a********y 发帖数: 474 | 1 问卷里给 4样产品打分 (ABCD),从0到10代表从低到高。
想看,不同人群(例如在女性里)是否认为A比B更好,而且A的分数significantly
different in rating score from 其他产品BCD。
请问这个用什么方法做?如何adjust for stratification and weight? |
a********y 发帖数: 474 | 2 自己顶一下,谢谢!
【在 a********y 的大作中提到】 : 问卷里给 4样产品打分 (ABCD),从0到10代表从低到高。 : 想看,不同人群(例如在女性里)是否认为A比B更好,而且A的分数significantly : different in rating score from 其他产品BCD。 : 请问这个用什么方法做?如何adjust for stratification and weight?
|
s***h 发帖数: 357 | 3 嗯,算anova。其实就直接做个线性回归,X变量里有产品(你的ABCD indicator),
gender,产品*gender |
a********y 发帖数: 474 | 4 需要GLM吗?如果有strata 和weight 的话?
【在 s***h 的大作中提到】 : 嗯,算anova。其实就直接做个线性回归,X变量里有产品(你的ABCD indicator), : gender,产品*gender
|
w*******9 发帖数: 1433 | 5 Need to consider the subject effect by including a subject random effect
term in your linear model. |
p********6 发帖数: 1339 | 6 有人用ANOVA,因为比较简单。但是很多人不提倡这样做,因为scale data不是
continuous,而且是truncated,样本不够大的时候理用ANOVA偏差很大。比较正确的做
法是用ordinal model来做。 |
y*****w 发帖数: 1350 | 7 As suggested above, use ordinal logistic regression model for the scale data
. In terms of independent variables, create dummy variables for ABCD as well
as for the interaction term of gender*ABCD. |
w*******9 发帖数: 1433 | 8 Good point, but there are 11 possible outcomes, which would require almost a
hundred parameters. Do you think it makes sense to conduct a sensitivity
analysis by adding a Unif(-0.5,0.5) term to each response and see how this
affects the estimates?
data
well
【在 y*****w 的大作中提到】 : As suggested above, use ordinal logistic regression model for the scale data : . In terms of independent variables, create dummy variables for ABCD as well : as for the interaction term of gender*ABCD.
|
y*****w 发帖数: 1350 | 9 I don't think you have to have 11 outcomes. You can categorize the scales
somehow, for example 0-3, 4-7, 8-10.
a
【在 w*******9 的大作中提到】 : Good point, but there are 11 possible outcomes, which would require almost a : hundred parameters. Do you think it makes sense to conduct a sensitivity : analysis by adding a Unif(-0.5,0.5) term to each response and see how this : affects the estimates? : : data : well
|
w*******9 发帖数: 1433 | 10 Then why bother to have the score range from 0 to 10 rather 1,2,3? You're
wasting too much information.
【在 y*****w 的大作中提到】 : I don't think you have to have 11 outcomes. You can categorize the scales : somehow, for example 0-3, 4-7, 8-10. : : a
|
|
|
y*****w 发帖数: 1350 | 11 Note it's not me who bother to have the score range from 0 to 10. It's the
investigator(s) of the study who did that.
【在 w*******9 的大作中提到】 : Then why bother to have the score range from 0 to 10 rather 1,2,3? You're : wasting too much information.
|
c**d 发帖数: 104 | 12 1. Scoring (0-10) data are very common and treated as a continuous variable
instead of an ordinal variable.
2. Weiwei pointed out that categorizing will lost a lot of information.
Furthermore, you have to use proportion odds model to fit an ordinal outcome
. So you make thing be complicated and difficult to interpret results
such as odds ratios across the levels of the ordinal outcome.
3. People still want to know how big difference (effect size) between A and
B by using unadjusted and/or adjusted mean difference (95% CI) or median
difference (95% CI).
4.GLM model still is your first choice here. If residuals look good, please
go head to report adjusted least square mean difference (95% CI).Otherwise,
nonparametric method.
5.If you want to be consecutive, please adjust multiple comparisons to
control type I error.
6.If your sample is large, you also consider Weiwei suggestion to put a
random effect there.
7. Categorizing outcome is your last choice.
8. For me, if a simple way works, I don't choose complicated methods. |
y*****w 发帖数: 1350 | 13 Theoretically Scoring (0-10) data can be treated as a continuous variable,
but it all depends on how the data look like. If in the real data almost all
the scores fall within 0-2, 5-7, and 9-10, i.e. not continuous at all, you
would have to treat them as ordinal. |
y*****w 发帖数: 1350 | 14 Theoretically Scoring (0-10) data can be treated as a continuous variable,
but it all depends on how the data look like. If in the real data almost all
the scores fall within 0-2, 5-7, and 9-10, i.e. not continuous at all, you
would have to treat them as ordinal. |
c**d 发帖数: 104 | 15 Yes.I frequently have zero-inflated scoring data + repeated measurements. It
makes me headache. |
a********y 发帖数: 474 | 16 Thanks so much!
Thank you all for your thoughts.
variable
outcome
and
please
【在 c**d 的大作中提到】 : 1. Scoring (0-10) data are very common and treated as a continuous variable : instead of an ordinal variable. : 2. Weiwei pointed out that categorizing will lost a lot of information. : Furthermore, you have to use proportion odds model to fit an ordinal outcome : . So you make thing be complicated and difficult to interpret results : such as odds ratios across the levels of the ordinal outcome. : 3. People still want to know how big difference (effect size) between A and : B by using unadjusted and/or adjusted mean difference (95% CI) or median : difference (95% CI). : 4.GLM model still is your first choice here. If residuals look good, please
|