Sample Size Guidelines for Logistic Regression from Observational Studies with Large Population: Emphasis on the Accuracy Between Statistics and Parameters Based on Real Life Clinical Data.
Bujang MA, Sa'at N, Sidik TMITAB, Joo LC. · 2018
Read Sample Size Guidelines for Logistic Regression from Observational Studies with Large Population: Emphasis on the Accuracy Between Statistics and Parameters Based on Real Life Clinical Data. by Bujang MA, Sa'at N, Sidik TMITAB, Joo LC. (2018) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc6422534.txt.
Introduction
Bujang and co-authors offer sample size guidelines in the Malaysian Journal of Medical Sciences. Logistic regression is a workhorse for medical studies on yes-or-no outcomes. Choosing too small a sample makes its estimates stray from the truth.
Using real clinical data, the team tested how estimates settle as samples grow. At a sample of five hundred, sample values sat close to population values. Their audit of a mid-size population backed that rule. The guidelines target observational studies with large populations. They give researchers a simple floor for planning such studies.
Catalog record
| CREATOR | Bujang MA, Sa'at N, Sidik TMITAB, Joo LC. |
|---|---|
| YEAR | 2018 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 25 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Sample Size Guidelines for Logistic Regression from Observational Studies with Large Population: Emphasis on the Accuracy Between Statistics and Parameters Based on Real Life Clinical Data By Mohamad Adam Bujang, Nadiah Sa’at, Tg Mohd Ikhwan Tg Abu Bakar Sidik, Lim Chien Joo Published in: The Malaysian Journal of Medical Sciences : MJMS DOI: 10.21315/mjms2018.25.4.12 ABSTRACT Background Different study designs and population size may require different sample size for logistic regression. This study aims to propose sample size guidelines for logistic regression based on observational studies with large population. Methods We estimated the minimum sample size required based on evaluation from real clinical data to evaluate the accuracy between statistics derived and the actual parameters. Nagelkerke r-squared and coefficients derived were compared with their respective parameters. Results With a minimum sample size of 500, results showed that the differences between the sample estimates and the population was sufficiently small. Based on an audit from a medium size of population, the differences were within ± 0.5 for coefficients and ± 0.02 for Nagelkerke r-squared. Meanwhile for large population, the differences are within ± 1.0 for coefficients and ± 0.02 for Nagelkerke r-squared. Conclusions For observational studies with large population size that involve logistic regression in the analysis, taking a minimum sample size of 500 is necessary to derive the statistics that represent the parameters. The other recommended rules of thumb are EPV of 50 and formula; n = 100 + 50i where i refers to number of independent variables in the final model. INTRODUCTION Logistic regression is one of the most utilised statistical analyses in multivariable models especially in medical research. Beside the fact that most clinical outcomes are defined as binary form (e.g. survived versus died or poor outcome versus good outcome), logistic regression also requires less assumptions as compared to multiple linear regression or Analysis of Covariance (ANCOVA). In observational studies, logistic regression is commonly used to determine the associated factors with or without controlling for specific variables and also for predictive modelling (1–4). Since the purpose of most of statistical analyses is for inference, determination of sample size requirement is necessary before the analysis is conducted. The sample size requirement for logistic regression has been discussed in the literature. Earlier on, Hsieh (5) proposed a sample size table for logistic regression but limited the estimation for only one covariate. According to the paper, adjustment needed to be made for the sample size tables such as dividing the estimated sample size with a factor of (1–p2) when sample size need to be estimated for logistic regression. Another famous sample size guideline proposed that the minimum required sample size should be based on the rule of event per variable (EPV) (6). According to Concato et al. and Peduzzi et al., the concept of EPV of 10 is acceptable for both logistic regression and cox regression (6–7). Based on EPV, researchers need to estimate the proportion for the outcome in the least category and divide it by 10 in order to determine the number of independent variables which can be studied. The concept of EPV with 10 received some critics (8) and hence, Austin and Steyerberg recommended EPV of 20 instead (9). Besides that, studies with small to moderate samples size such as less than 100 usually overestimate the effect measure. Nemes and colleagues from their simulation study, showed that large sample size preferably 500 will increase the accuracy of the estimates (10). The rules of thumb with 500 subjects were also been recommended by other studies (11–12). In these studies, sample size with 500 and above yielded statistics which represented the parameters in the targeted population. The results were derived after evaluating few populations and were analysed based on various statistical tests. The present study did an evaluation using real patient data derived from an observational study to evaluate the extent of different sample sizes used in affecting the discrepancy between the sample statistics and the actual parameters in the target population. The purpose of this comparison is to estimate a minimum sample size required for a research study which is able to yield the closest estimate for the coefficients and also r-squared. This is to determine a sizeable sample size for logistic regression that can produce the statistics which is able to be inferred to the larger population particularly for observational studies. Sample size for experimental studies are usually calculated using sample size softwares. In experimental studies, the confounders are usually controlled at study design stage and this made the calculation is feasible based on univariate analysis. The researcher only need to estimate of effect sizes in order to calculate the minimum requirement of sample size. Very often, observational studies will involve multivariable analysis with many parameters and various effect sizes. Therefore, in the present study, we propose a simple rule of thumb as a basis for sample size estimation for logistic regression particularly for observational studies. In the perspective of observational studies, the findings obtained from the validation of real data were used as the basis for sample size recommendation for logistic regression. MATERIAL AND METHODS Validation was conducted to verify the accuracy between statistics and parameters. The validation was performed using real patient data from “An Audit of Diabetes Control and Management (ADCM) 2009”, which included all data collection (at a national-level) of patients with diabetes mellitus from all government health clinics in Malaysia in 2009. The methodology of this data collection process was explained in a previous paper and published elsewhere (13). We selected one government health clinic which had a relatively high number of patients with a total population of 1,595, and re-analysis was done by using different sub-samples (n = 30, 50, 100, 150, 200, 300, 500, 700 and 1,000). We tested a multivariable model by using eight explanatory (or independent) variables and one outcome (or dependent variable). The dependent variable was glycemic control (HbA1c) in binary form (< 7.0 versus ≥ 7.0) while a set of independent variables included gender, age, body mass index, diabetes treatment, duration of diabetes mellitus, systolic blood pressure, status of co-morbidity and low-density lipoprotein level. Since data was not collected in a prospective manner, the model developed could only be used to test for an association between the independent variables and the outcome; rather than to identify and determine the risk factors or determinants for HbA1c (14–15). The findings obtained from the validation were then analysed. The statistics such as r-squared and coefficients derived from the samples were compared with the respective true values (parameters) in the targeted population. The analysis was conducted using logistic regression where the sample sizes (n = 30, 50, 100, 150, 200, 300, 500, 700 and 1,000) were selected at random. From the results, guidelines of sample size estimation for logistic regression based on the concept of event per variable (EPV) and sample size formula (n = 100 + xi, where x is integer and i represents the number of independent variables in the final model) were introduced. After the guidelines of the sample size were identified, these guidelines (based on EPV and sample size formula) were re-evaluated based on another extremely large population with total population of 70,899 records. This population was also