Prevalence of Visual Acuity Loss or Blindness in the US

Key Points Question How many people in the US are living with uncorrectable visual acuity loss or blindness? Findings This bayesian meta-analysis generated an estimate that, in 2017, there were 7.08 million people living with visual acuity loss, of whom 1.08 million were living with blindness. Meaning Per this study, uncorrectable visual acuity loss and blindness are even larger drivers of health burden in the US than was previously known.

Initially we estimated the numbers of cases of visual acuity loss using only PBS data. We then added NHANES data as reference with PBS data for the estimation, and then merged NSCH data with NHANES data and PBS data in the model and generated estimates of visual acuity loss. We then added ACS data for older ages and group quarters sequentially (denoted ACSolder and ACS-gq in Table 1 below). After that we used ACS data for state-specific prevalence and for year 2017 to generate estimates (denoted ACS-state and ACS-2017 in Table 1 below). In the final model we added interaction terms along with all the datasets to generate visual acuity loss estimates. driven by the PBS that were used in our model as compared to the earlier VPUS. The studies we included as PBS are described in the data section above. They can be compared to those used by VPUS by reviewing the VPUS Methods and Sources page. In general, the studies that we included reported a lower prevalence of visual impairment and blindness than the studies included in the VPUS and this lower prevalence is reflected in our model estimate.
We believe specific differences between our estimates are attributable to the inclusion of

MICE-Multiple imputation by chained equation
Missing data in meta-analysis can jeopardize inferences from the study and examination data, and possibly bias the estimates. The vision examination data from NHANES is a key input in our meta-analytic estimates of the prevalence of uncorrectable visual acuity loss or blindness in the United States, but when we mapped this data to dichotomous outcomes for visual acuity loss or blindness, we were unable to determine a value for 11.51% of respondents. (1) Visual acuity loss ∼ C(age_group) + sex + C(race_eth) + vidrva + vidlva, We considered a range of regression equations for our outcomes of interest (visual acuity loss or blindness): (in the formulas below, the notation (var1 + var2)**2 means the variables var1 and var2 are included as main effects and all pair-wise interactions are also included) 11. outcome ~ C(age_group) + sex + C(race) + vidrva + vidlva + vidrova + vidlova 12. outcome ~ C(age_group) + sex + C(race) + (vidrva + vidlva + vidrova + vidlova)**2 We tested each equation for each outcome for two alternative perturbation methods: (a) the Gaussian perturbation method, which samples from a multivariate normal distribution derived from the fit of the regression equation; and (b) the bootstrap perturbation method, which samples from the conditional model fitted to a bootstrapped version of the data set.
For our testing approach, we used an out-of-sample cross-validation approach from machine learning, where we withheld the presenting and best corrected visual acuity measurements for a randomly selected subset of data (our "test dataset") and compared the model predictions for these individuals to the true values of blindness and visual acuity loss. (3) To be precise,

MICE validation results
Missing values were best imputed using the variables age, sex, race, vidrva and vidlva, as shown in Table 3.  where indexes the specific measurement, is the prevalence count of those with visual acuity loss (or blindness) in measurement , eff is the effective sample size from which the count was taken (and so = eff is the prevalence rate typically reported in a PBS), is the prevalence rate predicted by the model, is the over-dispersion parameter of the negative binomial distribution (assumed to be the same for all measurements). DisMod-MR 1. In the systematic component of the model, we included fixed effects for sex ( sex ), age ( for = 0, … , ), race ( for = white, Black, Hispanic, and other races), and data source ( for =EDPRG, CHES, LALES, ACS, BPEDS, MESA, and NSCH; we coded NHANES as the reference category), as well as sex/race and race/age interaction fixed effects. We also include random effects for state (50 states and Washington, D.C.; for = 1, … ,51). Our formulation includes a piece-wise linear spline model on age, with spline knots (knot 1 ,knot 2 ,…) indexed by for = 1, … , , as well as intercept shifts for additional key covariates to account for differential sex ratios and age patterns by race as follows: where sex is the sex measured in measurement ; age 0 and age are the start and end of the age group measured in measurement ; race is the race/ethnicity group measured in measurement ; source is the data source for measurement ; location is the state location of measurement ; and notation [variable = value] represents an indicator function which takes value 1.0 if the variable is equal to the value and 0.0 otherwise and notation (value) + represents the value in the parenthesis if it is positive, and takes value zero, otherwise. We used evenly spaced spline knots at ages (0, 20, 40, 60, 80, 100).
We also included interaction terms to capture the possibility that the sex and age effects were different by race/ethnicity and group quarters status: where ′ is the effect coefficient for race interacted with sex (for = 1,…4), ″ is the effect coefficient for race interacted with age, ‴ is the effect coefficient for race interacted with age > 50, and age mid = age 0 +age 1 2 is the midpoint of the age group measured in measurement .
Together ′′′′ and ′′′′′ constitute an age-dependent spline for the group quarters population with knots at 0 and 50. We used a Bayesian framework for inference with weakly informative priors for model parameters other than state to assist in regularization, which primarily allowed the data to inform the model estimates. standard deviation 1.0. To capture state variation in prevalence, we used informative priors for state random effects, which were informed by the state-to-state variation in age-/sex-/racestandardized endorsement rates of the respondent-reported visual acuity loss question in ACS.
This prior took the form where, in short, is the log of the ratio of the standardized state prevalence to the national prevalence, and 2 is the standard deviation of the log-ratio of the crude state prevalence to the national prevalence for all strata with at least 500 individuals, and is normal distribution with mean and standard deviation . To be precise, to estimate and 2 , we first estimated the crude prevalence rate of respondent-reported visual acuity loss in the US in 2017 as observed in ACS data. We next estimated the prevalence rate of respondent-reported visual acuity loss for each state from the same data source, stratified by age-group, sex, race/ethnicity, and group quarters status. We then used the state-specific stratified estimates to calculate a standardized prevalence rate, standardizing with the age group, sex, race/ethnicity, and group-quarters weights at the national level, to control for demographic difference between states. We then used these data to estimate the ratio of the standardized state prevalence to the national prevalence. For each state , we set the prior on to be normally distributed with mean = log�ratio �, where ratio is the ratio of standardized prevalence in state to the national prevalence. We obtained an informative standard deviation for this prior as follows: analogous to the ratio use used for the mean, we constructed a ratio of the national prevalence rate to the state-specific prevalence rate for each strata (stratified by age-group, sex, race/ethnicity, and group quarters status) and calculated the standard deviation of the log-ratio for all strata with at least 500 individuals in the © 2021 Flaxman AD et al. JAMA Ophthalmology.
state, age, sex, and race/ethnicity-specific strata. We put a cap on the predicted prevalence at 0.25 to ensure that we are not over estimating our results.
Only data from ACS was available to inform the group-quarters effect coefficients, but unlike the state-to-state variation, we used this ACS data in the likelihood instead of constructing an informative prior. To guard against inferring with more confidence than appropriate from the ACS group quarters data, we used aggregate measurements for broad age groups (25-year intervals) and did not stratify them by sex or by race/ethnicity. We also included fixed effects for each age group, which effectively treated each comparison of the prevalence in the institutional group quarters population and the free-living population as a separate study in the metaregression.
We included all other data sources (NHANES, PBS, NSCH) in the likelihood with the assumption that they applied to a nation-level estimate, which we assumed to be constant over time. This includes data from NSCH, NHANES, and BPEDS for children under the age of 18, and also ACS data on free-living adults over the age of 85.

Validation and Verification
We assessed the cross-model validation of our model by setting its population parameters