When to use crosswalks
The most widely used health utility measures include the EuroQol-5D (EQ-5D), Health Utilities Index (HUI), Six-Dimensional Health State Short Form (SF-6D), and the Quality of Well-being Index (QWB). PROPr has benefits over these older measures which are described in the FAQ. Unfortunately, having multiple health utility measures available makes it difficult to compare results across studies that have used different measures. When different studies use different measures, crosswalks provide a second-best approach to combine results from such studies.
Though the EQ-5D, HUI, SF-6D, QWB, and PROPr all measure “health utility,” they vary in the range of health described, the method of health valuation, and the time and place where valuation was completed. Directly comparing scores from different systems is therefore limited. In general, it is recommended that economic analyses use a single scoring system as the source of health utility scores. If scores from multiple measures must be compared or combined, crosswalks are used to put all values on a common metric.
How these crosswalks were developed
We use data from large, US nationally representative online surveys with n>1000 that co-administered PROPr with other measures. Links to the datasets are at the end of this page. The respondents were randomly split into a two-thirds estimation sample and one-third validation sample. When possible, crosswalks were also validated in other surveys which co-administered the same measures.
The PROPr summary scores form a bell-shaped curve and the data fulfilled all major ordinary least squares (OLS), regression criteria, so was used as the primary model for crosswalks to PROPr. We tested alternative models (Tobit and beta regression) that did not show improvement in model fit over OLS. For models with multiple predictors (i.e., models with domain level data), Akaike information criterion (AIC) was used to select the best subset of predictor variables. The domain scores were evaluated as both ordinal and continuous variables and the best model was selected using AIC.
We evaluated models using the validation sample using mean error, mean squared error, product-moment correlations, and intraclass correlation coefficients between predicted and observed scores. We also compared estimate and observed scores by age- and gender- subgroups and by number of self-reported chronic conditions.
These crosswalks generally performed well. Minimally important differences for most health utility measures are between 0.03 and 0.05 and the mean error for most of the subgroup analyses is <0.02. There are important exceptions to this generalization: the crosswalks perform poorly for populations with >3 chronic conditions. The crosswalks work better in healthier subgroups because the sample used in this report is from the general US population and is relatively healthy. This pattern of results is similar to crosswalks developed in other studies and is a reminder that crosswalking is always imprecise, but crosswalking to a common metric is better than combining un-crosswalked scores from several sources.
All analyses were performed using SAS 9.4. OLS analyses were performed using PROC GLMSELECT, Tobit analyses were performed using PROC QLIM, and beta regression analyses were performed using PROC GLIMMIX.
Crosswalks from domain level data are preferred
EQ-5D-5L domain responses to PROPr summary score
This function assumes all domains are scored such that 1 is best health and 5 is worst health for the domain.
PROPrScore = 0.671 + [-0.072(if EQ_Mobility = 2 or EQ_Mobility = 3), -0.091 (if EQ_Mobility=4 or if EQ_Mobility=5)] + [-0.092(if EQ_UsualAct = 2), -0.154 (if EQ_UsualAct = 3 or if EQ_UsualAct = 4), -0.174(if EQ_UsualAct = 5)] + [-0.050(if EQ_Pain = 2), -0.114(if EQ_Pain = 3), -0.151(if EQ_Pain = 4), -0.153(if EQ_Pain = 5)] + [-0.099(if EQ_AnxDep = 2), -0.181(if EQ_AnxDep = 3), -0.268(if EQ_AnxDep = 4), -0.270(if EQ_AnxDep = 5)]
HUI Mark 2 domain responses to PROPr summary score
This function assumes all domains are scored such that 1 is best health and 4 or 5 is worst health for the domain.
PROPrScore = 0.639 + [-0.074 (if HUI2_mobility=2), -0.091 (if HUI2_mobility=3), -0.124 (if HUI2_mobility=4 or if HUI2_mobility=5)] + [-0.116 (if HUI2_cognition=2 or if HUI2_cognition=3), -0.215 (if HUI2_cognition=4)] + [-0.081 (if HUI2_pain=2), -0.167 (if HUI2_pain=3), -0.201 (if HUI2_pain=4), -0.266 (if HUI2_pain=5)] + [-0.053 (if HUI2_emotion=2), -0.120 (if HUI2_emotion=3), -0.157 (if HUI2_emotion=4), -0.237 (if HUI2_emotion=5)]
HUI Mark 3 domain responses to PROPr summary score
This function assumes all domains are scored such that 1 is best health and 5 or 6 is worst health for the domain.
PROPrScore =0.651 + [-0.061 (if HUI3_amb=2), -0.072 (if HUI3_amb=3), -0.099 (if HUI3_amb=4), -0.139 (if HUI3_amb=5), -0.248 (if HUI3_amb=6)] + [-0.058 (if HUI3_emotion=2), -0.097 (if HUI3_emotion=3), -0.145 (if HUI3_emotion=4), -0.179 (if HUI3_emotion=5)] + [-0.095 (if HUI3_cognition=2 or if HUI3_cognition=3), -0.152 (if HUI3_cognition=4 or if HUI3_cognition=5 or if HUI3_cognition=6)] + [-0.066 (if HUI3_pain=2), -0.174 (if HUI3_pain=3), -0.200 (if HUI3_pain=4), -0.211 (if HUI3_pain=5)]
SF-6D domains to PROPr summary score
This function assumes the data are coded as:
sf12_1 = 1=”excellent” to 5=”poor”
sf12_3, sf12_4 = 1=”limited a lot” to 3=”not limited at all”
sf12_7 = 1=”all the time” to 5 =”none of the time”
sf12_8 = 1=”not at all” to 5=”extremely”
sf12_10, sf12_11 = 1=”all the time” to 6=”none of the time”
sf12_12 = 1=”all the time” to 5=”none of the time”
PROPrScore = 0.234 + sf12_1* -0.032 + sf12_3 * 0.019 + sf12_4 * 0.030 + sf12_7 * 0.063 + sf12_8 * -0.047 + sf12_10 * -0.028 + sf12_11 * 0.033 + sf12_12 * 0.033
Crosswalks from overall scores are available if needed
The domain scores always perform better than the summary scores but many studies only report summary scores; if domains scores are available, use those instead of overall scores.
EQ-5D-5L US summary score to PROPr summary score
PROPrScore = -0.008 + 0.623*EQ5DScore
HUI Mark 2 summary score to PROPr summary score
PROPrScore = -0.079 + 0.696*HUIMark2Score
HUI Mark 3 summary score to PROPr summary score
PROPrScore = 0.127 + 0.491*HUIMark3Score
SF-6D summary score to PROPr summary score
PROPrScore = -0.454 + 1.218*SF6DScore
Would you prefer a different model or sample?
All data used to estimate these crosswalks are publicly available. The data used for these crosswalks are from samples meant to represent the US general population, so analysts who are concerned that their population differs greatly from the general population could replicate these analyses with a subset of the data that more accurately represents their population of interest. The data sets have other information such as age, sex, and presence of common chronic conditions.
Analysts could also consider using the publicly available data for multiple imputation.
These data include PROPr, EQ-5D-5L, HUI, and PROMIS-Global.
These data include PROPr, HUI, and PROMIS-Global.
These data include PROPr, EQ-5D, HUI, and the SF-6D.