Welch’s ANOVA Test

When the homogeneity of variances assumption is not met, especially with unequal sample sizes, Welch’s Test is a good approach for performing an ANOVA analysis.

Property 1: If F is defined as follows:



Example 1: Repeat Example 1 of Kruskal-Wallis using the data in range E19:G29 of Figure 1 by performing Welch’s Test.

Welch's test

Figure 1 – Welch’s Test

We see from row 33 of Figure 1 that the variances of the three groups are 16.2, 86.5 and 265.6, and so we suspect there is a significant difference between the variances. This is confirmed by using Levene’s test (on the medians) since Levene(E20:G29,1) = 0.005478. Thus the normal one-way ANOVA is not the correct test to use. We employ Welch’s test instead, as shown in Figure 1.

We see from Figure 1 that the p-value = .041355 < .05 = α, and so we conclude that there is a significant difference between the means of the three groups.

Note that if we had used ANOVA (see Figure 2) we would have come to a completely different conclusion (since p-value = .14 > .05 = α).

Anova instead of Welch's

Figure 2 – ANOVA on the same data

Real Statistics Function: The Real Statistics Resource Pack contains the following array function where R1 is the data without headings, organized by columns:

WELCH_TEST(R1, lab): outputs a column range with the values F, df1, df2 and p-value for Welch’s test for the data in range R1.

If lab = TRUE a column of labels is added to the output, while if lab = FALSE (default) no labels are added.

For Example 1, the result of WELCH_TEST(E20:G29,TRUE) is similar to range D40:E43 of Figure 1. The main difference is that this function uses the Real Statistics F_DIST function instead of the Excel function F.DIST (or FDIST) to calculate the p-value and so obtains a more accurate result.

Real Statistics Data Analysis Tool: The Real Statistics Resource Pack provides access to Welch’s test via the One Factor Anova data analysis tool, as described in the following example.

Example 2: Repeat Example 1 using the Real Statistics data analysis tool.

Enter Ctrl-m and double click on Analysis of Variance, and select Anova: one factor on the dialog box that appears. Now fill in the dialog box that appears as shown in Figure 3.

Welch's test dialog box

Figure 3 – Dialog box for Welch’s data analysis tool

The output is shown in Figure 4.

Welch's Anova data analysis

Figure 4 – Welch’s test data analysis tool

Note that the results shown in Figure 4 agrees with those in Figure 1 except that the p-value is slightly lower. The reason for this is that Figure 1 uses the formula = FDIST(E40, E41, E42), which is equivalent to =1–F.DIST(E40,E41,E42, TRUE). Both of these formulas truncate the value in E42 down to an integer value, i.e. to =FDIST(4.315278,2,11). The calculation in Figure 4 is more exact and uses F_DIST instead of F.DIST and so the full value of df2 = 11.69964 is used.

As can be seen from Figure 3, data for Welch’s test can be organized in standard format. The first 10 of the 27 rows of the data for Example 1 in standard format is shown in Figure 5.

Welch's test stacked format

Figure 5 – Data in standard format

50 Responses to Welch’s ANOVA Test

  1. Valerie says:


    What happens if I have outliers? My data is normal and non homogeneous.


    • Charles says:

      First things first. Are these outliers representative of what is really going on or are they errors (typing mistakes, poor instrument readings, etc.)? Also how far away from the mean/median are they (a little out of what is expect or a lot)?
      One approach to dealing with outliers is to a test which is more forgiving of outliers (e.g. Krusal-Wallis since it uses ranks) or to run the test twice: once with the outliers and one without and report both results. You can also use bootstrapping techniques.

  2. Jeff says:

    Hi Charles,
    I’ve encountered a bit of a problem. I need to run Welch’s and Games-Howell on my data. Because I use a Mac, I’ve had to do this on a Windows computer in another area of our campus. I saved the Excel file after doing so, but when I open the file on my Mac I get a linking error. In the Welch’s and Games-Howell results boxes I get this sort of thing in the formula bar in front of certain formulas: AppData/Roaming/Microsoft/AddIns/RealStats.xlam’!QCRIT(COUNT(X3:X5),W9,Y1,2). It’s obviously trying to refer back to the computer on which I ran the analyses.
    Any solution for this?
    Best wishes,

    • Charles says:

      Hi Jeff,
      If you have the Real Statistics software installed on the Mac, all you need to do is erase AppData/Roaming/Microsoft/AddIns/RealStats.xlam’! so that the formula becomes =QCRIT(COUNT(X3:X5),W9,Y1,2).
      You can also use Excel’s Replace capbility (found at Home>Editing|Find&Select) to replace all instances of the string “AppData/Roaming/Microsoft/AddIns/RealStats.xlam’!” by blank

  3. Jeff says:

    I really like your package. I use it on Mac. A few months ago I did some analyses that required Welch’s ANOVA with Games-Howell and Levene’s testing. However, today I cannot do that as the options are not available in the dialogue window. Have you changed the Mac version? (Or, has Excel changed so yours doesn’t work the way it used to?)

    • Jeff says:

      Apologies. I saw an answer you gave to another user in a different topic that says the Mac version works differently than the Windows version. I have the output I got earlier in my spreadsheet, so I must have done it on a Windows computer.
      Best wishes,

  4. Ken says:


    In figure 1, you show a p-value of 0.041355 based on the formulas in Excel.
    In figure 4, the output from using your add-in shows a p-value of 0.039466. This is confirmed by running the same one-way anova (with Welch) in Minitab.

    Can you help explain why the p-value in figure 1 is different than what your add-in and Minitab return?

    • Charles says:

      The results shown in Figure 4 agrees with those in Figure 1 except that the p-value is slightly lower. The reason for this is that Figure 1 uses the formula = FDIST(E40, E41, E42), which is equivalent to =1–F.DIST(E40,E41,E42,TRUE). Both of these formulas truncate the value in E42 down to an integer value, i.e. to =FDIST(4.315278,2,11). The calculation in Figure 4 is more exact and uses the Real Statistics function F_DIST instead of the Excel function F.DIST and so the full value of df2 = 11.69964 is used in the calculation.

  5. Asnicar says:

    Hi Charles,

    with your example I can see that there is a significantly difference between the three groups New, Old and Control. But, how can I investigate the difference in pairwise comparisons? Like New compare to Old and Old compare to Control and New compare to Control. Is it the Games Howell the follow up test to be used after Welch’s? Alternatively, is it correct to do the Welch’s Test in two groups per time? First New-Old, then Old-Control and then New-Control, for example.

    And in the case my data don’t respect both the homoscedasticity and the normality assumption, is it good to use the Welch’s Test or is it to prefer the Kruskal-Wallis?

    Thank you very much.

    • Charles says:

      1. You can follow up Welch’s Anova with multiple Welch’s t tests, but this will increase the experimentwise error, which you can compensate for by using a Bonferroni correction factor. Games-Howell is a commonly used post-hoc test, which automatically takes care of experimentwise error.
      2. Kruskal-Wallis requires homoscedasticity and Welch’s requires normality. If the data is close to normality use Welch’s. If both assumptions are strongly violated then the only approach that I know is resampling, which is explained on the website.

  6. Pingback: Le test de Welch – lemakistatheux

  7. AbiAbboud says:

    IN ANOVA TEST IF levene’s test was 0.05 what should we do?
    1) we continue the test normaly and we use Tukey’s value in our report
    2) we stop the report and we write: ANOVA TEST CAN’T BE DONE”!!!

    • Charles says:

      As usual this is a judgement call. I wouldn’t say that if p-value = .051 then use ANOVA, but if p-value = .049 then don’t use ANOVA. Most likely I would use ANOVA and point out that the homogeneity of variance assumption may be marginal.

  8. Sili says:

    Hi Charles,
    Beautiful work!

    I now understand the process of calculating the Welch ANOVA. I wonder what is the difference between the result from Welch ANOVA and general linear regression (GLR) model. In the variance matrix in the general linear regression model, I assume that the covariance terms are zero and the variances are different for each group. I think this GLR model should be equivalent with Welch ANOVA. But I get different p value from them. Do you know the reason?

    • Charles says:

      Hi Sili,
      I have never investigated this, but I don’t think Welch’s ANOVA is equivalent to GLR, and so it is not surprising that the p-values are different.

  9. knk says:

    Guys, homogeneity of variances almost never happens in real data if we speak about economics/social sciences. So no need to test it, cause 90% chance you’ll find your variances are not equal.

    Just always proceed with Cramer-Welch’s / non-parametric…

  10. Alexandra says:

    Dear Charles,

    I have one question for you if that’s ok. I don’t have extended knowledge and hence comprehension of statistics, but still I’m trying to find my way with it for my current Master thesis. I tried to do a Factorial Anova, but Levene’s test showed to be non-significant. I then tried via an ANOVA to check for Welch and Brown-Forsythe. Both show non-significant as well. I can’t seem to find any explanation of what this actually means (whether this is good or bad), and what should/can be done as a result. Is there any way I could send you my data and explain what it is I’m trying to check? Maybe that would be easier for you to advise me. I leave it up to you, but thank you for attention in any case!

    Thanks and kind regards,


  11. Alessandro Zarpellon says:

    Beautiful website Charles, lots of nice and clear explanations! I have a couple of questions. First, I did not have time yet to check all the website, but does your add- also runs post-hoc tests for datas that are not normal nor homoscedastatic? Something like the Dunnet-T3. I was thinking first to run a Welch`s test, but I don`t know what to do after, to do post-hoc test. Second: I do biochemical analysis and we take the SAME cells suspension, we divided in 3, we use 1 aliquot as control, and 2 others for 2 different pharmacological treatments. I may be wrong, but in a set-up like this it seems to me the data are NOT independent, thus I cannot run an ANOVA, am I right?

    • Charles says:


      Thank you for your kind words about the website.

      The Real Statistics software supports Dunnett’s post-hoc test, but I am not sure whether this is the Dunnet-T3 test that you are referring to. It is common to use the Games-Howell post-hoc test after Welch’s ANOVA. This test is also supported by the Real Statistics software. You can find information about both of these tests at Unplanned Comparisons.

      I can’t tell from your description whether or not your data are independent. In any case, Repeated Measures ANOVA handles data groups that are not independent. You can read more about this on the website. It too is supported by the software.


  12. R says:

    In Excel, the function ‘t-test’ allows you to perform a wWelch corrected test, just type ‘3’ where the formula requires ‘type’ .

  13. Valerie says:


    I originally conducted a one-way ANOVA between 3 separate groups, but found that equality for variance was not met -does that mean I need to run a Welch test instead?


  14. Hamid A says:

    Dear Charles

    thanks again for the amazing job you have been doing, you are helping of hundreds, maybe thousands of people like me.
    My question is, when you said “I believe that Welch’s is not so good with skewed data” above, do you think you can supply the reference for that please. I am using the same argument but cant find anything to backup my argument.

    Many thanks

  15. Raulatu says:

    My results show that homogeneity of variances is not met. My question is using the Games -Howell recommended i can’t see any significant difference between the groups. I have a sample sample size (3) and its the same for all the groups. The mean of the groups look different.
    What is wrong ?

    Thank you.

    • Charles says:

      Are you saying that Welch’s test shows there is a significant difference between the groups, but based on Games-Howell there is no significant difference between the groups? If you send me an Excel file with your data and the tests that you have performed I will try to figure out what is going on. See Contact Us for my email address.

  16. Kevin Bluxome says:

    Hi Mr. Zaiontz,

    I understand how to actually perform the Welch ANOVA, thanks to the clear instructions given here. I do like to understand what goes on “under the hood,” so to speak, and I have yet to find any easy to understand info that explains exactly how the Welch test “works.” If the “weight” for each group is defined as the (group size/group variance), then Based on the calculations, I can see the numerator of Welch’s F being sort of a “weighted average” of the total variance, more or less. What in the world does the denominator of the Welch F ratio represent? I think I read somewhere that the denominator was somewhat equivalent to a correction factor based on the “expected value” of the variance, or something similar. Do you have any ideas about what the denominator might represent, and if it is related to the numerator seeming to be a kind of weighted average? Thanks for any insight!

    • Charles says:

      Hi Kevin,

      Sorry, but I haven’t had the time to research how the denominator was derived, but you can look at the original paper, which you can find online.

      Welch, B.L. (1951) On the comparison of several mean values: an alternative approach. Biometrika.


      • Kevin Bluxome says:

        Thank you, sir! After looking at that paper, I’m a bit sorry I asked…looks like what’s going on involves just a little more calculus than I took in college! I have a whole new respect for the people like Welch who figured these things out for the rest of us!

  17. Renee says:

    Hello Charles,

    Thank you so much for this article. However, even with my Data Analysis Tool pack add in on, Figure 3 shows that there is an ‘option’ window where one can select a Welchs Test. I am not able to do this in the most current version of excel. I tried downloading and installing the add-in provided by this site, but I’m having problems. Any advice or help is greatly appreciated.

    • Charles says:

      Welch’s Test is not part of Excel’s Data Analysis Toolpak. You need to install the Real Statistics Resource Pack to use Welsh’s Test.
      You said “I tried downloading and installing the add-in provided by this site, but I’m having problems”. What sort of problems are you having?

  18. Renee says:

    Hello Charles,

    Thank you so much for publishing this article. However, when I activate the data analysis pack add in, to run a one way ANOVA I don’t see all the options as shown in Figure 3. I cant choose Welch’s vs. Scheffe vs. ect. Is there a different add-in I need to use? I have the most current version of MSO.


    • Charles says:

      Welch’s Test is not part of Excel’s Data Analysis Toolpak. You need to install the Real Statistics Resource Pack to use Welsh’s Test.

  19. Mohammed GHALIP says:

    What is cutoof P-value in levene test in One way Anova, does it depend on the sample size
    what about variance ratio ,when it used.

  20. Pablo Mancheño says:

    Dear Dr Zaiontz.

    I am not sure if my question is correct in this area.
    I have data from three different groups and I would like to know if there are significant differences between each group.

    Could you show me the correct test that I have to aply if my data are not normaliced.

    Thank you very much and congratulations for your website.

    • Charles says:

      The usual test to use to determine whether there is a significant difference between three groups is ANOVA, provided the assumptions for this test are met.

  21. Tara says:

    Hi Charles,

    I love the work you have put into this site!

    I am trying to do post tests following a Welch’s ANOVA (my data has unequal variances). Do you recommend Games-Howell following a Welch’s ANOVA? Over a Dunnett’s? In the past I have used a Dunnett’s, which I think doesn’t require an additional familywise error correction, but if I use a Games-Howell, do I need an additional correction for family wise error?

    If I want to correct for family wise error, do I need to change the alpha value in the table? I think the software makes the correction automatically for the contrast method, but not for the Games-Howell. I realize I can adjust the “alpha” value in the Games-Howell table to make my p cut off more stringent, but would you suggest I use a Bonferroni correction or is there something other correction I should use, especially if I am only comparing each group to control?

    Thank you,

    • Charles says:

      Generally Games-Howell would be my recommendation after Welch’s test. Games-Howell corrects for familywise error and so no additional correction factor is needed.

  22. Melanie says:

    What if I need to run a two way ANOVA but my HOV is violated? I believe the Welch is only for one way, is that correct?

    Thank you!

    • Charles says:


      Welch is only one-way. I don’t know of any two-way test, although some sort of bootstrap approach might work. I saw on the Internet the following articles as well:


      Some seem to normalize the ranks of the data, using a formula such as NORMSINV((r-3/8)/(n+1/4)) where nn= the sample size and r = the rank of the ith element, and then use some other technique.


  23. Eliza says:

    Thank you Charles! Very much appreciated.

  24. Eliza says:

    Thank you very much for putting up a very informative site…

    I would really appreciate it if you could help me on this one. For the data set I’m working on, I used the Welsch’s test since the p value of the means after performing Levene’s test was less than 0.05 and after transforming the data to log, the p value of the means after performing Levene’s test was still less than 0.05. Were my procedures correct? Also, is the Games Howell Test applicable as the post-hoc test for Welch’s test, specifically the since the data I’m working on have unequal sample size? Thanks very much.

    • Charles says:

      Your procedures seem correct to me. Some might use an alpha value less than .05 for Levene’s test, but the approach you used (alpha = .05) is generally the one people use. Games Howell is generally a good choice after Welch’s test.

  25. Matt says:

    I am happy I found your description of Welch’s Anova as I work with environmental bacteria samples with a large degree of variance.

    I do have on question though: would it be acceptable to go through my data and perform both normal Anovas (where there is not significant variance) and Welch’s anova (where the difference is significant) and then report whichever P value was used in the same table? I would probably indicate with an asterisk which anova was used. Basically should you JUST use one form of anova for all your data or can you pick which to use dependent on what your variances are? I have many groups of observations (6 simultaneous samples, 3 each from 2 locations) from different days (10 days) and some have equal variance and some don’t. I want to look at individual days but am not sure that the P values gained from each anova are comparable.

    I hope that made some sense.

    Thank you, great article.

    • Charles says:

      If the assumptions for ANOVA are reasonably met, you only need to use ANOVA. If the group variances are quite different then you might want to use Welch’s test (or resampling, etc.). If you do use both ANOVA and Welch’s test, I don’t see any problem in reporting both. The thing to avoid is deciding after the fact which result you like better and then using that result and ignoring the other result.

      • Matt says:

        Thank you for the quick response.

        I will pick the P value given by an ANOVA unless unequal variance is significant in which case I will use Welch’s ANOVA.

        I further analyzed my data today and realized I am also encountering many cases of significant non-normality as well. I know the Kruskall-Wallace ANOVA is an option (as well as transformations) but does Welch’s still do well against non-normality in addition to unequal variance? I am not such a fan of Kruskall-Wallace ANOVA but does that do well against both unequal variance and non-normality?

        Thanks again,

Leave a Reply

Your email address will not be published. Required fields are marked *