[R] help in interpreting paired t-test

[Pedro Mardones]

Thanks for all the replies and comments. I've followed Marc's
suggestion of using the Bland-Altman's approach which I found pretty
clarifying for comparing data collected on the same subjects.

BR,
PM

On Wed, Sep 21, 2011 at 1:39 PM, Marc Schwartz <marc_schwartz at me.com> wrote:
> Jeremy,
>
> Correlation alone is irrelevant when comparing two separate sets of measurements on the same specimen. Correlation does not mean good agreement, but good agreement tends to infer high correlation.
>
> T1 <- rnorm(50, mean = 100)
>
>> mean(T1)
> [1] 99.80257
>
> T2 <- T1 * 1.5
>
>> mean(T2)
> [1] 149.7039
>
> The two measures are off by a systematic 50%, but:
>
>> cor(T1, T2)
> [1] 1
>
>
> The key here, as I noted in my reply yesterday and as Greg noted in his this morning regarding Bland-Altman, is whether or not the two measures agree within an acceptable margin of error and whether or not there is systematic bias in the measures, either overall or perhaps one measure tends to be low at one end of the range, while high at the other.
>
> HTH,
>
> Marc Schwartz
>
>
> On Sep 21, 2011, at 11:20 AM, Jeremy Miles wrote:
>
>>> cor(A, B)
>> [1] 0.9986861
>>
>> The data are very, very highly correlated. The higher the correlation,
>> the greater the power of the t-test to detect the same difference
>> between the means.
>>
>> Jeremy
>>
>> On 20 September 2011 10:46, Pedro Mardones <mardones.p at gmail.com> wrote:
>>> Dear all;
>>>
>>> A very basic question. I have the following data:
>>>
>>> ************************************************************************************
>>>
>>> A <- 1/1000*c(347,328,129,122,18,57,105,188,57,257,53,108,336,163,
>>> 62,112,334,249,45,244,211,175,174,26,375,346,153,32,
>>> 89,32,358,202,123,131,88,36,30,67,96,135,219,122,
>>> 89,117,86,169,179,54,48,40,54,568,664,277,91,290,
>>> 116,80,107,401,225,517,90,133,36,50,174,103,192,150,
>>> 225,29,80,199,55,258,97,109,137,90,236,109,204,160,
>>> 95,54,50,78,98,141,508,144,434,100,37,22,304,175,
>>> 72,71,111,60,212,73,50,92,70,148,28,63,46,85,
>>> 111,67,234,65,92,59,118,202,21,17,95,86,296,45,
>>> 139,32,21,70,185,172,151,129,42,14,13,75,303,119,
>>> 128,106,224,241,112,395,78,89,247,122,212,61,165,30,
>>> 65,261,415,159,316,182,141,184,124,223,39,141,103,149,
>>> 104,71,259,86,85,214,96,246,306,11,129)
>>>
>>> B <- 1/1000*c(351,313,130,119,17,50,105,181,58,255,51,98,335,162,
>>> 60,108,325,240,44,242,208,168,170,27,356,341,150,31,
>>> 85,29,363,185,124,131,85,35,27,63,92,147,217,117,
>>> 87,119,81,161,178,53,45,38,50,581,661,254,87,281,
>>> 110,76,100,401,220,507,94,123,36,47,154,99,184,146,
>>> 232,26,77,193,53,264,94,110,128,87,231,110,195,156,
>>> 95,51,50,75,93,134,519,139,435,96,37,21,293,169,
>>> 70,80,104,64,210,70,48,88,67,140,26,52,45,90,
>>> 106,63,219,62,91,56,113,187,18,14,95,86,284,39,
>>> 132,31,22,69,181,167,150,117,42,14,11,73,303,109,
>>> 129,106,227,249,111,409,71,88,256,120,200,60,159,27,
>>> 63,268,389,150,311,175,136,171,116,220,30,145,95,148,
>>> 102,70,251,88,83,199,94,245,305,9,129)
>>>
>>> ************************************************************************************
>>>
>>> plot(A,B)
>>> abline(0,1)
>>>
>>> At a glance, the data look very similar. Data A and B are two
>>> measurements of the same variable but using different devices (on a
>>> same set of subjects). Thus, I thought that a paired t-test could be
>>> appropriate to check if the diff between measurement devices = 0.
>>>
>>> t.test(A-B)
>>>
>>> ************************************************************************************
>>>
>>> One Sample t-test
>>>
>>> data:  A - B
>>> t = 7.6276, df = 178, p-value = 1.387e-12
>>> alternative hypothesis: true mean is not equal to 0
>>> 95 percent confidence interval:
>>>  0.002451622 0.004162903
>>> sample estimates:
>>>  mean of x
>>> 0.003307263
>>>
>>> ************************************************************************************
>>> The mean diff is 0.0033 but the p-value indicates a strong evidence to
>>> reject H0.
>>>
>>> I was expecting to find no differences so I'm wondering whether the
>>> t-test is the appropriate test to use. I'll appreciate any comments or
>>> suggestions.
>>>
>>> BR,
>>> PM
>>>
>>> ______________________________________________
>>> R-help at r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-help
>>> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
>>> and provide commented, minimal, self-contained, reproducible code.
>>>
>>
>> ______________________________________________
>> R-help at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-help
>> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
>> and provide commented, minimal, self-contained, reproducible code.
>
>