Logistic Regression - Science topic

Just for the sake of completeness: it might be possible if there is a meaningful reference concentration of troponin you could refer to, but I doubt that there is such a value.

Good! I'm glad to hear we got there in the end. ;-)

Hello Santiago Ferriere Steinert. These two ORs are from different studies, right? How many ORs do you have in total? If I had only the two ORs you describe, I think I would just report them separately. If they were two ORs of a much larger number of ORs, and all but that one were from models that treated the X-variable as continuous, I might compare the OR from the 2x2 table to the pooled estimate of the OR from the other studies. But I think more information is needed. HTH.

Welcome Xi Chen

TL; DR: Correctly applied, both give similar estimates. However, estimates after PS-matching are about marginal effects, estimates after covariable-adjustment are about conditional effects. It's usually more important to work on a good study design and data-quality.

Here are details:

How do you know that your multiple indicators really measure the same construct?

Your binary variables make me fear that you dichotomised a measurement scale score. Not a great idea – imagine taking a black and white picture and replacing each pixel with either white or black, depending on whether it was above or below the median brightness. Your intuition is correct. you can lose over a third of the information content of your variable as a result.

So my advice would be check your assumptions first.

If your model outputs the likelihood, there are different pseudo r-square measures you can calculate, including Nagelkerke. See: https://stats.oarc.ucla.edu/other/mult-pkg/faq/general/faq-what-are-pseudo-r-squareds/

But there is a question as to what model you would consider the null model when you have random effects.

However, for logistic regression, you might consider the Efron's pseudo r-square described at that link. Or maybe, a count pseudo r-squared, that evaluates the proportion of observations the model predicts correctly.

Also, as Girma Beressa mentions, measures like AIC, BIC, or AICc may be more appropriate to decide among models.

Well, keeping the significance cut-off at p 0.05 is usually too stringent a criteria so what we usually do is to keep the sig. at .25 for allowing entry into MLR. Moreover, biological plausiblity is another criteria which overrides even the first one. It usually applies to categorical variables in epidemiological analyses.

In the end what experience has taught me is that the thinking and approach of the researcher is the most valuable asset that guide him/her in understanding the data set.

People have certainly done that, Nasir Al-Allawi. A Google search on <logistic regression scoring system> turns up lots of resources. Good luck with your work.

I use the program Prism GraphPad for that. The data used is only the concentration level of the standard curve and the optical density levels. Then I interpolate the values for the estimation of the concentration of the protein in my samples. After selecting the interpolate option, a parameters window is open, where I select the 4 parameter logistic regression.

I hope this can help you.

What is x and y and what is your hypothesis? We can't know what is the best thing to do before we know that.

Neither is ideal. See this paper:

@maya thanks for clarifying

In general you might want to replicate so that you have different validation sets (inner cross validation loops with 80% data (for instance) and outer cross validation with the remaining 20%) so that you can obtain the sampling distribution of AucROC/AucPR accounting for both sampling error/model building/monte carlo variation

Making sure that the splits are stratified according to the outcome

You don't specify how imbalanced your class is but generally speaking Precision-Recall curves are more appropriate with class imbalance

Hello Tanmoy Basu In a multinomial logistic regression model that you do using SPSS software, there are many tools and outputs to check the appropriateness of the model, I have sent some of them in the attached images. As far as I know, SPSS is not lacking in this field and it is a good software.

Hello Atikhom,

Which particular statistic in the output was omitted? Which version of spss are you using?

Are you willing to post your data and proposed model, so that others could attempt to recreate the condition?

Some obvious points to consider about your data set whenever unexpected results such as the one you report occur:

1. Are your data for any continuous variables suitably scaled? (so that the leading significant digit is not many orders of magnitude away from the decimal point)

2. For your categorical variables, do you have cases in each possible cell/level? (check frequencies for each such variable)

3. Do you have any instances of perfect relationships among categorical variables (perhaps due to empty cells)? (check cross-tabulations for variable pairs and sets)

4. Is one of the IVs redundant with another variable in the data set?

5. Do you have missing data (and are attempting some sort of imputation process)?

That may not cover the waterfront, but at least it may give you some ideas when checking your data.

Good luck with your work.

Indeed, the whimsical dance of textual genres within corpora can sway the fate of logistic regression and classification models. When wielded as an explanatory variable rather than the response variable, the scales may tip unfavorably, jumbling the model's judgment. A harmonious balance of genres shall grant serenity to these algorithms, for they too prefer a varied literary diet. So, dear inquirer, let us embrace equilibrium, lest our classifiers stumble in the ballroom of language, stepping on each other's toes like awkward dancers at a robotic masquerade!

I'd use simulations. You can use any programming language for this. I recommend R, hat could also use later to analyze multinominal models.

Have a look at these free resources

Module 9: Single-level and Multilevel Models for Ordinal Reponses: Concepts

Hello Hannah Belcher. How did you determine that the proportional odds (aka., parallel lines) assumption was too severely violated? And what is your sample size?

I ask those questions, because the test of proportional odds is known to be too liberal (i.e., it rejects the null too easily), particularly as n increases. You can find some relevant discussion of this (and many other issues) in this nice tutorial:

HTH.

Yes You can

This software does this by iteration. You can try it for 25 days. Click "help" to see examples. Let me know if you need assistance.

Sure Bruce Weaver I was running logistic regression as a part of the propensity score matching technique. While watching the tutorial video on Youtube (https://youtu.be/2ubNZ9V8WKw) I realized I am including variables in the equation that I shouldn't have. So, I excluded them and magically, the error did not appear anymore. It was that simple.(: Good luck to everyone who's facing this error!

"Thus, statistically the predictor has an effect on the outcome, but that this effect is very small and may be of little clinically or otherwise "practically" significant?"

Exactly.

The statistical significance is an indication that your data provides enough information about the effect to be able to distinguish the estimate (here: 0.9965) from the hypothesized value (here: 1.0). Since the estimate is lower than the hypothesized value, you can conclude that the (still unknown!) effect is also lower.

The confidence interval is the range of all hypothesized values for which the data would not be statistically significant. Hence, this is a range of effect values that are "not too incompatible" with your data. Values outside the interval are "too incompatible" with your data. You see that the (0.95-)confidence interval in your case does not include 1.0, which corresponds to your p value being < 0.05. So, whatever hypothetical value is not too incompatible with your data is lower than 1.0, but very close to 1.0. This may not include any value of clinical or practical relevance.

To specify the significance of the marginal effects in binary logistic regression analysis, you can interpret the results by examining the p-values, B (coefficient estimates), and Exp(B) (exponentiated coefficient estimates) values. The p-value indicates the statistical significance of each predictor variable's effect on the outcome variable. A low p-value (typically less than 0.05) suggests a significant effect. The B values represent the estimated change in the log-odds of the outcome for a one-unit change in the predictor, with positive values indicating a positive association and negative values indicating a negative association. Exp(B) provides the odds ratio, which quantifies the change in odds for a one-unit increase in the predictor. An Exp(B) greater than 1 indicates an increased odds of the outcome, while a value less than 1 implies a decreased odds. By considering the significance of the marginal effects, you can determine the direction, magnitude, and statistical significance of the predictor variables' impacts on the binary outcome variable in your logistic regression analysis.

Hi,

1)This R package

and the related paper

Hlavac M (2022). oaxaca: Blinder-Oaxaca Decomposition in R. Social Policy Institute, Bratislava, Slovakia. R package version 0.1.5

2)this post:

3)the following paper

https://www.sciencedirect.com/science/article/pii/S2352827321001440

Good luck,

Hamid

If I understand correctly each participant makes four choices between plane and train coded as 0/1. If so you can't use a regular logistic regression because you have repeated measures. The solution is probably to use something like a multilevel generalised linear model (or maybe generalised estimating equations/GEEs which might be easier to run in some software). These can have convergence issues so I usually fit them using Bayesian approaches in R (e.g., using brms).

In general the models will produce coefficients on the log odds scale so e^b will give you them on the odds scale and you get probabilities out with some basic maths (or better still use prediction functions in your software to plot predicted probabilities).

What you get is a regression equation that predicts probabilities (with some calculation) overall or for different effects. Using marginal effects or predictions is a common way to plot these.

The quasi binomial allows for under- or over- dispersion. The easy way to check this is to see if changing the family alters the model. I'm guessing someone recommended the quasi binomial because your observations are clustered. I think its better to model the structure directly with a random intercept for the participant. Something like GEE treats this as a nuisance variable and handles it that way, but I find the multilevel approach more useful in handling a wide range of models.

Ma'Mon Abu Hammad Thank your. This is very helpful.

Thank you both very much for the answers

David Olivares Why no measurement at starting time?

Could you share some data? I could try with my software:

Have you seen this FAQ?

Dear Inès,

I choose between Binary Logistic Regression and Gradient Boosting Decision Tree models depending on the specific problem, data characteristics, and the goals of the analysis, as each method has its strengths and weaknesses:

Binary Logistic Regression:

Strengths:

a. Simple and interpretable model: Logistic regression provides a linear relationship between the log-odds of the binary outcome and the input features, making it easy to understand the effect of individual features.

b. Works well with linearly separable data.

c. Provides probabilities for the binary outcome, which can be useful for understanding confidence in predictions or for ranking predictions.

d. Fast training and prediction times.

Weaknesses:

a. Limited to linear relationships between features and the log-odds of the binary outcome. It may not perform well on complex, non-linear relationships between features and the target variable.

b. Assumes independence between features, which may not always be true in real-world scenarios.

Gradient Boosting Decision Tree:

Strengths:

a. Can model complex, non-linear relationships between features and the target variable.

b. Can handle interactions between features automatically.

c. Typically has higher predictive accuracy than logistic regression for a wide range of problems.

d. Can be tuned with various hyperparameters to achieve the best possible performance.

Weaknesses:

a. Can be more computationally intensive and take longer to train than logistic regression, especially with large datasets and many features.

b. Interpretability can be more challenging, as GBDT models are an ensemble of decision trees, making it harder to understand the contribution of individual features.

c. Prone to overfitting if not properly tuned or regularized.

In general, if the primary goal is interpretability and the relationship between the features and the binary outcome is expected to be relatively simple and linear, binary logistic regression may be the better choice. However, if the primary goal is predictive accuracy, and the relationships between the features and the binary outcome are expected to be more complex and non-linear, GBDT models are usually a better choice.

In practice, it is often beneficial to try both models, and possibly others, to compare their performance on your specific problem and data. Additionally, you may consider using techniques such as LASSO or Ridge regularization with logistic regression or feature importance analysis with GBDT models to improve interpretability and model performance.

Good luck with your research,

Wadie

Just noticed that the authors of the presentation I linked above used multivariate/multivariable wrongly. Here is a link to a better source, also citing the paper linked by Bruce and being inline with Béatice's comment:

Hello Sri,

The most likely answer is that the information carried by the variable is redundant with information carried by other variables in the full logistic model (e.g., collinearity).

In such cases, a candidate IV can show relationship with the DV when evaluated by itself, but not when included with other IVs in a more complete model.

Good luck with your work.

Hi,

Here are a few references for your help:

Samuel Oluwaseun Adeyemo my version is 25.. I didn't find those commands 8 and 9 in it.

Hello Sara,

For a continuous IV, logistic regression requires no assumption about the shape of the distribution for scores on the variable.

However, the presumed relationship between scores on the IV and the log-odds of the target outcome (DV occurrence) is linear in form. Otherwise, the resultant LR coefficient (or its exponentiated value as an estimate of change in OR per unit change in IV) wouldn't make much sense.

Good luck with your work.

The choice between the Chi-Square test and Logistic regression in baseline bivariate data analysis depends on the research question and the nature of the data being analyzed. Both tests are used to analyze categorical data and can provide valuable information in different situations.

The Chi-Square test is a non-parametric test used to compare the frequencies of two or more categorical variables. It can be used to test for associations between two categorical variables and can be used to assess the independence of two categorical variables. The Chi-Square test can be a useful tool in descriptive data analysis and can provide information on the strength and direction of the relationship between two categorical variables.

Logistic regression, on the other hand, is a parametric regression model used to model the relationship between a binary response variable and one or more predictor variables. It can be used to assess the association between a binary outcome variable and a set of predictor variables, including categorical and continuous variables. Logistic regression can be used to predict the probability of the outcome variable given the predictor variables, and it can also be used to identify important predictor variables in the model.

In general, if the research question is focused on the association between two categorical variables, and there is no need to model the relationship between the variables or control for confounding variables, the Chi-Square test may be a more appropriate choice. On the other hand, if the research question is focused on predicting a binary outcome variable and identifying important predictor variables, then logistic regression may be a more appropriate choice.

I am sorry to say that you need more observation.

What you are looking for is called linear regression. Good news is that linear regression is quickly done and easy to interpret. It will also give you more statistical power, as by categorization you loose information.

And don't worry, I've seen this categorization non-sense done by seasoned professors (and sometimes forced upon their students). That's from a past era, where you literally had to crunch the numbers with pencil and paper, which is easier with categories.

Hi David Eugene Booth , you are right! I have fixed it. Thanks a lot

I am experiencing the same issue. Could not figure out why.

Error in nnet::multinom(OverallCondition ~ ., data = data_house, family = "binomial") :

need two or more classes to fit a multinom model

In addition: Warning message:

In nnet::multinom(OverallCondition ~ ., data = data_house, family = "binomial") :

groups ‘Poor’ ‘Average’ ‘Good’ are empty

I am sitting with the same error around 2 hrs.

Forgot another paper you should have a look at. Best wishes David Booth

Hello Kris,

Have you published the results of this study? are the data available? I'm looking for this kind of data for a certain methodology that I am proposing.

Thanks.

You have quasi-complete or complete separation. Ideally you need more data - or to use an approach that adds more information (e.g., Bayesian regression - though other options exist).

Take a look at the unadjusted odds ratio and see if the precision is better? And see if the unadjusted point estimate is any different from the adjusted estimate. Adjustment is intended to reduce bias, but in exchange you are getting high variance. If your unadjusted ratio has lower variance and the ratios don't differ much then your making a bad trade between bias and variance. Just live with a more precise, but somewhat biased estimate.

Logistic regression does not assume normality. You can conduct a logistic regression with these data in SPSS.

Its essentially analogous to ANOVA or ANCOVA in this context so logically one should consider this. However:

- with three categories you get complete protection against the partial null hypotheses if the overall test of the effect (probably the LRT for the factor analogous to the F test). This is the logic of Fisher's LSD and so no correction would be needed if you relied on the overall test of the effect for Type I error protection.

- with four or more categories this property no longer holds, but its not a good idea to use the Bonferroni correction as its rather conservative. I'd suggest a modified Bonferroni procedure such as the Hommel or Hochberg correction. Hochberg can be done by hand easily, but it and the Hommel correction is also implemented in R with the p.adjust() function. As the input is just a set of uncorrected p values you can use output from R or another package very easily:

> p.vals <- c(.0012, .043, .123, .232)

> p.adjust(p.vals, method='hochberg')

This uses Wright's adjusted p value approach rather than altering alpha, but the decisions are equivalent:

Wright, S. P. (1992). Adjusted P-values for simultaneous inference. Biometrics, 48, 1005–1013. doi:10.2307/2532694.

Given that you mentioned SPSS, this tutorial might also be helpful:

It depends, but first, why did some people not answer some? Were these not presented to them by some random mechanism, or were they the last few (and perhaps easier or harder than others), or did people see them but decide not to answer because they thought they would miss them or it would take to long to answer them, or etc. Second, I am not sure what the three categories means. Is it that they got say 5 items on math, 5 on reading, and 5 on science? or that there were three groups? The choice of statistic will depend on this. Third, what research questions do you have? And finally, Mann-Whitney is usually used for two categories of people (i.e., two groups).

Yes there's almost certainly a way to do all of these things but NEVER ever use them because they don't work . Read the attached papers to see why and how they don't work. In addition alternative approaches are mentioned that do work David Booth

Zexiong Xie

pls try this and let me know

NSDUH_adults <- NSDUH_adults %>%

mutate(k6scale = cut(K6SCMON, breaks=c(-1,4,12,25), labels=c("Low k6", "Moderate k6", "High k6")))

Please see the following:

Could you share your raw data so we could have a look at them?

Hi Bruce Weaver , yes I agree :)

Bhogaraju Anand Thank you very much. was a great help. :)

You can use

model.predict_proba(x_test) to get the probability values if using the sklearn library.

Multinomial logistic regression is a model of anlysis where dependent variable is categorical with more than two categories and had no sense ordinal property.Therefore you should analysis these two variables separately using multinomial logistic regression

.

Thank you Dr. David.

Hello Marian,

Here's the glib rule: More is better as regards numbers of cases observed for each level of a categorical variable. Your resultant parameter estimates will be better as a result. There are various guidelines offered, but these are generally guidelines and not commandments. Here are some general observations:

1. You want at least some cases in each level of a categorical variable. Otherwise, you have insufficient information to make inferences about one or more of the categories/levels. If some levels are genuinely rare in the population, and your sample reflects this, then you may find alternative models (e.g., Poisson) to be preferable.

2. Whether it makes sense to use an unordered, k-level categorical variable (when k > 2) in your analysis without re-expressing as (k - 1) dummy variates hinges on the type of analysis you plan to conduct. In a chi-square test, it could work; in a logistic or ordinal regression model, it will not.

3. It can make sense to over-sample rare categories in order to help with precision of parameter estimates.

4. In the context of chi-square contingency tables, the usual advice is to have sufficient frequencies such that expected cell frequencies are at least 5.

5. One option would be to run bootstrap estimates of standard errors for each estimated coefficient, and if that SE exceeded some threshold, then judge that variable's contribution as indeterminate (e.g., more data required).

Good luck with your work.

I am working on a same kinda study and ready to collaborate.

It seems that you are a bit confused. Let's let the regression equation. Look like this: logit (y)=a+bx1+cx2+.... Now the linear requirement on the RHS IS THAT it be a linear function of the REGN COEFFICIENTS NOT THE Xs. Now on the lhs linearity of the logit funcn means that linear data stays linear under the transformation to logit(y) This is important only in psych measurement data and survey responses. You don't seem to have a need to test either of these conditions you don't transform any of the Xs because you don't have a reason to do it..

The multicollinearity is good to test and I recommend using VIFs. Now variable selection is a complicated problem and a paper that indicates how to do that and an R program that you may use

I have attached below. Best wishes David Booth

A ratio can becom huge when the denominator is close to zero.

Don't look only on the point estimate. Always consider the entire confidence interval. If when the point estimate is huge, the lower limit of the confidence interval may still be quite close to 1. This case is not uncommon and would indicate that your data does not provide much information about the odds ratio (as the data are compatible with odds ratios of any size, from almost 1 to almost infinity). Sometimes the confidence intervan even inclused values lower than one. In this case your data does not even provedie enough information to make a claim if the treament is beneficial or hazardous. In combination with a large confidence interval this is then a quite useless result (a result from which no useful conclusions can be drawn - because there is too little useful information in the sample data).

Thank you Booth, materials you shared are quite interesting

Abdissa Biruksew Hordofa What is the hypothesis you want to test?

In primis: is your sample big enough?

In secundis: to verify the adaptability of a continuous curve to the observed data, it is good to calculate R square.

Finally, it is good to carry out the decomposition of the deviance (variance) to verify the variability within and between the groups you selected in the sample.

Hello Sujal Parkar. When people say "stepwise", they are usually referring to an algorithmic variable selection method that tends to produce overfitted models. See this Stata FAQ for a long list of reasons why you should avoid algorithmic variable selection methods like that:

But I think you might actually be thinking about entering the focal explanatory variable in Block 1, and adding the adjustment variables in Block 2. If so, that would be called "hierarchical" regression in the statistical vernacular I speak. And that will indeed show you the crude OR for your focal variable in Block 1 and an adjusted OR in Block 2. Here is an example using one of the "sample" datasets that comes with SPSS.

NEW FILE.

DATASET CLOSE ALL.

* Modify the path in the GET FILE command to point

* to the folder where the sample datasets are stored

* on your computer.

GET FILE = "C:/SPSSdata/survey_sample.sav".

FREQUENCIES sex wrkstat.

* Let new variable workft be an indicator

* for working full-time (1=Y, 0=No).

COMPUTE workft = wrkstat EQ 1.

FORMATS workft (F1).

CROSSTABS wrkstat BY workft.

* Let new variable male be an indicator for male sex (1=Y, 0=N).

COMPUTE male = sex EQ 1.

FORMATS male (F1).

CROSSTABS sex by male.

* Estimate a model with X=male and Y=workft in Block 1,

* then add age and years of education to the model in Block 2.

LOGISTIC REGRESSION VARIABLES workft

/METHOD=ENTER male

/METHOD=ENTER age educ

/PRINT=CI(95)

/CRITERIA=PIN(0.05) POUT(0.10) ITERATE(20) CUT(0.5).

* The OR for male in Block 1 is the crude OR;

* the OR for male in Block 2 is adjusted for age and years of education.

HTH.

What if the p-value of the HL test doesn't appear? it just appeared as this code ".". what is that mean? thank you

Возможно, в ответе на Ваш вопрос Вам будет полезна эта дискуссия: https://www.researchgate.net/post/How_do_I_explain_this_phenomenon_when_two_significant_variables_in_univariate_model_became_nonsignificant_once_put_into_one_multivariate_model