A comparison of approaches to studying relationships between lumbar spinal degeneration findings demonstrates advantages to interspace-level analyses over person-level analyses

Abstract

Purpose:

To examine associations between lumbar intervertebral disc degeneration (LDD) and type II Modic changes (MC) when retaining information at each interspace (“interspace-level analysis”), as compared to aggregating information across interspaces as is typically done in spine research (“person-level analysis”). The study compared results from (1) interspace-level analyses assuming a common relationship across interspaces (the “interspace-level, common-relationship” approach), (2) interspace-level analyses allowing for interspace-specific associations (an “interspace-level, interspace-specific” approach), and (3) a conventional person-level analytic approach.

Methods:

Adults in primary care (n=147) received lumbar spine magnetic resonance imaging (MRI) and neuroradiologist-evaluated assessments of prevalent disc height narrowing (DHN), type II MC, and other LDD parameters. Analyses examined associations between DHN and type II MC in interspace-level, common-relationship analyses, interspace-level, interspace-specific analyses, and conventional person-level analyses.

Results:

Cross-sectional, interspace-level, common-relationship analyses found large-magnitude DHN-type II MC associations (adjusted OR [aOR]=6.5, 95% confidence intervals (CIs) 3.3-13; p<0.001). The magnitude of this association was larger and more precise than that yielded by person-level analyses (aOR=2.9 [95% CI 1.2-7]), and substantially more precise than interspace-level, interspace-specific analyses which allowed the association between DHN and type II MC to vary across levels. Across exploratory analyses of disc signal intensity and other MC types, interspace-level, common-relationship analyses produced larger-magnitude and more precise associations than person-level analyses in most situations, and were more precise than interspace-level, interspace-specific analyses.

Conclusions:

Interspace-level analytic approaches offer some advantages to person-level analyses that may be useful in understanding relationships between spinal degeneration findings.

MINI-ABSTRACT

Analyses of relationships between different spinal degeneration findings commonly take a person-level analytic approach. This study of lumbar intervertebral disc degeneration (LDD) and type II Modic changes (MC) found that interspace-level analytic approaches offer some advantages to person-level analyses, and may be useful in understanding relationships between spinal degeneration findings.

INTRODUCTION

An active area of spine research is the study of how lumbar spinal degeneration findings are related to each other. Analyses of relationships between different imaging-detected lumbar intervertebral disc degeneration (LDD) findings commonly take a “person-level” approach, with the unit of analysis being the individual person. While person-level analyses of LDD are easy to conduct and easy to interpret, they also have potential limitations. First, person-level analyses of LDD lose information by summarizing the information contained in the 5 lumbar interspaces, which may differ from each other with regards to typical LDD prevalence, biomechanics, and biology. This loss of information can lead to decreased precision and effect sizes. Second, person-level analytic approaches to spine analyses do not provide a complete framework that allows modelling of complex interspace-specific interactions that may exist between different LDD findings. Therefore, in some situations, a person-level approach may impede the discovery of interactions between spinal levels, degenerative findings, alignment, and other factors.

“Interspace-level” analytic approaches are an alternative to or complement to person-level analytic approaches that may offer advantages in some scenarios when studying relationships between LDD findings. Interspace-level analyses use each spinal interspace as the unit of analysis, rather than the person. Historically, two simple yet flawed approaches to multivariable interspace-level analyses have been widely used. One such approach treats each interspace as the unit of analysis but assumes that different interspaces within a person are independent.^1–3 This assumption is clearly not true, given the interrelatedness of spinal levels within a person that exists due to biomechanics and biology, and can potentially affect estimates of precision.⁴ A second approach is to conduct multiple person-level analyses of LDD, each restricted to a specific interspace (e.g., L5-S1)^5–9 or region (e.g., L4-S1).¹⁰ This approach requires multiple statistical comparisons and consequently can increase type I error. Neither of these simple approaches allow for explicit statistical tests of more complicated questions such as whether different imaging findings have significantly different relationships at different interspaces or regions. Contemporary analytic methods such as generalized estimating equations (GEE)¹¹ allow better ways to conduct interspace-level analyses, such as (1) interspace-level analytic approaches that assume a common relationship across interspaces (in which the unit of analysis is the spinal interspace rather than the person, which we refer to here as the “interspace-level, common-relationship” approach) and (2) interspace-level analytic approaches that allow for associations between different imaging findings to vary between different spinal interspaces (which we refer to here as the “interspace-level, interspace-specific” approach). Fewer spinal imaging studies have used interspace-level, common-relationship¹² and interspace-level, interspace-specific approaches¹³ that account for the interrelatedness of spinal interspaces and their lack of independence.

In the current work, we compare and contrast the results of interspace-level vs. person-level analyses of spinal degeneration findings using an illustrative example of the relationship between two LDD findings, disc height narrowing [DHN] and disc signal intensity loss [DSI], and vertebral endplate or “Modic” changes (MC).^10,14 LDD and MCs are generally associated in cross-sectional imaging studies, with varying magnitudes of association between studies.^5,10,12,14 Given the low prevalence of type I and type III MCs,⁵ which limit power to detect associations with MCs, the current study focused on type II MCs.

The first study aim was to examine associations between LDD and type II MCs in a cross-sectional interspace-level analysis that assumes a common relationship across interspaces (the interspace-level, common-relationship approach). We hypothesized that LDD and type II MC would be significantly associated when using this approach. The second aim was to descriptively compare results from analyses of LDD and type II MC using different analytical approaches: (1) an interspace-level, common-relationship approach, (2) an interspace-level, interspace-specific approach that allows for interspace-specific associations, and (3) a conventional person-level approach. We hypothesized that the interspace-level, interspace-specific approach would be significantly different from the interspace-level, common-relationship approach.

METHODS

Overview

This was a secondary analysis of data from the Longitudinal Assessment of Imaging and Disability of the Back (LAIDBACK)[4,5]. Participants received lumbar spine magnetic resonance imaging scans (MRIs) at baseline and repeat MRIs at 3-year follow-up. DHN was the primary LDD independent variable, as we expected that a more advanced LDD change (such as DHN) would associate more strongly with MCs. Exploratory analyses were also conducted for other possible combinations of LDD and MC variables (Table 1). To obviate data mining, all analytic steps were publicly posted prior to analysis.¹⁵

Table 1.

Overview of the study aims, designs, primary analyses, and exploratory analyses.

Aim 1: To examine associations between LDD and type II MCs using an interspace-level, common-relationship analytic approach. Hypothesis: LDD and type II MC are significantly associated when using this approach.
Aim 2: To descriptively compare associations between LDD and type II MC when using (1) an interspace-level, common-relationship approach, (2) an interspace-level, interspace-specific approach, and (3) a person-level approach. Hypothesis: A model using an interspace-level, interspace-specific approach is significantly different from a model using an interspace-level, common-relationship approach.
Primary analysisa
Independent variable	Dependent variable
Disc height narrowing (DHN)	Type II Modic change (MC)
Exploratory analyses
Independent variable(s)	Dependent variable(s)
DHN, disc signal intensity (DSI)	Type I MC, type II MC, Type III MC, any MC

^aCross-sectional analyses of DHN and type II MC using generalized estimating equations (GEEs), adjusting for age, sex, interspace, and time point. One statistical comparison was made for each of the 2 Aims, each using a p-value significance threshold of 0.05.

^bAn analogous approach to the primary analysis was taken for the exploratory analyses, which included 2 predictor variables and 4 outcome variables, for a total of 7 exploratory statistical comparisons (not including the association of DHN with MC, which was the primary analysis)

Study sample

LAIDBACK participants were randomly sampled from primary care in the Veterans Affairs (VA) Puget Sound health system.¹⁶ Participants were excluded for having low back pain (LBP) in the last 4 months; prior lumbar spine procedures; serious comorbidities limiting study participation; or contraindications to MRI. The sample had a mean age of 53 years (range 36-71), 85% were White, 8% were Black/African-American, and 11% were also or solely from other race groups. Consistent with demographics of people with prior military service, only 12% were female. Extensive details of the sample have been previously reported [4,5]. This research involved a completely deidentified dataset and was determined to be non-human subjects research.

Evaluation of lumbar disc degeneration (LDD) and Modic changes (MCs)

All participants received standardized lumbar spine MRI protocols using a Philips 1.5 Tesla MRI system (see Supplemental Content). One of 2 experienced senior neuroradiologists interpreted imaging findings at each of the L1-S1 interspaces. DHN was defined as a decrease in disc height compared with the expected height of a hydrated disc at the same level (a binary variable).¹⁶ For grading interspaces at L4-L5 and superior, the disc height of the interspace immediately above was used in deciding whether an interspace exhibited DHN. If the disc space above the interspace of interest appeared to be narrowed, and at the L5-S1 level where normal disc height can be more variable, the neuroradiologist’s clinical judgement was exercised. DSI was graded as “mild” if there was a mild diffuse decrease in signal, or a focal moderate or severe decrease involving < 1/3 of the disc area on sagittal images; “moderate’ if there was a moderate diffuse decrease in signal or a focal moderate or severe decrease involving 1/3-2/3 of the disc area ; and “severe” if there was a severe diffuse decrease (no residual high T2 signal) involving >2/3 of the disc area.¹⁶ Interrater reliability was moderate for DHN (k=0.56) and almost perfect for DSI (weighted k=0.84).

Type I MC was defined as an MC with low T1 and high T2 signal intensity. Type II MC was defined as an MC with high T1 and low to intermediate T2 signal intensity. A type III MC was defined as an MC with low T1 and low T2 signal intensity.¹⁶ Mixed MC types were not evaluated. Reliability was not evaluated for MCs due to their expected low prevalence.

Statistical Analysis

The primary analysis for the first aim was a cross-sectional interspace-level, common-relationship analysis of the DHN and type II MC association using GEEs adjusting for age, sex, interspace, and MRI time point.¹¹ All methodologic details and models are provided in the Supplemental Content. We calculated adjusted odds ratios (aORs) and 95% confidence intervals (CIs), using a p-value <0.05 to determine statistical significance. For the second aim, we also calculated aORs and 95% CIs for the DHN-type II MC association in an interspace-level, interspace-specific analysis using a model identical to that from the interspace-level, common-relationship analysis but including interaction terms for each interspace with DHN. We then conducted person-level analyses for the DHN-type II MC association adjusting for age, sex, and time point, using GEE. Other variables were not adjusted for, due to concerns about inducing collider bias. We then descriptively compared aORs and 95% CIs from the interspace-level, common-relationship vs. interspace-level, interspace-specific vs. person-level analyses, using a logarithmic scale as is appropriate.¹⁷ We expected wider confidence intervals and smaller-magnitude associations in the person-level analysis. We used an omnibus statistical test to examine for significant differences between the interspace-level, interspace-specific model and the interspace-level, common-relationship analyses. Importantly, while LAIDBACK was a longitudinal cohort, the current work conducted cross-sectional analyses only, analyzing both time points while adjusting for time point, as (1) the low incidence of new MRI findings over 3 years¹⁸ was expected to provide limited power for longitudinal analyses, and (2) word limits precluded additional presentation of longitudinal analysis results. We then repeated the analytic steps above in exploratory analyses for the other 7 possible combinations of LDD independent variables and MC dependent variables (Table 1). We did not conduct power calculations as the sample size was fixed, given well-known concerns regarding post hoc power analysis.¹⁹

RESULTS

Of 147 participants, 57% had DHN at ≥1 lumbar interspaces and 95% had at least mild DSI at ≥1 interspaces (Table 2). The prevalence of DHN and severe DSI was generally higher at the L4-S1 interspaces than the L1-L4 interspaces. Type II MCs were present at ≥1 interspaces in 27% of participants, with the lowest prevalence at the L1-L3 interspaces (4-7%) and the highest at L4-S1 (12-14%).

Table 2.

Prevalence of lumbar disc degeneration (LDD) and Modic changes (MC) findings^*

Prevalence at baseline (n=147).
LDD at baseline
	L1	L2	L3	L4	L5	Any interspace
DHN	20 (14%)	17 (12%)	23 (16%)	41 (28%)	45 (31%)	84 (57%)
DSI
None	60 (41%)	56 (38%)	46 (31%)	33 (22%)	35 (24%)	8 (5%)
Mild	44 (30%)	37 (25%)	32 (22%)	37 (20%)	19 (13%)	92 (63%)
Moderate	24 (16%)	36 (24%)	35 (24%)	36 (24%)	34 (23%)	99 (67%)
Severe	19 (13%)	18 (12%)	34 (23%)	41 (28%)	59 (40%)	83 (56%)
Any LDD	87 (59%)	91 (62%)	101 (69%)	114 (78%)	112 (76%)	139 (95%)
Modic changes at baseline
	L1	L2	L3	L4	L5	Any interspace
Type I MC	1 (1%)	1 (1%)	1 (1%)	0	2 (1%)	5 (3%)
Type II MC	7 (5%)	6 (4%)	11 (7%)	17 (12%)	20 (14%)	39 (27%)
Type III MC	0	1 (1%)	0	0	1 (1%)	2 (1%)
Any MC	8 (5%)	8 (5%)	12 (8%)	17 (12%)	23 (16%)	43 (29%)

^*There was no missing data

In the primary interspace-level, common-relationship analysis, DHN was strongly associated with type II MC (aOR=6.5 [95% CI 3.3-13], p= 4.0 x 10⁻⁸; Supplemental Table S1). Associations between DHN and type II MC in interspace-level, interspace-specific analyses were largest at the L1-2 and L5-S1 interspaces (aOR=13 [95% CI 2.5-73] and aOR=15 [95% CI 4.1-56], respectively), and relatively smaller at the L2-L5 interspaces (aOR at L2-3=2.6 [95% CI 0.44-16], L3-4=4.0 [95% CI 1.1-15], and L4-5=3.8 [95% CI 1.2-16], respectively) (Supplemental Tables S2–3). Associations between DHN and type II MC in person-level analyses were smaller (aOR=2.9 [95% CI 1.2-7.2]; Supplemental Table S4).

Figure 1 summarizes the contrasts between DHN-type II MC associations with the 3 types of analyses, using a logarithmic scale appropriate for odds ratios.¹⁷ It shows that the DHN-type II MC association point estimate is larger and 95% CI narrower in the interspace-level, common-relationship analysis (shown at the top of the Forest plot) as compared to the person-level analysis (shown at the bottom of the plot). The magnitude of DHN-type II MC association point estimates from interspace-level, interspace-specific analyses are similar to those from the person-level analysis (L2-L5 levels) or larger (L1-2 and L5-S1 levels), but all 95% CIs are substantially wider. When comparing interspace-level, common-relationship vs. interspace-level, interspace-specific analyses as part of the second aim, there was no significant difference between the two (interaction p=0.53, Supplemental Table S2). This indicates that, in this context, the interspace-level, interspace-specific DHN-type II MC associations are not significantly different than the interspace-level, common-relationship DHN-type II MC association (aOR=6.5), with no advantage to the more complex interspace-level, interspace-specific analysis. In both the interspace-level, common-relationship and interspace-level, interspace-specific analyses, no individual interspace (e.g., L5-S1) was significantly associated with type II MC once other factors such as DHN, age, and sex had been accounted for (Supplemental Tables S1, S2, and S4).

Figure 1. Odds ratios for the DHN-type II MC association.

DHN=disc height narrowing, MC=Modic change

Results of exploratory analyses for DHN and DSI with other MC variables are provided in Supplemental Figures S1–3 and, Supplemental Tables S1–S4. These showed -a consistent pattern of larger-magnitude LDD-MC associations with narrower confidence intervals in interspace-level, common-relationship analyses as compared to person-level and interspace-level, interspace-specific analyses. There were no significant differences between any interspace-level, common-relationship and. interspace-level, interspace-specific analysis (Supplemental Table S2; all interaction p-values>0.05). As in the primary analysis, no individual interspace was associated with any MC independent of other factors.

DISCUSSION

This study found that DHN and type II MC were strongly associated in an interspace-level, common-relationship analysis, with a larger-magnitude and more precise association than found in a person-level analysis (Figure 1). The same pattern was seen across all exploratory analyses (Supplemental Figures S1–3). These findings illustrate how, in some contexts, interspace-level, common-relationship analyses may yield larger and more precise associations than person-level analyses. While the magnitude of DHN-type II MC associations from interspace-level, interspace-specific analyses differed by interspace in the current study, the absence of statistically significant differences across interspaces suggest that these differences may reflect random variability and did not indicate an advantage to interspace-level, interspace-specific analyses in this context.

To our knowledge this is the first study to comprehensively compare the results of interspace-level and person-level analyses of spinal degeneration findings. Our finding of larger-magnitude associations with interspace-level, common-relationship analyses than person-level analyses within the same study comport with isolated findings scattered across prior reports. A previous cross-sectional, interspace-level, common relationship analysis found a large-magnitude association of any MC with LDD (aOR=6.2 , p<0.001).¹² This contrasts with much smaller-magnitude any MC-LDD associations in a person-level analysis from another study (OR for DHN=1.3 and OR for DSI=1.2, both p-values <0.001).¹⁴ The current findings suggest that such between-study differences in effect size are more likely due to the different analytic methods involved, and less likely due to differences in study populations,

Interspace-level analyses may also have other advantages. In contrast to person-level analyses, interspace-level analyses allow evaluation of the relationship between an interspace and a specific lumbar degeneration finding that is independent of lumbar degeneration findings at the same interspace. An example of this is the current study’s finding that spinal interspace was not significantly associated with MCs once the LDD variable at the same interspace was adjusted for. This is notable, as most prior studies find differences in MC prevalence by interspace, with lower prevalences at L1-L4, and higher prevalences at L4-S1.^10,12,20 These findings taken together with points from preceding paragraphs suggest that spinal interspace is not associated with MCs once LDD at the same level has been accounted for because LDD at an interspace may largely explain the subsequent development of an MC at the same interspace. This is an example of the type of inference permitted by interspace-level analytic approaches that can be obscured by person-level analyses. On the other hand, the use of interspace-level analyses shifts interpretation to the spinal level rather than the person. This may be more appropriate for certain research questions such as when examining a biological relationship that credibly has its predominant effect at the same interspace. An example of this is a setting of experimental disc injury, in which disc injury would more likely cause early disc degeneration at the involved interspace, rather than simultaneously causing degeneration at multiple interspaces. In other situations, however, a person-level analysis may be more appropriate. Researchers should consider interspace-level analyses as potential alternatives or complements to person-level analyses.

A study strength is that we publicly posted our analysis plan a priori, to prevent data mining.¹⁵ While the study’s abundant statistical power for evaluating the DHN-type II MC relationship in an interspace-level, common-relationship analysis is shown by the highly significant association found (p=4.0 x 10⁻⁸), a potential limitation is that the study may have been underpowered when comparing interspace-level, interspace-specific vs. interspace-level, common-relationship models. Given the conceptual advantages of interspace-level, interspace-specific models in emulating the complexity of the lumbar spine, in which biomechanics, disc nutrition, and other factors are thought to affect different interspaces differently, we suggest that future studies use such models in the larger cohorts now available.^7,21,22

In summary, this study illustrates how interspace-level analytic approaches may offer some advantages to person-level analyses that can be useful in understanding relationships between spinal degeneration findings.