19

I have a table with four groups (4 BMI groups) as the independent variable (factor). I have a dependent variable that is "percent mother smoking in pregnancy".

Is it permissible to use ANOVA for this or do I have to use chi-square or some other test?

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
drew
  • 191
  • 1
  • 1
  • 3

4 Answers4

23

There is a difference between having a binary variable as your dependent variable and having a proportion as your dependent variable.

  • Binary dependent variable:

    • This sounds like what you have. (i.e., each mother either smoked or she did not smoke)
    • In this case I would not use ANOVA. Logistic regression with some form of coding (perhaps dummy coding) for the categorical predictor variable is the obvious choice if you are conceptualising the binary variable as the dependent variable (otherwise you could do chi-square).

  • Proportion as dependent variable:

    • This does not sound like what you have. (i.e., you don't have data on the proportion of total waking time that a mother was smoking during pregnancy in a sample of smoking pregnant women).
    • In this case, ANOVA and standard linear model approaches in general may or may not be reasonable for your purposes. See @Ben Bolker's answer for a discussion of the issues.
Jeromy Anglim
  • 42,044
  • 23
  • 146
  • 250
  • For a binary dependent variable, in the case that I only have summary data for the binary proportions (ie # in the A, B, and C groups, and the # of successes in the A, B, and C group), and not the actual raw data, how can we go about using logistic regression? I am only familiar with using it with the raw data. – Bryan Jan 09 '15 at 05:39
19

It depends on how close the responses within different groups are to 0 or 100%. If there are a lot of extreme values (i.e. many values piled up on 0 or 100%) this will be difficult. (If you don't know the "denominators", i.e. the numbers of subjects from which the percentages are calculated, then you can't use contingency table approaches anyway.) If the values within groups are more reasonable, then you can transform the response variable (e.g. classical arcsine-square-root or perhaps logit transform). There are a variety of graphical (preferred) and null-hypothesis testing (less preferred) approaches for deciding whether your transformed data meet the assumptions of ANOVA adequately (homogeneity of variance and normality, the former more important than the latter). Graphical tests: boxplots (homogeneity of variance) and Q-Q plots (normality) [the latter should be done within groups, or on residuals]. Null-hypothesis tests: e.g. Bartlett or Fligner test (homogeneity of variance), Shapiro-Wilk, Jarque-Bera, etc.

Ben Bolker
  • 34,308
  • 2
  • 93
  • 126
11

You need to have the raw data, so that the response variable is 0/1 (not smoke, smoke). Then you can use binary logistic regression. It is not correct to group BMI into intervals. The cutpoints are not correct, probably don't exist, and you are not officially testing whether BMI is associated with smoking. You are currently testing whether BMI with much of its information discarded is associated with smoking. You'll find that especially the outer BMI intervals are quite heterogeneous.

Frank Harrell
  • 74,029
  • 5
  • 148
  • 322
  • 2
    @Frank - why is it "not correct" to group BMI? this seems perfectly reasonable, so long as the results are appropriately interpreted. You could well be testing, for example, whether being "underweight" "healthy weight" "overweight" and "obese" are associated with smoking, where these terms are defined by the ranges of BMI. I see no "wrong" here. – probabilityislogic May 29 '11 at 13:42
  • I believe that the OP is working with a common instructional data set and may not have the raw BMI. While it's generally not ideal to discretize continuous regressors it's not "incorrect". It can even be helpful to resort to this when we suspect the measurements are noisy and there's no other recourse. Indeed, the real hypothesis we'd want to test is whether obesity is related to smoking; BMI is just one way to measure obesity (and has its problems from what I understand). – JMS May 29 '11 at 15:22
  • 4
    Even when measurements are noisy, analyzing variables as continuous is superior. Categorizing BMI creates more problems than different choices of analysis can fix. In fact the estimates upon categorization no longer have a scientific interpretation. A scientific quantity is one having meaning outside the current experiment. You'll find that group estimates (e.g., log odds that Y=1 for high vs low intervals of X) are functions of the entire set of observed BMIs. For example, if you were to add more extremely high or extremely low BMIs to the sample, the "effects" would get stronger. – Frank Harrell May 29 '11 at 19:20
  • For those who have installed R and RStudio, an interactive demonstration may be found at http://biostat.mc.vanderbilt.edu/BioMod - see the green NEW marking. You have to load the script into RStudio and also install the Hmisc package. – Frank Harrell May 29 '11 at 22:45
  • "Even when measurements are noisy, analyzing variables as continuous is superior" This is just incorrect (the generality of it, that is - usually it's true). Imagine you have a continuous covariate where the error in its measurement increases with its magnitude, for example. Of course the best thing to do is model the error, or get better measurements, etc. But to say that it's incorrect is simply too strong a statement to make. – JMS May 30 '11 at 15:47
  • Also: "For example, if you were to add more extremely high or extremely low BMIs to the sample, the "effects" would get stronger." That's a problem with your sample, not the method. Or I guess you could cast it as a problem with the method in that it might require larger samples. And I still don't understand why categories based on the BMI are all of a sudden meaningless and invalid - it's exactly the same quantity, just measured at a different resolution. Whether those categories are meaningful is a separate question entirely. – JMS May 30 '11 at 15:51
  • What makes categorization bad when there is measurement error is the high likelihood of classifying to the wrong category, which can be a 100% error for the binary case. I couldn't agree less with the second comment. This is not a problem with the sample; it is a problem of incomplete conditioning (on X > c rather than X = x). – Frank Harrell May 30 '11 at 17:17
  • Let me add that if you want to analyze BMI (which may not even be the right summary of ht and wt for the problem at hand) then by all means analyze BMI. But analysis of categorized BMI is not an analysis of BMI. [As an aside people don't even agree on the category boundaries.] One could say that analyzing BMI categories is analyzing a crude approximation of BMI when the full-information variable was available for analysis. The result is residual confounding, badness of fit (predictions will be inaccurate when BMI is not near the middle of an interval), and artifactual discontinuities. – Frank Harrell May 30 '11 at 17:42
  • "What makes categorization bad when there is measurement error is the high likelihood of classifying to the wrong category, which can be a 100% error for the binary case." Usually, sure. But think about the example I gave where measurement error depends on the magnitude of $X$; you could imagine that discretizing by sign(X) would give you conservative inferences in that superfluous high-leverage points are "damped". It might be the best of two bad options. I don't mean to disagree with the thrust of your argument, just the generalization to all possible cases - I'm willing to accept that... – JMS May 30 '11 at 19:02
  • ...discretizing BMI is a bad idea, since you're clearly more familiar with the subject matter than I am. Your point about incomplete conditioning is a good one - offhand I think I had your point backwards in my head. – JMS May 30 '11 at 19:22
  • "But analysis of categorized BMI is not an analysis of BMI." This, I think, we all agree on (my earlier sloppy comment aside). Maybe I've misinterpreted what you have been saying. – JMS May 30 '11 at 19:26
  • Thanks for the good discussion. Yes, if measurement error is not independent of X or Y, I may be in trouble. But categorization is very sensitive to the cutpoint in that case also. One other point: for a given number of parameters (regression degrees of freedom; no. of dummy variables in the categorical case), smooth regression fits explain a higher proportion of variance in Y than piecewise flat relationships (categorizing X). I think the interactive simulation I posted above may help explore some of the issues (although it assumes completely random measurement error). – Frank Harrell May 30 '11 at 19:33
4

If you choose to do an ordinary ANOVA on proportional data, it is crucial to verify the assumption of homogeneous error variances. If (as is common with percentage data), the error variances are not constant, a more realistic alternative is to try beta regression, which can account for this heteroscedasticity in the model. Here is a paper discussing various alternative ways of dealing with a response variable that is a percentage or proportion: http://www.ime.usp.br/~sferrari/beta.pdf

If you use R, the package betareg may be useful.

Will Townes
  • 333
  • 1
  • 7