I am currently researching if some fund characteristics such as (fund size, fund family size, capital flows, and fund age) explains fund performance measured (monthly alpha).

Therefore, I am using a fixed effect panel regression where the dependant variable is monthly fund performance; the independent variables are lagged monthly fund characteristic. I perform a fixed effect regression and cluster by (ID) to adjust standard errors for cross-sectional dependence and heteroscedastic residuals.

My sample size is 350 funds with 120 observation for each fund (10 years). In total, I have around 42,000 observations.

Are there any disadvantages and/or benefits for using a large panel dataset?

You're help is appreciated.

• By monthly fund performance you mean $R_{it}$ is the return of fund $i$ in month $t$? And to compute standard errors, you're clustering by fund? – Matthew Gunn Aug 4 '17 at 18:56
• @MatthewGunn I estimate fund performance in two ways. 1) return of fund i in month t. 2) four-factor alpha of fund i in month t. Then, I cluster standard errors by fund. The monthly alpha is calculated as (realised return - expected return) similar to how abnormal returns are calculated in event studies. – user28909 Aug 4 '17 at 20:02
• The big issue when computing standard errors is cross-sectional correlation of returns within a time period. Eg. firms in the same industry will have correlated error terms in either model (1) or (2). You should cluster by date (in your case, month). – Matthew Gunn Aug 4 '17 at 20:02
• @MatthewGunn I appreciate your help sir. Is it an issue, however, to have a large panel dataset? – user28909 Aug 4 '17 at 20:09
• Generally speaking, the more data the better. What can go wrong? You may have too small of standard errors if you make a mistake and treat error terms as uncorrelated even though they aren't. In some sense, you'd be computing standard errors as if you had way more data than you actually do. (Speaking loosely, correlated error terms is similar to having less data.) Eg. if you cluster by time, I'd expect it may change your standard errors quite a lot. – Matthew Gunn Aug 4 '17 at 20:20

In general, more data is better than less data.

On the topic of your specific scenario, you want to cluster by date or use some other procedure to produce consistent standard errors in the presence of cross-sectional correlation.

Monthly returns are basically uncorrelated over time but exhibit significant cross-sectional correlation.

# With large quantities of data, treating correlated error terms as uncorrelated can massively understate standard errors!

### Example:

Let's say I have $i=1,\ldots,50$ people recording the results of me flipping a coin 20 times ($t=1,\ldots,20$). Let $y_{it}$ be person $i$'s recorded result of flip $t$. I have $20 \cdot 50 = 1000$ observations.

My model is: $$y_{it} = \mu + \epsilon_{it}$$

If I treat each $\epsilon_{it}$ as uncorrelated, I'm going to massively understate my standard errors. In reality, I have basically 20 independent observations, not 1,000. For each time $t$, the $\epsilon_{it}$ will be significantly correlated.

### Basically the same thing happens with returns

For any time period $t$, returns ${R}_{it}$ are correlated. There's huge cross-sectional correlation.

Hence, you'd want to cluster by date. There are other methods of course to deal with cross-sectional correlation.

The same logic is behind: - forming portfolios and using time-series variation in portfolio returns - the Fama-Macbeth procedure of running $T$ cross-sectional correlations and taking the time-series average and standard deviation to compute estimates and standard errors.

See this comment from the wiki page on fixed effects models:

In statistics, a fixed effects model is a statistical model in which the model parameters are fixed ... Such models assist in controlling for unobserved heterogeneity when this heterogeneity is constant over time.

Emphasis on the time constancy of unobserved heterogeneity is my own. Do you believe that the unobserved heterogeneity in mutual fund characteristics is constant over time? This could be a reasonable dimension on which less is more.