0) Summary
We have previously shown a core model capturing how ARCHES participants perceive the dementia risk system, and with an expanded number of models, shown that neighborhood risk (mean ADI of a model-building group) significantly affects the ways that participants perceive and build these models.
Here we ask: how can SDOH deepen this understanding?
So far:
1) mean SDOH of group can be quantified as significantly relating to the models that groups report (p~0.01, r-squared ~ 0.20)
2) I’ve run comparison charts showing which model components are particularly prevalent in high, mid, and low SDOH groups, and which are most different in prevalence between high and low SDOH groups
Still to do:
3) run a multivariate model asking how much of the resisual variance (after taking into account ADI) SDOH explains
Fitting SDOH along wtih other characteristics of groups
We now add SDOH to our fit of various characteristics of model-building groups to the distances of models from one another in this NMMDS ordination space.
In the table below Pr(>r) is the fit pvalue, with asterisks indicating significance thresholds. r2 is the amount of variation explained by the factor. You can ignore NMDS1 and NMDS2 for now – this gives the coordinates in the NMMDS ordination where the centroid of a factor (or the middle of a vector) would be.
In summary:
We see that group means of ADI National and SDOH count both have significant (pval<0.05) and moderately strong (r-squared ~ 0.20 - 0.30) fit with where a group’s model falls in the ordination space.
Next steps:
This model is looking independently at the effecrts of each. It would be nice to see them in an additive models, as we show above that these two variables are not correlated (as so we might expect them to show different facets of variation). I’m currently working on this.
***VECTORS
NMDS1 NMDS2 r2 Pr(>r)
n_people -0.25072 -0.96806 0.0288 0.601
mean_age -0.14531 0.98939 0.1425 0.097 .
percent_female 0.67518 0.73765 0.0643 0.352
mean_education_years -0.81925 0.57344 0.0642 0.357
mean_ADI_N 0.59858 0.80106 0.2967 0.002 **
mean_SDOH_count 0.24054 -0.97064 0.2051 0.018 *
round 0.99774 -0.06713 0.0332 0.580
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Permutation: free
Number of permutations: 999
If we break SDOH count (a continuous variable) into quantile levels (lowest 25%, middle 50%, highest 25%), we can see they each occupy distinct areas of the ordination space, although there is significant overlap (as expected, given that SDOH as a continuous variable only explains about 20% of the variation in models — that is, that there is still agreement on a core model across SDOH levels).
Here, each point is one group’s model, and distances between each pair of points are defined by the dis-similarity in the edges (indirect and direct) in those models. So, points that are relatively close together are groups that had similar models, points that are widely spaced apart had more dis-similar models. I’ve colored and drawn outlines around groups that fell into each of the three SDOH levels.
We may also characterize which components are relatively more common in SDOH categories, (i.e. what the differences in the actual models are that are driving the dissimilarity):
_____
Every edge that occurred in >75% of one or more SDOH level
This shows every edge that had high frequency (was reported by 75% or more of the groups) in at least one of the SDOH levels. Edges at the right were reported with high frequency across all levels. Edges with long great bars are ones with a large difference between the frequency with which they are reported by groups of the different SDOH levels.
_____
Edges with the greatest difference (at least 40%)
This shows edges that were maximally different (at least 40% difference) among the high and low SDOH levels.
TESTING THE ABOVE WITH 20-60-20 instead of 25-75-25
quantile levels (lowest 20%, middle 60%, highest 20%)
_____
Every edge that occurred in >75% of one or more SDOH level
This shows every edge that had high frequency, with 20-60-20 split of SDOH levels
_____
Edges with the greatest difference (at least 40%)
This shows edges that were maximally different (at least 40% difference) , with 20-60-20 split of SDOH levels.
