Ibrahim Q, Iorio A, Curtis R, Nichol M, Noone D, Stonebraker J, Skinner M, Germini F, and the PROBE Investigators. A Comparative Study of Different Methods of Handling Missing Data in Patient Reported Outcomes, Burdens and Experiences (PROBE) Score Algorithm among People with Hemophilia. (2022), Abstratc. THSNA. 2022.
Background
The Patient Reported Outcomes, Burdens and Experiences (PROBE) questionnaire measures quality of life (QoL) in people with hemophilia (PWH) and people with no bleeding disorder (NoBD). A score is calculated as the average of nine core item score (0=worst and 1=best reported health status). There is currently no validated method for calculating the PROBE score when some item scores are missing. Grouping highly correlated question scores into a domain and then treating missing values within a domain could be an effective strategy.
Objective
Our objectives were to identify domains within PROBE and compare four strategies of estimating the score as an average of the available item scores when the availability of scores were:
- ≥50% item scores within a domian,
- only one item score within a domain,
- ≥50% item scores irrespective of any domain,
- 8 out of 9 item scores.
Method
The observational PROBE phase 3 study data (2018/10/10 – 2021/10/29) were used. Item scores with intra-class correlation (ICC) ≥0.5 were grouped into a domain. We created 36 data sets with artificially generated missing PROBE and item scores from each combination of i) 3 types of hypothetical missing data: Missing Completely At Random (MCAR), Missing At Random (MAR): missing among aged >45 years, and Missing Not At Random (MNAR): missing from the lower quartile of the score, ii) 3 percentages of missing values: 10, 15, and 20% , and iii) the 4 scenarios of missing item scores. A strategy with mean of absolute errors (MAE) (Standard Deviation (SD)) <0.05, and calibration intercepts and slopes not systematically different from 0 and 1, respectively, was considered acceptable.
Results
Among 3217 participants, 48% were PWH, 9% were hemophilia carriers, and 43% were NoBD. 20% of the participants were female, and the mean (SD) age was 41 (15) years. Chronic pain score had ICC ≥0.5 with each of acute pain; pain medication; and difficulty of activities of daily living (ADLs). These four item scores were grouped into a domain. Internal consistency within the domain was shown (Cronbach’s alpha=0.8). For MCAR and MAR data, the closest estimates of PROBE were observed for strategies 1 and 4 (MAE ± SD: 0.02 ± 0.02), followed by strategies 2 and 3. Strategy 1 estimated PROBE score accurately for MCAR and MAR data, and slightly underestimated the score in case of MNAR data. Strategies 2 and 4 slightly underestimated the score for MCAR and MAR data and considerably underestimated the score for MNAR data. Strategy 3 substantially underestimated the score for all missing data types.
Conclusion
Chronic pain was correlated with acute pain, pain medication, and difficulty of ADLs for participants of the PROBE study, indicating that pain is an important contributer to QoL and the PROBE score. We recommend estimating PROBE scores as a simple average of available item scores if at least one item within pain/difficulty in ADLs domain is available or 8 out of 9 items are available irrespective of the domain. This algorithm allows for calculation of the PROBE score for those with missing data.
View Poster: Missing Data Algorithm
