Background The use of propensity scores (PS) to balance covariates in the groups of treatment in order to address the indication bias in treatment effect evaluation in longitudinal studies, are more and more frequently used. However, no consensus exists regarding the selection of covariates to include in the PS.
Objectives To compare two approaches for variable selection in the construction of a PS for TNF alpha blockers (TNFb) effectiveness assessment in a cohort of patients (pts) with inflammatory back pain (IBP).
Methods Prospective, multi-centre cohort (DESIR). 708 pts with early IBP suggestive of ax-SpA. Data of the first 2 years of follow-up were included in this analysis. Statistical analysis: a) Exposure: receiving at least 1 TNFb during the first 2 years of follow-up. b) Outcome: ASAS40 response at the last available visit still on treatment (mean 74±30.9w). Data to assess such outcome was available in 197 pts c) Construction of 2 PS:PScomplete: logistic regression was calculated to predict exposure by including all covariates available at baseline visit (except for comorbidities) in the model. No variable selection (no univariate testing or correlation testing) was performed. PSselect: association of each covariate to the outcome was tested by univariate analysis, and only variables with a significant (p<0,10) association were selected. Among these, correlations were tested; in case of two variables were highly correlatied the most clinically informative was selected to be included in the model. e) Outcome comparison: the 197 pts on TNFb were matched (1:1 nearest neighbour technique) to controls according to both PS; the outcome was estimated by logistic regression.
Results PScomplete included 52 variables. PSselect included only 9 variables. Variables selected in PSselect with a significant association with outcome were: gender (p=0,0024), synovitis (p=0.0297), uveitis (p=0.0138), highest education level (p=0.0859), BASGweek (p=0.0174), disease activity (physician) (p=0.0034), HLA-B27 (p=0.0087), ESR (p <0.0001), CRP (p=0.0002), Xray sacroiliitis (p=0.0072) and MRI sacroiliitis (p=0.0001); among these, some were highly correlated and were eliminated. The final PSselect model included 9 covariates: gender, synovitis, disease activity, highest level of education, uveitis, MRI sacroiliitis, X-ray sacroiliitis, CRP and HLAB27. Model comparison: Likelihood test p<0,0001 (PScomplete fitted the data significantly better); AUC=0,8759 (PScomplete) vs. 0.7614 (Psselect); Response to anti-TNF was similar using both PS techniques: ASAS 40 OR =2.75 [1.66–4.54] (p=0,004) in favour of the TNFb treatment with PScomplete, and OR =2.29 [1.26–4.142] for PSselect (p=0.006).
Conclusions Variable selection for construction of PS did not show any advantage compared to include all available covariates: the selective model failed to better fit data and the outcome estimation was comparable with either model. Furthermore in both cases, TNFb lead to a doubled chance of ASAS 40 response at 2 years. The effectiveness of TNFb may be sufficiently robust to “erase” methodological differences. Simulation studies should allow confirming our findings.
Acknowledgements Anna Molto has been granted with a bursary from Société Français de Rhumatologie
Disclosure of Interest : None declared
Statistics from Altmetric.com
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.