Recent content by hlsmith

  1. hlsmith

    Help analyzing nonnormal data

    How can a Kruskal Wallis test - "didn't look good"? Based on what criteria?
  2. hlsmith

    Bernoulli distribution

    Can you give an example of what you are thinking about - say like the existence of a black swan when one has never been seen?
  3. hlsmith

    What is teh difference between a linear regression and a correlation between 2 variables?

    We have no idea what kind of data you are working with. By predict - what do you mean, just estimating the y-hat for new data, which would be called "scoring". Also, take a look at the terms "interpolation" and "extrapolation" and reply back. Thanks.
  4. hlsmith

    correlations

    If you just look at Aces and Games won, you almost get into an ecologic paradox issue. Meaning you are correlating two macro numbers, but not linking them directly. So player A had 10 aces in a game they lost and 0 in a game they won, but on average given all the games you find a spurious...
  5. hlsmith

    comparing test results between subgroups with those between main groups

    I was thinking of something closer to a forest plot. I am not extremely versed in multilevel models, but it seems people throw out numbers like you need at least 40-60 clusters, which you seem to have two clusters for As and 1 for B. I can't seem to process how it would fit your data, but I may...
  6. hlsmith

    Meta-analysis using revman

    I haven't really used revman, but I am curious what "quadas" represents.
  7. hlsmith

    Is it permissible to aggregate results from multiple cross sectional regressions?

    Well, I will start off by asking why are you doing this and what is the purpose? Second, significance is always a poor metric. You could have a pvalue of 0.0501 and you are saying no effect, when it could just be a statistical power issue for that month. Have you considered time series...
  8. hlsmith

    Help analyzing nonnormal data

    Can you provide a qqplot and scatterplot of the residuals? Could the outcome be better defined as a negative binomial regression or other count model. the reason I asked about controlling for other variables was that the kruskal wallis test is another option.
  9. hlsmith

    Help analyzing nonnormal data

    What is your sample size? Do you need to control for other variables? Tell us more about the context if possible.
  10. hlsmith

    comparing test results between subgroups with those between main groups

    Agreed, I was just getting a feeling for what you had done so far. I am not sure the project's context, but the use of pvalues is considered a faux pas these days and may be apparent in your setting. With two sided tests, you could have a test statistic that completely flips directions though...
  11. hlsmith

    comparing test results between subgroups with those between main groups

    Did you make corrections for false discovery?
  12. hlsmith

    How to get the odds of an event, based on its linear regression analisy?

    Performance is likely your variable of interest, dependent variable. Graph shows performance is expected to decrease as anxiety increases and there is a sex difference. Meaning males have lower performance on average. You may also have the makings of Simpson's paradox here as well.
  13. hlsmith

    Logistic Regression help

    Depends on which outcome group you are predicting. In general, the bigger the absolute value the bigger the estimate for the outcome or against. Negative value results in odds ratios below 1 and if positive above 1, you should be looking at the coefficients.
  14. hlsmith

    Binary logistic regression

    Well you don't need to use the 0.5 probability cutoff for outcome classification. You can try different ones based on revising confusion matrices. You could also split predicted probabilities into 3 groups, yes, no, maybe.
  15. hlsmith

    Possible dumb question

    So the engineer is finding the SD for a single test across time, that is how it read to me. Or is it the SD of multiple tests for that time point, which seems more appropriate. my suggestion would be to collect data from multiple test and fit a model. It sounds like the rate is not constant, so...