Rapid Responses to:

LETTERS: Petri E Voutilainen Assessment of grouping variable should have been blind in trial of dementia BMJ 2001; 322: 1491b [Full text]

Rapid Responses: Submit a response to this article

Rapid Responses published:

Stratifying for Variables Influenced by Treatment

Vance W Berger, Ellen X Yi, Lauren Edukat   (26 July 2001)

Stratifying for Variables Influenced by Treatment 26 July 2001
Vance W Berger, National Cancer Institute National Cancer Institute, University of Maryland, Ellen X Yi, Lauren Edukat Send response to journal: Re: Stratifying for Variables Influenced by Treatment	Stratifying for Variables Influenced by Treatment EDITOR - Voutilainen [1] noted that in a study [2] of the effect of intensive geriatric rehabilitation on dementia patients with hip fracture, the degree of dementia was determined subsequent to randomization. In addition, evaluators were not masked to the treatment assignment, and different evaluators may have been used for the two groups. Clearly, this opens the door for the possibility that the treatment assignment influenced the classification of degree of dementia. In fact, the patients in the intervention group did tend to have a deeper level of dementia in comparison to the control group. The degree of dementia was used as a stratification variable for the analysis presented in Table 2 of [2]. The concern is that if the apparent between-group difference in level of dementia is actually an artifact created by differential scoring conventions across treatment groups, then the wrong comparisons are being made. This could, even in the absence of any real treatment effect, account for the significant between-group p-values in length of hospital stay within strata (0.042 for the stratum defined by mini mental state examination scores 12-17 and 0.002 for in the stratum defined by scores 18 -23). We agree with Voutilainen [1] that masked assessments would prevent such biased classification. However, this step alone would not remove our concern because it would still allow for the possibility that the treatment had a true effect on the stratification variable. This is known to lead to unreliable results [3]. Hence, we feel that masking would not be sufficient. No adjustment should be made for variables measured subsequent to randomization [3]. Vance W. Berger, Ph.D., National Cancer Institute, University of Maryland, Baltimore County Ellen X. Yi, University of Maryland, Baltimore County, National Cancer Institute Lauren Edukat, University of Maryland, College Park, National Cancer Institute References 1 Voutilainen, P. Assessment of Grouping Variable Should Have Been Blind in Trial of Dementia. BMJ 2001; 322:1491 (16 June.) 2 Huusko TM, Karppi P, Avikainen V, Kautiainen H, Sulkava R. Randomised, Clinically Controlled Trial of Intensive Geriatric Rehabilitation in Patients with Hip Fracture: Subgroup Analysis of Patients with Dementia. BMJ 2000;321:1107-1111 [Abstract/Full Text]. (4 November.) 3 Andersen PK, Armitage P, Colton T. Time-Dependent Covariate. Encyclopedia of Biostatistics; Vol 6: 4523-4525. (1998)