Lilliefors Test Table

These tables give the critical values Dn,α for the Lilliefors test for normality, as described in Lilliefors Test. The first table is the original table from Lilliefors, while Table 2 is a revised version from Abdi and Molin.

Table 1

Lilliefors table original version

Table 2

Lilliefors Table revised part1

Lilliefors Table revised part2

where image158z

Download Table

Click here to download the Excel workbook with the above table.

References

Abdi, H. & Molin, P. (2007) Lilliefors/Van Soest’s test of normality. Encyclopedia of measurement and statistics. Neil Salkind (Ed.). Sage, Thousand Oaks, California
https://www.utdallas.edu/~herve/Abdi-Lillie2007-pretty.pdf

Lilliefors, H. W. (1967) On the Kolmogorov-Smirnov Test for Normality with Mean and Variance Unknown, Journal of the American Statistical Association, Vol. 62, No. 318, pp. 399-402.
https://pdfs.semanticscholar.org/4aad/1756e88dba86399a75891895e00b160f5460.pdf

39 thoughts on “Lilliefors Test Table”

  1. Excuse me sir, how about if my sample only use 3 data? Because the littlest point is only 4. Is my data wrong?
    Or, how to calculate critical value of liliefors table?

    Reply
    • If your sample only has 3 elements you can use Shapiro-Wilk test to test for normality, but no matter what test you use, the result will be very problematic with such a small sample. This doesn’t mean that your data is “wrong”, only that it will be difficult to perform some of the tests with such a small sample.
      Charles

      Reply
  2. For n=1000

    My program
    2E4 samples
    ___________1%_____________5%_______
    _________0.0335__________0.0287
    _________0.0332__________0.0286
    _________0.0334__________0.0286
    _________0.0331__________0.0285
    _________0.0333__________0.0286

    2E5 samples
    ________0.0333__________ 0.0286
    ________0.0333__________0.0286

    Since Abdi 0.0327(1%), 0.0283(5%)
    I am fully satisfied in what concerns 5%, not so much for 1% significance.
    Note
    on request, I would be happy to display here the program listing,
    corrected now, in order to be cross- validated.

    Luis

    Reply
  3. _______My_____________#samples
    _______________________/size
    _n___alpha=.01____.05________________.01______.05__

    After bug corrected my values and those from Abdi [1] are in accordance
    (I will provide my routine on request)

    _____________MY______________________Abdi & Molin____1E5
    __________.01____.05_____#samples/size_.01______.05___alpha
    _50______.1450___.1245______4E6______.1457____.1246_
    _100_____.1036___.0889______________ 1026____.0888_
    _150_____.0850___.0729______________.0840____.0727_
    _200____ .0738___.0633______________ .0729____.0630_
    _250____ .0662___.0567____1E6_______ .0652____.0564_
    _300____.0605___.0519____4E5_______ .0596____.0515_
    _400____.0525___.0450_____________ .0516____.0447_
    _500____.0471___.0404_____________ .0462___ .400_

    [1]- Hervé Abdi, Paul Molin, Lilliefors/Van Soest test of normality

    Reply
  4. I apologize to all.
    Either the explicit CV´s as the fitting formulas Abdi display are correct. My critics concerning the latter are wrong. In fact, a “bug” was caught at the ratio u(j+1)/u(j) affecting the conversion normal rv into p-value.
    My soundest apologies (I feel ashamed, true)
    Forget me
    Luis

    Reply
  5. Could we have an idea how much fine work was unduly spoiled owing to the erroneous extreme CV´s Abdi formulas provide?
    It is urgent to find somebody, not “Anonymous”, influent at Papers Editors, available to warn the Scientific/Technical society about this issue . . . (I would say).

    Reply
  6. Charles

    The problem had neve been the incorrectness in the CV values
    Abdi displayed. They are correct as long I know, in the range 4(1)50.
    However, when the Authors, willing to expand the values, they provide
    the formulas at the bottom of each column (significance) they were
    absolutely ridiculous because the fitting fails soon or latter (in case for n>100)
    when we go to further values of the independent variable n.
    Did I be clear this time?

    Reply
  7. What seems to have happened is that Abdi & Morin [1], apparently following what Kolmogorov-Smirnov did, after obtained the simulated C.V.´s, 4(1)100, fitted them through a smoothing equation that they supposed it be true for whatever high size. Unfortunately, they were wrong as I could find through a Monte Carlo routine I had made myself. No harm, if we limit to “not greater than 100” the sample sizes.

    Reply
  8. In fact
    __________Simulation___Abdi&Morin
    100________.0890_______.0890__
    105________.0869_______.0868__
    110________.0851_______ .0849__
    115________.0834_______ .0830__
    120________.0819_______.0813__
    125________.0806_______.0797__
    130________.0795_______.0782__
    135________.0785_______.0767__
    140________.0775_______.0753__
    145________.0768_______.0740__
    150________.0761_______.0728__

    Reply
    • If I understand correctly, I believe that you are saying that the Simulation values are more accurate than the ones labelled Abdi&Morin. Is this correct?
      Charles

      Reply
  9. WARNING!
    Lilliefors Test for Normality
    To all Lilliefors Test users evaluating CV (95% confidence) through
    f_n = (.83+n)/√n-.01 , CV(n=100,.95)=.896/f_n
    Comparing with simulated values
    Critical Values
    n Simulation Formula
    50 .1245 .1248
    100 .0890 .0890
    150 .0761 .0728
    200 .0728 .0631
    250 .0721 .0565

    DO NOT use for sizes n>100, it goes worse and worse whith growing sizes . . .

    Reply
  10. Charles (not to be published if you think so)

    Indirect “proof”

    Through the window, at home, I cannot see the drops falling. However I observe that people their umbrellas unfolded, if they do not carry them walking unusually fast.
    Then, I assume that most likely, is raining.

    Data processing uses similar procedure, by “symptoms”, to find out the statistical procedure fitted to data treatment. Very rarely we have a deductive mathematical inference. This leads me to the conclusion “Statisticians ARE NOT Mathematicians”.

    Luis

    Reply
  11. Charles

    When a person gets old tends to turns out his beliefs to certainty, I dread. Let be the Jacob Cohen “The Earth is Round (p<0.05)” specifically of a syllogism the author found (Pollard, Richardson, 1987).

    __1(A)- If a person is an American, then is probably not a member of the Congress,
    __2(A)- This person is a member of the Congress,
    __3(A)- Therefore, he is probably not an American.
    The conclusion is risible, of course, so J. C. put in confront with the following basic clauses “validating” the NHST,

    Conclusion: J.B. states that NHST is invalid . . .

    Let´s make comparison

    (B)
    __If H0 is true then the result (statistical significance) will probably not occur,
    __This result has occurred,
    __Then H0 is probably not true and therefore formally invalid.

    I intend to present two notes:

    Note 1
    In my opinion (A) is invalid because one condition to fulfil to be a member is precisely to be an American citizen. The conclusion 3(A) is completely senseless.

    Note 2
    I feel that (B) cannot anyway to be assimilated to a Syllogism. We can get certitude from random events, (B) is simply, not more, not less common sense, I dare.

    Would you concede to put this matter under discussion? If you consider it uninteresting/noisy, please do not publish. I do understand. Thanks in advance

    Luis

    Reply
  12. Charles

    It could be interesting to you, I guess

    influentialpoints.com/…/Kolmogorov-Smirnov_test_.

    “ Both the one- and two-sample Kolmogorov-Smirnov and related tests are widely used in all disciplines. Unfortunately, the one-sample Kolmogorov-Smirnov test is commonly misused to test normality when the parameters of the normal distribution are estimated from the sample rather than specified a priori. The result is that the test is far too conservative, and distributions that are clearly not normal are wrongly classified as such*. This practice is perhaps reinforced by a sometimes unconcealed desire to demonstrate normality so that subsequent parametric tests can be carried out **. The situation is not helped by various software packages being unclear about which test is being used. The correct test to use to test for normality when the parameters of the normal distribution are estimated from the sample is Lilliefors test.*** “.

    Luis
    My notes
    * Because the Critical Values are much larger than the corresponding (size, level) ones relative to the Lilliefors Test.
    ** Psychology = Normal against abnormal
    *** The POINT!

    Reply
  13. Charles

    In preference it was wiser to ask to a genuine expert to comment your text. The only thing I am able to provide you is my opinion, which I do with pleasure.
    Though I was my duty I would eliminated the term improvement
    (I know, it appears persistently in the context). Why? In my opinion, you know, one-sample KS and Lilliefors are two different tests in spite of using the same (test) statistics, namely the maximum difference between the Normal Distribution Function and the Empirical one provided by the data.
    In preference I would stress the differences rather than similarities. Such that

    ___ KS test intends to find if data is likely to follows a normal law to which the parameters (mean, standard deviation) are given beforehand, contrarily to Lilliefors test, aiming the same goal but both parameters are estimated from data.

    Dear Charles, what I said is not but babbling, you know?

    Luis

    Reply
  14. Lilliefors test for normality

    “When the population mean and standard deviation for the One Sample Kolmogorov-Smirnov Test for Normality is estimated from the sample mean and standard deviation, the results are not very accurate. The Lilliefors Test corrects the KS Test in such cases, and so provides a much more accurate test for normality. The calculation of the test statistic is the same as for the KS test, but the Lilliefors test uses a different table of critical values.”

    ______My nightmare comes true . . .
    I will not repeat that they are different tests (with similar test statistics).

    Reply
    • Luis,

      Is it correct to say the following?

      The Lilliefors Test is an improvement over the KS Test in such cases, and so provides a much more accurate test for normality. The calculation of the test statistic is the same as for the KS test, but the Lilliefors test uses a different table of critical values.”

      Charles

      Reply
  15. Risking to be taken as a know-all guy
    (appealing to your amicable patience Charles, this is a strictly personal cry)

    Things that I hope to find no more from Statistical literature:

    ___The Lillefors Test is a Kolmogorov-Smirnov´s correction. They use the same test statistics, sure, but are based on different assumptions namely the latter suppose that the normal parameters are known, the former, contrarily, should be estimated from data.

    journals.ametsoc.org › JAM6282 › December 1975

    ___ Critical values displayed with more than 3 decimal places when simulated data derived from 100´000 samples each size or less.

    Reply
  16. Just above the table show a misprint
    _____________n=30, alpha = 0.05, Abdil & Molin cv=0.1590 (my value 0.159)

    Reply
    • Luis,
      Thanks for finding this typo. I will update the table shortly.
      I appreciate your help in making the website (and the software) more accurate.
      Charles

      Reply
  17. Charles,

    I dare . . . to suggest that people using the Table should be warned that the 4th decimal places (Abdi and Molin) strongly problematic. Do you agree?
    Note that they used 100000 samples, I two times that number, and the values varied from those I found mainly by +/- 0.001 (or less). Sure, sufficient evidence?

    Luis

    Reply
  18. Charles says:
    March 24, 2016 at 8:58 am

    Charles

    I´m very pleased by your concern towards my (unpublished) paper: #1
    and #2 are two independent trials, 100´000 samples for each size. Then I performed a LS fitting using this data, then the “adjusted” from the polynomial. Sorry I am unable to provide detailed account because the job was concluded more than 30 years ago, and sorry I do not keep any notes.
    Not harm at all. You can refresh your values through

    https://www.utdallas.edu/…/Abdi-Lillie2007-pretty.p.

    The authors found values that are very close to I did, so you can be absolutely confident about. I invite you to the link I posted:

    https://goo.gl/photos/tGo6G64qMVp5CRX79

    Luis

    Reply
    • Luis,
      The Lilliefors distribution function LDIST that is already included in the Real Statistics Resource Pack implements the procedure described in the Abdi and Molin paper that you referenced.
      Charles

      Reply
  19. Mr. Charles Zaiontz

    How could I desagree: the Table you present (with an error) is plainly the oroginal 1967 Lilleifors´one
    with “1000 or more” samples for each sample size !!! Could I?
    I am waiting you make a revision, please, using a lot larger data. . . is your job!

    Reply
  20. In complement
    As I said two series of experiments were performed. We could state that 60% of the values, for the same size and level, have the three decimals equal, 38% differ from 1 unit at the third decimal, and only one pair out of 115, do have a 2E-3 difference, one pair 3E-3.

    Reply
  21. Thank you for your reply.
    I am not contesting the results displayed under your name, at all. I ask pardon, as you can see my English is rather imperfect
    At a time (1990´s) that only original data was available we review the values not basedon “1000 or more samples by size” as in 1967, but twice 100´000, in order to obtain more reliability.
    As I could not publish my paper because the Review I sent refused to.
    Retired since 1999 I completely loss contact with this business. Days ago, under your name, I did read with joy that the matter did return to concern. Possibly my 1995 values would be possibly outdated, then unuseful.
    However if, by chance you are curious, I will send my values I found.
    Luís

    Reply
    • Thank you very much for sharing this history. It is interesting how these things have evolved. Do you know whether the results you got many years ago are more accurate than the table that I put in the website?
      Charles

      Reply
  22. Mr. Charles Zaiontz

    Dear Sir

    By the Dvoretzky-Kiefer- Wolkovitz Inequality in order to attain 3 decimal places exacts with a probability “p” or less are
    __________________p=0.05 ________________7.4 million simulated samples
    ____________________0.04________________ 7.8
    ____________________0.03________________ 8.4
    ____________________0.02________________ 9.2
    ____________________0.01_______________ 10.6

    I think that all people is aware that tables were constructed with a lot fewer data
    than that required to state all 3 decimal places.
    Would you so kind to send me your opinion about this point?

    Thanks in advance

    L. Amaral Afonso
    From Portugal, L. Amaral Afonso former senior Investigator Instituto Nacional de Engenharia e Tecnologia Industrial, Lisbon.

    Reply
    • Thank you for your observation. Based on your comment, I believe that you are saying that the table included on the website may not be accurate to 3 decimal places (please correctly me if this was not your intention). If you are aware of a table which is more accurate, I would welcome hearing about this.
      I have no knowledge about this issue, and in fact don’t use the Lilliefors test. I included this table because some users have asked for it.
      Charles

      Reply

Leave a Comment