly associated with cognition, the next step is to see if this differs by:
1. Sex
2. Whether you include cases in the clinical range for iron deficiency or not.

Within sex effects of Hgb on cognition.

This is to see if hgb is dimensionally associated with cog irrespective of sex.
First, test for an age-by-sex interaction.
Then, test males and females separately.

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## NAR_Overall_Accuracy ~ oSex + s(Hgb, k = 4) + s(COG_AGE) + s(Hgb, 
##     by = oSex, k = 4)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)  
## (Intercept)  0.01944    0.02343   0.830   0.4067  
## oSex.L       0.06106    0.03347   1.825   0.0683 .
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                     edf Ref.df      F  p-value    
## s(Hgb)            1.000  1.000 13.549 0.000242 ***
## s(COG_AGE)        4.762  5.843 90.714  < 2e-16 ***
## s(Hgb):oSexFemale 1.023  1.046  0.069 0.816528    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =   0.33   Deviance explained = 33.4%
## GCV = 0.62323  Scale est. = 0.61903   n = 1302

Hgb and cognition separated by sex.

term	edf	ref.df	statistic	partialr2	p.value
Males
s(COG_AGE)	2.9	3.6	74.1		0.000
s(Hgb)	1.0	1.0	6.1	0.011	0.014
Females
s(COG_AGE)	3.8	4.7	53.9		0.000
s(Hgb)	1.0	1.0	11.2	0.015	0.001

Mediation analysis

This is looking at sex differences in cognition with and without accounting for hemoglobin differences. The hypothesis is that females cognitive performance is negatively impacted by a greater likelihood of low Hgb levels relative to males.

There is a significant age-by-sex interaction for overall cognitive performance that essentially disappears when controlling for hemoglobin. This suggests a mediation effect.

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## NAR_Overall_Accuracy ~ s(COG_AGE) + oSex + s(COG_AGE, by = oSex)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)  
## (Intercept) 0.038253   0.022458   1.703   0.0887 .
## oSex.L      0.002182   0.031776   0.069   0.9453  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                         edf Ref.df     F  p-value    
## s(COG_AGE)            4.188  5.187 83.07  < 2e-16 ***
## s(COG_AGE):oSexFemale 1.000  1.000 13.12 0.000303 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.324   Deviance explained = 32.7%
## GCV = 0.62883  Scale est. = 0.62536   n = 1302

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## NAR_Overall_Accuracy ~ s(Hgb, k = 4) + s(COG_AGE) + oSex + s(COG_AGE, 
##     by = oSex)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)
## (Intercept)  0.02884    0.02241   1.287    0.198
## oSex.L       0.05012    0.03349   1.496    0.135
## 
## Approximate significance of smooth terms:
##                         edf Ref.df      F   p-value    
## s(Hgb)                1.000  1.000 18.289 0.0000207 ***
## s(COG_AGE)            4.543  5.597 65.614   < 2e-16 ***
## s(COG_AGE):oSexFemale 1.000  1.000  5.175    0.0231 *  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.333   Deviance explained = 33.7%
## GCV = 0.62094  Scale est. = 0.61686   n = 1302

We know that sex effects for hemoglobin emerge during puberty, and the above plots suggest that accounting for hemoglobin differences reduces sex differences in cognition post-puberty. These patterns suggest a moderated mediation effect. I.e. The mediating effect of Hgb on cogntion~sex changes with age.

We can do this in two ways. First, we can examine the mediation effect at age 10y and age 18y. Second, we can directly test the difference in the mediation effect at ages 10y and 18y.

All mediation testing is done using non-parametric bootstrapping (sims = 10000) and bias corrected and accelerated confidence intervals (boot.ci.type = "bca").

Mediation effect at age 10y:

## 
## Causal Mediation Analysis 
## 
## Nonparametric Bootstrap Confidence Intervals with the BCa Method
## 
## (Inference Conditional on the Covariate Values Specified in `covariates')
## 
##                Estimate 95% CI Lower 95% CI Upper p-value    
## ACME            0.00685     -0.01151         0.04    0.51    
## ADE             0.29284      0.13520         0.45  <2e-16 ***
## Total Effect    0.29969      0.13988         0.46  <2e-16 ***
## Prop. Mediated  0.02285     -0.04372         0.13    0.51    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Sample Size Used: 1302 
## 
## 
## Simulations: 10000

Mediation effect at age 18y:

## 
## Causal Mediation Analysis 
## 
## Nonparametric Bootstrap Confidence Intervals with the BCa Method
## 
## (Inference Conditional on the Covariate Values Specified in `covariates')
## 
##                Estimate 95% CI Lower 95% CI Upper p-value   
## ACME             -0.099       -0.184        -0.02  0.0220 * 
## ADE              -0.129       -0.286         0.03  0.1074   
## Total Effect     -0.228       -0.375        -0.09  0.0018 **
## Prop. Mediated    0.434        0.119         1.43  0.0238 * 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Sample Size Used: 1302 
## 
## 
## Simulations: 10000

Difference between the magnitude of ACME at 18y and 10y:

## 
##  Test of ACME(covariates.1) - ACME(covariates.2) = 0
## 
## data:  estimates from med.init
## ACME(covariates.1) - ACME(covariates.2) = -0.075055, p-value = 0.0196
## alternative hypothesis: true ACME(covariates.1) - ACME(covariates.2) is not equal to 0
## 95 percent confidence interval:
##  -0.13935660 -0.01159855
## 
## 
##  Test of ADE(covariates.1) - ADE(covariates.2) = 0
## 
## data:  estimates from med.init
## ADE(covariates.1) - ADE(covariates.2) = -0.2985, p-value < 2.2e-16
## alternative hypothesis: true ADE(covariates.1) - ADE(covariates.2) is not equal to 0
## 95 percent confidence interval:
##  -0.4539088 -0.1457751

SES

Does Hgb also mediate the association between SES and cognition?

## 
## Causal Mediation Analysis 
## 
## Nonparametric Bootstrap Confidence Intervals with the BCa Method
## 
##                Estimate 95% CI Lower 95% CI Upper p-value    
## ACME            0.01027      0.00358         0.02  0.0032 ** 
## ADE             0.28981      0.25177         0.33  <2e-16 ***
## Total Effect    0.30008      0.26202         0.34  <2e-16 ***
## Prop. Mediated  0.03422      0.01156         0.07  0.0032 ** 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Sample Size Used: 1302 
## 
## 
## Simulations: 10000

Non-clinical associations

Look at only participants that do not meet clinical cutoffs for iron deficiency. This is to see whether Hgb levels are associated with cognition even when individuals do not meet thresholds for iron deficiency/anemia (including visits with Hgb>=12g/dL).

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## NAR_Overall_Accuracy ~ s(COG_AGE, k = 6) + sex + s(Hgb, k = 4)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)
## (Intercept) -0.03851    0.08276  -0.465    0.642
## sex          0.07252    0.05103   1.421    0.156
## 
## Approximate significance of smooth terms:
##              edf Ref.df      F p-value    
## s(COG_AGE) 4.346  4.806 95.900 < 2e-16 ***
## s(Hgb)     1.000  1.000  9.681 0.00191 ** 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.342   Deviance explained = 34.6%
## GCV = 0.5839  Scale est. = 0.57976   n = 1036

## Partial R2 for Hgb = 0.01

Subclinical cognitive effects

term	edf	ref.df	statistic	p.value
s(COG_AGE)	4.3	4.8	95.9	0.000
s(Hgb)	1.0	1.0	9.7	0.002

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## NAR_Overall_Accuracy ~ s(COG_AGE, k = 6) + oSex + s(Hgb, k = 4) + 
##     s(Hgb, by = oSex)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)   
## (Intercept)  0.07088    0.02579   2.748   0.0061 **
## oSex.L       0.04781    0.03662   1.306   0.1919   
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                     edf Ref.df      F p-value    
## s(COG_AGE)        4.309  4.786 93.446  <2e-16 ***
## s(Hgb)            1.000  1.000  4.946  0.0264 *  
## s(Hgb):oSexFemale 1.344  1.620  0.164  0.7487    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.342   Deviance explained = 34.7%
## GCV = 0.58487  Scale est. = 0.57999   n = 1036

Figure 3 panels:

Part 3: PCA analysis of iron metrics

Fit the PCA models

We match all blood data by time of blood draw and perform 2 different PCAs. The first is done is a subset of the data that also includes C-reactive protein (inflammatory marker). The second uses only Hgb, Ferritin, and Transferrin. The CRP model is used as a validation of the solution from the 3-variable model. After deriving the PCs, we look at developmental effects, and then filter by time from cog visit for the cognitive analysis.

Look at how the iron status index changes with development.

Now that we have created iron status index, let’s next look how it changes with age. We see the same sort of age*sex interaction we saw when looking at the individual measures above.

## Partial R2 = 0.01

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## PC2 ~ s(Age) + oSex + s(Age, by = oSex)
## 
## Parametric coefficients:
##             Estimate Std. Error t value       Pr(>|t|)    
## (Intercept)  0.18442    0.05226   3.529       0.000446 ***
## oSex.L      -0.48327    0.07392  -6.538 0.000000000122 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                     edf Ref.df     F p-value   
## s(Age)            2.682  2.682 3.789 0.00835 **
## s(Age):oSexFemale 1.000  1.000 4.210 0.04056 * 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.0961   
## lmer.REML = 1709.5  Scale est. = 0.46569   n = 690

## [1] "show data is on"

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## PC2 ~ s(Age) + oSex + s(Age, by = oSex) + s(envSES)
## 
## Parametric coefficients:
##             Estimate Std. Error t value       Pr(>|t|)    
## (Intercept)  0.18874    0.05064   3.727        0.00021 ***
## oSex.L      -0.45853    0.07185  -6.381 0.000000000324 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##                     edf Ref.df      F  p-value    
## s(Age)            2.456  2.456  3.485 0.014696 *  
## s(Age):oSexFemale 1.000  1.000  3.256 0.071621 .  
## s(envSES)         1.495  1.495 14.283 0.000126 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.141   
## lmer.REML = 1695.9  Scale est. = 0.46904   n = 690

## Partial R2 = 0.050

PC scores and cogntive variables

Now fit cognitive models using the principle component score. The second principle component, which reflects iron status, is robustly associated with the accuracy factor scores from the CNB.

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## NAR_Overall_Accuracy ~ s(COG_AGE) + sex + s(PC2)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)  
## (Intercept)   0.5206     0.3031   1.718   0.0976 .
## sexFemale    -0.2902     0.3412  -0.851   0.4027  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##              edf Ref.df     F p-value  
## s(COG_AGE) 1.000  1.000 7.037  0.0134 *
## s(PC2)     1.712  2.135 5.179  0.0114 *
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.334   Deviance explained = 41.6%
## -REML = 35.287  Scale est. = 0.50018   n = 31

## Partial R2 = 0.311

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## NAR_Overall_Accuracy ~ sex + s(COG_AGE) + s(PC1, k = 4)
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)  
## (Intercept)   0.7803     0.3568   2.187   0.0376 *
## sexFemale    -0.6121     0.4011  -1.526   0.1386  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##            edf Ref.df     F p-value
## s(COG_AGE)   1      1 2.356   0.136
## s(PC1)       1      1 0.109   0.743
## 
## R-sq.(adj) =  0.0619   Deviance explained = 15.6%
## -REML = 39.436  Scale est. = 0.70441   n = 31

## Partial R2 = 0.004

Cognitive summary

Summarizing effects across individual cogntive subdomains.

eTable 7. Association between the iron status index (PC2) and cognitive subdomains.

Response var.	edf	F	R2partial	P	PFDR
Executive	1.72	4.8	0.32	.016	.024
Social	1.66	2.0	0.16	.148	.148
Memory	1.00	9.9	0.27	.004	.012
Note:
abbr. edf: Effective degrees of freedom; FDR: False discovery rate.

Interaction with sex

Does the association depend on sex, or is it true across sexes?
(It does not).

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## NAR_Overall_Accuracy ~ s(COG_AGE) + PC2 * oSex
## 
## Parametric coefficients:
##             Estimate Std. Error t value Pr(>|t|)  
## (Intercept)   0.2545     0.2251   1.131   0.2686  
## PC2           0.5620     0.2488   2.259   0.0325 *
## oSex.L       -0.1013     0.3180  -0.319   0.7526  
## PC2:oSex.L   -0.1278     0.3427  -0.373   0.7123  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##            edf Ref.df     F p-value  
## s(COG_AGE)   1      1 6.731  0.0154 *
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.277   Deviance explained = 37.3%
## GCV = 0.64755  Scale est. = 0.54311   n = 31

Figure 4

eFigure 3

Part 4: Effects on the brain.

Hemoglobin and fractional anisotropy.

We really only have enough data to look at hemoglobin in relation to the brain scans after matching up scans to blood tests that occurred within 6 months. Below, we can see the effect of hemoglobin on fractional anisotropy from each tract from the JHU atlas. Hgb is significantly associated with SLF and uncinate anisotropy.

Demographics table for white matter FA models

N	Male/Female	Age range	Mean (SD)
185	67/118	8.58-22.58	15.34 (3.23)

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## dti_fa_slf ~ sex + s(SCAN_AGE, k = 4) + s(Hgb, k = 4)
## 
## Parametric coefficients:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  0.3621542  0.0045424  79.727   <2e-16 ***
## sex         -0.0004419  0.0026840  -0.165    0.869    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##               edf Ref.df     F   p-value    
## s(SCAN_AGE) 2.329  2.700 9.868 0.0000284 ***
## s(Hgb)      1.000  1.001 6.246    0.0133 *  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.171   Deviance explained =   19%
## -REML = -485.77  Scale est. = 0.00024219  n = 185

## 
## Family: gaussian 
## Link function: identity 
## 
## Formula:
## dti_fa_uf ~ sex + s(SCAN_AGE, k = 4) + s(Hgb, k = 4)
## 
## Parametric coefficients:
##              Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  0.381852   0.005964  64.029   <2e-16 ***
## sex         -0.001821   0.003524  -0.517    0.606    
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Approximate significance of smooth terms:
##               edf Ref.df     F p-value  
## s(SCAN_AGE) 2.433  2.779 3.067  0.0736 .
## s(Hgb)      1.000  1.000 4.892  0.0282 *
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## R-sq.(adj) =  0.0702   Deviance explained = 9.26%
## GCV = 0.0004295  Scale est. = 0.00041689  n = 185

eTable 8. Association between hemoglobin and white matter fractional anisotropy.

Response var.	edf	F	R2partial	P	PFDR
ATR	1.00	2.8	0.02	.097	.259
CGC	1.56	0.4	0.01	.592	.592
CGH	1.00	0.5	0.00	.475	.543
CST	1.07	1.2	0.01	.251	.385
IFO	1.00	1.1	0.01	.289	.385
ILF	1.00	1.6	0.01	.212	.385
SLF	1.00	6.2	0.03	.013	.107
UF	1.00	4.9	0.03	.028	.112
Note:
abbr. edf: Effective degrees of freedom; FDR: False discovery rate.
Tracts
ATR: Anterior thalamic radiation
CGC: Cingulum (cingulate gyrus)
CGH: Cingulum (hippocampus)
CST: Cerebrospinal tract
IFO: Inferior fronto-occipital fasciculus
ILF: Inferior longitudinal fasciculus
SLF: Superior longitudinal fasciculus
UF: Uncinate fasciculus