# A practical evaluation of statistical methods for the analysis of patient reported outcomes in an observational pharmaceutical study

## TL;DR

This paper compares different statistical methods for analyzing patient-reported outcomes in HIV studies to find better ways to understand treatment effects.

## Contribution

The study evaluates multivariable regression approaches as superior for handling missing data and providing detailed insights in observational PRO analyses.

## Key findings

- Paired difference tests showed significant increases in PRO scores but assumed missing data were random.
- LMMs and wGEEs revealed significant increases in PRO scores with a steep initial rise followed by a plateau.
- Multivariable regression approaches outperformed simplistic methods in handling missing data and controlling for confounding factors.

## Abstract

Patient-reported outcomes (PROs) provide a unique opportunity to tailor clinical care to patients’ needs. Observational pharmaceutical industry analyses of PROs in the HIV field often utilise simplistic pairwise comparisons of pre-defined follow-up periods to baseline, making inappropriate missing data assumptions and yielding limited information on the nature of the change in PRO. Our aim was to evaluate different statistical approaches for PRO analyses.

Paired difference tests, Friedman’s ANOVAs (F-ANOVA), linear mixed models (LMMs) and weighted generalised estimating equations (wGEEs) were applied to the analysis of the Short Form 36 (SF-36) mental component score (MCS) and physical component score (PCS) from treatment-naïve patients in an observational cohort of people living with HIV. Changes in MCS and PCS were assessed to compare the benefits of each approach.

The paired difference test demonstrated statistically significant increases in MCS and PCS from baseline to every follow-up, assuming however, data were missing completely at random. Use of the F-ANOVA was limited due to unbalanced data, leading to non-responder bias. While controlling for covariates, the LMMs and wGEEs illustrated a statistically significant increase in MCS and PCS with a steep increase over the first few months, followed by a plateau.

Relative to paired difference tests, multivariable regression approaches can better handle missing data, control for confounding factors, and provide information on the timing and magnitude of PRO changes. Regression methods therefore facilitate more informative conclusions in observational PRO analyses, and thus provide more detailed evaluations of treatment regimens from the patient’s perspective.