[Nigel Moore]

And that, Dave, is the answer that've been looking for! You identified the point that I was overlooking...correction for FWER!

Thank you, so much!

[Joseph Coveney]

There's no conflict between testparm and margins or regress's regression table. Take a look at what testparm says it's doing. They're testing different things: a joint null hypothesis (likewise with ANOVA) versus individual comparisons to the omitted baseline. What did your protocol say you were going to do?

[Marcos Almeida]

First things first. Your doubts, hereby a couple of times clarified, relate to simple regression's as well as one-way ANOVA's core principles. The explanation is usually found in the first pages of any decent textbook. I believe everybody in this Forum will agree that we must have a sound knowledge of both strategies of estimation so as to proceed with the "real-world" analysis, which will usually demand more complex models. Please take some time to grasp information on the basic concepts before jumping to the statistical analysis. Rest assured the reward will be great.

[Nigel Moore]

There's no conflict between testparm and margins or regress's regression table. Take a look at what testparm says it's doing.

To be honest, I wasn't sure what is was doing. Even after reading the State manual.

They're testing different things: a joint null hypothesis (likewise with ANOVA) versus individual comparisons to the omitted baseline. What did your protocol say you were going to do?

I wish to compare the levels of the endogenous metabolite in the treatment groups against the control.

But, in case you think I'm being obtuse (and I admit that I'm no statistician), here's another example from the same data set. This time, the dose-response is clear and significant:

* Example generated by -dataex-. To install: ssc install dataex
clear
input int dose double response2
  0 .003192307692307692
  0 .006386363636363636
  0  .00450381679389313
 10  1.2737819025522044
 10  1.2633228840125392
 10  1.2972972972972974
 10  2.8341708542713566
 10   5.055555555555555
 50  12.352941176470587
 50   4.388571428571429
 50   7.451737451737452
 50   5.728476821192054
 50   6.473684210526317
100  12.569269521410579
100  10.555555555555555
100  10.587301587301589
100  10.833333333333334
100   9.278350515463917
200             42.1875
200  39.075630252100844
200  38.983050847457626
200  43.333333333333336
200   40.28776978417267
end

Now, regression:

. regress  response2 i.dose 

      Source |       SS           df       MS      Number of obs   =        23
-------------+----------------------------------   F(4, 18)        =    335.68
       Model |  5117.72301         4  1279.43075   Prob > F        =    0.0000
    Residual |  68.6057438        18  3.81143021   R-squared       =    0.9868
-------------+----------------------------------   Adj R-squared   =    0.9838
       Total |  5186.32875        22  235.742216   Root MSE        =    1.9523

------------------------------------------------------------------------------
   response2 |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
        dose |
         10  |   2.340132    1.42575     1.64   0.118    -.6552579    5.335521
         50  |   7.274388    1.42575     5.10   0.000     4.278999    10.26978
        100  |   10.76007    1.42575     7.55   0.000     7.764679    13.75546
        200  |   40.76876    1.42575    28.59   0.000     37.77337    43.76415
             |
       _cons |   .0046942   1.127154     0.00   0.997    -2.363369    2.372757
------------------------------------------------------------------------------

. pwcompare dose, mcompare(dunnett)

Pairwise comparisons of marginal linear predictions

Margins      : asbalanced

---------------------------
             |    Number of
             |  Comparisons
-------------+-------------
        dose |            4
---------------------------

--------------------------------------------------------------
             |                                  Dunnett
             |   Contrast   Std. Err.     [95% Conf. Interval]
-------------+------------------------------------------------
        dose |
  10 vs   0  |   2.340132    1.42575     -1.475284    6.155547
  50 vs   0  |   7.274388    1.42575      3.458972     11.0898
 100 vs   0  |   10.76007    1.42575      6.944652    14.57548
 200 vs   0  |   40.76876    1.42575      36.95335    44.58418
--------------------------------------------------------------
Note: The dunnett method requires balanced data for proper
      level coverage. A factor was found to be unbalanced.

If -testparm- is comparing differences from reference, and the values before were all zero because there are no differences, now I should have some non-zero values:

. testparm i.dose

 ( 1)  10.dose = 0
 ( 2)  50.dose = 0
 ( 3)  100.dose = 0
 ( 4)  200.dose = 0

       F(  4,    18) =  335.68
            Prob > F =    0.0000

Nope, not even close. So I have no clue as to what this is telling me. The Stata manual wasn't that intuitive.

[Dave Airey]

Testparm is comparing the joint null hypothesis that 10.dose = 0, 50.dose = 0, 100.dose = 0, and 200.dose = 0. If this is the only factor in your model this will be the same as the overall F omnibus test, like from your ANOVA. In fact you can see the same F value twice in your output as you should in this case.

[Nigel Moore]

Testparm is comparing the joint null hypothesis that 10.dose = 0, 50.dose = 0, 100.dose = 0, and 200.dose = 0. If this is the only factor in your model this will be the same as the overall F omnibus test, like from your ANOVA. In fact you can see the same F value twice in your output as you should in this case.

The light is dawning...thank you, Dave.