The input is the same as for the other variables, with (1) and without (0) inputs.

　　

------------------------------
TS016685054
Megumi IZUMIOSAWA（JAPAN）
17-JUL-24
------------------------------

Hi @恵 泉澤

I looked at your data and I see that for each level of your  "NumberofDrugs" variable, the "PolyPharmacy" variable is a constant.  So there is no variation (in "PolyPharmacy") at each stratum.

This is why the COXREG procedure was ignoring the "PolyPharmacy" variable.  If you drop that variable from the analysis, you get the exact same results (minus the footnotes that concerned you).

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

Hi @恵 泉澤

I looked at your data and I see that for each level of your  "NumberofDrugs" variable, the "PolyPharmacy" variable is a constant.  So there is no variation (in "PolyPharmacy") at each stratum.

This is why the COXREG procedure was ignoring the "PolyPharmacy" variable.  If you drop that variable from the analysis, you get the exact same results (minus the footnotes that concerned you).

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

Hi @恵 泉澤

I will readily defer to the actual Statisticians in this Community.  Here is my understanding:

• You are asserting the "PolyPharmacy" variable is the dependent variable, yet you are treating it as a covariate.  You are treating the "Outcome" variable (0=survive, 1=death) as your dependent; and that seems appropriate for Cox Regression.
  
• As a covariate, the "PolyPharmacy" variable contributes no new information to the analysis.  "PolyPharmacy is linearly dependent upon "NumberofDrugs".  That is,  if you know the value of "NumberofDrugs" then you already know the value of "PolyPharmacy"
• Literally, PolyPharmacy is equal to the condition (NumberofDrugs > 4). When that condition is false, Polypharmacy=0.  When that condition is true, PolyPharmacy=1.  The presence of the "NumberofDrugs" variable contributes exactly the same information to the analysis as having both "NumberofDrugs" and "PolyPharmacy" in the analysis.

Hi @恵 泉澤

I looked at your data and I see that for each level of your  "NumberofDrugs" variable, the "PolyPharmacy" variable is a constant.  So there is no variation (in "PolyPharmacy") at each stratum.

This is why the COXREG procedure was ignoring the "PolyPharmacy" variable.  If you drop that variable from the analysis, you get the exact same results (minus the footnotes that concerned you).

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

Hi,

Adding more "ones" to the polypharmacy column fixed the problem with the "polypharmacy" variable not generating results in your Cox regression analysis. The issue you're encountering suggests that the problem is likely related to the number of events (e.g., deaths) in the categories of this variable. Some possible explanations are,

• Sparse Data: If the initial number of events in the polypharmacy group is too low, the Cox regression model might not have enough information to estimate the hazard ratio reliably. By adding more "ones," you increase the number of events, allowing the model to generate a result.

• Complete Separation: Complete separation occurs when the outcome variable perfectly separates the groups defined by the predictor variable. For example, if all patients with polypharmacy survived and all patients without polypharmacy died, the model cannot estimate the effect of polypharmacy because there is no variability in the outcome within each group. By adding more "ones," you introduce more variability, which helps the model estimate the hazard ratio.

• Event Distribution: Cox regression relies on the distribution of events across the levels of a covariate. If the distribution is heavily skewed (e.g., almost all events occur in one group), the model might struggle to estimate hazard ratios for less frequent groups. Adding more cases can balance the distribution.

• Power and Sample Size: Statistical power and sample size are crucial for reliable estimates in regression models. If the polypharmacy group initially had too few cases, the model might lack the power to detect an effect. Increasing the sample size improves the power and the reliability of the estimates.
  
  
  

------------------------------
Aruna Saraswathy
Statistician
SPSS Statistics
IBM
------------------------------
-------------------------------------------
Original Message:
Sent: Wed July 17, 2024 03:01 AM
From: 恵 泉澤
Subject: TS016685054：I don't understand the COX regression analysis results -> why the polypharmacy variable does not come up with a hazard ratio.

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

------------------------------
TS016685054
Megumi IZUMIOSAWA（JAPAN）
17-JUL-24
------------------------------

Hi,

Adding more "ones" to the polypharmacy column fixed the problem with the "polypharmacy" variable not generating results in your Cox regression analysis. The issue you're encountering suggests that the problem is likely related to the number of events (e.g., deaths) in the categories of this variable. Some possible explanations are,

• Sparse Data: If the initial number of events in the polypharmacy group is too low, the Cox regression model might not have enough information to estimate the hazard ratio reliably. By adding more "ones," you increase the number of events, allowing the model to generate a result.

• Complete Separation: Complete separation occurs when the outcome variable perfectly separates the groups defined by the predictor variable. For example, if all patients with polypharmacy survived and all patients without polypharmacy died, the model cannot estimate the effect of polypharmacy because there is no variability in the outcome within each group. By adding more "ones," you introduce more variability, which helps the model estimate the hazard ratio.

• Event Distribution: Cox regression relies on the distribution of events across the levels of a covariate. If the distribution is heavily skewed (e.g., almost all events occur in one group), the model might struggle to estimate hazard ratios for less frequent groups. Adding more cases can balance the distribution.

• Power and Sample Size: Statistical power and sample size are crucial for reliable estimates in regression models. If the polypharmacy group initially had too few cases, the model might lack the power to detect an effect. Increasing the sample size improves the power and the reliability of the estimates.
  
  
  

------------------------------
Aruna Saraswathy
Statistician
SPSS Statistics
IBM
------------------------------
-------------------------------------------
Original Message:
Sent: Wed July 17, 2024 03:01 AM
From: 恵 泉澤
Subject: TS016685054：I don't understand the COX regression analysis results -> why the polypharmacy variable does not come up with a hazard ratio.

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

------------------------------
TS016685054
Megumi IZUMIOSAWA（JAPAN）
17-JUL-24
------------------------------

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

Hi @Megumi IZUMIOSAWA

• Initially, there were too few events (e.g., deaths) in the polypharmacy group for the Cox regression model to estimate the hazard ratio reliably. Sparse data can lead to unreliable or non-estimable results in regression analysis.
• By adding more "1"s, you increased the number of events in the polypharmacy group, providing the model with enough data to estimate the hazard ratio.

• The distribution of events across the levels of the polypharmacy variable might have been heavily skewed, meaning there was a significant imbalance between the groups (polypharmacy present vs. absent). This lack of variability can prevent the model from effectively estimating the hazard ratio.

Stratification:

• By stratifying by the names of five drugs, you controlled for their confounding effects, ensuring that these drugs did not overlap with the covariates in your model.

Model Inclusion:

• You included polypharmacy and ten other covariates as covariates in your Cox regression model. This approach allowed the model to isolate the effect of polypharmacy on mortality.

Results:

• After performing the stratified analysis, polypharmacy emerged as a significant risk factor for mortality, with an HR of 1.455. This means that, after adjusting for other covariates and the stratified drug names, patients with polypharmacy have a 45.5% higher risk of mortality compared to those without polypharmacy.

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

Hi @Meni Berger

While there isn't a strict, universally applicable thumb rule for event distribution in Cox regression, a commonly cited guideline is the "10 events per variable" rule.

This rule suggests that for reliable model performance, you should have at least 10 events for each predictor variable included in your model. However, it's important to note that:

• This is a general guideline and not a rigid rule.
• The performance of the model can be influenced by various factors beyond the number of events, such as the distribution of event times, the number of censored observations, and the complexity of the model.
• Recent research suggests that this rule might be overly conservative in some cases.

Some tests to consider:

1. Frequency Counts: Ensure that each category of a categorical variable has a reasonable number of cases. A rule of thumb is that each category should have at least 10 events (e.g., deaths) for reliable estimates.
2. Proportional Distribution: Check the proportion of events across the levels of your variable. Ideally, there should not be extreme imbalances (e.g., one category having 95% of the events while the others have very few).
3. Chi-Square Test: Use the chi-square test for independence to check if there is a significant association between your predictor and outcome variables. A significant result might indicate an imbalance that could affect your regression results.
4. Log-Rank Test: This test can compare the survival distributions of two samples. It helps ensure that there is a sufficient number of events across different strata or groups.
5. Visual Inspection: Plot the Kaplan-Meier survival curves for different levels of your covariates. If the curves are extremely separated or show little overlap, it could indicate issues with event distribution.
   
   
   And regarding plans for upgrading SPSS power analysis features, we are likely to expand them in the future based on user feedback and market trends.

------------------------------
Aruna Saraswathy
Statistician
SPSS Statistics
IBM
------------------------------
-------------------------------------------
Original Message:
Sent: Wed July 17, 2024 03:01 AM
From: 恵 泉澤
Subject: TS016685054：I don't understand the COX regression analysis results -> why the polypharmacy variable does not come up with a hazard ratio.

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

------------------------------
TS016685054
Megumi IZUMIOSAWA（JAPAN）
17-JUL-24
------------------------------

Hi @Meni Berger

While there isn't a strict, universally applicable thumb rule for event distribution in Cox regression, a commonly cited guideline is the "10 events per variable" rule.

This rule suggests that for reliable model performance, you should have at least 10 events for each predictor variable included in your model. However, it's important to note that:

• This is a general guideline and not a rigid rule.
• The performance of the model can be influenced by various factors beyond the number of events, such as the distribution of event times, the number of censored observations, and the complexity of the model.
• Recent research suggests that this rule might be overly conservative in some cases.

Some tests to consider:

1. Frequency Counts: Ensure that each category of a categorical variable has a reasonable number of cases. A rule of thumb is that each category should have at least 10 events (e.g., deaths) for reliable estimates.
2. Proportional Distribution: Check the proportion of events across the levels of your variable. Ideally, there should not be extreme imbalances (e.g., one category having 95% of the events while the others have very few).
3. Chi-Square Test: Use the chi-square test for independence to check if there is a significant association between your predictor and outcome variables. A significant result might indicate an imbalance that could affect your regression results.
4. Log-Rank Test: This test can compare the survival distributions of two samples. It helps ensure that there is a sufficient number of events across different strata or groups.
5. Visual Inspection: Plot the Kaplan-Meier survival curves for different levels of your covariates. If the curves are extremely separated or show little overlap, it could indicate issues with event distribution.
   
   
   And regarding plans for upgrading SPSS power analysis features, we are likely to expand them in the future based on user feedback and market trends.

------------------------------
Aruna Saraswathy
Statistician
SPSS Statistics
IBM
------------------------------
-------------------------------------------
Original Message:
Sent: Wed July 17, 2024 03:01 AM
From: 恵 泉澤
Subject: TS016685054：I don't understand the COX regression analysis results -> why the polypharmacy variable does not come up with a hazard ratio.

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

------------------------------
TS016685054
Megumi IZUMIOSAWA（JAPAN）
17-JUL-24
------------------------------

Note that nowadays, Frank Harrell and others recommend a minimum of 20 events (not 10) per explanatory variable degree of freedom.  See point 3 here (or search for 20:1):

Author Checklist

Datamethods Discussion Forum		remove preview
Author Checklist Statistical Problems to Document and to Avoid Checklist for Authors References | Ten Simple Rules | Checklist to Avoid p-Hacking | Research Reliability and Publishing | Guidelines for Reporting of Statistics | Guidelines for Figures and Tables | Statistical Myths | Video | Common Pitfalls This list is not all-inclusive, and nominations for additional entries are welcomed. View this on Datamethods Discussion Forum >

It is important to note that these rules are about explanatory variable degrees of freedom, not explanatory variables per se.  E.g., if you include a categorical variable with 5 categories, it eats up 4 explanatory df. 

Note too that such rules of thumb are concerned with reducing the likelihood of overfitting your model, not with ensuring adequate power to detect some smallest effect size of interest. 

HTH.

Hi @Meni Berger

While there isn't a strict, universally applicable thumb rule for event distribution in Cox regression, a commonly cited guideline is the "10 events per variable" rule.

This rule suggests that for reliable model performance, you should have at least 10 events for each predictor variable included in your model. However, it's important to note that:

• This is a general guideline and not a rigid rule.
• The performance of the model can be influenced by various factors beyond the number of events, such as the distribution of event times, the number of censored observations, and the complexity of the model.
• Recent research suggests that this rule might be overly conservative in some cases.

Some tests to consider:

1. Frequency Counts: Ensure that each category of a categorical variable has a reasonable number of cases. A rule of thumb is that each category should have at least 10 events (e.g., deaths) for reliable estimates.
2. Proportional Distribution: Check the proportion of events across the levels of your variable. Ideally, there should not be extreme imbalances (e.g., one category having 95% of the events while the others have very few).
3. Chi-Square Test: Use the chi-square test for independence to check if there is a significant association between your predictor and outcome variables. A significant result might indicate an imbalance that could affect your regression results.
4. Log-Rank Test: This test can compare the survival distributions of two samples. It helps ensure that there is a sufficient number of events across different strata or groups.
5. Visual Inspection: Plot the Kaplan-Meier survival curves for different levels of your covariates. If the curves are extremely separated or show little overlap, it could indicate issues with event distribution.
   
   
   And regarding plans for upgrading SPSS power analysis features, we are likely to expand them in the future based on user feedback and market trends.

------------------------------
Aruna Saraswathy
Statistician
SPSS Statistics
IBM
------------------------------
-------------------------------------------
Original Message:
Sent: Wed July 17, 2024 03:01 AM
From: 恵 泉澤
Subject: TS016685054：I don't understand the COX regression analysis results -> why the polypharmacy variable does not come up with a hazard ratio.

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

------------------------------
TS016685054
Megumi IZUMIOSAWA（JAPAN）
17-JUL-24
------------------------------

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

Hi @恵 泉澤

A small percentage doesn't automatically mean there's a problem. The critical factor is how these deaths are distributed across the categories of the covariate. Even if the percentage is low, as long as the absolute number (701 deaths) provides enough data in each relevant category, it's generally sufficient. For Cox regression, it's important to look at how events (deaths) are spread across different levels of your covariate (e.g., different levels of polypharmacy). If there are many deaths but they are all concentrated in one category, it might skew results. Review the distribution of deaths across categories of polypharmacy. For example, if the majority of the deaths are in one category, make sure that the other categories also have a sufficient number of events.

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

Hi @恵 泉澤

A small percentage doesn't automatically mean there's a problem. The critical factor is how these deaths are distributed across the categories of the covariate. Even if the percentage is low, as long as the absolute number (701 deaths) provides enough data in each relevant category, it's generally sufficient. For Cox regression, it's important to look at how events (deaths) are spread across different levels of your covariate (e.g., different levels of polypharmacy). If there are many deaths but they are all concentrated in one category, it might skew results. Review the distribution of deaths across categories of polypharmacy. For example, if the majority of the deaths are in one category, make sure that the other categories also have a sufficient number of events.

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

Dear Dr  Aruna

Thank you for your guidance.
As for the distribution among categories, I thought I would attempt to confirm this by drawing a scatter plot. However, I cannot see the shape of the distribution because the events are 0, 1 (alive, dead) and polypharmacy (1 with, 0 without).
So I created a forest plot of reported odds ratios with death and each factor.
Here we ran into a problem.
For the factors that Kaplan-Meier found to have significant negative hazard ratios, the reported odds ratio calculations were all positive and significant, the opposite result.

Polypharmacy remained a significant positive factor in the reported odds ratio. Hypertension, however, was a significant negative hazard ratio in Kaplan-Meier .

So the question is.
Q1, Is the reported odds ratio (calculated by χ2 test) not accurate because it confirms the risk ratio only between the two factors?
However, can we consider the reported odds ratio as a form of distribution?
Q2, When considering the outcome of death, is it common to use the multi-factor Kaplan-Meier calculation? Forest plots are also commonly used.
HR values for heart failure and hypertension are negative.
Sorry to keep repeating this, but it doesn't feel right to me that the positive and negative relationship between HR and OR values are reversed.

Sorry to hear back from IBM who will close this thread on the 14th.
However, I would appreciate your guidance as it is necessary to properly understand how to use SPSS.
Thank you in advance for your cooperation

 ※*Reported odds ratios are attached in the attachment.

Best Regard 

M.IZUMISAWA10-AUG-24

Hi @恵 泉澤

A small percentage doesn't automatically mean there's a problem. The critical factor is how these deaths are distributed across the categories of the covariate. Even if the percentage is low, as long as the absolute number (701 deaths) provides enough data in each relevant category, it's generally sufficient. For Cox regression, it's important to look at how events (deaths) are spread across different levels of your covariate (e.g., different levels of polypharmacy). If there are many deaths but they are all concentrated in one category, it might skew results. Review the distribution of deaths across categories of polypharmacy. For example, if the majority of the deaths are in one category, make sure that the other categories also have a sufficient number of events.

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

Hi,

You're absolutely right to question the discrepancy between odds ratios and hazard ratios.

Odds ratio is a measure of association between two categorical variables. It's suitable for case-control studies but less informative for survival data. Hazard ratio is specific to survival analysis. It measures the instantaneous risk of an event at a specific time, considering time-dependent covariates.   The key point is that odds ratios and hazard ratios measure different things. While an odds ratio of greater than 1 suggests an increased risk, a hazard ratio of less than 1 suggests a decreased risk. This difference can lead to seemingly contradictory results.

Can we consider the reported odds ratio as a form of distribution?
No, an odds ratio is a single value, not a distribution. It represents the ratio of odds between two groups.

You're correct in using Kaplan-Meier curves and forest plots for survival data.

Kaplan-Meier curves visualize survival probabilities over time, allowing comparison between groups. Forest plots are excellent for presenting hazard ratios and their confidence intervals from multiple models.  Multi-factor Kaplan-Meier calculations are not typically used. Instead, Cox proportional hazards models are employed to analyze the effects of multiple factors on survival.  In short, Odds ratios and hazard ratios measure different aspects of the data. Hazard ratios are more appropriate for survival analysis. Kaplan-Meier curves and forest plots are valuable tools for visualizing and interpreting survival data.

Dear Dr  Aruna

Thank you for your guidance.
As for the distribution among categories, I thought I would attempt to confirm this by drawing a scatter plot. However, I cannot see the shape of the distribution because the events are 0, 1 (alive, dead) and polypharmacy (1 with, 0 without).
So I created a forest plot of reported odds ratios with death and each factor.
Here we ran into a problem.
For the factors that Kaplan-Meier found to have significant negative hazard ratios, the reported odds ratio calculations were all positive and significant, the opposite result.

Polypharmacy remained a significant positive factor in the reported odds ratio. Hypertension, however, was a significant negative hazard ratio in Kaplan-Meier .

So the question is.
Q1, Is the reported odds ratio (calculated by χ2 test) not accurate because it confirms the risk ratio only between the two factors?
However, can we consider the reported odds ratio as a form of distribution?
Q2, When considering the outcome of death, is it common to use the multi-factor Kaplan-Meier calculation? Forest plots are also commonly used.
HR values for heart failure and hypertension are negative.
Sorry to keep repeating this, but it doesn't feel right to me that the positive and negative relationship between HR and OR values are reversed.

Sorry to hear back from IBM who will close this thread on the 14th.
However, I would appreciate your guidance as it is necessary to properly understand how to use SPSS.
Thank you in advance for your cooperation

 ※*Reported odds ratios are attached in the attachment.

Best Regard 

M.IZUMISAWA10-AUG-24

Hi @恵 泉澤

A small percentage doesn't automatically mean there's a problem. The critical factor is how these deaths are distributed across the categories of the covariate. Even if the percentage is low, as long as the absolute number (701 deaths) provides enough data in each relevant category, it's generally sufficient. For Cox regression, it's important to look at how events (deaths) are spread across different levels of your covariate (e.g., different levels of polypharmacy). If there are many deaths but they are all concentrated in one category, it might skew results. Review the distribution of deaths across categories of polypharmacy. For example, if the majority of the deaths are in one category, make sure that the other categories also have a sufficient number of events.

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

Hi,

You're absolutely right to question the discrepancy between odds ratios and hazard ratios.

Odds ratio is a measure of association between two categorical variables. It's suitable for case-control studies but less informative for survival data. Hazard ratio is specific to survival analysis. It measures the instantaneous risk of an event at a specific time, considering time-dependent covariates.   The key point is that odds ratios and hazard ratios measure different things. While an odds ratio of greater than 1 suggests an increased risk, a hazard ratio of less than 1 suggests a decreased risk. This difference can lead to seemingly contradictory results.

Can we consider the reported odds ratio as a form of distribution?
No, an odds ratio is a single value, not a distribution. It represents the ratio of odds between two groups.

You're correct in using Kaplan-Meier curves and forest plots for survival data.

Kaplan-Meier curves visualize survival probabilities over time, allowing comparison between groups. Forest plots are excellent for presenting hazard ratios and their confidence intervals from multiple models.  Multi-factor Kaplan-Meier calculations are not typically used. Instead, Cox proportional hazards models are employed to analyze the effects of multiple factors on survival.  In short, Odds ratios and hazard ratios measure different aspects of the data. Hazard ratios are more appropriate for survival analysis. Kaplan-Meier curves and forest plots are valuable tools for visualizing and interpreting survival data.

Dear Dr  Aruna

Thank you for your guidance.
As for the distribution among categories, I thought I would attempt to confirm this by drawing a scatter plot. However, I cannot see the shape of the distribution because the events are 0, 1 (alive, dead) and polypharmacy (1 with, 0 without).
So I created a forest plot of reported odds ratios with death and each factor.
Here we ran into a problem.
For the factors that Kaplan-Meier found to have significant negative hazard ratios, the reported odds ratio calculations were all positive and significant, the opposite result.

Polypharmacy remained a significant positive factor in the reported odds ratio. Hypertension, however, was a significant negative hazard ratio in Kaplan-Meier .

So the question is.
Q1, Is the reported odds ratio (calculated by χ2 test) not accurate because it confirms the risk ratio only between the two factors?
However, can we consider the reported odds ratio as a form of distribution?
Q2, When considering the outcome of death, is it common to use the multi-factor Kaplan-Meier calculation? Forest plots are also commonly used.
HR values for heart failure and hypertension are negative.
Sorry to keep repeating this, but it doesn't feel right to me that the positive and negative relationship between HR and OR values are reversed.

Sorry to hear back from IBM who will close this thread on the 14th.
However, I would appreciate your guidance as it is necessary to properly understand how to use SPSS.
Thank you in advance for your cooperation

 ※*Reported odds ratios are attached in the attachment.

Best Regard 

M.IZUMISAWA10-AUG-24

Hi @恵 泉澤

A small percentage doesn't automatically mean there's a problem. The critical factor is how these deaths are distributed across the categories of the covariate. Even if the percentage is low, as long as the absolute number (701 deaths) provides enough data in each relevant category, it's generally sufficient. For Cox regression, it's important to look at how events (deaths) are spread across different levels of your covariate (e.g., different levels of polypharmacy). If there are many deaths but they are all concentrated in one category, it might skew results. Review the distribution of deaths across categories of polypharmacy. For example, if the majority of the deaths are in one category, make sure that the other categories also have a sufficient number of events.

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

Aruna Saraswathy wrote: 

"Odds ratio is a measure of association between two categorical variables. It's suitable for case-control studies but less informative for survival data." 

To be clear, the odds ratio can also be used as a measure of association between a quantitative explanatory variable and a categorical outcome variable.  E.g., the following code estimates a binary logistic regression with Age in years as the only explanatory variables, and Exp(B) for Age is an odds ratio.

* Modify path to point to where the "sample" datasets are stored on your computer.
GET FILE='C:\SPSSdata\bankloan.sav'.
LOGISTIC REGRESSION VARIABLES default
  /METHOD=ENTER age 
  /PRINT=CI(95)
  /CRITERIA=PIN(0.05) POUT(0.10) ITERATE(20) CUT(0.5).

Hi,

You're absolutely right to question the discrepancy between odds ratios and hazard ratios.

Odds ratio is a measure of association between two categorical variables. It's suitable for case-control studies but less informative for survival data. Hazard ratio is specific to survival analysis. It measures the instantaneous risk of an event at a specific time, considering time-dependent covariates.   The key point is that odds ratios and hazard ratios measure different things. While an odds ratio of greater than 1 suggests an increased risk, a hazard ratio of less than 1 suggests a decreased risk. This difference can lead to seemingly contradictory results.

Can we consider the reported odds ratio as a form of distribution?
No, an odds ratio is a single value, not a distribution. It represents the ratio of odds between two groups.

You're correct in using Kaplan-Meier curves and forest plots for survival data.

Kaplan-Meier curves visualize survival probabilities over time, allowing comparison between groups. Forest plots are excellent for presenting hazard ratios and their confidence intervals from multiple models.  Multi-factor Kaplan-Meier calculations are not typically used. Instead, Cox proportional hazards models are employed to analyze the effects of multiple factors on survival.  In short, Odds ratios and hazard ratios measure different aspects of the data. Hazard ratios are more appropriate for survival analysis. Kaplan-Meier curves and forest plots are valuable tools for visualizing and interpreting survival data.

Dear Dr  Aruna

Thank you for your guidance.
As for the distribution among categories, I thought I would attempt to confirm this by drawing a scatter plot. However, I cannot see the shape of the distribution because the events are 0, 1 (alive, dead) and polypharmacy (1 with, 0 without).
So I created a forest plot of reported odds ratios with death and each factor.
Here we ran into a problem.
For the factors that Kaplan-Meier found to have significant negative hazard ratios, the reported odds ratio calculations were all positive and significant, the opposite result.

Polypharmacy remained a significant positive factor in the reported odds ratio. Hypertension, however, was a significant negative hazard ratio in Kaplan-Meier .

So the question is.
Q1, Is the reported odds ratio (calculated by χ2 test) not accurate because it confirms the risk ratio only between the two factors?
However, can we consider the reported odds ratio as a form of distribution?
Q2, When considering the outcome of death, is it common to use the multi-factor Kaplan-Meier calculation? Forest plots are also commonly used.
HR values for heart failure and hypertension are negative.
Sorry to keep repeating this, but it doesn't feel right to me that the positive and negative relationship between HR and OR values are reversed.

Sorry to hear back from IBM who will close this thread on the 14th.
However, I would appreciate your guidance as it is necessary to properly understand how to use SPSS.
Thank you in advance for your cooperation

 ※*Reported odds ratios are attached in the attachment.

Best Regard 

M.IZUMISAWA10-AUG-24

Hi @恵 泉澤

A small percentage doesn't automatically mean there's a problem. The critical factor is how these deaths are distributed across the categories of the covariate. Even if the percentage is low, as long as the absolute number (701 deaths) provides enough data in each relevant category, it's generally sufficient. For Cox regression, it's important to look at how events (deaths) are spread across different levels of your covariate (e.g., different levels of polypharmacy). If there are many deaths but they are all concentrated in one category, it might skew results. Review the distribution of deaths across categories of polypharmacy. For example, if the majority of the deaths are in one category, make sure that the other categories also have a sufficient number of events.

The input is the same as for the other variables, with (1) and without (0) inputs.

　　

The input is the same as for the other variables, with (1) and without (0) inputs.