How handling extreme C-reactive protein (CRP) values and regularization influences CRP and depression criteria associations in network analyses

TLDR: These associations suggest that fatigue and changes in appetite might be particularly receptive to anti-inflammatory treatments and recommend considering the use of nonregularized networks in immunopsychiatry and inclusion of cases with CRP values >10 mg/L when testing the association between CRP and depression criteria, unless contraindicated by the research question being tested.

Abstract: Increasingly, it has been recognized that analysis at the symptom, rather than diagnostic, level will drive progress in the field of immunopsychiatry. Network analysis offers a useful tool in this pursuit with the ability to identify associations between immune markers and individual symptoms, independent of all other variables modeled. However, investigation into how methodological decisions (i.e., including vs. excluding participants with C-reactive protein (CRP) >10 mg/L, regularized vs. nonregularized networks) influence results is necessary to establish best practices for the use of network analysis in immunopsychiatry. In a sample of 3,464 adult participants from the 2015–2016 National Health and Nutrition Examination Survey dataset, this study found consistent support for associations between CRP and fatigue and changes in appetite and some support for additional CRP—criterion associations. Methodologically, results consistently demonstrated that including individuals with CRP >10 mg/L and estimating nonregularized networks provided better estimates of these associations. Thus, we recommend considering the use of nonregularized networks in immunopsychiatry and inclusion of cases with CRP values >10 mg/L when testing the association between CRP and depression criteria, unless contraindicated by the research question being tested. Additionally, results most consistently suggest that CRP is uniquely related to fatigue and changes in appetite, supporting their inclusion in an immunometabolic phenotype of depression. Finally, these associations suggest that fatigue and changes in appetite might be particularly receptive to anti-inflammatory treatments. However, future research with more nuanced measures is necessary to parse out whether appetite increases or decreases drive this association. Further, longitudinal research is an important next step to test how these relationships manifest over time.

Figures

Figure 1. Regularized, two covariate network models of CRP and depression criteria 887 888 Note: Green edges in the networks depict positive associations, red edges represent negative 889 associations, and thicker/more saturated edges depict stronger associations. 890 891

Figure 4. Nonregularized, total covariate network models of CRP and depression criteria 905 906 Note: Green edges in the networks depict positive associations, red edges represent negative 907 associations, and thicker/more saturated edges depict stronger associations. 908 909

Figure 3. Nonregularized, two covariate network models of CRP and depression criteria 899 900 Note: Green edges in the networks depict positive associations, red edges represent negative 901 associations, and thicker/more saturated edges depict stronger associations. 902 903

Table 1. Summary of Sample Characteristics 872

Figure 2. Regularized, total covariate network models of CRP and depression criteria 893 894 Note: Green edges in the networks depict positive associations, red edges represent negative 895 associations, and thicker/more saturated edges depict stronger associations. 896 897

Table 2. Node Descriptives and Item Descriptions 876

Full text

1
2
3
4
5
6
7
This paper has recently been accepted in Brain, Behavior, and Immunity. Please do not cite the
8
pre-print. Thank you!
9
10

How handling extreme C-reactive protein (CRP) values and regularization influences CRP and
11
depression criteria associations in network analyses
12
13
14
15
Daniel P. Moriarity
1
, Sarah R. Horn
2
, Marin M. Kautz
1
, Jonas M. B. Haslbeck
3
, Lauren B. Alloy
1
16
17
1
Department of Psychology, Temple University, Weiss Hall, 1701 N. 13th St., Philadelphia, PA
18
19122, United States of America
19
2
Department of Psychology, University of Oregon, 1227 University St, Eugene, OR 97403,
20
United States of America
21
3
Psychological Methods Group, University of Amsterdam
22
23
Correspondence concerning this article should be addressed to Daniel P. Moriarity,
24
Temple University, Weiss Hall, 1701 N. 13
th
St., Philadelphia, PA 19122, United States of
25
America. Fax: 1-215-204-5539. E-mail: Daniel.moriarity@temple.edu.
26
27
Funding: Daniel P. Moriarity was supported by National Research Service Award
28
F31MH122116. Marin M. Kautz was supported by National Science Foundation Graduate
29
Research Fellowship 1650457. Lauren B. Alloy was supported by National Institute of Mental
30
Health grant R01MH101168.
31
Word count: 7,324
32
Abstract
33

Increasingly, it has been recognized that analysis at the symptom, rather than diagnostic,
34
level will drive progress in the field of immunopsychiatry. Network analysis offers a useful tool
35
in this pursuit with the ability to identify associations between immune markers and individual
36
symptoms, independent of all other variables modeled. However, investigation into how
37
methodological decisions (i.e., including vs. excluding participants with C-reactive protein
38
(CRP) > 10 mg/L, regularized vs. nonregularized networks) influence results is necessary to
39
establish best practices for the use of network analysis in immunopsychiatry. In a sample of
40
3,464 adult participants from the 2015-2016 National Health and Nutrition Examination Survey
41
dataset, this study found consistent support for associations between CRP and fatigue and
42
changes in appetite and some support for additional CRP—criterion associations.
43
Methodologically, results consistently demonstrated that including individuals with CRP > 10
44
mg/L and estimating nonregularized networks provided better estimates of these associations.
45
Thus, we recommend the use of nonregularized networks in immunopsychiatry and inclusion of
46
cases with CRP values > 10 mg/L when testing the association between CRP and depression
47
criteria in preliminary analyses, unless contraindicated by the research question being tested.
48
Additionally, results most consistently suggest that CRP is uniquely related to fatigue and
49
changes in appetite, supporting their inclusion in an immunometabolic phenotype of depression.
50
Finally, these associations suggest that fatigue and changes in appetite might be particularly
51
receptive to anti-inflammatory treatments. However, future research with more nuanced
52
measures is necessary to parse out whether appetite increases or decreases drive this association.
53
Further, longitudinal research is an important next step to test how these relationships manifest
54
over time.
55

Introduction
56
Systemic low-grade inflammation is an established correlate of, and is gaining evidence
57
as a potential causal risk factor for, depression symptoms (Dhabhar et al., 2009; Dowlati et al.,
58
2010; Howren et al., 2009; Moriarity et al., 2020a). In fact, elevated inflammation is associated
59
with treatment-resistant depression (Yang et al., 2019). In particular, the acute phase reactant C-
60
reactive protein (CRP) arguably is the most researched inflammatory index in depression
61
research (Horn et al., 2018; Howren et al., 2009). CRP is a non-specific pentameric protein
62
synthesized in the liver that is upregulated during the acute phase of inflammation in response to
63
stimulation from other proinflammatory proteins (e.g., interleukin (IL)-6 and IL-1 ; Du Clos,
64
2000).
65
Meta-analytic evidence generally has indicated that CRP is higher in individuals with
66
Major Depressive Disorder (MDD) than those without a diagnosis (Haapakoski et al., 2015;
67
Howren et al., 2009). Another meta-analysis found that higher CRP is associated with more
68
depression symptoms in cross-sectional and longitudinal studies (Valkanova et al., 2013).
69
However, the effect sizes vary (Horn et al., 2018), and, in some samples, relationships are not
70
evident (e.g., Khandaker et al., 2014; Miller & Cole, 2012). One plausible explanation for this
71
inconsistency, supported by Horn and colleagues (2018), is variability in covariates included
72
across studies. Elevations in CRP also are associated with a diverse range of conditions and
73
environmental stimuli, several of which also are confounded with depression (e.g., obesity,
74
smoking, chronic health conditions; Kushner et al., 2006). Further, elevated inflammation only is
75
seen in a subset of individuals with depression (Raison & Miller, 2011). One potential
76
explanation for this is that CRP is not associated equally with all depression symptoms, which
77
has inspired calls for investigations into the inflammatory phenotyping of depression (Felger et
78

al., 2018; Krishnadas & Harrison, 2016) and expanding the scalar variety of psychopathological
79
constructs analyzed in immunopsychiatry research (Moriarity & Alloy, in press).
80
Initial work examining the association between CRP and specific depression symptoms
81
and symptom subtypes supports this argument. For example, several studies have found
82
associations between CRP and neurovegetative symptoms such as sleeping problems, fatigue,
83
and changes in appetite (Jokela et al., 2016; Lamers et al., 2018; White et al., 2017). In fact, out
84
of all inflammatory proteins, a recent review concluded that CRP was the most consistently
85
associated with neurovegetative symptoms (Majd et al., 2020). However, it is important to note
86
that some studies find that CRP is associated with neurovegetative profiles including increased
87
appetite (Glaus et al., 2014; Hickman et al., 2014; Lamers et al., 2018; see Milaneschi et al.,
88
2020 for a review), and others find it is associated with scores featuring decreased appetite
89
(Duivis et al., 2015; Elovainio et al., 2009). CRP also has been associated with depressed mood
90
and negative attention biases (Boyle et al., 2017; White et al., 2017), in line with theory that
91
inflammation might be particularly associated with exaggerated responsivity to negative stimuli
92
(Dooley et al., 2018). There also is support for an association between CRP and blunted reward
93
processing in terms of self-report anhedonia (Felger et al., 2016; Moriarity et al., 2019),
94
decreased functional connectivity within reward circuitry (Felger et al., 2016), and high basal
95
ganglia glutamate (Haroon et al., 2016). Although there is some support that CRP is associated
96
with cognitive difficulties seen in depression, this association might be dependent on
97
neurovegetative symptoms (Krogh et al., 2014). In sum, there is the strongest evidence for an
98
association between CRP and neurovegetative symptoms, affective symptoms, and reward
99
abnormalities.
100

Frequently asked questions

Q1. What are the contributions in this paper?

In a sample of 40 3,464 adult participants from the 2015-2016 National Health and Nutrition Examination Survey 41 dataset, this study found consistent support for associations between CRP and fatigue and 42 changes in appetite and some support for additional CRP—criterion associations. One potential 76 explanation for this is that CRP is not associated equally with all depression symptoms, which 77 has inspired calls for investigations into the inflammatory phenotyping of depression ( Felger et 78 al., 2018 ; Krishnadas & Harrison, 2016 ) and expanding the scalar variety of psychopathological 79 constructs analyzed in immunopsychiatry research ( Moriarity & Alloy, in press ). 100 Recently, Fried and colleagues ( 2019 ) applied network analysis ( a statistical approach 101 designed to investigate unique, pairwise associations in multivariate data ) in a sample of 2,321 102 adults from the Netherlands Study of Depression and Anxiety ( NESDA ) to extend this line of 103 inquiry. Network analysis is a powerful tool for this research question because it results in 104 estimates of the unique association between two variables, controlling for all other variables in 105 the network. Thus, replication is crucial, especially because CRP—symptom associations 118 were small in this study. 134 The Present Study 135 This study attempts to replicate and extend Fried et al. ’ s ( 2019 ) models using data from 136 the National Health and Nutrition Examination Survey ( NHANES ), a study of physical and 137 mental health in the United States. This is beneficial 140 because network analyses require a lot of power given the amount of parameters estimated. 157 This study also will extend Fried et al. ( 2019 ) by re-estimating the original models with 158 different analytic choices. 160 However, there is evidence that nonregularized ( estimated without this procedure ) network 161 models are preferable to regularized models for psychopathology data ( Williams et al., 2019 ). This will be addressed in this study. Thus, this study will estimate all models with and without 331 170 participants with CRP values > 10 mg/L. Through a rigorous test of replication and the addition 171 of novel statistical methodologies, this study aims to enhance their understanding of the 172 relationship between CRP and specific depression criteria and to guide future network analyses 173 in immunopsychiatry. 48 Additionally, results most consistently suggest that CRP is uniquely related to fatigue and 49 changes in appetite, supporting their inclusion in an immunometabolic phenotype of depression. 50 Finally, these associations suggest that fatigue and changes in appetite might be particularly 51 receptive to anti-inflammatory treatments. 53 Further, longitudinal research is an important next step to test how these relationships manifest 54 over time. 55 Introduction 56 Systemic low-grade inflammation is an established correlate of, and is gaining evidence 57 as a potential causal risk factor for, depression symptoms ( Dhabhar et al., 2009 ; Dowlati et al., 58 2010 ; Howren et al., 2009 ; Moriarity et al., 2020a ). Further, elevated inflammation only is 75 seen in a subset of individuals with depression ( Raison & Miller, 2011 ). Although the overlap 114 between these findings and the others described above is promising, the replicability of network 115 analyses has been called into question due to concerns about single-item measurement properties 116 and high power requirements ( Forbes et al., 2017 ; see Borsboom et al., 2017 and Jones, et al., 117 2019 for responses ). 141 Further, estimates of associations between self-report and biological variables are subject to 142 downward biasing due to measurement-domain specific error variance, further attenuating 143 power. 

Q2. How many edges were negative in the two-covariate subsample model?

24 out of 66 possible 956 edges were negative in the two-covariate subsample model (compared to only 16 in the complete 957 sample) and 56 out of a possible 136 were negative in the full-covariate subsample model 958 (compared to 54 in the complete sample). 

Q3. What is the effect of the negative edges in the two-covariate model?

The higher proportional increase in negative edges in the two-961 covariate model could explain why that model was more discrepant with the other models than 962 the subsample model with all covariates. 

Q4. How many CRP-criterion edges were recovered in both models?

all three CRP-criterion edges that were recovered in both models were 922 less stable compared to the model estimated in the full sample. 

Q5. What is the effect size of the two models?

both models have substantially less 941 power than the models with the total sample, which might be particularly detrimental to the 942 CRP—criterion associations, given the small effect sizes seen in both Fried et al. (2019) and the 943 models using the total sample in this study. 

Q6. How stable were the CRP-criterion edges in the models with the full sample?

Almost all CRP—criterion edges also were less 944 stable than those observed in the models with the entire sample, reducing confidence in this 945 approach. 

Q7. What is the significance of the CRP-criterion edges?

they need to be considered in the context that they are at 938 greater risk of bias due to conditioning the sample on a measured and tested characteristic (i.e., 939 level of depression).